diff --git a/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-privileged--cfh2oo7.json b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-privileged--cfh2oo7.json new file mode 100644 index 0000000000000000000000000000000000000000..e535aaf77f5b67426292941a6c0630372c7ce8f0 --- /dev/null +++ b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-privileged--cfh2oo7.json @@ -0,0 +1,173 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-balance-hard-i00-privileged--cfh2oo7", + "taskKey": "humanoidbench/loco_balance_hard", + "dag_sha256": "419d18f9a6fdc6ba23a8ef8263e0e2a03d0e684ca7c5204c63fb59f163363f46", + "outcome": { + "success": 0, + "reward": 7.171364379042991e-08, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 205, + "final_reward": 7.171364379042991e-08, + "sim_steps": 205, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 2.452, + "peak_processes": 12, + "peak_disk_gb": 0.832, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-balance-hard-i00-privileged--cfh2oo7/replay.mp4", + "duration_s": 2.15, + "fps": 20.0, + "frames": 43, + "sha256": "3795024022d58c5a0bfbe4a800964ed30e9c3b023e86f0f615fe59d39b79632a" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "board_balance_sustained": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:本次视频0.00 s机器人双脚承重于平衡板,板下球体支承板面;视频0.05–1.50 s机器人抬臂并保持倾斜站姿,双脚仍在板上。开局支撑和短暂站立不满足本节点的累计奖励800验收。" + }, + { + "t": 1.9, + "log_step": null, + "text_zh": "观察:逐帧检查本次视频1.50–2.10 s,视频1.70–1.85 s机器人躯干倾斜加剧,视频1.90–2.05 s双脚先后离开板面,板与球滚向画面左侧;视频2.10 s末帧机器人已在板旁,未恢复板上支撑。" + }, + { + "t": 2.1, + "log_step": null, + "text_zh": "推断:评分器字段success=0、deterministic=1、sim_steps=n_actions=205,与本次视频末段失去板上支撑一致;按本节点每步至多1的归一化乘积奖励,205个控制步即使全部满分也不足累计800,因此本事件全程未达成。reward.final_reward=7.171364379042991e-08未标为累计回报,不能据此填episode_return。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频0.00–1.85 s机器人双脚实际接触平衡板承重,腿和躯干随板调整姿态;视频1.90–2.05 s双脚离开板面,末段落到板旁地面。双臂抬起但未见接触板、球或其他物体。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "board_balance_sustained", + "t": 1.8, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:本次视频1.70–1.85 s机器人在板上保持平衡时躯干倾斜继续增大,视频1.90–2.05 s失去双脚在板上的支撑;视频2.10 s末帧没有恢复。评分器记录success=0、sim_steps=205,失败发生于持续平衡步骤。" + }, + "unmet": { + "board_balance_sustained": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 1.8, + "log_step": null, + "evidence_zh": "观察:本次视频1.70–1.90 s机器人向板外倾倒,视频1.90–2.05 s脚与板分离,板球向画面左侧滚开,机器人落到板旁;视频2.10 s仍未回板。可见的是失去板上支撑,录像结束前躯干尚未完全倒地。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "反馈控制", + "evidence_zh": "推断:log step 155、161、177的生成器代码使用板与躯干俯仰、质心和滚动球运动、骨盆高度及腿关节状态生成关节目标,并加入定时和后段反馈修正;log step 195更新后段反馈矩阵并重新生成最终205步交付轨迹,评分器deterministic=1。本次视频1.50–2.05 s仍出现逐渐加剧的倾斜和脱板,说明该反馈生成的动作未能抑制失稳;日志中的较早测试仅作控制方案证据,不作为本次视频状态或时间。", + "source_zh": "日志 / 视频 / 评分器字段" + } + ], + "gap_zh": "没有状态日志或原生累计episode_return字段;reward.final_reward仅以该字段名报告,缺少证据证明其为累计回报,故原生指标值留null。视频足以判定板上支撑丧失,但不能精确恢复接触力、实际质心轨迹及最先触发的原生终止谓词,也不能进一步确定哪个反馈增益导致失稳。" + }, + "summary_zh": "本次评分轨迹实际尝试了板上平衡,双脚最初由平衡板支撑,但累计奖励800的唯一节点未完成,最终进度0/1。视频约1.80 s机器人倾斜加剧,随后双脚脱板,至视频2.10 s未恢复;结合生成器日志,根因归为反馈控制未能抑制失稳。缺少状态日志和明确的累计回报字段,无法精确判定接触力与原生终止触发项。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "视频帧与标注一致:开局双脚在板上,约1.5–1.8 s倾斜加剧,1.9–2.05 s脱板,节点未达成,断点、平衡现象、反馈控制根因均有证据支持。未发现需要改动的内容。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:52:28Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-privileged--cfh2oo7/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-privileged--cfh2oo7/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "41bec359f6ce8cde51e6dcceab8b7980e1f77f4806fa8ec8a196852143a7bf2b", + "video_sha256": "3795024022d58c5a0bfbe4a800964ed30e9c3b023e86f0f615fe59d39b79632a", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-privileged--cxpenpw.json b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-privileged--cxpenpw.json new file mode 100644 index 0000000000000000000000000000000000000000..b0e622116c1c332e46d9f5e3a93efdb19d361e78 --- /dev/null +++ b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-privileged--cxpenpw.json @@ -0,0 +1,178 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-balance-hard-i00-privileged--cxpenpw", + "taskKey": "humanoidbench/loco_balance_hard", + "dag_sha256": "419d18f9a6fdc6ba23a8ef8263e0e2a03d0e684ca7c5204c63fb59f163363f46", + "outcome": { + "success": 0, + "reward": 9.17032743676796e-08, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 107, + "final_reward": 9.17032743676796e-08, + "sim_steps": 107, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 7.539, + "peak_processes": 45, + "peak_disk_gb": 0.24, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-balance-hard-i00-privileged--cxpenpw/replay.mp4", + "duration_s": 1.15, + "fps": 20.0, + "frames": 23, + "sha256": "5a714126468c1f1efb381664b8d8033ec7aaf1cca82eff5564a8d170db9b280c" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "board_balance_sustained": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00 s 机器人双脚站在球体支承的平衡板上;视频 0.10–0.80 s 机器人反复摆臂、屈腿,平衡板转动、倾斜,身体随之偏移。开局站姿仅是开始尝试,尚不满足累计奖励达到800的节点判据。" + }, + { + "t": 0.9, + "log_step": null, + "text_zh": "观察:逐帧检查视频 0.00–1.10 s 的全部23帧;视频 0.90–1.00 s 双脚与板面分离,机器人向板外下落;视频 1.05–1.10 s 身体明显后仰、下降,未恢复板上站立。" + }, + { + "t": null, + "log_step": 75, + "text_zh": "观察:log step 75 中 agent 将种子1的动作文件选为最终交付,重放工具报告 deterministic=true、sim_steps=107;评分器同样报告 n_actions=107、sim_steps=107、success=0、deterministic=1。" + }, + { + "t": null, + "log_step": null, + "text_zh": "推断:按任务的逐步奖励乘积上限1,评分器记录的107个控制步累计最多107,低于节点要求的800,因此该事件始终未达成;此结论结合短回合及可见失稳,并非仅由零成功分推断未尝试。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.50 s 机器人双脚在平衡板上承重,双腿调节站姿,双臂在空中摆动;视频 0.60–0.85 s 支撑姿态明显改变,视频 0.90 s 起双脚离开板面。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "board_balance_sustained", + "t": 0.9, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频 0.50–0.85 s 机器人倾斜并改变脚下支撑,视频 0.90 s 双脚脱离板面,随后持续后仰下落,至视频 1.10 s 未恢复。失稳发生于板上平衡尝试过程中;断点以视频 0.90 s 的明确失去板面支撑为导航锚点。" + }, + "unmet": { + "board_balance_sustained": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 0.9, + "log_step": null, + "evidence_zh": "观察:视频 0.75–0.85 s 平衡板倾转、机器人身体偏离原站姿;视频 0.90–1.00 s 双脚离板,视频 1.05–1.10 s 机器人躯干继续后仰下降,板上支撑未恢复。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志 / 视频", + "evidence_zh": "推断:log step 34、44、51 的生成代码以模拟预测和随机采样选择腿、臂关节目标来抑制板倾角、身体倾斜与水平漂移;log step 75 选定最终动作文件并确定性重放107步,log step 79 又确认重新生成的107×19动作数组与交付文件完全相同。结合视频 0.10–1.10 s 的持续调姿后失去支撑,支持归因为所生成的平衡控制动作不能稳定该系统;证据不足以把原因进一步确定为某个反馈环节、参数或接触设计。" + } + ], + "gap_zh": "无状态日志,不能精确确认终止由骨盆过低还是哪一对碰撞触发,也不能从画面还原接触力或控制器内部失效环节。评分器仅报告 final_reward=9.17032743676796e-08,未明确报告 episode_return,故累计回报指标留空;不把该字段直接当作累计值。agent 关于踝伺服过软及静态重心偏置的解释(log step 80)涉及会话诊断,未据此认定最终交付的具体物理根因。" + }, + "summary_zh": "机器人已用双脚支撑在球体上的平衡板上并摆臂、调腿尝试平衡,但视频约0.90 s 双脚离板,随后后仰下落,唯一节点“板上持续站稳”未达成,最终0/1。最终交付可确定性重放107步,结合可见失稳,将现象标为全身·平衡,根因保留为控制。缺少状态日志及明确的累计回报字段,无法确定精确终止接触和控制失效细节。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "逐帧核对:0.90 s 双脚离板、之后后仰下落,唯一节点未达成(107步≤107<800)、断点/现象/根因与证据相符,时钟均为视频或log步数,无需修改。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:49:03Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-privileged--cxpenpw/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-privileged--cxpenpw/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "a201a4cff502394afe391b1f6cf40466d7bbef223b22b284fff885b6de43b801", + "video_sha256": "5a714126468c1f1efb381664b8d8033ec7aaf1cca82eff5564a8d170db9b280c", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-privileged--lrmhxam.json b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-privileged--lrmhxam.json new file mode 100644 index 0000000000000000000000000000000000000000..90ebe277f2fc9f90508de6538c1055da6e4f6e4e --- /dev/null +++ b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-privileged--lrmhxam.json @@ -0,0 +1,169 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-balance-hard-i00-privileged--lrmhxam", + "taskKey": "humanoidbench/loco_balance_hard", + "dag_sha256": "419d18f9a6fdc6ba23a8ef8263e0e2a03d0e684ca7c5204c63fb59f163363f46", + "outcome": { + "success": 0, + "reward": 0.47801494026554986, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 41, + "final_reward": 0.47801494026554986, + "sim_steps": 41, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-balance-hard-i00-privileged--lrmhxam/replay.mp4", + "duration_s": 0.5, + "fps": 20.0, + "frames": 10, + "sha256": "efa558c7342e8df9d520b8c3d116528bc6e79fe442f8fc244c21b5e4fa6f4202" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "board_balance_sustained": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 11, + "text_zh": "观察:视频 0.00–0.30 s,机器人双脚承重于浅灰平衡板,板下有深色球体,躯干起初接近竖直;视频 0.35–0.45 s,板面倾斜加剧,机器人身体随之倾斜、头部下降。已逐帧检查全部10帧,初始站姿不等于持续平衡达标。" + }, + { + "t": 0.45, + "log_step": 11, + "text_zh": "观察:log step 11 的脚本反复发送由初始关节角归一化得到的固定动作,保存轨迹并重放,返回 deterministic=true、scored=false、sim_steps=41;本次评分器同样报告 sim_steps=41、n_actions=41、deterministic=1、success=0。推断:受评录像对应这份固定姿态基线,后续会话中的闭环控制实验未替换交付;41个逐步奖励各不超过1,累计不可能达到800,因此该事件从未完成。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.45 s,机器人双脚踩在浅灰平衡板上承重,腿和躯干随板倾斜;双手伸在身体前方,未见接触板或球。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "board_balance_sustained", + "t": 0.35, + "log_step": 11, + "where": "操作中", + "evidence_zh": "观察:视频 0.30 s 后,机器人脚下的平衡板倾斜明显加剧;视频 0.35–0.45 s,机器人躯干倾斜、头部下沉,未恢复球上平衡支撑姿态,录像随即结束。log step 11 已保存并验证41步固定姿态基线,评分器确认本次仅41步且 success=0;失衡发生在持续站稳步骤执行中。" + }, + "unmet": { + "board_balance_sustained": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 0.35, + "log_step": null, + "evidence_zh": "观察:视频 0.00 s 板面近乎水平且机器人直立;视频 0.35–0.45 s 板与球的相对支撑位置发生明显偏移,板面急倾,机器人随板倾斜、头部降低,直到末帧没有恢复。这里可见的是失衡过程,录像未展示完整倒地。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "log_step": 11, + "source_zh": "日志", + "evidence_zh": "观察:log step 11 保存并验证的生成脚本在 reset 后读取初始关节角,只计算一次 a0,随后循环调用 s.step(a0),不根据板、球或机器人姿态更新动作;该次重放41步即结束且 scored=false,与评分器的41步和视频 0.35–0.45 s 失衡相符。推断:固定关节目标未能补偿滚动球与板的支撑漂移,交付控制器不能维持全身平衡;现有证据不足以进一步锁定某个关节增益或力矩参数。" + } + ], + "gap_zh": "观察:inputs/ 中无状态日志;评分器仅提供 final_reward=0.47801494026554986,未提供原生 episode_return,故该指标保留 null,不能把末步奖励当累计回报。视频最后一帧为视频 0.45 s,能确认失衡,但不能精确确认触发终止的碰撞对或高度条件。log step 101 为模型过滤提示,发生在 log step 11 保存并重放验证交付之后;按 D37,已固定的物理失败仍正常标注,后续会话实验不计入节点进度。" + }, + "summary_zh": "受评交付为 log step 11 保存并验证的41步固定关节姿态基线,双脚实际在平衡板上承重,但“板上持续站稳”未完成,最终0/1。视频 0.35–0.45 s 显示板面急倾和机器人失衡,固定姿态控制未维持滚动支撑系统的平衡;成功参考视频 0–9 s 的采样帧则持续保持板上站姿。无状态日志及累计回报数值,终止接触对无法确认;log step 101 的模型过滤发生在交付已保存验证之后。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "节点未完成、断点(板上持续站稳,视频约0.35 s失衡)、平衡现象和控制根因均与视频及评分器41步结果相符,未发现需要修改的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:50:37Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-privileged--lrmhxam/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-privileged--lrmhxam/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "8bfb72287fe573430c54b63660b176029e6f2323c45e93d736a34631d06a885e", + "video_sha256": "efa558c7342e8df9d520b8c3d116528bc6e79fe442f8fc244c21b5e4fa6f4202", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-privileged--r5t8nay.json b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-privileged--r5t8nay.json new file mode 100644 index 0000000000000000000000000000000000000000..53c75a898c42eac3b7c73691d5b259ce56c56c30 --- /dev/null +++ b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-privileged--r5t8nay.json @@ -0,0 +1,180 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-balance-hard-i00-privileged--r5t8nay", + "taskKey": "humanoidbench/loco_balance_hard", + "dag_sha256": "419d18f9a6fdc6ba23a8ef8263e0e2a03d0e684ca7c5204c63fb59f163363f46", + "outcome": { + "success": 0, + "reward": 1.1718730002009324e-09, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 102, + "final_reward": 1.1718730002009324e-09, + "sim_steps": 102, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 7.261, + "peak_processes": 43, + "peak_disk_gb": 0.239, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-balance-hard-i00-privileged--r5t8nay/replay.mp4", + "duration_s": 1.1, + "fps": 20.0, + "frames": 22, + "sha256": "ff262e09b88a0d368141a9c324c909e2738b393c477d3876145bca062352b457" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "board_balance_sustained": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00 s 机器人双脚位于浅灰色平衡板上,深色球体在板下承重;视频 0.05–0.65 s 机器人屈膝、后仰并摆臂调节,确已尝试板上平衡,初始站姿不代表累计奖励达标。" + }, + { + "t": 0.8, + "log_step": null, + "text_zh": "观察:逐帧检查视频 0.00–1.05 s;视频 0.80–1.05 s 机器人后仰不断加剧,头和躯干下坠,平衡板由近水平转为明显翻斜,未恢复站稳。" + }, + { + "t": null, + "log_step": 103, + "text_zh": "观察:log step 103 的生成及重放输出均为 102 个控制步;inputs/run.json 的评分器字段 sim_steps=102、n_actions=102、deterministic=1、final_reward=1.1718730002009324e-09、success=0。" + }, + { + "t": 1.05, + "log_step": null, + "text_zh": "推断:结合视频末段失衡及评分器 final_reward 远低于累计奖励阈值 800,板上持续站稳节点始终未达成;不把开局短暂直立计为完成。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.80 s 机器人双脚在浅灰色平衡板上承重,腿与躯干调节站姿;视频 0.80–1.05 s 板和机器人一起倾斜下坠。可见手臂摆动,未见手部接触板或球。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": 1.1718730002009324e-09, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "board_balance_sustained", + "t": 0.8, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频 0.65–0.80 s 机器人在继续板上平衡时后仰加剧;视频 0.90–1.05 s 平衡板明显翻斜,机器人躯干接近横卧并下坠,之后没有恢复。失效发生在唯一的板上持续站稳步骤中。" + }, + "unmet": { + "board_balance_sustained": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 0.8, + "log_step": null, + "evidence_zh": "观察:视频 0.80–1.05 s 机器人连同平衡板倾倒,头部和躯干明显下降,躯干从后仰发展到接近水平;摆臂与腿部调节未能恢复站立平衡。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "反馈控制", + "t": 0.8, + "log_step": 101, + "evidence_zh": "推断:log step 101 的生成脚本以机器人质心相对球心的误差、相对速度、板倾角和身体高度反馈调节踝、髋、膝与肩关节;log step 102 更新参数,log step 103 将该控制器产生的 102 步动作保存并重放,log step 111 核对交付参数一致。该动作在视频 0.05–0.65 s 已使身体明显后仰,视频 0.80–1.05 s 后仰继续扩大直至倾倒,说明这套反馈控制未稳定住身体与板的俯仰运动。评分器 deterministic=1,未见交付重放失配证据。", + "source_zh": "日志 / 视频 / 评分器字段" + } + ], + "gap_zh": "没有交付回合的状态日志或逐步奖励、关节限位与接触力记录;仅凭视频不能确定最先触发的原生终止谓词。agent 在 log step 112 自述踝关节饱和后板失稳,但该限位机制未在交付回合中独立核实,不据此标参数调试或具体限位根因。" + }, + "summary_zh": "机器人确实用双脚支撑于球上平衡板并调节身体,但累计回合评分仅为 1.1718730002009324e-09,未达到 800,节点完成 0/1。约视频 0.80 s 后后仰持续扩大,至视频 1.05 s 板明显翻斜、身体下坠;结合生成及重放日志,归因为反馈控制未能稳定身体和板的俯仰运动。缺少交付回合状态日志,无法核实踝关节饱和或最先触发的终止条件。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "标签与视频一致:机器人起始站在板上,视频约0.8 s后后仰加剧、板翻斜倒下,节点未达成,断点与平衡/反馈控制归因有日志支撑。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:47:45Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-privileged--r5t8nay/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-privileged--r5t8nay/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "f2a83b33df5428b1a248dc22eda7536f2beb4e68ca671f60f456304c48c2470e", + "video_sha256": "ff262e09b88a0d368141a9c324c909e2738b393c477d3876145bca062352b457", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-privileged--vhtndkz.json b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-privileged--vhtndkz.json new file mode 100644 index 0000000000000000000000000000000000000000..f73caca9f32b03d6803a2dc103b8bbeaa3e93ee3 --- /dev/null +++ b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-privileged--vhtndkz.json @@ -0,0 +1,179 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-balance-hard-i00-privileged--vhtndkz", + "taskKey": "humanoidbench/loco_balance_hard", + "dag_sha256": "419d18f9a6fdc6ba23a8ef8263e0e2a03d0e684ca7c5204c63fb59f163363f46", + "outcome": { + "success": 0, + "reward": 1.8271339750586608e-06, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 290, + "final_reward": 1.8271339750586608e-06, + "sim_steps": 290, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 1, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 21.43, + "peak_processes": 21, + "peak_disk_gb": 0.258, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0, + "over_memory_gb": 1 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-balance-hard-i00-privileged--vhtndkz/replay.mp4", + "duration_s": 3.0, + "fps": 20.0, + "frames": 60, + "sha256": "cd3cebeb212e7ab02dbf35895a177bf2fc4d7f793cb6322e7331cfccde17706e" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "board_balance_sustained": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.0 s 机器人双脚站在由深色球体支承的浅灰色板上;视频 0.05–0.20 s 机器人举起双臂,此后在板上维持明显倾斜的躯干姿态。初始站在板上仅表示开始尝试,不代表累计奖励达标。" + }, + { + "t": 2.65, + "log_step": null, + "text_zh": "观察:视频 2.50–2.80 s 板面逐渐大幅倾斜,机器人上身随之摆动;视频 2.85–2.95 s 机器人屈膝下沉、躯干向前倒至近水平,未恢复站立。成功参考视频 0.0、2.0、4.0、6.0、8.0、10.0 s 均保持站立高度、双脚在球上方的板面承重,仅用于辨认平衡状态。" + }, + { + "t": null, + "log_step": 295, + "text_zh": "观察:log step 288–289 的最终生成脚本录制并保存最多290个动作;log step 295 对最终交付文件的重放返回 sim_steps=290、deterministic=true、scored=false,评分器 reward 同样记录 success=0、sim_steps=290。" + }, + { + "t": null, + "log_step": 295, + "text_zh": "推断:DAG所述原生逐步奖励是取值不超过1的各奖励项的乘积,最终交付回合仅有290个控制步,累计回报至多290,无法达到800;因此整个视频中节点均未完成。此判断结合交付回合步数、奖励定义和视频末段失衡,不由零分单独推断。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–2.50 s 机器人双脚接触浅灰色平衡板并承重,板由下方深色球体支承;双腿和躯干调整站姿,双臂举起但未见手接触板或球。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "board_balance_sustained", + "t": 2.65, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频 2.50–2.80 s 机器人正在双脚承重、举臂维持平衡,板面开始明显倾斜,随后视频 2.85–2.95 s 躯干前倒、膝部塌低,至结束未恢复。失效发生于唯一的板上平衡步骤执行中。" + }, + "unmet": { + "board_balance_sustained": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 2.65, + "log_step": null, + "evidence_zh": "观察:视频 2.65–2.95 s 板面倾角增大,机器人失去稳定站姿,躯干从接近直立转为前倾至近水平,双膝弯曲、身体明显下沉。末帧仍处于塌落姿态,无平衡恢复。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "反馈控制", + "evidence_zh": "推断:log step 288 的最终生成脚本按板与躯干的姿态、速度及重心相对球体的偏移计算踝、膝、髋目标,随后录制动作供开环重放;log step 287、290、295 均验证290步交付轨迹可一致重放但 scored=false。视频 2.50–2.95 s 板面持续倾斜、机器人最终前倒塌低,说明生成这些动作的状态反馈控制未能稳定该支承系统;没有证据把失稳进一步归到某一个增益或执行器饱和。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "无状态日志,无法精确确认末段足底接触变化、执行器力或触发终止的具体碰撞/高度谓词。评分器只给出 final_reward=1.8271339750586608e-06,没有报告原生累计 episode_return,故该指标值留空;不将末步奖励或会话训练的自定义分数充作累计回报。控制根因由最终脚本和视频失稳过程推断,具体失稳机理尚不能确定。" + }, + "summary_zh": "机器人已在球支承的平衡板上实际尝试站稳,但交付回合仅290个控制步,未达到累计奖励800的验收条件,节点最终为0/1。视频约2.65 s起板面明显倾斜,随后机器人前倒、屈膝塌低,断点为板上持续站稳,归为全身平衡失败,控制证据支持反馈控制未能稳定。缺少状态日志及累计回报数值,不能精确判定终止接触或失稳参数。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "节点未达成、断点(板上平衡约视频2.65 s失稳前倒)、全身·平衡现象与控制·反馈控制根因均与视频和log一致;290步交付回合不可能达到累计800,推断有评分器步数支撑。未发现需修改的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:50:30Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-privileged--vhtndkz/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-privileged--vhtndkz/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "508e9b1f32556ecb94ca33b5efb73af1930c3286e7a8e0bff91a92222a2cb962", + "video_sha256": "cd3cebeb212e7ab02dbf35895a177bf2fc4d7f793cb6322e7331cfccde17706e", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-privileged--w3mutx9.json b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-privileged--w3mutx9.json new file mode 100644 index 0000000000000000000000000000000000000000..8f6c63c9da0cb2386c388d137ac98a3115d81a4e --- /dev/null +++ b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-privileged--w3mutx9.json @@ -0,0 +1,173 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-balance-hard-i00-privileged--w3mutx9", + "taskKey": "humanoidbench/loco_balance_hard", + "dag_sha256": "419d18f9a6fdc6ba23a8ef8263e0e2a03d0e684ca7c5204c63fb59f163363f46", + "outcome": { + "success": 0, + "reward": 1.145564995584152e-11, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 23, + "final_reward": 1.145564995584152e-11, + "sim_steps": 23, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.614, + "peak_processes": 10, + "peak_disk_gb": 0.951, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-balance-hard-i00-privileged--w3mutx9/replay.mp4", + "duration_s": 0.3, + "fps": 20.0, + "frames": 6, + "sha256": "676a49c4e5543dea051546d08eae8b400c01bb59ffb68134c1feb9740825d52b" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "board_balance_sustained": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:受评视频 0.00 s,机器人双脚踩在由深色球体支承的浅灰平衡板上;这只是初始支撑姿态,尚未完成累计奖励达到800的持续平衡事件。" + }, + { + "t": 0.1, + "log_step": null, + "text_zh": "观察:已检查受评视频全部6帧(视频 0.00、0.05、0.10、0.15、0.20、0.25 s);视频 0.05 s 一脚抬离板面,视频 0.10 s 双脚均离板,视频 0.15–0.25 s 身体翻转、下沉,未恢复双脚承重的站姿。" + }, + { + "t": 0.25, + "log_step": 197, + "text_zh": "观察:评分器 reward.sim_steps=23、reward.n_actions=23、reward.final_reward=1.145564995584152e-11、reward.success=0、reward.deterministic=1;log step 197 对最终交付文件重放两次,报告23步且deterministic=true。累计奖励未达到800,节点未完成;日志中的其他稳态测试不作为本受评节点的进度。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:受评视频 0.00 s 双脚在平衡板上承重;视频 0.05 s 仍有一脚踩板,另一脚抬起;视频 0.10 s 双脚均离开板面,此后身体翻倒,未恢复站立支撑。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": 1.145564995584152e-11, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "board_balance_sustained", + "t": 0.1, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:受评视频 0.00 s 机器人已在板上开始平衡;视频 0.05 s 一脚离板,视频 0.10 s 双脚脱离板面、身体倾斜,视频 0.15–0.25 s 翻倒下沉,持续站稳步骤失败且未恢复。评分器仅记录23个控制步,累计奖励远低于800。" + }, + "unmet": { + "board_balance_sustained": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 0.1, + "log_step": null, + "evidence_zh": "观察:受评视频 0.10 s 机器人双脚悬空、失去平衡板的足底支撑;视频 0.15–0.25 s 躯干进一步翻倒至近水平并下沉,直至受评视频末帧仍未恢复站立。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "反馈控制", + "evidence_zh": "推断:log step 173 的生成脚本用姿态和速度误差dx计算u=cr-((1-a)*K_init+a*K)@dx,跟踪从初始姿态到稳态的参考路径;log step 177 修改前馈与初态增益计算后生成23步轨迹,log step 197 再生成并验证最终23步交付可确定性重放。结合受评视频 0.05–0.25 s 足底脱离支撑、身体翻倒,说明生成动作的反馈调节未能稳定初始过渡过程;稳态求解本身没有提供从固定初态进入稳态的稳定控制。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "无状态日志,视频仅6帧,不能精确确定最先触发的是骨盆过低、板或机器人触地,还是球体违规接触;也不能恢复各关节力与误差。agent 在log step 199–200 自述交付回合以板端触地结束,但本标注只将可见的脱离支撑和翻倒作为现象;反馈失稳的更细机制未确认,未把会话测试中的参数问题直接套到最终回合。" + }, + "summary_zh": "机器人从双脚踩板的初态开始尝试平衡,但在受评视频 0.05–0.10 s 迅速离板,并于视频 0.15–0.25 s 翻倒下沉,板上持续站稳节点未完成(0/1)。交付的23步轨迹可确定性重放,累计奖励仅约1.15×10⁻¹¹,日志与画面支持初始过渡过程的反馈控制失稳;没有状态日志,无法确认最先触发的精确终止条件。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "节点未达成、断点在 board_balance_sustained、现象(全身·平衡)和根因(控制·反馈控制)均有视频和日志支持;时间都使用受评视频时钟,未发现需要修改的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:47:19Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-privileged--w3mutx9/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-privileged--w3mutx9/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "e58c8c9e96cd304acd026b8c2e58f790d489cc5dc529e56b3b26aa0f6ee549d2", + "video_sha256": "676a49c4e5543dea051546d08eae8b400c01bb59ffb68134c1feb9740825d52b", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard--l4egvtw.json b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard--l4egvtw.json new file mode 100644 index 0000000000000000000000000000000000000000..c0623e44e6115f079f785746b858e2f32a93da44 --- /dev/null +++ b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard--l4egvtw.json @@ -0,0 +1,187 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-balance-hard-i00-standard--l4egvtw", + "taskKey": "humanoidbench/loco_balance_hard", + "dag_sha256": "419d18f9a6fdc6ba23a8ef8263e0e2a03d0e684ca7c5204c63fb59f163363f46", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 62, + "sim_steps": 2705, + "calls": 884, + "rows": 2705, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1599.1, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-balance-hard-i00-standard--l4egvtw/replay.mp4", + "duration_s": 0.7, + "fps": 20.0, + "frames": 14, + "sha256": "414fbfc0369097c0b7bf7a1b9e96688fec30b289f66c095811de2c2f19a026be" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "board_balance_sustained": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 149, + "text_zh": "观察:视频 0.00 s 机器人双脚承重于浅灰色平衡板,深色球体支承板底;机器人已开始板上平衡,但初始站姿不代表累计奖励达标。log step 149 的服务记录显示最后一次重置开始 episode=51,并执行 --final-balance 控制器。" + }, + { + "t": 0.65, + "log_step": 150, + "text_zh": "观察:逐帧检查视频 0.00–0.65 s,机器人双脚随板运动,平衡板逐渐倾斜,视频 0.50–0.65 s 近镜头侧板端明显下沉,末帧机器人躯干仍接近直立。log step 150 的最终服务元数据为 episode=51、control_steps=62、done=true、reason=\"failed\";评分器 live_success=0。" + }, + { + "t": 0.65, + "log_step": 150, + "text_zh": "推断:本节点要求累计奖励至少800,而该最终回合在62个控制步即判负,结合每步奖励至多1,可知累计奖励不可能达到800,故整个视频均未达到节点。评分器的 sim_steps=2705 是全会话总步数,不能作为本回合持续平衡的证据;未将仿真 1.24 s 当成视频时间。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.65 s,机器人双脚踩在浅灰色平衡板上承重,腿和躯干随板倾斜调节姿态;没有可见手部接触。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "board_balance_sustained", + "t": 0.55, + "log_step": 150, + "where": "操作中", + "evidence_zh": "观察:视频 0.45–0.65 s,机器人仍以双脚在板上承重,板的倾角继续增大,近镜头侧板端降至地面附近,未恢复球上悬空平衡;log step 150 的服务记录确认最终回合在62个控制步结束,reason=\"failed\"。推断:失败发生在持续站稳的过程中,导航锚点取视频 0.55 s。" + }, + "unmet": { + "board_balance_sustained": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 0.55, + "log_step": 150, + "evidence_zh": "观察:视频 0.00–0.65 s,双脚下的平衡板由较平缓姿态逐渐转为明显倾斜,视频 0.55–0.65 s 近镜头侧板端下沉至地面附近,球体相对板中心偏向另一侧;机器人直到末帧仍站在板上,没有出现身体倒地。推断:机器人与板、球组成的承重系统失去原有平衡,未能保持板两端悬空;这是支撑失稳,不能描述为已观察到机器人摔倒。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "观察:log step 147 中 agent 自述要使用板面视觉估计并配合低增益俯仰、横滚修正;log step 149 实际运行更新后的 --final-balance 控制器,log step 150 显示提交正常返回 ok=true、n_clipped=0,而最终回合仅执行62个控制步即判负。视频 0.45–0.65 s 显示腿部姿态变化未阻止平衡板继续倾斜。推断:本次站立控制未能稳定板球支撑系统;证据不足以把原因进一步确定为某个增益、视觉估计偏差或特定反馈项。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "没有状态日志或接触对记录,服务仅报告 reason=\"failed\",因此末段板端触地是结合视频与提前终止的推断,无法确认精确接触时刻或排除其他原生终止谓词。评分器未报告 episode_return,原生指标留 null;视频也不能恢复执行器力和累计回报。日志中的最终控制代码与执行输出部分有截断,无法定位具体控制参数或反馈失效机制;此前调试回合仅作会话背景。" + }, + "summary_zh": "最终受评的是第51回合:机器人双脚已在板上承重并尝试调节姿态,但“板上持续站稳”累计达标节点未完成,最终为0/1。视频 0.45–0.65 s 平衡板明显倾斜、近镜头侧板端下沉,服务在62个控制步判负;根因保守归为控制未稳定板球支撑系统。无状态日志、精确接触谓词和累计回报数值,不能进一步确定具体反馈或参数问题。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "节点未达成(62个控制步即判负,累计不可能达800)、断点、平衡现象和保守的控制根因与0.7 s视频及log一致,未发现需改动之处。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:50:30Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-standard--l4egvtw/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-standard--l4egvtw/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "1df31939e79420da72289954ca63765c0cdb047a474b1394df481e95d7b417b5", + "video_sha256": "414fbfc0369097c0b7bf7a1b9e96688fec30b289f66c095811de2c2f19a026be", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard--lmn9nh4.json b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard--lmn9nh4.json new file mode 100644 index 0000000000000000000000000000000000000000..7ba355253c110fe481322e833f701169b6a8c131 --- /dev/null +++ b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard--lmn9nh4.json @@ -0,0 +1,181 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-balance-hard-i00-standard--lmn9nh4", + "taskKey": "humanoidbench/loco_balance_hard", + "dag_sha256": "419d18f9a6fdc6ba23a8ef8263e0e2a03d0e684ca7c5204c63fb59f163363f46", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 49, + "sim_steps": 2733, + "calls": 2757, + "rows": 2733, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 829.3, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.157, + "peak_processes": 10, + "peak_disk_gb": 0.24 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-balance-hard-i00-standard--lmn9nh4/replay.mp4", + "duration_s": 0.55, + "fps": 20.0, + "frames": 11, + "sha256": "a9cc361a193fd290976e9f3223abb9193dd392fcbc12b5ea09ef5be7d507a4ee" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "board_balance_sustained": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00–0.35 s,机器人双脚在浅灰色平衡板上承重,板下球体接地;机器人伸展膝关节并调整手臂,已开始平衡尝试。开局站姿不满足本节点累计奖励达到800的事件判据。" + }, + { + "t": 0.45, + "log_step": 47, + "text_zh": "观察:视频 0.40–0.50 s,平衡板倾斜加剧,机器人身体前倾,末帧一只脚已明显离开板面、原有双脚支撑失去。log step 47 执行最后的控制器后,基准服务 Robot.status(在线回合状态接口)返回 episode=51、control_steps=49、done=True、reason=failed、resets_left=0;评分器 live_success=0、replay_success=0、deterministic=1。" + }, + { + "t": 0.5, + "log_step": 47, + "text_zh": "推断:最后受评回合仅有49个控制步,按DAG所述各项不超过1的逐步乘积奖励,累计值不可能达到800,因此整个录像没有达标区间;此判断结合最终回合服务记录及视频中的支撑失稳,而非只由总评分为零推断。前面被重置的50个回合不计入本节点。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.40 s,机器人双脚直接接触并承重于平衡板,腿部伸展和躯干姿态随板面变化;视频 0.50 s 一只脚明显脱离板面。手臂摆动期间未见手部接触平衡板或球体。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "board_balance_sustained", + "t": 0.45, + "log_step": 47, + "where": "操作中", + "evidence_zh": "观察:视频 0.40–0.50 s,机器人仍在板上调整身体,但平衡板进一步倾斜,末帧脚部失去原有支撑,未恢复;log step 47 的最后回合由基准服务返回 control_steps=49、reason=failed。失败发生在板上持续平衡步骤中。" + }, + "unmet": { + "board_balance_sustained": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 0.45, + "log_step": 47, + "evidence_zh": "观察:已逐帧检查视频 0.00–0.50 s 的全部11帧;视频 0.40–0.50 s 板面倾斜加剧,机器人躯干前倾,末帧一只脚明显悬离板面,失去原有双脚支撑。录像结束前未见恢复,也未见身体完全倒地;这是支撑失稳。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "反馈控制", + "t": 0.45, + "log_step": 47, + "source_zh": "日志", + "evidence_zh": "推断:最后受评回合的姿态反馈未能稳住身体与支撑板。log step 43–47 的最终脚本用骨盆IMU、关节角及角速度构造踝、髋和肩的反馈;log step 47 最终回合打印的控制循环索引24至40中,骨盆俯仰估计 phi 从0.043增至0.179 rad,脚本合成的板倾角估计 beta 从−0.154变为−0.507 rad,左踝实测角从−0.425变为−0.881 rad,随后该回合失败;beta 是agent脚本的估计,非评分器提供的板角。视频 0.40–0.50 s 的持续倾斜和脚部失去支撑与这些记录一致,支持反馈失稳;不据此断言具体增益错误或力矩不足。" + } + ], + "gap_zh": "没有状态日志,评分器也未报告原生 episode_return,故该指标保持null,不能用catalog reward=0.0替代。基准服务仅返回failed,256×256录像末帧无法精确确认触发终止的接触对或骨盆高度;可以观察到支撑失稳,但不能证实力矩饱和、摩擦参数或某个反馈增益是更深层原因。日志控制循环索引没有换算成视频时间,视频锚点均来自直接逐帧检查。" + }, + "summary_zh": "机器人已用双脚在球支承的平衡板上承重并调整姿态,但最后受评回合仅执行49个控制步,板上持续站稳节点未达成,最终进度为0/1。视频约0.45 s 开始出现加剧的倾斜并在末帧失去原有双脚支撑,结合最终控制器日志,归因为反馈控制未能稳定姿态。缺少原生累计回报和接触状态日志,无法确定具体终止接触及更细的动力学原因。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "逐帧核对11帧:板在球上倾斜加剧,末帧一脚离板,与标注一致;最终回合仅49控制步,累计800不可能达成,节点未达、进度0/1、断点与平衡/反馈控制归因均有证据支持且标注为推断。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:50:07Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-standard--lmn9nh4/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-standard--lmn9nh4/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "d044e4da6e45344f4d6fe3427b6cd8388af83e9141fdd206a222c6ed6e5029d5", + "video_sha256": "a9cc361a193fd290976e9f3223abb9193dd392fcbc12b5ea09ef5be7d507a4ee", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard--ntgiprt.json b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard--ntgiprt.json new file mode 100644 index 0000000000000000000000000000000000000000..9d642636f736f62c6e88c09d37aa99b81e5bc244 --- /dev/null +++ b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard--ntgiprt.json @@ -0,0 +1,211 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-balance-hard-i00-standard--ntgiprt", + "taskKey": "humanoidbench/loco_balance_hard", + "dag_sha256": "419d18f9a6fdc6ba23a8ef8263e0e2a03d0e684ca7c5204c63fb59f163363f46", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 45, + "sim_steps": 1315, + "calls": 1423, + "rows": 1276, + "resets": 32, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3310.0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-balance-hard-i00-standard--ntgiprt/replay.mp4", + "duration_s": 0.55, + "fps": 20.0, + "frames": 11, + "sha256": "b5e9fa0512f4c200f80efd9dd2b6cdf199fe3d529c180639f26a7b0fe57c2a23" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "board_balance_sustained": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频0.00–0.35 s中,机器人双脚承重于球体支承的平衡板,膝、髋与身体姿态随控制动作变化;视频0.40–0.50 s中球体偏到板下侧边,板面明显翻斜,未恢复。已检查完整11帧,初始短暂站立不等于累计奖励达标。" + }, + { + "t": 0.5, + "log_step": 75, + "text_zh": "观察:log step 75运行最后一次重置后的ctrl2.py g17.json,输出“DONE step 44 failed”和“steps: 45”;inputs/run.json的评分器字段为graded_by_live=1、n_actions=45、live_success=0、infra_error=0,任务文字规定只评最后一回合。" + }, + { + "t": 0.5, + "log_step": 75, + "text_zh": "推断:最后回合仅执行45个控制步便失败,结合视频0.40–0.50 s的平衡板翻斜,不能满足本节点累计奖励至少800的验收;初始站姿不记为节点已完成。评分器未报告episode_return,不能把二值失败分数0当作累计回报0。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频0.00–0.50 s,机器人双脚踩在浅灰色平衡板上承重,腿和躯干随板面变化调姿;视频0.40–0.50 s板面翻斜时脚仍随板运动。未见手部接触板或球。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "board_balance_sustained", + "t": 0.4, + "log_step": 75, + "where": "操作中", + "evidence_zh": "观察:机器人正在板上调姿保持平衡;视频0.35 s板面尚接近水平,视频0.40 s球体已偏向板下侧边,视频0.45–0.50 s板面明显翻斜且未恢复。log step 75所运行的最后回合在控制循环索引44报告failed,共45步;此处视频锚点直接取自画面,未将agent工作时间或控制步索引换算成视频时间。" + }, + "unmet": { + "board_balance_sustained": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 0.4, + "log_step": 75, + "evidence_zh": "观察:视频0.40–0.50 s,机器人脚下平衡板由近水平向一侧翻斜,球体偏出板下中心,双腿随板明显改变姿态,原先的球上平衡支承未能保持。录像结束前未恢复;未将其描述为机器人已经全身倒地。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "反馈控制", + "t": 0.4, + "log_step": 69, + "evidence_zh": "观察:log step 32建立的ctrl2.py把reset返回保存为res,每次submit结果保存为res_last;log step 69改为逐步视觉更新时却读取vision.detect(res.last[\"room_camera\"]),log step 70–75未修正这一引用,log step 75最终运行该控制器并输出球位置估计约[-0.005,-0.013]。视频0.40–0.50 s可见球体偏离板下中心、板面翻斜。推断:读取重置图像会使球位置反馈不能响应实际滚动,支持反馈控制缺陷可能参与此次失衡,但不能单独解释失败。观察:在log step 69引入该引用前,log step 68的配置仍通过get_obs获取新图像,末尾球估计为[0.076,0.214],也在控制循环索引43报告failed;log step 75在索引44失败。推断:这些回合配置并不相同,且没有最终配置仅修复图像引用的对照,不能认定过期视觉是必要或主导原因;新图像配置的估计准确性也未经状态真值核验。不能据agent在log step 76自述的“物理不可控”断定所有合法控制方案均不可行。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "没有状态日志,评分器与服务输出未报告原生episode_return,故该指标保留null;只能确认最后回合失败、未达800,不能恢复精确累计回报。末段画面可见板面翻斜,但没有接触对、骨盆高度或具体终止谓词,无法确定是哪一条原生终止检查首先触发。过期图像引用是日志代码支持的反馈缺陷,但使用新图像的log step 68配置也在控制循环索引43失败;缺少最终配置仅修复该引用的对照,无法确定其独立因果贡献。" + }, + "summary_zh": "受评的是log step 75启动的最后回合:机器人双脚在球上平衡板承重并尝试调姿,但仅执行45个控制步即失败,“板上持续站稳”未完成,节点最终0/1,视频0.40–0.50 s可见球体偏出板下中心、板面翻斜。日志代码支持过期图像反馈这一控制缺陷可能参与失衡,但此前使用新图像的log step 68配置也在控制循环索引43失败,故不能单独用过期视觉解释失败;因关键因果贡献仍属推断,conf由8降为6。没有状态日志或原生累计回报记录,也无法精确恢复奖励和最先触发的终止谓词。", + "conf": 6, + "review": { + "status": "agreed", + "rounds": 2, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "fix", + "conf": 6, + "summary_zh": "节点、断点、平衡现象与时间锚点均与视频一致;但根因“过期视觉反馈”的因果归属证据偏弱,conf 8 偏高。", + "n_issues": 1, + "n_major": 0 + }, + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 2, + "verdict": "agree", + "conf": 7, + "summary_zh": "节点未达成、断点、平衡现象、时间锚点均与视频一致;根因已按上轮意见降级并写明反向证据,无新增问题。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:20:53Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-standard--ntgiprt/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-standard--ntgiprt/04_review_r1" + }, + { + "stage": "reconcile", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-standard--ntgiprt/05_reconcile_r1" + }, + { + "stage": "traj_review", + "round": 2, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-standard--ntgiprt/04_review_r2" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "2de5e94fcf1c2a37dc3e115d02f7fae24bdee7a68dc14da041329fe5387a6c74", + "video_sha256": "b5e9fa0512f4c200f80efd9dd2b6cdf199fe3d529c180639f26a7b0fe57c2a23", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard--phmae2f.json b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard--phmae2f.json new file mode 100644 index 0000000000000000000000000000000000000000..f141bc6f7eaafb6185d13a59057294945d0300f0 --- /dev/null +++ b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard--phmae2f.json @@ -0,0 +1,177 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-balance-hard-i00-standard--phmae2f", + "taskKey": "humanoidbench/loco_balance_hard", + "dag_sha256": "419d18f9a6fdc6ba23a8ef8263e0e2a03d0e684ca7c5204c63fb59f163363f46", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 0, + "sim_steps": 1976, + "calls": 2017, + "rows": 1976, + "resets": 39, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3569.2, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-balance-hard-i00-standard--phmae2f/replay.mp4", + "duration_s": 0.1, + "fps": 20.0, + "frames": 2, + "sha256": "45422a1aef61814f07743c88058dbd4d79c0de1fcfa79dd4dfff141b84b5c56e" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "board_balance_sustained": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:已检查本次视频全部两帧(视频 0.00 s、0.05 s);机器人双脚位于浅灰色板面,板下球体支承板面,躯干直立,两帧姿态相同,未见主动调节、板端触地或跌倒。这里只呈现初始站姿,没有持续平衡过程;初始支撑未经过操作,故不填 contact_actual。" + }, + { + "t": 0.05, + "log_step": 134, + "text_zh": "观察:评分器记录 n_actions=0、success=0、live_success=0、replay_success=0、deterministic=1,视频全长仅0.1 s;log step 134 的最后发言仍称此前最佳试验仅维持1.6秒且均未通过。推断:受评末回合停留于初态,未开始持续平衡,累计800奖励条件未达成;这不是把开局站直当作节点完成,也不是仅凭零分判断未尝试。" + }, + { + "t": null, + "log_step": 133, + "text_zh": "观察:log step 133 回报四个会话试验分别在80、48、66、53控制步以 END failed 结束;run.json 全部 steps[].sim.submit 合计1976、reset 合计39,恰与评分器 sim_steps=1976、resets=39 相符。推断:这些是多回合会话累计数据,不能转换成受评视频的平衡时长或受评跌倒事件。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "board_balance_sustained", + "t": 0.0, + "log_step": 134, + "where": "首步前", + "evidence_zh": "观察:本次视频 0.00–0.05 s 两帧完全停留在机器人板上初始站姿,评分器 n_actions=0;log step 134 最后发言仍在描述会话调试计划,没有已完成的平衡回合。推断:最终受评回合没有启动持续平衡控制,断在板上持续站稳这一步开始之前;此前重置掉的试验不属于这段受评轨迹。" + }, + "unmet": { + "board_balance_sustained": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 134, + "evidence_zh": "观察:受评视频 0.00 s 与0.05 s 中机器人、板和球的位置相同,没有运动或倒下,评分器 n_actions=0;全段只是初态。推断:要求累计800奖励的持续平衡步骤在受评回合尚未启动,因此记没尝试;不把会话试验中的跌倒标为本段平衡失败。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志", + "evidence_zh": "推断:会话中始终未得到可持续站稳的控制器,是最终没有执行有效平衡回合的有证据支持的能力缺口;仅归为控制,不确定具体子类。观察:agent 脚本在 log step 129、130、131 的三个实际服务试验分别执行79、45、38控制步后均回报 END failed;log step 132–133 对同一控制器作四个小修正,结果仍分别在80、48、66、53控制步失败;log step 134 的 agent 自述也明确说此前无一次通过。这些只用于解释会话中的控制困难,不作为受评视频发生跌倒的证据。" + } + ], + "gap_zh": "观察:inputs/ 无状态日志,评分器未给出 episode_return,故原生累计奖励值保留 null;本次完整视频虽有但仅两帧,不能分析主动反馈或力学失稳过程。log step 134 的 sim 汇总另含214次提交、6次重置,但 calls/out 没有这些末尾操作的明细;最终为何留在无动作回合的具体操作顺序不可复原,不能据此追加交付或规划根因。控制根因来自此前会话试验的反复失败,为推断,无法确定参数、反馈或接触设计中的具体缺陷。" + }, + "summary_zh": "受评视频 0.00–0.05 s 仅显示机器人在板上的初始站姿,最终回合 n_actions=0,“板上持续站稳”未启动也未达到累计800奖励,节点最终0/1。断点为首步前,现象为没尝试;会话控制器多次短程试验均失败,支持控制能力不足的推断,但此前跌倒不属于本次受评轨迹。无状态日志和累计奖励数值,且末尾六次重置仅有汇总,无法还原最终停在初态的具体操作顺序。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "受评视频仅两帧(0.00/0.05 s)且 n_actions=0,节点未达成、首步前/没尝试,与 FAIL-22 同型的写法一致;控制根因标为会话证据推断,未当作受评跌倒。未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:50:40Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-standard--phmae2f/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-standard--phmae2f/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "3aaac5b6a5e5a8568fe9b9c4fa62af79479138cb203ce4ea3da9de736d0d0014", + "video_sha256": "45422a1aef61814f07743c88058dbd4d79c0de1fcfa79dd4dfff141b84b5c56e", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard--uz4vhcl.json b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard--uz4vhcl.json new file mode 100644 index 0000000000000000000000000000000000000000..f4685a9e5d769439a0d30759db44a6ea5e147517 --- /dev/null +++ b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard--uz4vhcl.json @@ -0,0 +1,189 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-balance-hard-i00-standard--uz4vhcl", + "taskKey": "humanoidbench/loco_balance_hard", + "dag_sha256": "419d18f9a6fdc6ba23a8ef8263e0e2a03d0e684ca7c5204c63fb59f163363f46", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 50, + "sim_steps": 1816, + "calls": 1946, + "rows": 1816, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 2308.8, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.319, + "peak_processes": 12, + "peak_disk_gb": 0.998, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-balance-hard-i00-standard--uz4vhcl/replay.mp4", + "duration_s": 0.6, + "fps": 20.0, + "frames": 12, + "sha256": "bc575bf5d14e6302c4f0e99b61f3aca8644f969c9b1dafc5e27b3552878e4175" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "board_balance_sustained": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00 s,黑色人形机器人双脚踩在浅灰色平衡板上,板下球体承重,机器人已开始板上平衡尝试;初始站姿本身不是累计奖励达标事件。" + }, + { + "t": 0.5, + "log_step": null, + "text_zh": "观察:逐帧检查视频 0.00–0.55 s;视频 0.30–0.45 s 机器人身体向画面右侧倾斜,板面明显翻转;视频 0.50–0.55 s 平衡板已滑到一旁,机器人屈膝下沉、脚部落到裸露球体处,没有恢复板上站立。" + }, + { + "t": null, + "log_step": 130, + "text_zh": "观察:log step 130 中 agent 的最终控制脚本运行后,机器人服务 status 返回 episode=51、control_steps=50、done=True、reason=failed、resets_left=0;inputs/run.json 的评分器字段 live_success=0、replay_success=0、deterministic=1。此前重置掉的调参回合不属于本次受评轨迹。" + }, + { + "t": null, + "log_step": 130, + "text_zh": "推断:DAG 的逐步奖励为不超过1的各质量项之积,log step 130 确认最终回合仅执行50个控制步,累计不可能达到800;结合视频中未修复的失衡,板上持续站稳事件从未完成。评分器未报告原生累计奖励的具体数值。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.40 s,机器人双脚直接在平衡板上承重,腿部姿态随板面倾斜而改变;视频 0.50–0.55 s 板已从脚下滑离,脚部落到裸露球体处。手臂未见接触支承物。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "board_balance_sustained", + "t": 0.35, + "log_step": 130, + "where": "操作中", + "evidence_zh": "观察:视频约 0.35 s,机器人在板上平衡过程中身体明显倾斜,随后视频 0.40–0.50 s 平衡板翻转并从脚下滑离,机器人屈膝下沉,直至视频 0.55 s 未恢复板上支撑;log step 130 的最终回合服务状态为 failed。" + }, + "unmet": { + "board_balance_sustained": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 0.35, + "log_step": 130, + "evidence_zh": "观察:视频 0.30–0.55 s,机器人身体倾斜加大,平衡板翻转、滑离,双脚失去原有板面支撑,身体下沉;最后可见姿态未恢复。这里可确认失去平衡,录像没有显示身体完全倒地。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "反馈控制", + "evidence_zh": "推断:log step 124、127、129 记录最终脚本使用关节角和骨盆IMU反馈,通过踝、膝、髋目标调节板倾角、身体俯仰和相对重心,并在重置后运行;log step 130 确认该脚本执行了最终回合。视频 0.20–0.50 s 仍出现持续扩大的身体偏斜和板面翻转,直至失去板上支撑,支持该平衡反馈控制未能稳定支撑系统;不能进一步确定是哪项增益、符号或状态估计误差导致。", + "source_zh": "日志、视频" + } + ], + "gap_zh": "inputs/ 中没有状态日志,评分器也未给出 episode_return 数值,故原生指标留空;没有终止谓词或碰撞对记录,不能区分板触地、脚碰球或其他条件中哪个首先触发终止。已检查全部12帧,视频末帧为 0.55 s;时间锚点直接取录像,未把 log step 130 的控制步或 agent 工作时间直接换算为视频时间。具体控制参数或状态估计误差的因果作用缺少独立验证。" + }, + "summary_zh": "最后受评回合从双脚在板上承重开始,已尝试板上平衡,但未完成累计奖励达到800的唯一节点,最终为0/1。视频约 0.35 s 起机器人明显失衡,至视频 0.50 s 平衡板滑离脚下;结合最终控制脚本,根因支持反馈控制未能稳定身体和支撑板。无状态日志及原生累计回报数值,不能确定最先触发终止的接触条件,也不能进一步锁定具体控制参数错误。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "标签与视频和日志一致:最终回合50个控制步,视频约0.3–0.5 s板翻转滑离,节点未达成,反馈控制根因有日志与视频支持,未发现需修改之处。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:59:32Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-standard--uz4vhcl/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-standard--uz4vhcl/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "6a05bdd150c085db91bc022154e0e3e808176bd693459f312aab2ab566d9b654", + "video_sha256": "bc575bf5d14e6302c4f0e99b61f3aca8644f969c9b1dafc5e27b3552878e4175", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard--vbqjngx.json b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard--vbqjngx.json new file mode 100644 index 0000000000000000000000000000000000000000..76a9878dfa7fba799a2730dcf6b92bf756fb15d0 --- /dev/null +++ b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard--vbqjngx.json @@ -0,0 +1,173 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-balance-hard-i00-standard--vbqjngx", + "taskKey": "humanoidbench/loco_balance_hard", + "dag_sha256": "419d18f9a6fdc6ba23a8ef8263e0e2a03d0e684ca7c5204c63fb59f163363f46", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 57, + "sim_steps": 2010, + "calls": 1757, + "rows": 2010, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1138.3, + "video_rendered": 1 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-balance-hard-i00-standard--vbqjngx/replay.mp4", + "duration_s": 0.65, + "fps": 20.0, + "frames": 13, + "sha256": "40d3adb761abf6471441b7d7c5f16db10d08bbf3349ecf01715a05e3d35b0296" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "board_balance_sustained": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00 s 机器人双脚站在浅灰色平衡板上,板下深色球体支承,机器人双手未接触板或外物;这只是初始站姿,并非已完成累计奖励验收。" + }, + { + "t": 0.4, + "log_step": 222, + "text_zh": "观察:逐帧检查视频 0.00–0.60 s 的全部13帧,机器人逐渐后仰;视频 0.40–0.60 s 后倾加剧、头部下降,末帧身体接近横倒且板面大幅倾斜,未恢复站稳。log step 222 记录最后一次 episode 51 在 control_steps=57 时 done=True、reason=failed;评分器字段 live_success=0、replay_success=0。" + }, + { + "t": 0.6, + "log_step": 222, + "text_zh": "推断:结合该回合仅执行57个控制步即失败、视频可见的失衡及评分器未通过,累计奖励800的板上持续站稳节点从未达标;不能把开局短暂直立记为完成。日志中的2010总控制步跨越多次重置,不属于本受评回合。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.50 s 可见机器人双脚接触平衡板并承重,双腿和躯干随控制动作调整;视频 0.55–0.60 s 身体后倒、脚板支撑随平衡板倾覆而失稳。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "board_balance_sustained", + "t": 0.4, + "log_step": 222, + "where": "操作中", + "evidence_zh": "观察:视频 0.35–0.40 s 机器人在板上调节姿态时出现持续增大的后仰,视频 0.50–0.60 s 躯干进一步横倒、平衡板急剧倾斜,直至录像结束均未恢复;log step 222 的最终 episode 51 在57控制步结束,服务报告 reason=failed。断点定位于正在执行的板上持续站稳步骤。" + }, + "unmet": { + "board_balance_sustained": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 0.4, + "log_step": 222, + "evidence_zh": "观察:视频 0.40–0.60 s,机器人由后仰站姿持续倒向身后,头部和躯干下降;视频 0.60 s 身体已接近横向,平衡板随之大幅翻斜,未能恢复直立支撑。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "反馈控制", + "log_step": 222, + "source_zh": "日志", + "evidence_zh": "推断:log step 219 的最终控制脚本按骨盆IMU重力方向、角速度和关节状态逐步修正髋踝目标,目标是板上直立平衡;log step 222 的该脚本执行记录显示,从control step 7起髋修正hp已到+0.180 rad限幅,control step 50和57的踝修正ak也为+0.200 rad限幅,而脚本读取的重力方向gx由control step 15的-0.2205继续增至control step 50的-0.7730、control step 57的-0.9513。结合视频 0.40–0.60 s 的持续后倒,反馈调节未能抑制姿态发散,归为反馈控制;未将脚本估计的板角作为真实板角,也未据此断言某个增益或反馈符号错误。" + } + ], + "gap_zh": "观察:inputs/没有状态日志;run.json评分器字段与log step 222的最终服务状态均未报告原生episode_return,因此该指标留null。视频可确认失衡,但无法精确区分终止是板触地、骨盆过低还是其他接触谓词;日志未给出与每一视频帧的精确对应,故所有视频锚点直接取自已检查帧,未换算agent时间或控制步。具体反馈增益、模型估计或执行器跟踪各自的影响缺少进一步测量。" + }, + "summary_zh": "观察:受评的是log step 222执行的最后一次episode 51,机器人双脚承重于球上平衡板并进行了姿态调节,但视频 0.40–0.60 s 持续后倒,板上持续站稳节点未完成,最终进度0/1。推断:最终控制日志中的髋踝修正限幅伴随后倾继续增大,支持反馈控制未能稳定姿态,失败现象为全身失衡。观察:无状态日志且原生累计奖励未报告,无法精确确定终止接触谓词或具体控制参数问题。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "标签与视频、日志一致:单节点板上持续站稳未达成(最终回合57控制步 reason=failed,视频 0.40–0.60 s 后倒、板翻),断点、平衡现象与反馈控制根因均有证据支持,时间均用视频时钟。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:53:16Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-standard--vbqjngx/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-standard--vbqjngx/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "bd76da01a955fc1f30ad5d265b88d646832f265272d4957158c1dd7670d18e93", + "video_sha256": "40d3adb761abf6471441b7d7c5f16db10d08bbf3349ecf01715a05e3d35b0296", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard--ybdshzy.json b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard--ybdshzy.json new file mode 100644 index 0000000000000000000000000000000000000000..042f70ff21bba48c2ff9b8f0e7cfd5c5901c1809 --- /dev/null +++ b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard--ybdshzy.json @@ -0,0 +1,186 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-balance-hard-i00-standard--ybdshzy", + "taskKey": "humanoidbench/loco_balance_hard", + "dag_sha256": "419d18f9a6fdc6ba23a8ef8263e0e2a03d0e684ca7c5204c63fb59f163363f46", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 69, + "sim_steps": 452, + "calls": 436, + "rows": 452, + "resets": 7, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 2878.9, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-balance-hard-i00-standard--ybdshzy/replay.mp4", + "duration_s": 0.75, + "fps": 20.0, + "frames": 15, + "sha256": "54d3e44c5d0fe068a0159561c21fc8207ce8fcd78eb5c7dbb0328a3836ca7b47" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "board_balance_sustained": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00–0.40 s,机器人双脚承重于浅灰色平衡板,板下深色球体接地,机器人腿部与板面有小幅运动;初始站立本身不满足累计奖励达到800的事件判据。" + }, + { + "t": 0.6, + "log_step": null, + "text_zh": "观察:逐帧检查视频 0.45、0.50、0.55、0.60、0.65、0.70 s,机器人躯干逐渐向画面右侧倾倒,头和骨盆下降,平衡板随之倾斜,直到末帧没有恢复站稳。" + }, + { + "t": 0.7, + "log_step": 53, + "text_zh": "推断:log step 5 的服务说明规定只评最后回合;log step 52 的脚本在先前回合结束后重置并执行反馈控制,log step 53 为最后一次机器人动作,执行69个控制步后服务返回 done=True、reason=failed,后续 log step 54–70 均为图像查看或离线分析。该回合的倾倒趋势与视频一致,结合评分器 live_success=0、replay_success=0,判定累计达标事件从未发生;reward.sim_steps=452是跨回合总量,不作为视频时钟或本回合进度。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.70 s,机器人双脚接触平衡板并通过双腿承重,板由下方球体支承;末段机器人随板倾斜而倾倒,未见双手接触板或球。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "board_balance_sustained", + "t": 0.6, + "log_step": 53, + "where": "操作中", + "evidence_zh": "观察:视频 0.55–0.70 s,机器人仍以双脚接触板面,但身体持续倾倒、头部和骨盆明显下降,没有恢复直立;log step 53 的最后回合服务返回 reason=failed。失败发生在板上保持平衡过程中。" + }, + "unmet": { + "board_balance_sustained": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 0.6, + "log_step": 53, + "evidence_zh": "观察:视频 0.50–0.70 s,机器人从较直立姿态加速向画面右侧倾倒,双腿斜伸,躯干和头部下降,板面同时倾斜;末帧仍在倾倒,未出现恢复。此处可见失衡过程,未声称身体已落地。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "反馈控制", + "t": 0.6, + "log_step": 52, + "evidence_zh": "推断:log step 52 的控制脚本使用本体姿态、角速度和关节角构造板倾角反馈、踝关节积分补偿及髋关节姿态修正,每次提交一个控制步。log step 53 中脚本打印的骨盆俯仰估计从该回合控制步30附近的 +0.0140 rad 发展为控制步60附近的 -0.4366 rad,最终打印 -0.8601 rad,服务随后报失败;视频 0.50–0.70 s 独立显示相符的持续倾倒。控制器虽持续反馈修正,仍未抑制姿态偏差增长,支持反馈稳定性不足;这些日志控制步未换算为视频时间。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "无状态日志,评分器未报告 episode_return,故原生累计回报保留null;不能从二值失败分数恢复精确回报。视频末帧止于倾倒过程中,日志只给出笼统的failed,不能确定骨盆高度、板端触地或球体接触中哪一项触发终止;也不足以锁定反馈控制器的具体增益、模型误差或接触力问题。" + }, + "summary_zh": "受评最后回合中,机器人双脚在球支承的平衡板上尝试维持站立,但视频 0.55–0.70 s 持续倾倒,“板上持续站稳”未完成,最终进度0/1。log step 52–53 的反馈控制代码及该回合姿态偏差增长记录支持反馈控制未能稳定身体;先前7次回合不计入本次节点进度。缺少状态日志和原生累计回报,无法确定精确回报与终止触发条件。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "节点未达成、断点在板上平衡、全身·平衡现象和控制·反馈控制根因均有视频帧和 log step 53 的数值支持;时钟未换算,未发现需要修改的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:53:24Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-standard--ybdshzy/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-standard--ybdshzy/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "6dd404bdbe4a217c037bc5413878eec458686bbf735fcee92e99acbde8f3b524", + "video_sha256": "54d3e44c5d0fe068a0159561c21fc8207ce8fcd78eb5c7dbb0328a3836ca7b47", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard-realtime--8kdzhfv.json b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard-realtime--8kdzhfv.json new file mode 100644 index 0000000000000000000000000000000000000000..e95e7f00b4965e700a5eab37908b28a138e2abac --- /dev/null +++ b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard-realtime--8kdzhfv.json @@ -0,0 +1,196 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-balance-hard-i00-standard-realtime--8kdzhfv", + "taskKey": "humanoidbench/loco_balance_hard", + "dag_sha256": "419d18f9a6fdc6ba23a8ef8263e0e2a03d0e684ca7c5204c63fb59f163363f46", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 58, + "sim_steps": 2629, + "calls": 13818, + "rows": 14725, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1170.1, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 51.683, + "max_lag_s": 0.0028, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 0, + "late_steps": 0, + "requests_refused": 49, + "request_busy_s": 4.662, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.181, + "peak_processes": 10, + "peak_disk_gb": 0.243, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-balance-hard-i00-standard-realtime--8kdzhfv/replay.mp4", + "duration_s": 0.65, + "fps": 20.0, + "frames": 13, + "sha256": "6d84ca097ffe75d9fec9addbfeb7ee4d95b952b3507660ca7dae9bb687027679" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "board_balance_sustained": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频0.00–0.60 s全部13帧中,机器人双脚踩在浅灰平衡板上,腿部屈伸、踝部调整;板由深色球体支承。视频0.40–0.60 s板面明显倾斜,球体相对板向画面左侧滚移,末帧板的右下端接近地面,未形成持续稳定的板球支承状态。" + }, + { + "t": 0.6, + "log_step": 84, + "text_zh": "观察:log step 84启动最后的受评回合,服务状态给出episode=51、control_steps=58、done=True、reason=failed、resets_left=0;评分器live_success=0、replay_success=0、deterministic=1。此前重置掉的调参回合不是本视频的进度。" + }, + { + "t": 0.6, + "log_step": 84, + "text_zh": "推断:依据任务DAG的逐步奖励乘积公式,每步最高为1,受评回合仅58个控制步,累计回报不可能达到800,因此该事件节点从未达成;开局站在板上不等于完成。成功参考视频0–9 s的逐秒画面持续显示板球支承及站立姿态,仅用于比较可见形态。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频0.00–0.60 s机器人双脚踏在平衡板上承重,膝踝和躯干随板运动调节;双手悬空,未见手与板或球接触。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "board_balance_sustained", + "t": 0.45, + "log_step": 84, + "where": "操作中", + "evidence_zh": "观察:视频0.40–0.60 s机器人仍在板上执行平衡调节,但板球支承关系持续偏移,板一端抬高、另一端下落,未恢复稳定;log step 84中服务确认最后回合在control_steps=58时以failed终止。断点为正在执行的板上持续站稳步骤。" + }, + "unmet": { + "board_balance_sustained": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 0.45, + "log_step": 84, + "evidence_zh": "观察:视频0.40、0.45、0.50、0.55、0.60 s中,机器人脚下的平衡板逐渐倾转,深色球体滚向板的一侧,板的低端下落至接近地面,支承系统失稳;机器人躯干尚未完全倒地。log step 84的服务终止状态为failed。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "反馈控制", + "log_step": 85, + "evidence_zh": "推断:log steps 80–82记录最终控制器保留关节姿态PI和踝部板倾角PD,关闭踝部积分,髋部及侧倾反馈增益为0;log step 84确认该控制器执行最后回合。log step 85读取该回合轨迹,脚本计算的板倾角估计thb由control step 5的−0.098变化到control step 46的−0.547,踝关节角同时由−0.583变化到−0.870;视频0.40–0.60 s可见板球支承偏移及板端下落。结合控制律和执行结果,反馈控制未能抑制支承系统的倾转与漂移;这些估计不当作真实板角或特定增益错误的证明。", + "source_zh": "日志" + } + ], + "gap_zh": "没有模拟器状态日志;视频仅有单一斜侧视角且末帧仍近乎站立,不能精确区分终止由板端触地、球体碰机器人或其他原生终止谓词触发。run.json未报告episode_return,agent在log step 86自述的“约58分”不是原生评分值,故指标保留null。log step 85的thb是控制器由本体感觉计算的估计,缺少独立真值,不能据此断定反馈参数、稳态误差或积分设计中的具体错误;视频时间均直接取帧,未将agent时间或控制步换算为视频时间。" + }, + "summary_zh": "受评的是最后的第51回合:机器人双脚在平衡板上承重并主动调节,但板球支承系统在视频约0.40–0.60 s明显失稳,“板上持续站稳”完成0/1。服务在log step 84报告该回合仅58个控制步即failed,结合最终控制律与轨迹,根因归为反馈控制未能稳定支承系统。没有状态日志或原生累计回报,具体终止接触及更细的控制缺陷无法确认。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "单节点任务,节点未达成、断点、平衡现象和反馈控制根因与视频及log step 84/85一致;时间均取视频帧,未换算agent时间,推断已标注。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:10:46Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-standard-realtime--8kdzhfv/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-standard-realtime--8kdzhfv/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "6d72ef4b2e97b788129e34e9876bb9f0ad5affc253c13557f94a3b7b1ba1c5a9", + "video_sha256": "6d84ca097ffe75d9fec9addbfeb7ee4d95b952b3507660ca7dae9bb687027679", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard-realtime--bgbx8gb.json b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard-realtime--bgbx8gb.json new file mode 100644 index 0000000000000000000000000000000000000000..5dc8c255a7f289bfa72c5db2dc0dac411aab8a86 --- /dev/null +++ b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard-realtime--bgbx8gb.json @@ -0,0 +1,200 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-balance-hard-i00-standard-realtime--bgbx8gb", + "taskKey": "humanoidbench/loco_balance_hard", + "dag_sha256": "419d18f9a6fdc6ba23a8ef8263e0e2a03d0e684ca7c5204c63fb59f163363f46", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 75, + "sim_steps": 2989, + "calls": 3080, + "rows": 2977, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 2359.2, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 58.9, + "max_lag_s": 0.0047, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 323, + "late_steps": 0, + "requests_refused": 50, + "request_busy_s": 2.445, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.33, + "peak_processes": 19, + "peak_disk_gb": 0.31, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-balance-hard-i00-standard-realtime--bgbx8gb/replay.mp4", + "duration_s": 0.85, + "fps": 20.0, + "frames": 17, + "sha256": "0e1c75463ddb5291480399180baae1c9fe49f271f5830f490e36c884e4112d43" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "board_balance_sustained": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00–0.10 s,机器人双脚踩在球体支承的浅灰色平衡板上,躯干接近竖直;这只是初始站姿,不代表累计奖励已达到800。" + }, + { + "t": 0.6, + "log_step": null, + "text_zh": "观察:已逐帧检查视频 0.00–0.80 s。视频 0.15–0.35 s 机器人后仰,视频 0.45–0.55 s 转为明显前俯;视频 0.60–0.75 s 平衡板迅速倾斜,机器人再次后仰,末帧躯干接近水平,没有恢复板上站稳。" + }, + { + "t": null, + "log_step": 104, + "text_zh": "观察:log step 104 启动最终默认控制脚本后,机器人服务报告 episode=51、control_steps=75、done=True、reason=failed、resets_left=0;评分记录 success=0、live_success=0、replay_success=0。该回合是受评的最后一回合,不将此前重置掉的测试并入视频进度。" + }, + { + "t": null, + "log_step": 104, + "text_zh": "推断:受评回合仅执行75个控制步,且视频显示失衡终止;结合本任务每步奖励至多1、累计至少800的判据,该事件节点始终未达成。评分字段没有报告原生累计奖励的具体数值。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.55 s,机器人左右脚直接踩压平衡板承重,双腿与躯干随板前后摆动;视频 0.60–0.75 s 板面倾斜并失去稳定支承。双手未见接触板或球。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "board_balance_sustained", + "t": 0.6, + "log_step": 104, + "where": "操作中", + "evidence_zh": "观察:视频 0.60 s 起,机器人正在板上调节站姿,平衡板明显倾转;视频 0.65–0.75 s 机器人向后倾倒,最终躯干接近水平且未恢复。log step 104 的最终回合服务状态为 episode=51、75个控制步后 failed。" + }, + "unmet": { + "board_balance_sustained": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 0.6, + "log_step": 104, + "evidence_zh": "观察:视频 0.60–0.75 s,机器人脚下的板由小幅摆动变为大幅倾斜,机器人由屈膝站立转为明显后仰,末帧躯干接近水平,无法维持球支承板上的站立平衡;log step 104 记录该最终回合以 failed 结束。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "反馈控制", + "evidence_zh": "推断:log step 39 的控制脚本根据IMU俯仰、角速度及腿关节状态计算线性反馈,log step 58 加入偏置及积分选项,log step 102 将默认策略设为髋、踝反馈,log step 104 明确运行该默认脚本。视频 0.15–0.55 s 显示机器人在后仰与前俯之间摆动,随后在视频 0.60–0.75 s 板体倾转、机器人后仰失去支撑;说明这次闭环姿态调节没有稳定住机器人与平衡板。现有证据不足以确定某个具体增益、估计误差或力矩限制造成失稳。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "没有状态日志;视频可见失衡,但不能精确确定终止由板端触地、机器人触地或骨盆过低中的哪一原生谓词触发,也不能恢复执行器力与累计奖励数值。最终回合只有服务终止状态,缺少可读取的逐控制步本体反馈数据,故不进一步归因于特定增益或模型参数。" + }, + "summary_zh": "受评的是最后一回合(log step 104,episode 51),机器人双脚在球支承的平衡板上承重并尝试反馈调姿,但“板上持续站稳”累计奖励事件未达成,最终节点为0/1。视频约 0.60–0.75 s 平衡板大幅倾转、机器人后仰失去支撑,结合最终脚本记录,归因为反馈控制未能稳定姿态。没有状态日志或原生累计奖励数值,无法确定精确的终止接触与失稳参数。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "节点未达成、断点、平衡现象与反馈控制根因均与视频一致(视频 0.45–0.55 s 前俯,0.60–0.85 s 板倾转、机器人后仰倒下);未发现需要修改的内容。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:09:10Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-standard-realtime--bgbx8gb/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-standard-realtime--bgbx8gb/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "d407a954fd99c2cc338a3f3252f78e9fe3c067782b32c413c51b3dc46aaf85ed", + "video_sha256": "0e1c75463ddb5291480399180baae1c9fe49f271f5830f490e36c884e4112d43", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard-realtime--hrwms4v.json b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard-realtime--hrwms4v.json new file mode 100644 index 0000000000000000000000000000000000000000..4e6a864ad8678ba780068e7946a0dc4ce57d05d0 --- /dev/null +++ b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard-realtime--hrwms4v.json @@ -0,0 +1,195 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-balance-hard-i00-standard-realtime--hrwms4v", + "taskKey": "humanoidbench/loco_balance_hard", + "dag_sha256": "419d18f9a6fdc6ba23a8ef8263e0e2a03d0e684ca7c5204c63fb59f163363f46", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 66, + "sim_steps": 2825, + "calls": 2847, + "rows": 2741, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 992.1, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 55.578, + "max_lag_s": 0.0112, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 84, + "late_steps": 0, + "requests_refused": 51, + "request_busy_s": 2.681, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-balance-hard-i00-standard-realtime--hrwms4v/replay.mp4", + "duration_s": 0.75, + "fps": 20.0, + "frames": 15, + "sha256": "d2a39963a38dccddc326902ed60870510ffe0b0186c7acad69965e1d069bd80f" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "board_balance_sustained": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00 s 机器人双脚已踩在球体支承的浅灰色平衡板上;视频 0.05–0.40 s 机器人伸直双腿并抬起双臂。初始站姿与短暂站立不能证明累计奖励已达到800。" + }, + { + "t": 0.65, + "log_step": 114, + "text_zh": "观察:视频 0.45–0.65 s 机器人与平衡板持续向后倾,头部下降;视频 0.65–0.70 s 身体已大幅倾倒,末帧没有恢复。log step 114 的服务终态为 episode=51、control_steps=66、done=true、reason=failed,与这段先站立后后倾的录像相符。" + }, + { + "t": 0.7, + "log_step": 114, + "text_zh": "推断:受评的最后回合只执行66个控制步便失衡结束,逐步奖励为不超过1的各项乘积,故不能累计达到800;结合视频 0.45–0.70 s 的失衡及评分器 success=0,该事件节点始终未达成。log step 116 的 agent 自述也明确此前74步的试验已重置,并非受评回合;此处不将日志控制步或 agent 时间换算为视频时间。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.60 s 机器人双脚接触并承重于球体上的平衡板,腿和躯干随姿态调整;视频 0.65–0.70 s 双脚仍在倾斜板面附近,身体向后倒。双臂抬起但未见手部接触板或球。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "board_balance_sustained", + "t": 0.45, + "log_step": 114, + "where": "操作中", + "evidence_zh": "观察:视频 0.45 s 起机器人在板上站立调整时向后倾斜,视频 0.50–0.65 s 后倾不断加剧,视频 0.70 s 末帧仍未恢复;log step 114 的服务确认最后第51回合在66个控制步后以 failed 结束。失败发生于持续平衡操作中。" + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 0.55, + "log_step": 114, + "evidence_zh": "观察:视频 0.45–0.70 s 机器人双腿和躯干整体向后倒,头部明显降低,平衡板随之大幅倾斜;末帧机器人已无法维持直立,且没有恢复。log step 114 的服务终态为 done=true、reason=failed。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "反馈控制", + "evidence_zh": "推断:log step 114 记录 agent 的最后回合控制程序使用板角相关的踝、髋反馈和肩部姿态反馈;同一调用中程序打印的 pitch 从控制步40的约−0.003 rad,变为控制步50的−0.183 rad、控制步60的−0.713 rad,pitch_rate 的绝对值亦增大。结合视频 0.45–0.70 s 持续加剧的后倾,说明该闭环姿态调节未能稳定身体与板的耦合运动,归为反馈控制失败;现有记录不能证明某个特定增益、反馈符号或板角估计就是错误来源。", + "source_zh": "日志" + } + ], + "gap_zh": "无状态日志;inputs/run.json 的评分字段未报告 episode_return,故原生累计奖励值留空。视频已逐帧检查至视频 0.70 s,但末端视角不能精确区分骨盆过低、板端触地或其他原生终止接触,服务仅报告 failed;未另行判定约束。日志中的姿态量由 agent 程序计算,缺少独立状态和控制目标记录,不能进一步确定反馈失稳的具体机械或参数原因。", + "unmet": { + "board_balance_sustained": false + } + }, + "summary_zh": "受评的是最后第51回合:机器人双脚在球体支承的板上尝试站稳,但“板上持续站稳”节点未达成,最终0/1。视频 0.45–0.70 s 显示机器人持续向后倾倒,结合 log step 114 的反馈控制记录及66控制步后失败的服务结果,根因判断为反馈控制未能稳定姿态。没有状态日志或累计奖励数值,具体触发终止的接触及失稳参数无法核实。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "标注与视频和日志一致:受评的是最后第51回合,视频 0.00–0.40 s 站立、0.45 s 起后倾、0.75 s 视频末尾已倒下,节点未达成,断点、平衡现象和反馈控制根因均有证据支持,未发现需要修改的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:11:34Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-standard-realtime--hrwms4v/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-standard-realtime--hrwms4v/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "c78addb39dfd40647e146b3e6a23bdf1adc8ce0e53b53e807b259f5012ba02f1", + "video_sha256": "d2a39963a38dccddc326902ed60870510ffe0b0186c7acad69965e1d069bd80f", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard-realtime--khsv6t6.json b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard-realtime--khsv6t6.json new file mode 100644 index 0000000000000000000000000000000000000000..42e3662d6b1555444f69e724f1d5fe4e7b035249 --- /dev/null +++ b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard-realtime--khsv6t6.json @@ -0,0 +1,202 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-balance-hard-i00-standard-realtime--khsv6t6", + "taskKey": "humanoidbench/loco_balance_hard", + "dag_sha256": "419d18f9a6fdc6ba23a8ef8263e0e2a03d0e684ca7c5204c63fb59f163363f46", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 38, + "sim_steps": 2247, + "calls": 24100, + "rows": 23996, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 433.9, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 44.039, + "max_lag_s": 0.0021, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 0, + "late_steps": 0, + "requests_refused": 52, + "request_busy_s": 6.488, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.407, + "peak_processes": 10, + "peak_disk_gb": 0.997, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-balance-hard-i00-standard-realtime--khsv6t6/replay.mp4", + "duration_s": 0.45, + "fps": 20.0, + "frames": 9, + "sha256": "078ccb5a50f81ecd502a7183a9d66db52d03c5ce8071ec9033d62a1ffc433b2c" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "board_balance_sustained": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频0.00–0.30 s中,机器人双脚承重于球体上方的浅灰平衡板,腿部和躯干随板晃动;初始站在板上不等于完成累计奖励800的持续平衡节点。" + }, + { + "t": 0.35, + "log_step": null, + "text_zh": "观察:已逐帧检查视频0.00–0.40 s全部9帧;视频0.35 s双脚离开板面,腿向前伸,板明显倾斜;视频0.40 s仍未恢复板上支撑。" + }, + { + "t": null, + "log_step": 61, + "text_zh": "观察:log step 59的search2.py最后实际执行配置为f3(kh_e=6、ka_e=-2);随后重置被拒。log step 61的机器人服务状态返回episode=51、control_steps=38、done=true、reason=failed、resets_left=0,log step 72再次确认;该最后回合是受评回合,其他调参回合不计进度。" + }, + { + "t": 0.4, + "log_step": 72, + "text_zh": "推断:受评回合仅执行38个控制步即失去板上支撑并失败,结合评分器success=0及节点累计奖励800的判据,节点始终未完成;不把评分器的二值结果reward=0.0当作原生episode_return。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频0.00–0.30 s机器人双脚接触平衡板并承重,板下球体支承板;视频0.35–0.40 s双脚离板,板上支撑丢失。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "board_balance_sustained", + "t": 0.35, + "log_step": 59, + "where": "操作中", + "evidence_zh": "观察:视频0.30 s机器人还以双脚在板上承重;视频0.35 s双脚已离开倾斜板面,腿向前伸,至视频0.40 s末帧没有恢复。log step 61的机器人服务确认最后回合episode=51在38个控制步后以reason=failed结束;失败发生在板上保持平衡过程中。" + }, + "unmet": { + "board_balance_sustained": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 0.35, + "log_step": null, + "evidence_zh": "观察:视频0.25–0.30 s板和机器人仍一起晃动;视频0.35–0.40 s平衡板翻斜,机器人双脚悬离板面、腿向前伸,失去原有承重支撑。录像终止前尚未显示机器人全身倒地,不将其描述为已躺倒。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "反馈控制", + "t": 0.35, + "log_step": 59, + "source_zh": "日志 / 视频", + "evidence_zh": "推断:log step 43的控制代码以骨盆俯仰、板角度等构造反馈,log step 58将更新改为每控制步一次;受评最后配置在log step 59为kh_e=6、ka_e=-2,未启用板角度、角速度及侧向反馈,脚本打印其末次估算骨盆俯仰约0.066 rad而板角约-1.047 rad。该估算仅作脚本诊断证据;结合视频0.30–0.40 s板翻斜、双脚离板,判断该闭环未能协调身体与板的运动以维持支撑。尚不能证明某个增益数值或符号本身错误;早先陀螺反馈振荡已在log step 58修改,不作为本次根因。" + } + ], + "gap_zh": "没有状态日志;评分器和最后回合服务状态未报告episode_return,故原生累计奖励值为null。视频只到0.40 s末帧,能确认失去板上支撑,但不能精确区分终止由板端触地、机器人触地还是其他原生终止谓词触发;没有接触对及受评回合完整本体感知轨迹,反馈控制归因为日志与画面的综合推断,不能进一步确定动力学误差或具体错误增益。所有时间锚直接取视频帧,未把agent时间或50 Hz控制步直接当作视频时间。" + }, + "summary_zh": "受评的是最后在线回合episode=51:机器人起初双脚承重于球支撑的平衡板,视频约0.35 s双脚离板、板明显翻斜,至视频0.40 s未恢复,板上持续站稳节点完成0/1。日志显示该回合使用骨盆俯仰反馈控制并在38个控制步后失败,综合画面判断控制未能稳定身体与板的支撑关系;此前调参回合不计进度,重置预算耗尽仅解释未再重试。缺少状态日志和原生累计奖励,无法确定精确终止接触或具体错误增益。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "标签与视频、日志一致:视频0.30 s双脚仍在板上,0.35 s双脚离板、板翻斜,节点未完成,断点与平衡/反馈控制归因有证据支持,时间锚均为视频时间。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:09:55Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-standard-realtime--khsv6t6/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-standard-realtime--khsv6t6/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "68ea3e2de16783c90b837d755022c5e8676d58331faaf1babeda6809e2b1e5b3", + "video_sha256": "078ccb5a50f81ecd502a7183a9d66db52d03c5ce8071ec9033d62a1ffc433b2c", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard-realtime--mg6fyec.json b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard-realtime--mg6fyec.json new file mode 100644 index 0000000000000000000000000000000000000000..334e5b8e9b9e499dc475d488f6e7a5675892499c --- /dev/null +++ b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard-realtime--mg6fyec.json @@ -0,0 +1,195 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-balance-hard-i00-standard-realtime--mg6fyec", + "taskKey": "humanoidbench/loco_balance_hard", + "dag_sha256": "419d18f9a6fdc6ba23a8ef8263e0e2a03d0e684ca7c5204c63fb59f163363f46", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 43, + "sim_steps": 2405, + "calls": 2371, + "rows": 2263, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1416.9, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 47.217, + "max_lag_s": 0.0024, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 373, + "late_steps": 0, + "requests_refused": 55, + "request_busy_s": 1.186, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.242, + "peak_processes": 7, + "peak_disk_gb": 0.258, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-balance-hard-i00-standard-realtime--mg6fyec/replay.mp4", + "duration_s": 0.5, + "fps": 20.0, + "frames": 10, + "sha256": "ff1acec7ef06a63bc3d73201c68679ead1d1b5df23b86e84bd5184e0aef2108a" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "board_balance_sustained": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:受评视频0.00–0.20 s,黑色人形机器人双脚踩在浅灰色平衡板上,板由下方球体支承;初始短暂直立不满足本节点累计奖励达到800的验收条件。" + }, + { + "t": 0.35, + "log_step": null, + "text_zh": "观察:受评视频0.25–0.45 s,机器人躯干逐渐后仰,视频0.35 s后倾倒明显加剧,视频0.40–0.45 s平衡板大幅倾斜;至末帧视频0.45 s未恢复直立。" + }, + { + "t": 0.45, + "log_step": 214, + "text_zh": "观察:log step 214中基准服务在最后一次重置后报告episode=51、resets_left=0,最终control_steps=43、done=True、reason=failed;log step 215再次确认同一终局,评分字段success=0、live_success=0、replay_success=0。推断:结合受评视频末段失衡和该回合服务判负,板上持续站稳未完成;没有把会话累计sim_steps=2405当作本回合步数,也未将agent时间或控制步数换算成视频时间。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:受评视频0.00–0.45 s,机器人双脚与平衡板接触承重,腿部姿态随板倾斜而变化;手臂悬空,未见手与平衡板或球体接触。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "board_balance_sustained", + "t": 0.35, + "log_step": 214, + "where": "操作中", + "evidence_zh": "观察:受评视频0.30–0.45 s,机器人在双脚承重于板面的平衡动作中持续向后倾倒,平衡板也明显翻斜,至末帧没有恢复;log step 214记录最后受评回合episode=51在43个控制步后由服务判为failed。推断:未修复的失败发生在板上持续站稳这一步,视频0.35 s为明显失衡的导航锚点。" + }, + "unmet": { + "board_balance_sustained": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 0.35, + "log_step": 214, + "evidence_zh": "观察:受评视频0.25–0.45 s,机器人头部和躯干越来越偏向脚后方,视频0.40–0.45 s身体大幅后仰、平衡板倾覆,末帧仍处于倾倒过程;log step 214的服务终态为reason=failed。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "反馈控制", + "evidence_zh": "推断:log step 213的最终控制脚本以起始姿态为参考,对踝、膝和髋关节施加平衡修正;log step 214确认该v10控制器实际执行于最后受评回合episode=51。受评视频0.25–0.45 s中后仰持续增大,控制器没有将机器人和支承板恢复到稳定状态,故归为平衡反馈控制失稳;现有证据不足以指定某个增益、符号或延迟为具体原因。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "没有状态日志;inputs/run.json的评分字段未报告episode_return原生累计奖励,因此该指标保留null,不能用目录reward=0.0代替。受评视频末帧为视频0.45 s,显示失衡但没有身体完全落地后的画面;服务仅报告failed,不能确定具体由骨盆过低、板端触地还是其他原生终止条件触发。log step 213的脚本记录中段有截断,且未提供最后回合逐控制步的姿态、接触和控制量,无法进一步核实反馈失稳的具体参数或动力学原因。" + }, + "summary_zh": "最后受评回合为第51回合:机器人实际以双脚支撑平衡板尝试保持平衡,但未完成“板上持续站稳”,最终进度0/1。受评视频0.25–0.45 s显示后仰不断加剧并伴随板面翻斜,log step 214确认43个控制步后失败,证据支持平衡反馈控制失稳。缺少原生累计奖励和状态日志,无法给出精确回报或确定具体终止接触。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "视频仅0.5 s(10帧):0.0–0.25 s直立,0.35 s起后仰,0.40–0.45 s板翻、身体倒下;节点未达成、断点、平衡现象与反馈控制根因均有证据支持,未发现需修改的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:10:46Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-standard-realtime--mg6fyec/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-standard-realtime--mg6fyec/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "7ab0e0ccdefde46e8955c9c487ad5127516cc4fe03dc93255e09bb8e04fdb944", + "video_sha256": "ff1acec7ef06a63bc3d73201c68679ead1d1b5df23b86e84bd5184e0aef2108a", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard-realtime--mnepqyu.json b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard-realtime--mnepqyu.json new file mode 100644 index 0000000000000000000000000000000000000000..e910c5a4bb3fa7fd44d1fc388ddc0f2234b4d9d0 --- /dev/null +++ b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard-realtime--mnepqyu.json @@ -0,0 +1,196 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-balance-hard-i00-standard-realtime--mnepqyu", + "taskKey": "humanoidbench/loco_balance_hard", + "dag_sha256": "419d18f9a6fdc6ba23a8ef8263e0e2a03d0e684ca7c5204c63fb59f163363f46", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 61, + "sim_steps": 515, + "calls": 741, + "rows": 714, + "resets": 8, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3505.1, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 10.137, + "max_lag_s": 0.0024, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 0, + "late_steps": 0, + "requests_refused": 9, + "request_busy_s": 0.437, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.21, + "peak_processes": 6, + "peak_disk_gb": 0.255, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-balance-hard-i00-standard-realtime--mnepqyu/replay.mp4", + "duration_s": 0.7, + "fps": 20.0, + "frames": 14, + "sha256": "04b650dd685e5abcc0d58fb981fc7352b0f10fd48b423ae9e817e4eeb4050887" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "board_balance_sustained": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00 s,机器人双脚已承重在球体支承的平衡板上,躯干大致直立;视频 0.05–0.15 s 机器人抬起双臂调节姿态。初始站姿不是累计奖励达标事件。" + }, + { + "t": 0.6, + "log_step": 94, + "text_zh": "观察:视频 0.35–0.65 s,球体逐渐滚向画面左侧,板面向画面右下侧倾斜并下沉,未恢复初始支承姿态;视频 0.65 s 机器人躯干仍大致直立。log step 94 的机器人服务状态为第9回合 control_steps=61、done=True、reason=failed,calls=741、rows=714、control_steps_total=515,与 inputs/run.json 的评分记录一致。推断:受评的是最后第9回合;按节点的逐步奖励乘积上限1,61个控制步累计不可能达到800,节点未完成。未将控制步或 agent 时间换算为视频时间。" + }, + { + "t": null, + "log_step": null, + "text_zh": "观察:成功参照 inputs/oracle/replay.mp4 的视频 0.00、2.00、4.00、6.00、8.00、10.00 和10.05 s,机器人维持板上站姿,球体保持在板下承重;本次录像视频 0.50–0.65 s 的球板相对位移和板面倾斜明显增大。参照仅用于比较可见姿态,不作为本回合完成证据。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.65 s,机器人双脚接触平衡板并承重,腿部随板面倾斜调节;双臂抬起但未见手部接触板、球体或地面。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "board_balance_sustained", + "t": 0.5, + "log_step": 94, + "where": "操作中", + "evidence_zh": "观察:视频 0.35–0.65 s,机器人正在用双腿承重并抬臂调节,球体向画面左侧滚动,板面持续倾斜、低端下沉,直到录像结束也未恢复;log step 94 服务确认该回合在61个控制步后以 failed 结束。推断:断点为正在进行的板上持续站稳,视频约 0.50 s 为支承明显失稳的导航锚点。" + }, + "unmet": { + "board_balance_sustained": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 0.5, + "log_step": 94, + "evidence_zh": "观察:视频 0.35–0.65 s,球体从板下中部向画面左侧移出,机器人脚下的板面倾斜加大,画面右下侧板端下落;视频 0.65 s 躯干仍大致直立,未见身体倒地。推断:机器人未能维持球板承重系统的平衡,支承失稳;log step 94 的服务状态确认本回合已经失败终止。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "反馈控制", + "log_step": 95, + "evidence_zh": "观察:log step 93 的最后回合控制脚本根据机器人本体关节、角速度与重力方向反馈调节踝、髋及肩部,并用计算出的板面角度和角速度作踝部反馈。log step 95 对该回合本体读数的脚本分析中,控制步42至48的双踝平均角由-0.754降至-0.880 rad,而脚本记录的踝目标为-0.220 rad;控制步56仍约-0.873 rad。视频 0.35–0.65 s 同时可见板面持续倾斜、球体向侧方滚出。推断:这套反馈控制未能使承重关节跟踪目标并抑制板面失稳;不据此断言某个增益或符号必然错误。这里的控制步只引用日志,不转换为视频时刻。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "没有状态日志,也未报告原生 episode_return 数值,故该指标为null;未把 success=0 当成累计回报0。视频仅有斜侧低分辨率视角,可见板面失稳但不能精确区分最终是哪一对几何体触发终止,服务只给出 failed,未给终止子原因。log step 95 的板角与重心量是 agent 脚本由本体读数计算的估计,不是独立状态真值;缺少同步接触力、执行器力矩和完整控制器记录,无法进一步确认关节跟踪偏差的具体动力学原因。" + }, + "summary_zh": "受评的最后第9回合已开始双脚承重和抬臂调节,但未完成“板上持续站稳”,最终节点为0/1;log step 94 的服务记录与评分计数一致,回合仅执行61个控制步后失败。视频 0.35–0.65 s 显示球体向侧方滚出、板面倾斜下沉,断点在平衡操作中,日志中的踝关节目标与实际角度偏差支持反馈控制未能稳住支承系统。缺少状态日志、累计回报及具体终止接触对,不能精确判定终止子原因。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "最后一回合在视频 0.35–0.65 s 内球体滚向左侧、板面倾斜下沉,节点未达成、断点、平衡现象与反馈控制根因均有视频和日志支持,未发现需要改动的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:11:12Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-standard-realtime--mnepqyu/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-standard-realtime--mnepqyu/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "af2684b46549042a7452a055f3a3bde5329d3ed3f34d4a71af64096b682f41cc", + "video_sha256": "04b650dd685e5abcc0d58fb981fc7352b0f10fd48b423ae9e817e4eeb4050887", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard-realtime--plwzmoa.json b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard-realtime--plwzmoa.json new file mode 100644 index 0000000000000000000000000000000000000000..23be7256e1c6c6d384171064c12bfbc2dab912d4 --- /dev/null +++ b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard-realtime--plwzmoa.json @@ -0,0 +1,199 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-balance-hard-i00-standard-realtime--plwzmoa", + "taskKey": "humanoidbench/loco_balance_hard", + "dag_sha256": "419d18f9a6fdc6ba23a8ef8263e0e2a03d0e684ca7c5204c63fb59f163363f46", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 65, + "sim_steps": 2513, + "calls": 9333, + "rows": 2459, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3433.9, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 49.404, + "max_lag_s": 0.0588, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 37, + "late_steps": 4, + "requests_refused": 2, + "request_busy_s": 4.667, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.263, + "peak_processes": 10, + "peak_disk_gb": 0.926, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-balance-hard-i00-standard-realtime--plwzmoa/replay.mp4", + "duration_s": 0.75, + "fps": 20.0, + "frames": 15, + "sha256": "17dc65801d30f02edf71ac23bf6f059d47cf3b366fcc1726e8000017f9e2ceda" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "board_balance_sustained": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00–0.40 s,机器人双脚踩在球体支承的浅灰色板上,双腿屈曲、上身基本直立;视频 0.45–0.65 s,板面倾斜逐渐增大,靠画面右下的板端下坠、另一端抬起,球体与脚下承重区域的相对位置明显改变;视频 0.65–0.70 s 停留在该失稳姿态。初始站在板上不等于累计奖励已达标。" + }, + { + "t": 0.65, + "log_step": 151, + "text_zh": "观察:log step 150 的脚本依次运行最后两个重置回合;log step 151 的服务状态显示最后 episode=51、control_steps=65、done=true、reason=failed、resets_left=0,与评分字段 n_actions=65、graded_by_live=1 一致,故受评的是最后回合而非前面的试验。" + }, + { + "t": 0.65, + "log_step": 151, + "text_zh": "推断:任务 DAG 规定逐步奖励是各个不超过1的奖励项之积、累计至少800才完成;最后受评回合在65个控制步即失败,累计值不可能达到800,结合视频末段板面失稳,节点从未完成。评分字段未报告 episode_return 的具体数值,不能把二元 reward=0 当作累计回报。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.60 s,机器人双脚接触平衡板承重,髋、膝、踝随板面变化调整姿态;末段板面明显倾斜。双手悬空,未见手与板或球接触。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "board_balance_sustained", + "t": 0.55, + "log_step": 151, + "where": "操作中", + "evidence_zh": "观察:视频 0.45–0.65 s,机器人正在板上承重调姿时,板端逐渐下坠、另一端抬起,球体相对脚下支撑区偏移,末帧未恢复;log step 151 确认受评 episode=51 在65个控制步以 reason=failed 结束。失败发生在持续保持板上平衡的过程中。" + }, + "unmet": { + "board_balance_sustained": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 0.55, + "log_step": 151, + "evidence_zh": "观察:视频 0.40 s 板面尚接近初始状态,视频 0.50–0.65 s 板面快速倾斜,靠画面右下的板端下坠至贴近地面,另一端翘起,机器人脚部姿态随之改变;视频 0.65–0.70 s 未恢复,身体尚未完整倒地。推断:结合 log step 151 的最后回合失败状态,这是脚下板球支撑系统失稳,归为平衡丧失。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "反馈控制", + "log_step": 150, + "evidence_zh": "推断:log step 10、26 的控制器代码读取机器人关节与IMU,计算姿态和板面估计,再修正髋、膝、踝目标;log step 150 最后选用的是先重放早期回合的10个控制步、随后启用模型反馈并叠加髋踝姿态增益的配置。该闭环在受评回合未能抑制板球相对运动,视频 0.45–0.65 s 板面仍持续倾斜,log step 151 确认回合失败,支持反馈稳定控制不足;未据此指定某个增益数值或模型误差为已证实原因。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "没有状态日志或精确接触对;视频末段可见板端接近地面,但256×256画面不足以确定首先触发终止的是板端触地、其他违规接触还是骨盆高度条件。评分与服务日志未报告原生累计回报,episode_return 保留null;视频时钟与log step 151 所报仿真1.3 s未建立逐步映射,所有视频锚点直接取自所看帧,未用会话时间换算。" + }, + "summary_zh": "受评的是最后一次在线回合(episode=51);机器人双脚在板上尝试承重调姿,但“板上持续站稳”节点未达到累计800奖励,最终0/1。视频约0.55 s起板端明显下坠、球板相对位置偏移,板上平衡失稳且未恢复;日志中的模型反馈控制未能稳定该回合。没有状态日志和累计回报数值,无法精确确定终止接触及具体控制误差。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "标签与视频、log一致:最后受评回合(episode=51,65控制步,reason=failed)板上平衡未达800累计,视频0.45–0.65 s板端下坠失稳;节点、断点、平衡现象与反馈控制根因均有证据支持,无需修改。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:11:24Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-standard-realtime--plwzmoa/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-standard-realtime--plwzmoa/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "687544cba8c9545d6c8d6486a49079e371ae28615f79126519da249c7d0b09e3", + "video_sha256": "17dc65801d30f02edf71ac23bf6f059d47cf3b366fcc1726e8000017f9e2ceda", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard-realtime--sw5l7i5.json b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard-realtime--sw5l7i5.json new file mode 100644 index 0000000000000000000000000000000000000000..15542903dd16ec9edd19401657a922d11d1d6db3 --- /dev/null +++ b/phase2/humanoidbench/loco_balance_hard/trajectories/humanoidbench-loco-balance-hard-i00-standard-realtime--sw5l7i5.json @@ -0,0 +1,198 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-balance-hard-i00-standard-realtime--sw5l7i5", + "taskKey": "humanoidbench/loco_balance_hard", + "dag_sha256": "419d18f9a6fdc6ba23a8ef8263e0e2a03d0e684ca7c5204c63fb59f163363f46", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 23, + "sim_steps": 1471, + "calls": 1437, + "rows": 1782, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 2182.7, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 28.533, + "max_lag_s": 0.0038, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 106, + "late_steps": 0, + "requests_refused": 50, + "request_busy_s": 0.807, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-balance-hard-i00-standard-realtime--sw5l7i5/replay.mp4", + "duration_s": 0.3, + "fps": 20.0, + "frames": 6, + "sha256": "84b69c1565f003bad8dbbad0480bbfd50459413512d3d84d2e6d6e1bc1e0a131" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "board_balance_sustained": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0, + "log_step": null, + "text_zh": "观察:视频 0.00 s,机器人双脚踩在浅灰色平衡板上,板由深色球体承托;视频 0.05–0.15 s 机器人摆臂并屈伸双腿,已开始平衡动作。初始站在板上不等于累计奖励达标。" + }, + { + "t": 0.2, + "log_step": null, + "text_zh": "观察:已检查本次视频全部六帧(视频 0.00、0.05、0.10、0.15、0.20、0.25 s);视频 0.15 s 机器人明显屈膝下沉,视频 0.20–0.25 s 平衡板快速翻倾、机器人身体后倾,末帧未恢复稳定承重姿态。" + }, + { + "t": 0.25, + "log_step": 43, + "text_zh": "推断:log step 43 的机器人服务状态明确记录最后第51回合 control_steps=23、done=true、reason=failed,评分器字段 live_success=0、graded_by_live=1;该回合每步奖励为不超过1的乘积,23步不可能累计达到800,因此节点未曾达成。评分器未给出原生累计奖励的数值;未把会话总 sim_steps=1471 当作本回合步数,也未将控制步或 agent 时间直接换成视频时间。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.15 s,机器人双脚在平衡板上承重,双腿屈伸、躯干和手臂调整姿态;视频 0.20–0.25 s 板面急剧翻倾。未见双手接触板或球体。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "board_balance_sustained", + "t": 0.2, + "log_step": 43, + "where": "操作中", + "evidence_zh": "观察:视频 0.15 s 机器人在板上屈膝下沉,视频 0.20–0.25 s 板面大幅翻倾、机器人随之失稳,直到末帧未恢复;log step 43 的服务记录最后第51回合在23个控制步后以 reason=failed 结束。失败发生在已开始的板上平衡操作中。" + }, + "unmet": { + "board_balance_sustained": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 0.2, + "log_step": 43, + "evidence_zh": "观察:视频 0.00 s 机器人站在球体支承的平衡板上;视频 0.15–0.25 s 双腿姿态迅速变化,板面由近水平翻为大角度倾斜,机器人后倾、稳定支撑丧失且未恢复。log step 43 服务记录该回合失败;画面未展示身体完全倒地的过程。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "推断:log step 41 的最终控制脚本用本体感知计算倾角,以髋膝、踝和肩关节反馈叠加迭代学习前馈来稳定机器人;log step 42 启动该脚本并反复执行最后阶段回合,log step 43 确认受评最后第51回合在23个控制步后失败。视频 0.15–0.25 s 可见板面翻倾持续扩大,说明实际执行的平衡控制未稳住身体与支承板;仅归到控制,不能据此证明某一增益、力矩上限或学习参数错误。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "无状态日志,评分器未报告 episode_return 数值;本次视频只有六帧,末帧为视频 0.25 s,未展示身体完全倒地,不能确认具体触发终止的是板端触地、机器人触地或其他原生终止条件。最终第51回合的逐步关节、接触和力矩记录未包含在输入中,无法细分控制根因;agent 在 log step 47 关于启动冲击与踝力矩不足的自述主要来自此前调试回合,未据此认定本回合的具体物理原因。" + }, + "summary_zh": "受评最后第51回合已尝试板上平衡,但“板上持续站稳”未达成,最终进度为0/1;服务在 log step 43 记录该回合仅运行23个控制步后失败。视频 0.15–0.25 s 显示平衡板快速翻倾、机器人失稳,根因保守归为控制未能稳定支撑系统。缺少状态日志和原生累计奖励数值,短片未展示完全倒地及具体终止接触,无法核实更细的力矩或参数解释。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "节点未达成、断点在板上平衡、平衡现象与控制根因均有视频六帧和服务记录(第51回合23控制步 reason=failed)支持;时钟未换算,gap说明诚实。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:13:03Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-standard-realtime--sw5l7i5/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-balance-hard-i00-standard-realtime--sw5l7i5/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "585d7ac456d42b23d0846b7bdb382f9b6be08f684709b229dabd1f371cc43949", + "video_sha256": "84b69c1565f003bad8dbbad0480bbfd50459413512d3d84d2e6d6e1bc1e0a131", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-privileged--dqvxr6t.json b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-privileged--dqvxr6t.json new file mode 100644 index 0000000000000000000000000000000000000000..16aa77069fed65371925eabff29a5ef907367ad2 --- /dev/null +++ b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-privileged--dqvxr6t.json @@ -0,0 +1,173 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-hurdle-i00-privileged--dqvxr6t", + "taskKey": "humanoidbench/loco_hurdle", + "dag_sha256": "ca09b7dc923221773af0078329063e8e11b1544a0f305c56d17055d1bf47910e", + "outcome": { + "success": 0, + "reward": 2.7921628563395863e-15, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 521, + "final_reward": 2.7921628563395863e-15, + "sim_steps": 521, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-hurdle-i00-privileged--dqvxr6t/replay.mp4", + "duration_s": 5.3, + "fps": 20.0, + "frames": 106, + "sha256": "bc9fee459d30961bef6c0c3593d1990e2276e3a9fa45916ac5351a6920a55e16" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "hurdle_run_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 2.25, + "log_step": null, + "text_zh": "观察:视频 0.0–2.25 s,机器人在跑道内向前迈步,第一道横栏从身前移到身后;这只证明已开始跨栏跑,不能证明累计奖励达到700。" + }, + { + "t": 2.5, + "log_step": null, + "text_zh": "观察:视频 2.40–2.65 s,机器人落地后骨盆迅速降至地面附近,腿向前伸出而躯干坐倒;视频 2.75–5.25 s 持续倒地摆腿、翻动,没有恢复直立跨栏跑。" + }, + { + "t": 5.25, + "log_step": 276, + "text_zh": "观察:评分器记录 success=0、all_success=0、deterministic=1,交付动作数与重放步数均为521;log step 276 对最终动作文件的重复重放也报告521步及 deterministic=true、scored=false。推断:结合未恢复的倒地过程与该累计奖励任务的失败判决,跨栏跑达标事件未发生。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–2.40 s,左右脚交替接触跑道地面承重并向前推进;视频约 2.50 s 后骨盆及躯干落向地面,腿仍摆动,失去正常双足跑动支撑。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "累计奖励", + "threshold": 700, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "hurdle_run_completed", + "t": 2.5, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频 2.25 s 机器人已来到第一道横栏另一侧,视频 2.40–2.65 s 落地时骨盆塌下并坐倒,随后翻倒;直到视频 5.25 s 末帧仍未恢复直立跑动。推断:未修复的失败发生在跨栏跑的落地与继续前进过程中。" + }, + "unmet": { + "hurdle_run_completed": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 2.5, + "log_step": null, + "evidence_zh": "观察:视频 2.45 s 机器人身体快速下降,视频 2.50–2.65 s 双腿伸到身前、骨盆接近地面,视频 2.75–3.50 s 躯干倒卧地面;视频 4.25–5.25 s 仍贴地翻动,未恢复双足直立支撑。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志", + "evidence_zh": "观察:log step 264 的最终生成脚本以三阶段周期关节目标加躯干姿态PD修正生成动作;log step 268 直接载入这份521步交付文件重放,记录控制步350、400的骨盆高度分别为0.29、0.25 m,头高分别为0.24、0.21 m,torso_upright分别为−0.07、−0.06,控制步520时 done=True。log step 276 确认最终文件可确定性重放;视频 2.50–5.25 s 也显示落地塌下后持续倒卧。推断:交付的运动控制未能在越栏落地后维持承重姿态并恢复跑动;证据不足以进一步指定某个增益、关节目标或反馈环节。" + } + ], + "gap_zh": "无状态日志,无法核实逐帧足底接触力、关节跟踪误差及具体失稳控制环节。评分记录只有 final_reward=2.7921628563395863e-15,未提供 episode_return;该字段不能直接当作累计奖励,故原生累计奖励值留空。log step 263 的104.35是agent自行计算的代理奖励,不替代评分器累计回报。" + }, + "summary_zh": "机器人启动向前跨栏跑,并在视频约 2.25 s 越过第一道横栏,但视频约 2.50 s 落地后塌下倒地,直到视频 5.25 s 末帧未恢复直立跑动,跨栏跑达标节点为0/1。交付文件重放可复现但评分失败;视频和最终文件的重放诊断支持控制未能维持落地平衡。缺少状态日志及评分器累计回报,不能定位具体控制子类或恢复精确累计奖励。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "视频显示机器人约2.0-2.25 s越过第一道栏,2.5 s后落地塌倒并持续倒卧至末帧;节点未达成、断点、平衡现象和控制根因与证据一致。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:45:59Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-privileged--dqvxr6t/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-privileged--dqvxr6t/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "86e3d6e0d2134ab22172d19ece5ff439caa1ef37df8c385efdf70b772a7f1aae", + "video_sha256": "bc9fee459d30961bef6c0c3593d1990e2276e3a9fa45916ac5351a6920a55e16", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-privileged--ev8v9za.json b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-privileged--ev8v9za.json new file mode 100644 index 0000000000000000000000000000000000000000..a625d3a5ddbabcd765527aba4c048ed1be414663 --- /dev/null +++ b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-privileged--ev8v9za.json @@ -0,0 +1,164 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-hurdle-i00-privileged--ev8v9za", + "taskKey": "humanoidbench/loco_hurdle", + "dag_sha256": "ca09b7dc923221773af0078329063e8e11b1544a0f305c56d17055d1bf47910e", + "outcome": { + "success": 1, + "reward": 0.8909936043798045, + "source": "catalog", + "raw": { + "success": 1, + "all_success": 1, + "deterministic": 1, + "n_actions": 1000, + "final_reward": 0.8909936043798045, + "sim_steps": 1000, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-hurdle-i00-privileged--ev8v9za/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "22732428f86850d4a191f23a15af3fb9fc5dd7947ff6034293ba2f89cb7ac7a7" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "hurdle_run_completed": { + "method": "visual", + "intervals": [ + [ + 10.0, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 1.4, + "log_step": null, + "text_zh": "观察:视频0.0–1.0 s机器人从直立姿态起跑,双脚交替接触跑道;视频1.2–1.6 s机器人抬腿越过第一道可见横栏,横栏从前方移至身后,随后继续向前跑。" + }, + { + "t": 8.3, + "log_step": null, + "text_zh": "观察:已按视频0.1 s间隔检查全段,机器人在视频2.0–8.3 s连续抬腿越过后续横栏并落地续跑,身体与两侧墙体保持可见间隔;视频8.3 s最后一道可见横栏已在机器人身后。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频9.0–10.05 s逐帧检查中,机器人仍交替迈腿前进,躯干保持直立,未出现骨盆塌落或倒地;视频10.0 s和10.05 s为相同的片尾姿态。" + }, + { + "t": null, + "log_step": 142, + "text_zh": "观察:log step 141的打包脚本选取最终动作序列,重新播放后保存交付轨迹;log step 142输出该序列含1000个动作,生成器执行后的state_sha与打包记录同为426189483946。该输出中的estimated_return=702.0316915689646明确标为估计值,不是评分器累计奖励。" + }, + { + "t": 10.0, + "log_step": null, + "text_zh": "推断:inputs/run.json的评分器字段success=1、all_success=1、deterministic=1、n_actions=1000,结合本段连续跨栏且片尾未倒地的画面,判定跨栏跑达标节点最终为真。累计奖励700门槛的逐步记录未提供,区间起点视频10.0 s仅取整段完成验收的片尾锚点,不表示已测得首次达到700的时刻;不把final_reward=0.8909936043798045换算为累计奖励。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频0.0–10.05 s机器人双脚交替接触跑道,完成起跑、越栏后的落地与续跑支撑;双臂摆动,未见手部操作物体。" + }, + { + "who": "腿/躯干", + "how": "避开", + "note_zh": "观察:视频0.0–8.3 s机器人在两侧墙体之间向前跑,腿和躯干与侧墙保持可见间隔,未见触墙;此项记录避开侧墙的通过行为。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "累计奖励", + "threshold": 700, + "direction": "higher_better" + } + ], + "failure": null, + "summary_zh": "机器人连续向前跨栏,视频8.3 s已通过最后一道可见横栏,视频10.05 s仍保持直立并继续迈步;评分器success=1、all_success=1、deterministic=1,唯一的跨栏跑达标节点最终完成。视频中的脚地支撑及避开侧墙行为可见,但没有状态日志或评分器累计奖励序列,episode_return保持null,首次达到700的精确时刻未知,使用视频10.0 s作为整段验收锚点。log step 142的约702累计回报为agent脚本估计,不用它替代官方指标。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "单节点跨栏跑:视频全段可见连续越栏、直立、未倒地,评分器success=1,final=true合理;区间起点10.0 s已明确为验收锚点而非首次达标时刻;腿/躯干支撑与避开接触合理;log步骤引用存在,未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:43:47Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-privileged--ev8v9za/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-privileged--ev8v9za/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "0db466e66bbbac5387bbe027300ef2cb8f57de3a0b5af08032ced5538eb10269", + "video_sha256": "22732428f86850d4a191f23a15af3fb9fc5dd7947ff6034293ba2f89cb7ac7a7", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-privileged--gwehgwq.json b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-privileged--gwehgwq.json new file mode 100644 index 0000000000000000000000000000000000000000..457e3e7ed60bc703757b83af94217d4b4205ab74 --- /dev/null +++ b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-privileged--gwehgwq.json @@ -0,0 +1,178 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-hurdle-i00-privileged--gwehgwq", + "taskKey": "humanoidbench/loco_hurdle", + "dag_sha256": "ca09b7dc923221773af0078329063e8e11b1544a0f305c56d17055d1bf47910e", + "outcome": { + "success": 0, + "reward": 6.200502551453507e-17, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 698, + "final_reward": 6.200502551453507e-17, + "sim_steps": 698, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 5.534, + "peak_processes": 25, + "peak_disk_gb": 0.056, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-hurdle-i00-privileged--gwehgwq/replay.mp4", + "duration_s": 7.05, + "fps": 20.0, + "frames": 141, + "sha256": "ededf464a828e3076290b55d6de20ec817c883c695054af01200ea611dddee48" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "hurdle_run_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.9, + "log_step": null, + "text_zh": "观察:视频 0.0–1.7 s 机器人由起点交替迈步前进;视频 1.8–2.3 s 抬腿通过第一道横栏并落地,之后视频 2.7–6.4 s 又连续通过多道横栏,确实开始了跨栏跑。" + }, + { + "t": 6.8, + "log_step": null, + "text_zh": "观察:视频 6.1–6.4 s 机器人躯干大幅侧转至接近横卧;视频 6.65–6.75 s 一度重新接近直立,但视频 6.8–7.0 s 再次倾倒,末帧骨盆及躯干降到地面附近,未恢复跑动。" + }, + { + "t": null, + "log_step": null, + "text_zh": "推断:inputs/run.json 的评分器记录 success=0、all_success=0、deterministic=1、grader_error=0,结合视频末段倒地,跨栏跑累计奖励达标事件未发生。评分记录只给出 final_reward=6.200502551453507e-17,未单列 episode_return,不能用该字段或会话中的自算代理奖励填充累计回报。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–6.9 s 机器人左右脚在跑道地面交替承重、起跳及落地;视频 7.0 s 倾倒后躯干和骨盆贴近地面。" + }, + { + "who": "腿/躯干", + "how": "避开", + "note_zh": "观察:视频 0.0–7.0 s 机器人沿两侧墙之间的跑道内部前进,起点后墙逐渐留在身后,身体与两侧墙保持可见间隔;此项只描述避开墙体,不据此断言脚腿与横栏完全无接触。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "累计奖励", + "threshold": 700, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "hurdle_run_completed", + "t": 6.8, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频 6.8 s 机器人仍在越栏后的向前跑动中,躯干开始再次侧倒;视频 6.9–7.0 s 骨盆迅速下沉、身体翻倒,直到视频结束没有恢复。此前视频 6.1–6.4 s 曾大幅侧转,视频 6.65–6.75 s 短暂接近直立;断点锚定最后未恢复的倾倒。" + }, + "unmet": { + "hurdle_run_completed": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 6.8, + "log_step": null, + "evidence_zh": "观察:视频 6.8–7.0 s 机器人躯干倾斜持续扩大,双腿未能继续承重,骨盆塌下、身体翻倒至地面附近;这是跑动中的失衡倒地。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "推断:log step 120 的控制脚本把逐栏空间调制叠加到 step_sprinter_v2 的关节目标上,再保存完整 rollout;视频 1.8–5.5 s 多次通过横栏,说明向前越栏动作已能执行,但视频 6.1–6.4 s 姿态大幅翻转、视频 6.8–7.0 s 最终失去落脚支撑,支持末段运动控制未能稳定身体。log step 124 的会话诊断也记录末段姿态恶化和骨盆下降;它仅作控制问题的佐证,不用来代替评分视频或正式累计回报。证据不足以区分轨迹、反馈或具体增益问题,保留在“控制”层级。", + "source_zh": "视频 / 日志" + } + ], + "gap_zh": "观察:inputs/ 内没有状态日志,评分记录未单列 episode_return;log step 124 的 dmc=369.8、str=339.3 是 agent 脚本在会话测试中累计的代理奖励,不是评分器报告的本段累计回报,因此 native_metrics 的 episode_return 留空。视频分辨率为 256×256,脚腿经过横栏时有重叠,不能逐一确认擦碰的部位、接触力,也不能证明某次横栏碰撞或某个控制参数导致最终倒地;控制根因只能定位到大类。" + }, + "summary_zh": "机器人确实向前跑并连续通过多道横栏,但视频 6.8–7.0 s 最终失衡倒地,“跨栏跑达标”未达到,最终节点为 0/1,与评分器失败判决一致。推断末段运动控制未能稳定身体,根因保留在“控制”大类。缺少状态日志和正式累计回报字段,无法细分具体接触或控制参数问题,episode_return 留空。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "节点未达成、断点(视频约6.8 s起最终倒地)、平衡现象与控制根因均与视频一致;未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:49:04Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-privileged--gwehgwq/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-privileged--gwehgwq/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "4fc775b0d703b4dc415e9177b84ba55c5e6dc87c423a750b33b1bd6985135622", + "video_sha256": "ededf464a828e3076290b55d6de20ec817c883c695054af01200ea611dddee48", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-privileged--jap2yhy.json b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-privileged--jap2yhy.json new file mode 100644 index 0000000000000000000000000000000000000000..734500d4335393c33c152f25b47f9cec0fb7b2e6 --- /dev/null +++ b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-privileged--jap2yhy.json @@ -0,0 +1,178 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-hurdle-i00-privileged--jap2yhy", + "taskKey": "humanoidbench/loco_hurdle", + "dag_sha256": "ca09b7dc923221773af0078329063e8e11b1544a0f305c56d17055d1bf47910e", + "outcome": { + "success": 0, + "reward": 1.50049780347006e-15, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 287, + "final_reward": 1.50049780347006e-15, + "sim_steps": 287, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-hurdle-i00-privileged--jap2yhy/replay.mp4", + "duration_s": 2.95, + "fps": 20.0, + "frames": 59, + "sha256": "44000706496b78e732c8f1115cbdc0f871349e172f958fa42278e6e44a7eb3dc" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "hurdle_run_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.0–2.4 s,机器人在围墙跑道内摆臂、反复抬脚并向前移动,第一道横栏始终仍在其前方;未见任何横栏由前方移至身后。" + }, + { + "t": 2.65, + "log_step": null, + "text_zh": "观察:视频 2.45–2.65 s,机器人躯干倾斜、骨盆下降,随后倒向地面;视频 2.70–2.90 s 躯干和骨盆贴近地面,未重新站起。" + }, + { + "t": null, + "log_step": 133, + "text_zh": "观察:log step 131 生成最终交付的 287 步动作,log step 133 与 log step 139 重放该交付,均报告 deterministic=true、scored=false、sim_steps=287,末态哈希一致;inputs/run.json 的评分器字段 success=0、all_success=0、deterministic=1。推断:结合未到第一栏即倒地的画面,跨栏跑达标节点从未达到。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–2.4 s,机器人双脚在跑道地面落地承重并推动身体前进;视频 2.65–2.90 s 失衡后躯干、骨盆落到地面。" + }, + { + "who": "腿/躯干", + "how": "避开", + "note_zh": "观察:视频 0.0–2.90 s,机器人沿跑道内部移动,身体与两侧墙保持可见间隔,未见触墙;没有到达或跨过前方横栏,故不记避开横栏。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "累计奖励", + "threshold": 700, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "hurdle_run_completed", + "t": 2.5, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频 2.4 s 机器人仍在第一道横栏前行进;视频 2.5–2.65 s 身体倾倒、骨盆下落,视频 2.70–2.90 s 倒地未恢复,跨栏跑未完成。" + }, + "unmet": { + "hurdle_run_completed": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 2.6, + "log_step": null, + "evidence_zh": "观察:视频 2.50–2.65 s,机器人双腿向前伸出、躯干向地面倾倒,失去直立支撑;视频 2.70–2.90 s 身体贴近地面,未起身,前方横栏仍未越过。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "推断:最终步态控制未能维持前进中的身体稳定。log step 100 展示生成脚本按步态时钟和骨盆姿态反馈计算关节目标,log step 127 固定参数,log step 131 重新生成最终 287 步动作,log step 133、139 验证相同末态;对应评分视频 2.4–2.7 s,机器人在横栏前的空地行进时失衡倒地,未见障碍物阻挡。现有记录支持控制失败,但不足以区分接触设计、轨迹、反馈或具体参数问题。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "无状态日志,无法核实足底接触力、关节跟踪误差及具体失稳机制,控制根因仅标通用项。评分记录只给出 final_reward=1.50049780347006e-15,未明确报告 episode_return 或该字段与累计奖励的等价关系,故原生累计奖励值留 null;未用 agent 自述的运动时长替代视频时钟。" + }, + "summary_zh": "机器人已起步并在跑道内前进,但始终未到达第一道横栏,“跨栏跑达标”最终为 0/1。视频约 2.5–2.7 s 机器人失衡倒地,至视频 2.90 s 未恢复;日志确认最终 287 步动作可确定性重放且评分未通过,证据支持通用控制失败。无状态日志及明确累计奖励数值,无法细分失稳机制或还原奖励构成。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "节点未达成、倒地约2.5 s、平衡现象与通用控制根因均与视频一致;时钟与 contact_actual 无误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:42:40Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-privileged--jap2yhy/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-privileged--jap2yhy/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "cf85697fdbaa89887b15522568dfc8d18cc9874d6dffe6f40d3dd634a65689ae", + "video_sha256": "44000706496b78e732c8f1115cbdc0f871349e172f958fa42278e6e44a7eb3dc", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-privileged--mit3ixh.json b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-privileged--mit3ixh.json new file mode 100644 index 0000000000000000000000000000000000000000..4eacb83cfb6d3ac2206e26631b71c435ab86a462 --- /dev/null +++ b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-privileged--mit3ixh.json @@ -0,0 +1,201 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-hurdle-i00-privileged--mit3ixh", + "taskKey": "humanoidbench/loco_hurdle", + "dag_sha256": "ca09b7dc923221773af0078329063e8e11b1544a0f305c56d17055d1bf47910e", + "outcome": { + "success": 0, + "reward": 0.0035784129888161457, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 155, + "final_reward": 0.0035784129888161457, + "sim_steps": 155, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-hurdle-i00-privileged--mit3ixh/replay.mp4", + "duration_s": 1.65, + "fps": 20.0, + "frames": 33, + "sha256": "d324c84e4693ed84942317f3c5b9ad40c27ff8264f928dc3a0a0fe9698c1e920" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "hurdle_run_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.3, + "log_step": null, + "text_zh": "观察:视频 0.00–1.60 s 的全部33帧中,机器人摆腿、迈步向前移动,第一道横栏始终在机器人前方,没有横栏从其前方移至身后的越栏过程。" + }, + { + "t": 1.6, + "log_step": null, + "text_zh": "观察:视频 1.20–1.60 s 中机器人腿部屈曲、身体高度下降,最后仍未到第一栏,未见躯干倒地;inputs/run.json 的评分器字段 success=0、all_success=0、deterministic=1、n_actions=155。" + }, + { + "t": null, + "log_step": 240, + "text_zh": "观察:agent 在 log step 240 将155步动作从初态重新录制并保存为最终 trajectory.npz;log step 242 确认文件动作形状为(155,19)。该交付与评分器的155步一致,不能把会话中继续试跑后的跌倒算入本次视频。" + }, + { + "t": null, + "log_step": null, + "text_zh": "推断:结合视频中只完成第一栏前的短距离起步和评分器 success=0,单一累计奖励达标节点未达到;评分器未提供 episode_return,不能从 final_reward 或 agent 的代理分数恢复累计奖励。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–1.60 s 中机器人左右脚先后接触跑道地面,承担起步、迈步和落脚支撑;未到横栏处,没有实际越栏接触可记。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "累计奖励", + "threshold": 700, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "hurdle_run_completed", + "t": 1.3, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频约 1.30 s 机器人在第一栏前继续迈腿时身体下沉,至视频 1.60 s 结束仍未越栏;这段交付仅含起步前进,跨栏跑步骤始终未完成,评分器 success=0。" + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 1.3, + "log_step": 239, + "evidence_zh": "观察:视频 1.10–1.60 s 中机器人在第一栏前迈腿后屈腿下沉,末段腿向前伸、骨盆降低,呈坐塌姿态,未越过第一栏。log step 239 对155步候选从初态重放,输出骨盆高度0.383 m、done=False;log step 240 的保存代码将该候选录制为最终交付,log step 242 确认动作形状为(155,19),评分器报告 n_actions=155、deterministic=1。推断:结合末段画面和对应候选的重放读数,机器人已失去有效承重支撑,归为平衡丧失;未见躯干倒地、done=False 均不足以排除此现象,且不能据此声称已触发跌倒终止。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "log_step": 239, + "evidence_zh": "推断:agent 的步行与落地控制未形成可持续前进的运动。log step 222–224 对最终轨迹所用前缀试验多组姿态反馈续接,均很快下沉并终止;随后优化的交付候选在 log step 239 的155步重放检查中仍为骨盆高度0.383 m、done=False,log step 240 保存该候选。视频 1.20–1.60 s 同样显示末段身体下沉且未到第一栏,因此支持控制能力不足;会话续接试验中的跌倒不作为本次交付的现象。", + "source_zh": "日志" + } + ], + "gap_zh": "输入没有逐帧状态日志,无法核实支撑力、关节跟踪误差或进一步区分控制子类;评分器只给 final_reward=0.0035784129888161457,未报告 episode_return,log step 239 的34.694是 agent 自定义代理分数,不能作为700门槛对应的累计奖励。跌倒终止的精确阈值未知;log step 239 对最终交付所用155步候选的重放报告 done=False,视频末段支持失去支撑,但未见躯干倒地,不能断言已触发跌倒终止。会话续接试验中的倒地不作为本次交付的现象。", + "unmet": { + "hurdle_run_completed": false + } + }, + "summary_zh": "机器人完成了起步和短距离前进,但视频 0.00–1.60 s 始终未越过第一栏,“跨栏跑达标”节点未达到,最终0/1。断点在跨栏跑操作中,末段屈腿坐塌、失去有效承重支撑,现象为平衡丧失;交付候选的重放读数和会话中的续接、落地试验支持控制问题。无逐帧状态日志和实际累计奖励,跌倒终止阈值未知,不能细分控制机制、认定已触发跌倒终止或用代理分数代替评分器累计奖励。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 2, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "fix", + "conf": 6, + "summary_zh": "节点未达到、final=0、断点指向 hurdle_run_completed 均成立。问题在现象:视频末段机器人屈腿坐塌、骨盆降至约0.38 m(log step 239 对交付轨迹的重放读数),标签自己也写“身体下沉”,却记为“步态”;TAXONOMY 定义步态是“没摔倒”的行进不足,失去支撑属平衡。", + "n_issues": 1, + "n_major": 1 + }, + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 2, + "verdict": "agree", + "conf": 7, + "summary_zh": "节点未达到(视频 1.65 s 内未到第一栏,评分器 success=0)、final=0、断点指向 hurdle_run_completed、平衡现象与控制根因均有画面和 log 支撑;未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:04:43Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-privileged--mit3ixh/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-privileged--mit3ixh/04_review_r1" + }, + { + "stage": "reconcile", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-privileged--mit3ixh/05_reconcile_r1" + }, + { + "stage": "traj_review", + "round": 2, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-privileged--mit3ixh/04_review_r2" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "d8ad19b518f9dd68bfd922ce365e54dec108ebf23f918f5e5ae2497a7e3acf6d", + "video_sha256": "d324c84e4693ed84942317f3c5b9ad40c27ff8264f928dc3a0a0fe9698c1e920", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-privileged--tqvmuf3.json b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-privileged--tqvmuf3.json new file mode 100644 index 0000000000000000000000000000000000000000..f9e6ddc5d79ac0a73211ead0a30503c6b5eb669e --- /dev/null +++ b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-privileged--tqvmuf3.json @@ -0,0 +1,173 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-hurdle-i00-privileged--tqvmuf3", + "taskKey": "humanoidbench/loco_hurdle", + "dag_sha256": "ca09b7dc923221773af0078329063e8e11b1544a0f305c56d17055d1bf47910e", + "outcome": { + "success": 0, + "reward": 7.796299497721263e-16, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 158, + "final_reward": 7.796299497721263e-16, + "sim_steps": 158, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 5.299, + "peak_processes": 25, + "peak_disk_gb": 0.239, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-hurdle-i00-privileged--tqvmuf3/replay.mp4", + "duration_s": 1.65, + "fps": 20.0, + "frames": 33, + "sha256": "46604b8b01db1ba480737e12185f3401989deb8193961419eefa862e8b461676" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "hurdle_run_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.3, + "log_step": null, + "text_zh": "观察:视频 0.00–0.90 s,机器人在跑道内迈步,躯干先明显后仰再前摆;首个横栏始终在机器人前方,未见越栏。" + }, + { + "t": 0.95, + "log_step": null, + "text_zh": "观察:视频 0.95–1.10 s,机器人躯干向前扑低;视频 1.20–1.60 s,机器人继续翻转,末帧躯干贴地、双腿弯曲,未恢复跨栏跑。" + }, + { + "t": 1.6, + "log_step": 132, + "text_zh": "观察:log step 130–132记录交付文件包含158个动作,重新生成动作一致,两次重放最终状态哈希均为42cfe9212304;评分器报告n_actions=158、sim_steps=158、deterministic=1、success=0、final_reward=7.796299497721263e-16。推断:结合视频中未越栏且倒地的过程,跨栏跑达标节点未达到;原生整段得分远低于700门槛。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.90 s,机器人双脚在跑道地面上承重、换步;视频 1.20–1.60 s翻倒后躯干落地。全段未见机器人跨过或碰到横栏,故不沿用任务方案中的越栏避开接触。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": 7.796299497721263e-16, + "unit": "累计奖励", + "threshold": 700, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "hurdle_run_completed", + "t": 0.95, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频 0.95 s机器人在首栏前迈步时躯干急剧前倾,视频 1.05 s上身接近地面,随后翻转,至视频 1.60 s末帧仍倒地;该步已开始但未完成。" + }, + "unmet": { + "hurdle_run_completed": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 0.95, + "log_step": null, + "evidence_zh": "观察:视频 0.95–1.10 s机器人失去直立支撑、向前扑低,视频 1.20–1.60 s翻转倒地,未恢复站立;全程尚未越过首栏。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "观察:log step 120、130记录agent用足步规划与关节位置目标控制器生成并保存158步动作,log step 131–132核验动作一致、双次重放状态一致;视频 0.30–0.55 s机器人明显后仰,视频 0.95–1.60 s前扑翻倒。推断:交付动作的动态稳定性不足,无法维持跨栏跑所需支撑;证据支持控制层归因,但不足以确定具体反馈、逆解或参数子类。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "输入没有状态日志或逐控制步奖励、关节跟踪误差及接触力记录,无法确定失稳对应的具体控制子类,也无法从画面精确恢复跌倒终止阈值。节点达标判定结合评分器最终整段得分与视频,不把会话搜索中的距离、速度或跌倒步数当作本段视频测量。" + }, + "summary_zh": "评分轨迹中机器人在首栏前迈步,未见越过横栏,“跨栏跑达标”最终为0/1,评分器原生得分约7.80e-16,低于700。视频约0.95 s机器人前扑失衡,随后翻转倒地;日志核验交付动作重放一致,支持动态稳定性不足的控制层归因。缺少状态、接触力和跟踪误差记录,无法进一步定位控制子类。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "视频仅1.65 s,机器人未越栏、很快失衡倒地,节点未达、final=0、评分器return≈7.8e-16均与帧一致;断点、平衡现象与控制根因有证据支持。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:42:54Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-privileged--tqvmuf3/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-privileged--tqvmuf3/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "f0fdbaf6936da77d238b5e955a215ff2e0fdb86326751f8d1a004595e10027a8", + "video_sha256": "46604b8b01db1ba480737e12185f3401989deb8193961419eefa862e8b461676", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-privileged--z93uelt.json b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-privileged--z93uelt.json new file mode 100644 index 0000000000000000000000000000000000000000..c5795af4ce15d33d32cf2df895c15ad8608d2cc7 --- /dev/null +++ b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-privileged--z93uelt.json @@ -0,0 +1,212 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-hurdle-i00-privileged--z93uelt", + "taskKey": "humanoidbench/loco_hurdle", + "dag_sha256": "ca09b7dc923221773af0078329063e8e11b1544a0f305c56d17055d1bf47910e", + "outcome": { + "success": 0, + "reward": 5.712020360063439e-17, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 260, + "final_reward": 5.712020360063439e-17, + "sim_steps": 256, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-hurdle-i00-privileged--z93uelt/replay.mp4", + "duration_s": 2.65, + "fps": 20.0, + "frames": 53, + "sha256": "a5a0ab1d89caa98868d2974da621342e6ffd6ebc044e388d97c52c529e13377b" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "hurdle_run_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.0–2.1 s,机器人从直立姿态沿两侧墙之间的跑道向第一道横栏前进,双脚交替着地;视频 2.2–2.6 s,机器人在横栏附近倾倒并落地,未继续连续跨栏跑。" + }, + { + "t": 2.6, + "log_step": 39, + "text_zh": "观察:视频 2.6 s 机器人仍倒在地面;inputs/run.json 的评分器字段 success=0、all_success=0、deterministic=1、n_actions=260、sim_steps=256。log step 37 和 log step 39 的交付文件重放均在控制步 256 结束、状态哈希一致,支持视频对应最终交付,而非其他会话试验。" + }, + { + "t": null, + "log_step": 33, + "text_zh": "推断:结合视频中首栏附近倒地、评分器 success=0 与 DAG 的累计奖励700验收规则,“跨栏跑达标”未达到。log step 33 的67.05是 agent 脚本计算的代理累计奖励,不是评分器原生累计奖励,不能用它填 episode_return。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–2.1 s 机器人的双脚交替接触跑道承重并向前行进;视频 2.4–2.6 s 机器人倒地后躯干落在跑道上。" + }, + { + "who": "腿/躯干", + "how": "避开", + "note_zh": "观察:视频 0.0–2.6 s 机器人沿跑道中部移动及倒地,身体与左右侧墙保持可见间隔,未退回起点后墙。此项仅指避开墙体,不表示成功避开第一道横栏。" + }, + { + "who": "腿/躯干", + "how": "其他", + "note_zh": "观察:视频 2.2–2.3 s 机器人的脚/小腿与第一道横栏的横杆重叠,身体随后向前倾倒并落地。推断:脚/小腿与横杆发生碰撞并绊阻,具体腿侧不可辨;候选接触方式为“碰撞”,属于意外接触阻断,现有支撑、推/扫、压/擦等方式均不能准确表达,故用其他。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "累计奖励", + "threshold": 700, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "hurdle_run_completed", + "t": 2.2, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频 2.1 s 机器人行进到第一道横栏附近;视频 2.2–2.4 s 躯干逐渐倾倒、骨盆下降,随后身体落地,至视频 2.6 s 未恢复站立或继续前进。失败发生在执行跨栏跑过程中。" + }, + "unmet": { + "hurdle_run_completed": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 2.2, + "log_step": null, + "evidence_zh": "观察:视频 2.2 s 起机器人在第一道横栏附近失去直立支撑,视频 2.35–2.6 s 躯干接近水平并倒在地面,未能以双脚支撑继续跨栏。" + }, + { + "cat": "运动", + "item": "被挡/碰撞", + "t": 2.3, + "log_step": null, + "evidence_zh": "观察:视频 2.2–2.3 s 机器人的脚/小腿在第一道横栏处与横杆重叠,腿部未完整抬过横杆,身体随后向前倾倒,视频 2.4–2.6 s 躯干落地而脚部仍邻近横栏。推断:该连续动作显示越栏时受到横杆接触阻断并绊倒;具体哪条腿发生接触无法分辨。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "推断:最终交付的关节目标序列未能在第一道横栏处维持支撑并完成越栏动作,属于运动控制失败。依据是视频 2.1–2.6 s 的失衡倒地,以及 log step 33 的脚本输出“committed prefix fell”、log step 37 和 log step 39 对同一260行动作文件的确定性重放确认;log step 38 中 agent 的说明也自述未找到能清栏的摆腿轨迹。证据不足以进一步区分轨迹求解、接触设计或反馈控制。", + "source_zh": "视频、日志、agent 自述" + } + ], + "gap_zh": "无状态日志或接触对记录;视频 2.2–2.3 s 的脚/小腿与横杆重叠及随后的前倾倒地支持碰撞绊阻,但斜俯视低分辨率画面不能可靠确认具体哪条腿、接触力或精确关节误差,因此不细分控制根因。评分器仅报告 success=0 和 final_reward=5.712020360063439e-17,未明确提供 episode_return;不能将 final_reward 或 agent 的代理累计奖励67.05当作原生累计奖励,故该指标留空。所有事件时间直接取自视频,未将会话时间或控制步数换算成视频时间。" + }, + "summary_zh": "机器人已起步并沿跑道前进到第一道横栏附近,但在视频约 2.2–2.3 s 脚/小腿受到横杆绊阻并失衡,至视频 2.6 s 倒地且未恢复,“跨栏跑达标”最终为0/1。最终交付的动作序列经日志确认确定性重放,失败表现为平衡丧失及被挡/碰撞,根因保守标为控制;缺少接触及关节状态日志,无法确认具体碰撞腿侧及控制子类型,评分器也未明确报告原生累计奖励。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 2, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "fix", + "conf": 6, + "summary_zh": "整体(未达标、首栏附近倒地、断点在 hurdle_run_completed、根因保守标控制)成立;但逐帧可见机器人脚/小腿在视频约 2.2–2.3 s 撞上横栏后向前绊倒,标签却称无法确认碰撞,漏标被挡/碰撞,contact_actual 的“避开”也与此不符。", + "n_issues": 2, + "n_major": 1 + }, + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 2, + "verdict": "agree", + "conf": 8, + "summary_zh": "标签成立:视频 1.8–2.6 s 逐帧可见脚/小腿撞上第一道横栏后向前绊倒;节点未达成、断点 hurdle_run_completed、平衡+被挡/碰撞、根因控制(保守)均有证据,时间钟清晰,上一轮问题已修正。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:08:31Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-privileged--z93uelt/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-privileged--z93uelt/04_review_r1" + }, + { + "stage": "reconcile", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-privileged--z93uelt/05_reconcile_r1" + }, + { + "stage": "traj_review", + "round": 2, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-privileged--z93uelt/04_review_r2" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "cb8305449713249ffb2ccd6a95228e6ca42f14d7a37d765a2db07e2cd8abedb3", + "video_sha256": "a5a0ab1d89caa98868d2974da621342e6ffd6ebc044e388d97c52c529e13377b", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard--cmii8wf.json b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard--cmii8wf.json new file mode 100644 index 0000000000000000000000000000000000000000..b30df03248ad4eb2065ee13852c89d3548a9ec60 --- /dev/null +++ b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard--cmii8wf.json @@ -0,0 +1,179 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-hurdle-i00-standard--cmii8wf", + "taskKey": "humanoidbench/loco_hurdle", + "dag_sha256": "ca09b7dc923221773af0078329063e8e11b1544a0f305c56d17055d1bf47910e", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 107, + "sim_steps": 16212, + "calls": 1894, + "rows": 16212, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 450.8, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-hurdle-i00-standard--cmii8wf/replay.mp4", + "duration_s": 1.15, + "fps": 20.0, + "frames": 23, + "sha256": "649e9bb1129e52a784a4bf1849879d9954647f102889087c9a7326a70822c76b" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "hurdle_run_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00–0.60 s,机器人在首栏前屈腿、摆动身体和双臂,脚仍在起点附近,未见横栏从身前移到身后。" + }, + { + "t": 0.9, + "log_step": 47, + "text_zh": "观察:视频 0.65–0.90 s,机器人躯干持续倾倒、骨盆下降,视频 0.90–1.10 s 已倒在跑道地面且未恢复;首栏始终在前方。log step 47 的服务状态为最后一个 episode=51、control_steps=107、done=true、reason=failed、resets_left=0;评分字段 live_success=0、replay_success=0,未达到本节点的累计奖励验收。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.60 s,机器人双脚接触跑道地面承重,屈腿并调整支撑;视频 0.90–1.10 s,失稳后躯干及下肢落在地面。未见腿或躯干越过横栏。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "累计奖励", + "threshold": 700, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "hurdle_run_completed", + "t": 0.65, + "log_step": 47, + "where": "操作中", + "evidence_zh": "观察:视频 0.60 s 机器人仍以双脚支撑,视频 0.65–0.85 s 躯干倾倒且骨盆迅速下沉,视频 0.90–1.10 s 倒地,首栏一直未通过。log step 47 确认最终评分 episode=51 已以 failed 结束,随后 log step 48 确认重置预算耗尽,未再修复此跨栏跑尝试。" + }, + "unmet": { + "hurdle_run_completed": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 0.65, + "log_step": 47, + "evidence_zh": "观察:视频 0.65–0.90 s,机器人屈腿后失去直立支撑,躯干向地面翻倒、骨盆塌下;视频 0.90–1.10 s 躯干及下肢留在地面,未重新站起。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志", + "evidence_zh": "推断:最终控制脚本存在可直接核对的动作构造错误,足以使所调步态与实际指令不一致。log step 4 的服务规格说明 joint_position 动作须按 range_rad 线性归一化,log step 6 已正确计算归一化站姿;但 log step 46 的最终脚本在 reset 后直接取以弧度表示的 proprio[:19] 作为 base,未转换即叠加步态并提交,例如初始约 0.8 rad 的膝角被当作约 0.8 的归一化目标。该脚本又令 p=dict(FAST),却用 p is FAST 判断是否回退,条件恒为假;log step 47 运行此脚本后最后一局失败,视频 0.65–1.10 s 的失稳倒地与控制未能维持支撑一致。缺少逐步状态记录,归因停在控制层,不将某个增益或回退失效单独认定为唯一致因。" + } + ], + "gap_zh": "未提供状态日志,评分字段也未报告 episode_return 数值,因此不能恢复逐步奖励、精确接触力或最终脚本各错误对失稳的相对影响;原生累计奖励填 null。视频时间直接取所检帧,未将 agent 时间或控制步计数换算为视频时间。" + }, + "summary_zh": "最后评分局中机器人只在首栏前屈腿调整支撑,未越过任何横栏,跨栏跑达标节点为 0/1;视频约 0.65 s 开始倾倒,视频 0.90–1.10 s 倒地,服务在 log step 47 确认最后一局失败。最终脚本将弧度关节角直接当作归一化动作基准,并有恒不触发的回退分支,根因标为控制;早期重置掉的试验不计入进度。缺少状态日志与累计奖励数值,无法分离各控制错误的具体动力学贡献。", + "conf": 9, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "节点未达成、断点在 hurdle_run_completed、平衡现象与控制根因均有帧和日志支持。我核对了 log step 4 与最终 gen.py:最终脚本直接以弧度 proprio 作 base,而此前实验都用归一化的 stand_target,推断成立。未发现需要修改的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:43:15Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-standard--cmii8wf/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-standard--cmii8wf/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "0a514e4acbb2b7f4d3cc19fe231534e2cfb741e3f7cc026edae3b47394cc9737", + "video_sha256": "649e9bb1129e52a784a4bf1849879d9954647f102889087c9a7326a70822c76b", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard--h8pcbpw.json b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard--h8pcbpw.json new file mode 100644 index 0000000000000000000000000000000000000000..f3bd9168cba11c9ee58d313a5853aba3fcbe4878 --- /dev/null +++ b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard--h8pcbpw.json @@ -0,0 +1,184 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-hurdle-i00-standard--h8pcbpw", + "taskKey": "humanoidbench/loco_hurdle", + "dag_sha256": "ca09b7dc923221773af0078329063e8e11b1544a0f305c56d17055d1bf47910e", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 227, + "sim_steps": 11834, + "calls": 11896, + "rows": 11834, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3535.1, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-hurdle-i00-standard--h8pcbpw/replay.mp4", + "duration_s": 2.35, + "fps": 20.0, + "frames": 47, + "sha256": "16a7c9e18cd07ed411370734c3ec7a17a0ba58a75f248d42226e49d2d314e742" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "hurdle_run_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.8, + "log_step": 173, + "text_zh": "观察:视频 0.0–1.4 s,机器人双脚交替落地并沿跑道向第一栏前进;log step 173 启动最后一次重置后的 final 控制程序,log step 175 的服务状态确认该次为第51局。此前重置掉的测试局不计入本轨迹。" + }, + { + "t": 1.7, + "log_step": null, + "text_zh": "观察:视频 1.45–1.65 s,机器人在第一栏处抬腿跨越;视频 1.70–1.80 s,躯干迅速倾倒、头部及上身降向地面;视频 1.85–2.30 s 持续倒地翻动双腿,没有恢复站立继续跑。" + }, + { + "t": 2.3, + "log_step": 175, + "text_zh": "推断:log step 175 的服务状态为 episode=51、control_steps=227、done=true、reason=failed;评分字段 live_success=0、replay_success=0、deterministic=1,且视频末段确实倒地。任务标准模式达标即结束,因此该局未达到DAG所定累计奖励700,跨到第一栏之后不能视为跨栏跑达标;节点全程未成立。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–1.65 s,机器人左右脚在跑道地面交替承重、起跑及跨步;视频 1.80–2.30 s,机器人躯干倒在跑道上,双腿继续翻动。横栏与腿部是否发生碰撞无法由该视角可靠分辨。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "累计奖励", + "threshold": 700, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "hurdle_run_completed", + "t": 1.7, + "log_step": 175, + "where": "操作中", + "evidence_zh": "观察:视频 1.65 s 机器人仍在第一栏处跨步,视频 1.70–1.80 s 身体快速翻倒,随后直到视频 2.30 s 末帧都未重新站立;log step 175 确认最后第51局以 failed 结束。失败发生在跨栏跑进行中,未有后续修复。" + }, + "unmet": { + "hurdle_run_completed": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 1.7, + "log_step": null, + "evidence_zh": "观察:视频 1.70 s 机器人躯干倾斜、支撑失稳,视频 1.75–1.85 s 上身倒向地面;视频 1.90–2.30 s 机器人横卧在跑道上并翻动腿部,未恢复双足站立。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "推断:log step 173 的最终控制程序采用跑步策略、近栏模型预测控制,并加入由躯干姿态和角速度驱动的髋部及肩部平衡补偿;log step 174–175 记录的是同一次 final 运行。视频 0.0–1.60 s 可见机器人完成助跑,但视频 1.70–1.85 s 在第一栏处失稳倒地,说明该控制组合未能维持越栏过程的全身稳定;证据不足以进一步归到反馈增益、模型误差或足部接触设计。", + "source_zh": "日志" + } + ], + "gap_zh": "未提供仿真状态日志;已检查完整47帧,第一栏附近腿与横栏在斜俯视画面中重叠,无法可靠确定是否绊栏、具体接触部位及力,因此不额外标被挡/碰撞,也不细分控制根因。评分字段未报告 episode_return,原生指标保留null,二元失败奖励不能当作累计奖励0。日志中的程序仿真时钟与视频时长不同,未将其数值换作视频时间;所有视频锚点取自所检查帧。" + }, + "summary_zh": "机器人完成助跑并在第一栏处尝试跨越,约视频 1.70–1.80 s 失衡倒地,直到视频 2.30 s 末帧都未恢复,跨栏跑达标节点最终为0/1。最后第51局的服务与重放均判失败,根因保守归为越栏控制未维持全身稳定。缺少状态接触数据和累计奖励数值,无法确认是否绊栏或进一步细分控制失效机制。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "标签与视频一致:机器人助跑,约视频1.5–1.7 s在第一栏处越栏,1.75–1.9 s起倒地并至2.30 s未恢复;节点未达成、断点、平衡现象与保守的控制根因均有帧和日志支持,未发现需改动的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:05:31Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-standard--h8pcbpw/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-standard--h8pcbpw/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "227238e40d0890bf253e93ac791cc1beaaa6fb48637c3a417c484f021068dc87", + "video_sha256": "16a7c9e18cd07ed411370734c3ec7a17a0ba58a75f248d42226e49d2d314e742", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard--hu23eab.json b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard--hu23eab.json new file mode 100644 index 0000000000000000000000000000000000000000..6c6e004a023d0f616ca0338a2a0c35ff843f1e4b --- /dev/null +++ b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard--hu23eab.json @@ -0,0 +1,189 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-hurdle-i00-standard--hu23eab", + "taskKey": "humanoidbench/loco_hurdle", + "dag_sha256": "ca09b7dc923221773af0078329063e8e11b1544a0f305c56d17055d1bf47910e", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 975, + "sim_steps": 16706, + "calls": 1481, + "rows": 16706, + "resets": 46, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3588.4, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-hurdle-i00-standard--hu23eab/replay.mp4", + "duration_s": 9.85, + "fps": 20.0, + "frames": 197, + "sha256": "7da1e02105e1b3e62144493496d04ab2c1b557959d3fec145093eaed6cbb835a" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "hurdle_run_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 412, + "text_zh": "观察:视频0.00 s机器人在跑道起点附近直立,首栏仍在前方;log step 412记录agent重置并开始最后一次运行,随后log step 413、415–420持续调用步行控制。此前试跳属于被重置的会话尝试,不计入本段进度。" + }, + { + "t": 2.9, + "log_step": null, + "text_zh": "观察:视频1.80–2.80 s机器人屈腿、身体逐渐降低;视频2.90–3.00 s双腿伸向前方,骨盆坐落地面,失去正常站立支撑。首栏始终在机器人前方,没有横栏移到身后的越栏过程。" + }, + { + "t": 9.8, + "log_step": null, + "text_zh": "观察:视频3.00–9.80 s机器人一直坐地摆腿,未恢复站立,末秒逐帧检查仍如此;inputs/run.json评分字段success、live_success、replay_success均为0。" + }, + { + "t": 9.8, + "log_step": 420, + "text_zh": "推断:本节点以累计奖励达到700验收;评分判失败,且视频全程未完成前进跨栏、坐地后未恢复,因此节点从未达成,终值为false。log step 420服务状态为episode 47、control_steps 975、done false,只表示该局当时未结束,不能证明机器人直立或任务达标。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频0.00–2.80 s机器人的脚接触跑道承重并尝试迈步;视频约2.90 s后骨盆坐落跑道地面,腿继续摆动直至视频9.80 s。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "累计奖励", + "threshold": 700, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "hurdle_run_completed", + "t": 2.9, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频2.80–3.00 s机器人在首栏前尝试迈步时下坐,双脚不再维持站立支撑;此后至视频9.80 s均未起身,跨栏跑未完成。" + }, + "unmet": { + "hurdle_run_completed": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 2.9, + "log_step": null, + "evidence_zh": "观察:视频2.90–3.00 s机器人骨盆降至地面、双腿向前伸出,随后视频3.00–9.80 s维持坐地姿态并摆腿,未恢复足部站立支撑。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "推断:log step 412–414显示agent在最后一次重置后执行站姿保持、交替步行和短暂平衡控制,log step 415–420继续相同步行指令;对应的最终视频中机器人在首栏前就于视频约2.90 s坐地,继续动作仍不能起身,说明所用步行控制未能维持承重和平衡。输入不能进一步区分轨迹生成、反馈控制或具体参数的问题。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "未提供状态日志或逐步累计奖励,episode_return未报告,不能精确恢复奖励数值、接触力或跌倒终止阈值;控制根因保留到大类。agent在log step 417自述“仍直立”,但视频3.00–9.80 s显示坐地,本次物理进度及现象以视频为准;不将该自述或服务done false当成直立证据。" + }, + "summary_zh": "最终受评运行尝试迈步,但未越过首栏,“跨栏跑达标”节点最终为0/1。机器人在视频约2.90 s失去站立支撑并坐地,直到视频9.80 s仍未恢复,失败现象为平衡,根因可归为控制。没有状态日志及累计奖励数值,无法细分控制问题;agent自述仍直立与视频不符,采用视频判断。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "视频帧显示机器人起步后很快后仰下坐,视频约2.9 s骨盆落地,之后一直坐地摆腿至9.8 s,跨栏节点未达成、断点与平衡现象、控制根因均有证据支持,未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:04:44Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-standard--hu23eab/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-standard--hu23eab/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "c599d13a785f76a04511ed98c83a9b0c2b9df611a9a022a53347287487b415e6", + "video_sha256": "7da1e02105e1b3e62144493496d04ab2c1b557959d3fec145093eaed6cbb835a", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard--kedt5y9.json b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard--kedt5y9.json new file mode 100644 index 0000000000000000000000000000000000000000..465f2e5d6cdb761efe3eefd14d465da0da88d3b0 --- /dev/null +++ b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard--kedt5y9.json @@ -0,0 +1,182 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-hurdle-i00-standard--kedt5y9", + "taskKey": "humanoidbench/loco_hurdle", + "dag_sha256": "ca09b7dc923221773af0078329063e8e11b1544a0f305c56d17055d1bf47910e", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 1000, + "sim_steps": 4209, + "calls": 408, + "rows": 4109, + "resets": 30, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 936.9, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.324, + "peak_processes": 11, + "peak_disk_gb": 1.03, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-hurdle-i00-standard--kedt5y9/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "4a89b63b5fa27ebbe3103820a4e972edc6073aeea7b8aa0f1ffac3f868f524dc" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "hurdle_run_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 101, + "text_zh": "观察:log step 101 的最后一局为 episode 31,agent 脚本将髋、膝、踝及髋外展摆动幅度和姿态反馈增益设为0,持续发送起始站姿目标;服务返回 control_steps=1000、done=true、reason=\"failed\"。评分字段 n_actions=1000、resets=30、deterministic=1,与该最后一局一致。" + }, + { + "t": 0.5, + "log_step": null, + "text_zh": "观察:视频 0.00–1.10 s 逐帧可见机器人起始姿态略有后倾调整,双脚仍在起点附近承重,未启动交替向前迈步;视频 1.5–9.5 s 每隔约0.5 s检查,机器人与前方首栏和地面格线的相对位置基本不变,无越栏。" + }, + { + "t": 10.05, + "log_step": 101, + "text_zh": "观察:视频 9.00–10.05 s 逐帧检查,机器人仍站在首栏之前,未倒地,也未前进越栏;inputs/run.json 的 success、live_success、replay_success 均为0,且未给出 episode_return 数值。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:结合全段未起跑的画面和评分器失败判决,累计奖励达到700的节点从未成立;没有把站立或早先测试局的动作记为跨栏跑完成。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "累计奖励", + "threshold": 700, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "hurdle_run_completed", + "t": 0.0, + "log_step": 101, + "where": "首步前", + "evidence_zh": "观察:视频 0.00–10.05 s 机器人只在起点附近维持支撑站姿,首栏一直在其前方,没有前进起跑或越栏。log step 101 明确将步态摆动幅度置0,最后一局持续至1000控制步后服务报告 failed;跨栏跑这一步没有启动,视频 0.0 s 作为导航锚点。" + }, + "unmet": { + "hurdle_run_completed": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 101, + "evidence_zh": "观察:视频 0.00–1.10 s 仅有站姿调整,随后至视频 10.05 s 始终停在首栏之前,没有交替前进或跳跃。log step 101 的评分局参数将 hpA、knA、anA、hrA、asA 全部设为0,只保持起始站姿;这不是一段已起跑但速度不足的轨迹。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "推断:会话中的步态控制尝试未建立稳定前进能力,最终退回站立基线。log step 93 的三组步态测试均由服务报告 failed,结束时 IMU 重力分量 gx 约0.97–0.99、gz约0.06–0.10;log step 97 的三组俯仰反馈测试也全部 failed;log step 99 的小幅步态在128控制步后 failed。随后 log step 101 改为全零摆动的站姿,视频 0.00–10.05 s 与此一致。上述测试只用于解释最终未起跑的控制能力缺口,不把测试局的倾倒当作评分局的现象;证据不足以细分到具体反馈项或参数。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "没有状态日志,评分字段未报告 episode_return,无法恢复累计奖励曲线或各奖励分量。控制根因依据会话测试的服务反馈及最终控制代码推断,无法确定具体控制失稳机制;评分视频本身只显示站立基线,没有测试局的跌倒。" + }, + "summary_zh": "评分采用最后的 episode 31:机器人在视频 0.00–10.05 s 仅维持起点附近的站姿,没有启动前进或越栏,跨栏跑达标节点未达到(0/1)。断点为跨栏跑首步前的没尝试;会话日志中多组步态与反馈测试失败后,agent 在 log step 101 关闭全部步态摆动,保守归因为控制能力不足。没有状态日志或累计奖励数值,具体控制机制及奖励分量无法确认。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "评分局(episode 31)全程站姿、未起跑越栏,节点未达、断点首步前/没尝试、根因控制(来自会话测试证据,符合 FAIL-22 的正确写法)均有证据支持;未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:45:37Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-standard--kedt5y9/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-standard--kedt5y9/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "00a50669711fff5774c587aa98464fe57faaaa26184d2aa41925f5a67cfd4f38", + "video_sha256": "4a89b63b5fa27ebbe3103820a4e972edc6073aeea7b8aa0f1ffac3f868f524dc", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard--pxdrvhe.json b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard--pxdrvhe.json new file mode 100644 index 0000000000000000000000000000000000000000..8cc9e5ad8b49dc88e8bb5bade3b22c8b8663559c --- /dev/null +++ b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard--pxdrvhe.json @@ -0,0 +1,187 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-hurdle-i00-standard--pxdrvhe", + "taskKey": "humanoidbench/loco_hurdle", + "dag_sha256": "ca09b7dc923221773af0078329063e8e11b1544a0f305c56d17055d1bf47910e", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 382, + "sim_steps": 3100, + "calls": 192, + "rows": 3100, + "resets": 25, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3159.1, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-hurdle-i00-standard--pxdrvhe/replay.mp4", + "duration_s": 3.9, + "fps": 20.0, + "frames": 78, + "sha256": "426472c512363f75feede8d1db1d93de4e308a3b99073647e0adebe202b2763c" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "hurdle_run_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.3, + "log_step": 55, + "text_zh": "观察:视频 0.0–0.3 s,机器人在起点附近交替抬腿、摆臂;第一道横栏仍在远处前方,没有横栏从机器人前方移到身后。log step 55 启动最终实时回合,log step 56 的服务状态确认评分回合为 episode 26。" + }, + { + "t": 0.7, + "log_step": null, + "text_zh": "观察:视频 0.4 s 机器人开始明显后仰,视频 0.5–0.7 s 躯干接近水平并跌落地面;视频 0.8–3.85 s 持续在地面摆腿、翻动,未重新站立,也未越过第一栏。" + }, + { + "t": 3.85, + "log_step": 56, + "text_zh": "推断:log step 55–56 中服务报告最终回合 control_steps=382、done=True、reason=failed,评分字段 live_success=0、graded_by_live=1;结合视频全程未越栏且早期倒地,跨栏跑累计奖励达700的事件未发生。评分未提供 episode_return 数值,不能把目录中的 reward=0.0 当作实际累计奖励。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–0.35 s 机器人双脚交替着地承重并抬腿;视频约 0.65–3.85 s 倒地后躯干和腿接触跑道地面。机器人未到达横栏,不据任务典型方案追加越栏避开接触。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "累计奖励", + "threshold": 700, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "hurdle_run_completed", + "t": 0.4, + "log_step": 55, + "where": "操作中", + "evidence_zh": "观察:视频 0.0–0.35 s 机器人已开始交替抬腿和摆臂;视频 0.4–0.7 s 后仰倒地,之后至视频 3.85 s 一直未恢复站立或越栏,故断点为正在执行的跨栏跑步骤。" + }, + "unmet": { + "hurdle_run_completed": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 0.4, + "log_step": null, + "evidence_zh": "观察:视频 0.4 s 机器人躯干向后倾斜,视频 0.5–0.7 s 双腿抬起、躯干落地,视频 0.8–3.85 s 持续躺地摆腿;该失衡发生于尚未到达第一栏时。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "推断:log step 50 的 gen.py 以双腿反相的关节正弦目标驱动机器人,并用骨盆IMU重力方向和角速度修正髋、踝目标;整批动作生成后才读取新本体感知。log step 55 的最终运行调用此控制器,chunk=25;视频 0.4–0.7 s 起步即后仰倒地,随后持续摆腿而未恢复支撑。该执行结果支持步态和姿态控制未能维持平衡;没有足够证据区分轨迹、反馈或具体参数哪个主导。", + "source_zh": "日志" + } + ], + "gap_zh": "没有状态日志或逐步奖励、接触力与关节跟踪记录,episode_return 未报告,不能确定控制失稳的具体参数或接触动力学机制。agent 在 log step 57 自述“行走382控制步后跌倒”,但视频 0.7 s 已倒地,之后至视频 3.85 s 均在地面;以画面判定跌倒时刻,不将382控制步或agent工作时间直接当作视频时间。" + }, + "summary_zh": "观察:评分视频中机器人起步抬腿后,在视频约 0.4 s 开始后仰、约 0.7 s 倒地,此后至视频 3.85 s 始终未恢复站立,未越过第一栏,最终节点为0/1。推断:最终实时回合的服务失败判决与画面一致,断点是跨栏跑步骤,现象为平衡失败,日志中的步态控制器未能维持支撑。观察:输入没有状态日志或实际累计奖励,无法细分控制根因;agent 关于持续行走的最终自述不符画面。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "标签与视频、log一致:视频 0.4–0.7 s 后仰倒地,之后躺地至 3.85 s,未越栏,节点未达成;断点、平衡现象和控制根因有证据支持,未把 382 控制步换算成视频时间。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:05:22Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-standard--pxdrvhe/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-standard--pxdrvhe/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "4bf59cb3312acf32c9d89230218fb532a34d1aeda54005e854fdafead8d5e565", + "video_sha256": "426472c512363f75feede8d1db1d93de4e308a3b99073647e0adebe202b2763c", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard--wcrvgen.json b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard--wcrvgen.json new file mode 100644 index 0000000000000000000000000000000000000000..26b09af45d82d37d5caa6eefcb1faed4dac7f7ab --- /dev/null +++ b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard--wcrvgen.json @@ -0,0 +1,191 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-hurdle-i00-standard--wcrvgen", + "taskKey": "humanoidbench/loco_hurdle", + "dag_sha256": "ca09b7dc923221773af0078329063e8e11b1544a0f305c56d17055d1bf47910e", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 493, + "sim_steps": 14037, + "calls": 14113, + "rows": 14037, + "resets": 49, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1292.2, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-hurdle-i00-standard--wcrvgen/replay.mp4", + "duration_s": 5.0, + "fps": 20.0, + "frames": 100, + "sha256": "6dad68047a0adb7b45cd279aa2162098b43209837f5ec84a1820248c5c3f4ce1" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "hurdle_run_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 2.75, + "log_step": null, + "text_zh": "观察:视频 0.0–2.75 s,机器人双脚交替支撑并沿跑道前进,接近第一道横栏;视频 2.90–3.15 s,其腿脚处在栏杆处,躯干随即侧倾。" + }, + { + "t": 3.4, + "log_step": 75, + "text_zh": "观察:视频 3.40 s 机器人躯干已落地,视频 4.00–4.95 s 仍在地面摆动四肢,未恢复站立或继续跨栏跑;log step 75 的最后一次实时运行输出 DONE failed、control_steps 为493,与评分记录 n_actions=493、live_success=0、replay_success=0 一致。" + }, + { + "t": 4.95, + "log_step": 75, + "text_zh": "推断:该节点以累计奖励达到700为验收;结合本次视频中的第一道栏处倒地、最后一次实时运行的失败状态以及评分器 success=0,跨栏跑达标事件未发生。评分器未报告 episode_return,不能把目录 reward=0.0 当作累计奖励。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–2.90 s,机器人左右脚交替接触跑道地面承重前进;视频约 3.40 s 后,躯干倒在跑道地面上,由地面承托。" + }, + { + "who": "腿/躯干", + "how": "其他", + "note_zh": "推断:候选“碰撞”。视频 2.95–3.15 s,机器人一侧腿脚与第一道横栏相接,腿脚在栏处受阻而躯干继续侧翻,随后倒地;具体左右侧与接触连杆不能可靠区分。该交互为越栏时的非预期碰撞,不属于抓握、推扫或成功避开等现有方式。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "累计奖励", + "threshold": 700, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "hurdle_run_completed", + "t": 3.05, + "log_step": 75, + "where": "操作中", + "evidence_zh": "观察:视频 2.90–3.15 s,机器人正在第一道横栏处抬腿通过,躯干开始侧倾并失去直立支撑;视频约 3.40 s 倒地,直至视频 4.95 s 未恢复跨栏跑。log step 75 是最后一次实时运行,输出 DONE failed,评分器 live_success 和 replay_success 均为0。" + }, + "unmet": { + "hurdle_run_completed": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 3.05, + "log_step": 75, + "evidence_zh": "观察:视频 3.00–3.30 s,机器人在第一道横栏处向侧方翻倒,身体由直立变为近水平;视频 3.40 s 躯干落地,视频 4.00–4.95 s 虽仍摆腿、撑动身体,但未重新站起。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "t": 3.05, + "log_step": 75, + "source_zh": "日志", + "evidence_zh": "推断:agent 的控制程序在 log step 55 按相机估计的位置切换高抬腿模式,在 log step 70 为该模式改变步态相位速度;log step 75 选定这套控制参数并实际执行最后一次实时运行,返回 DONE failed。结合视频 2.90–3.40 s 到达横栏后侧翻落地,所执行的步态和姿态控制未能稳定完成越栏;证据只支持控制层面,不能确定是轨迹、反馈还是某个参数不当。" + } + ], + "gap_zh": "没有状态日志或逐帧接触对、执行器力和目标跟踪误差,横栏接触由腿脚在栏处受阻及随后翻倒的连续画面推断,不能确定具体连杆或失稳的控制子因。评分记录未提供 episode_return,agent 在 log step 77 自述的约70奖励不是评分器实测值;本标签不据此填入原生指标,也不将其“700不可达”的判断作为任务或环境根因。" + }, + "summary_zh": "机器人沿跑道跑到第一道横栏处,但视频约 3.05 s 开始侧翻,约 3.40 s 倒地,此后至视频 4.95 s 未恢复站立;跨栏跑达标节点最终为0/1。最后一次实时运行(log step 75)和评分器均判失败,现象为平衡丧失,根因保留在控制层面。缺少状态日志与累计奖励实测值,不能进一步确定失稳的控制子因。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "节点0/1、断点 hurdle_run_completed(操作中)、现象全身·平衡、根因控制(日志)与视频一致:机器人在视频约 3.1 s 脚碰栏杆后前倾倒地,之后未站起。接触与推断均标注为推断,未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:45:14Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-standard--wcrvgen/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-standard--wcrvgen/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "c135c25b16dcca28e645b8bb9877fb56320329c9e6972858fe06b0de4006cf3e", + "video_sha256": "6dad68047a0adb7b45cd279aa2162098b43209837f5ec84a1820248c5c3f4ce1", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard--y8ukgsj.json b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard--y8ukgsj.json new file mode 100644 index 0000000000000000000000000000000000000000..7933d885dd769314e249b29ee48864b8cc423812 --- /dev/null +++ b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard--y8ukgsj.json @@ -0,0 +1,202 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-hurdle-i00-standard--y8ukgsj", + "taskKey": "humanoidbench/loco_hurdle", + "dag_sha256": "ca09b7dc923221773af0078329063e8e11b1544a0f305c56d17055d1bf47910e", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 220, + "sim_steps": 421, + "calls": 425, + "rows": 421, + "resets": 1, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3478.8, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-hurdle-i00-standard--y8ukgsj/replay.mp4", + "duration_s": 2.3, + "fps": 20.0, + "frames": 46, + "sha256": "22166f346dd858ba66dc4b9454c210ee9fe8fd5d18b0297ac4208eb8835cab56" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "hurdle_run_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 57, + "text_zh": "观察:视频 0.00–2.25 s 的全部46帧显示机器人在跑道内交替迈腿向第一道横栏前进;视频 1.25–2.25 s 仍未越过第一栏,末帧躯干保持竖直,未见倒地。log step 57 的脚本重置为 episode 2,执行220个实时控制步后正常退出循环,输出没有 Done 或 EpisodeOver;之后日志没有继续向实时服务提交动作。" + }, + { + "t": 2.25, + "log_step": null, + "text_zh": "推断:评分器字段 live_success=0、replay_success=0,结合视频全程未越过第一栏,本次“跨栏跑达标”未达到。累计奖励700的验收不能由站立或少量前进代替;评分字段未报告 episode_return,不能把本地仿真得分作为该段累计奖励。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–2.25 s,机器人的左右脚交替落在跑道地面承重并推进身体;末帧仍由脚部支撑。未见腿部接触或越过第一道横栏,不将任务级避栏方案计为已发生接触。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "累计奖励", + "threshold": 700, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "hurdle_run_completed", + "t": null, + "log_step": 57, + "where": "操作中", + "evidence_zh": "观察:视频 0.00–2.25 s 的全部46帧显示机器人交替迈腿前进,始终在第一栏前,未见越栏、碰栏或倒地。log step 57 重置为 episode 2,脚本以 range(220) 限定实时动作,输出 Finished 220 steps 并正常退出,未输出 Done 或 EpisodeOver;log step 58–102 未再调用实时动作提交,评分器 n_actions=220、live_success=0、replay_success=0。推断:已执行的跨栏跑步骤未达标,评分动作内容止于 log step 57 的平地步态测试;log step 103 的过滤消息没有切断正在执行的控制循环。没有单一可见的物理失败事件,故视频断点时刻保留 null。" + }, + "phenomena": [ + { + "cat": "全身", + "item": "步态", + "t": null, + "log_step": 57, + "evidence_zh": "观察:视频 0.00–2.25 s 全程机器人只以交替迈腿的步态向第一栏前进,横栏始终在前方,没有越栏动作;视频 1.25–2.25 s 仍未越过第一栏,末帧未倒地。log step 57 的实时脚本完成220步平地步态即退出,评分器 live_success=0、replay_success=0。推断:本次已执行运动的距离和内容不足以完成跨栏跑,符合 TAXONOMY_v2.md §3 中未倒地而行走、跑跳未达到步骤要求的“步态”,不将离线测试中的碰栏或跌倒记为本次现象。" + } + ], + "root_causes": [ + { + "cat": "认知", + "item": "推理/规划", + "log_step": 57, + "source_zh": "日志", + "evidence_zh": "观察:log step 56 测试的是 best_flat_theta,代码将带栏测试明确标为 no jump yet;log step 57 的实时脚本加载同一平地步态,仅按周期摆腿及姿态反馈生成动作,没有根据栏位切换起跳、收腿或落地控制,且只安排220步。log step 68、77、82、87、91、95、98 的越栏策略开发均在自建本地仿真中执行;直到 log step 101–102 仍只测试本地策略和打印参数,log step 58–102 没有实时动作提交。推断:评分执行与越栏控制的衔接策略不完整,实际提交只覆盖平地测试,任务所需的越栏控制始终未进入评分回合;归因于跨阶段的推理/规划,而非把停止或耗时本身当作根因。日志未直接记录最终保留该策略的意图,此归因为执行安排与代码所支持的推断。" + } + ], + "gap_zh": "观察:输入没有状态日志,评分器未报告 episode_return,无法给出该段精确累计奖励;视频没有单一可定位的跌倒或碰栏失败事件。log step 103 是会话末尾的模型提供方过滤消息,log step 57 的实时控制循环在它之前已经正常退出,之后没有续交实时动作。推断:依据 docs/03_trajectory.md Step 3.5 及 DECISIONS.md D37 的当前口径,对已提交并在 log step 58 检视的评分内容作常规失败分析;不因会话末尾过滤而清空两层。规划根因由实时脚本与后续离线开发的衔接缺失推断,日志未直接记录最终策略选择的意图;本地仿真回报和物理事件不能替代本次评分证据。", + "unmet": { + "hurdle_run_completed": false + } + }, + "summary_zh": "评分视频中机器人已起步并向第一栏前进,但视频 0.00–2.25 s 全程未越栏、未倒地,“跨栏跑达标”最终为假(0/1),失败现象为步态。log step 57 仅提交220步平地闭环测试,此后越栏控制开发始终停留在离线仿真,推断评分执行与越栏策略衔接不完整,根因为推理/规划;log step 103 的会话末尾过滤不作为本次外因截断。没有状态日志或该段累计奖励,规划意图和失败时刻也未直接记录,因此将置信度由8降为7。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 2, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "fix", + "conf": 6, + "summary_zh": "评分回合(episode 2)在 log step 57(agent 时间约1751 s)以220步平地步态跑完后即被agent自行放弃,之后约1700 s只做离线仿真;log step 103 的Gemini过滤中断发生在评分回合早已定型之后,并没有切断任何进行中的评分动作。把它当外因截断、现象与根因全部留空,不支持。视频只显示平地步态向栏前行进,没有任何越栏动作。", + "n_issues": 1, + "n_major": 1 + }, + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 2, + "verdict": "agree", + "conf": 7, + "summary_zh": "第2轮复核:视频 0–2.25 s 全程机器人仅以平地步态在第一栏前行进,未越栏未倒地,final=false、断点 hurdle_run_completed、步态现象与推理/规划根因(已标明为推断)与证据一致;上一轮的外因截断问题已修正,未发现新错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:07:19Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-standard--y8ukgsj/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-standard--y8ukgsj/04_review_r1" + }, + { + "stage": "reconcile", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-standard--y8ukgsj/05_reconcile_r1" + }, + { + "stage": "traj_review", + "round": 2, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-standard--y8ukgsj/04_review_r2" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "a778c1a4f467ba8229f7833b07fd595cc88cbc6cb26d7a73c3f1fbced776b656", + "video_sha256": "22166f346dd858ba66dc4b9454c210ee9fe8fd5d18b0297ac4208eb8835cab56", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard--zkhxakw.json b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard--zkhxakw.json new file mode 100644 index 0000000000000000000000000000000000000000..73ea485874e2429d958e0d7cb213455735229947 --- /dev/null +++ b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard--zkhxakw.json @@ -0,0 +1,184 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-hurdle-i00-standard--zkhxakw", + "taskKey": "humanoidbench/loco_hurdle", + "dag_sha256": "ca09b7dc923221773af0078329063e8e11b1544a0f305c56d17055d1bf47910e", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 548, + "sim_steps": 9498, + "calls": 1761, + "rows": 9498, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 2237.9, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-hurdle-i00-standard--zkhxakw/replay.mp4", + "duration_s": 5.55, + "fps": 20.0, + "frames": 111, + "sha256": "e870c005db1350f5e8ec2f24159b56b146235674f25e9e002df319648d8c950c" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "hurdle_run_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 381, + "text_zh": "观察:视频 0.0–4.8 s 中,机器人交替屈膝、抬脚和落脚,逐渐转向,但前方第一道横栏一直没有移到机器人身后;视频 5.2–5.5 s 中机器人倾倒并贴地。log step 381 明确重置到最后的 Episode 51,服务返回 control_steps=548、reason=failed;log step 382 确认该局已结束。" + }, + { + "t": 5.5, + "log_step": 382, + "text_zh": "推断:本节点验收是累计奖励达到700;本次评分字段 success、live_success 和 replay_success 均为0,且最后一局服务报告 failed,结合视频 5.5 s 的倒地终态,判定该事件未达成。评分未给出 episode_return,不能将目录中的归一化 reward=0.0 当作累计奖励值。" + }, + { + "t": 5.5, + "log_step": 387, + "text_zh": "观察:agent 在 log step 387 自述最后一局始终保持直立,但视频 5.3–5.5 s 清楚显示躯干横倒、骨盆降到地面;节点与失败现象按视频判定。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–5.1 s 中机器人左右脚交替接触跑道地面承重;视频 5.3–5.5 s 失衡后躯干及腿部落到地面。未观察到机器人越过横栏,故不照抄任务方案的越栏避开接触。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "累计奖励", + "threshold": 700, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "hurdle_run_completed", + "t": 5.2, + "log_step": 381, + "where": "操作中", + "evidence_zh": "观察:视频 5.1–5.2 s 中,仍在迈腿的机器人身体开始明显倾斜,随后视频 5.3–5.5 s 躯干横倒、骨盆落地,没有恢复站立;前方横栏尚未越过。log step 381 的最后一局 Episode 51 以 failed 结束,log step 382 确认 done=True。" + }, + "unmet": { + "hurdle_run_completed": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 5.2, + "log_step": 381, + "evidence_zh": "观察:视频 5.2 s 起机器人躯干倾斜迅速加大,双脚未能继续提供直立支撑,视频 5.3–5.5 s 身体横倒并落地;最后未起身。该失衡发生在跨栏跑动作执行中。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "推断:log step 379 的最终控制脚本以周期腿部关节位置目标驱动行进,并配置髋、踝和横滚反馈增益;log step 381 显示该脚本实际执行到最后一局,过程中姿态与转向读数变化,最终服务返回 failed。视频 4.8–5.5 s 显示机器人在尚未到达横栏时迈腿、倾倒并落地,支持行进中的全身稳定控制未能维持支撑;现有证据不足以在足端轨迹、反馈律或增益之间确定更细根因。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "没有状态日志,也没有逐步累计奖励、接触力或足底接触记录;评分字段未报告 episode_return,因此该指标为null。log step 379 中最终脚本的部分正文已被输入日志截断,不能完整复核其反馈实现;可确认跌倒和控制层失败,不能确定具体错误参数或触地机制。所有视频时刻均取自本次视频帧,未将agent时间或控制步数换算为视频时间。" + }, + "summary_zh": "最后一局机器人已开始交替迈腿,但视频 0.0–5.5 s 未见越过第一道横栏,“跨栏跑达标”最终为0/1,服务与评分均判失败。断点在视频约5.2 s 的跨栏跑操作中,机器人失衡倒地,证据支持控制层失败。没有状态日志和累计奖励数值,且最终控制脚本在输入日志中部分截断,无法确认具体失稳机制。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "节点未达成、断点(视频约5.2 s 失衡倒地)、平衡现象与控制根因均与视频帧和 log step 381/382 一致,未见需要修改的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:45:53Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-standard--zkhxakw/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-standard--zkhxakw/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "45790943474815d569d12580fc5353a7275dbdd14f5284c28b76317d5a5fec0c", + "video_sha256": "e870c005db1350f5e8ec2f24159b56b146235674f25e9e002df319648d8c950c", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard-realtime--3w2hhvp.json b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard-realtime--3w2hhvp.json new file mode 100644 index 0000000000000000000000000000000000000000..efedab3b48fa67500e5a2ed317ad47ea12735032 --- /dev/null +++ b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard-realtime--3w2hhvp.json @@ -0,0 +1,200 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-hurdle-i00-standard-realtime--3w2hhvp", + "taskKey": "humanoidbench/loco_hurdle", + "dag_sha256": "ca09b7dc923221773af0078329063e8e11b1544a0f305c56d17055d1bf47910e", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 90, + "sim_steps": 14789, + "calls": 7798, + "rows": 8274, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1938.5, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 294.922, + "max_lag_s": 0.0955, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 7628, + "late_steps": 0, + "requests_refused": 48, + "request_busy_s": 4.061, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-hurdle-i00-standard-realtime--3w2hhvp/replay.mp4", + "duration_s": 1.0, + "fps": 20.0, + "frames": 20, + "sha256": "bf2051a011c5ce6a57045e71bf1c6ab69798fb0424d6432804f09528be63d6c4" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "hurdle_run_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.25, + "log_step": null, + "text_zh": "观察:已逐帧检查视频0.00–0.95 s;机器人在起点附近屈伸双腿并摆臂,脚接触跑道地面,第一道横栏始终在其前方,未见越栏。" + }, + { + "t": 0.75, + "log_step": null, + "text_zh": "观察:视频0.65–0.85 s机器人躯干逐渐倾斜、骨盆下降,视频0.90–0.95 s倒在跑道上,未恢复站立。" + }, + { + "t": null, + "log_step": 339, + "text_zh": "观察:log step 339记录最终控制器启动Episode 51,服务返回control_steps=90、done=True、reason=failed;log step 340再次确认该结果。评分器success、live_success、replay_success均为0,未报告episode_return。" + }, + { + "t": 0.95, + "log_step": 340, + "text_zh": "推断:视频全程未见到达横栏,末段已倒地;结合最终评分局失败结果,累计奖励达700的事件未发生,跨栏跑达标节点为false。成功参考视频中可见机器人持续前进并让横栏移至身后,仅作为视觉判据对照。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频0.00–0.65 s双脚在跑道地面承重并随屈伸动作改变支撑;视频0.85–0.95 s机器人倒下,躯干贴近并落到地面。全程未到横栏,不记越栏避开接触。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "累计奖励", + "threshold": 700, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "hurdle_run_completed", + "t": 0.7, + "log_step": 339, + "where": "操作中", + "evidence_zh": "观察:视频0.00–0.65 s机器人已开始屈伸双腿的起步动作;视频约0.70 s起躯干失稳倾斜,视频0.90 s骨盆降至地面并倒下,仍未到第一道横栏。log step 339–340确认最后的Episode 51以failed结束。" + }, + "unmet": { + "hurdle_run_completed": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 0.7, + "log_step": null, + "evidence_zh": "观察:视频0.70–0.85 s机器人上身倾倒、双腿无法维持直立支撑;视频0.90–0.95 s机器人躺倒在跑道地面,横栏仍在前方。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "推断:log step 338的最终脚本用周期性髋、膝、踝目标驱动双腿屈伸,并根据重力方向和角速度修正髋踝、双膝及躯干;log step 339实际运行该脚本,服务以failed结束,未记录控制脚本崩溃。视频0.70–0.95 s显示这些控制动作未能维持身体支撑,且机器人尚未接近横栏;据此归于运动稳定控制失败,无法进一步区分轨迹、反馈或参数问题。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "无状态日志,评分字段未报告episode_return,故原生指标值保留null;缺少最终评分局的连续姿态、关节跟踪及接触力数据,不能确定具体失稳机制或参数。视频时间直接来自逐帧检查,未将log step 339的90个控制步或agent工作时间换算为视频时间;此前50次重置中的测试结果不作为本次节点进度。" + }, + "summary_zh": "最终评分局Episode 51只在起点附近屈伸双腿并尝试起步,未越过横栏,跨栏跑达标节点最终0/1。机器人在视频约0.70 s开始失稳并于视频0.90 s倒地,现象为全身·平衡,结合最终控制脚本归因于控制失败。无状态日志且未报告累计奖励,无法进一步确定失稳机制或具体参数。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "节点未达成(0/1)、视频0.7 s起失稳并在0.9 s倒地、断点与全身·平衡现象、控制类根因均有帧和服务读数(control_steps=90, reason=failed)支持,未发现需改动的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:06:28Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-standard-realtime--3w2hhvp/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-standard-realtime--3w2hhvp/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "bd0866646bfc0f1399388e39041530cbb52032de1e2e9094b33280fa67984e24", + "video_sha256": "bf2051a011c5ce6a57045e71bf1c6ab69798fb0424d6432804f09528be63d6c4", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard-realtime--kidexu7.json b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard-realtime--kidexu7.json new file mode 100644 index 0000000000000000000000000000000000000000..6c4991212cf63307759fcf887ff5cd45079adfe0 --- /dev/null +++ b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard-realtime--kidexu7.json @@ -0,0 +1,205 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-hurdle-i00-standard-realtime--kidexu7", + "taskKey": "humanoidbench/loco_hurdle", + "dag_sha256": "ca09b7dc923221773af0078329063e8e11b1544a0f305c56d17055d1bf47910e", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 295, + "sim_steps": 931, + "calls": 3858, + "rows": 3606, + "resets": 1, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 2114.8, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 18.584, + "max_lag_s": 0.0011, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 0, + "late_steps": 0, + "requests_refused": 1, + "request_busy_s": 1.198, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.21, + "peak_processes": 13, + "peak_disk_gb": 0.289, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-hurdle-i00-standard-realtime--kidexu7/replay.mp4", + "duration_s": 3.05, + "fps": 20.0, + "frames": 61, + "sha256": "46e487813ca81d6d9464b30406135a9fbad31913114f93846b9f7d8d0ed44df3" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "hurdle_run_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.0–2.35 s,机器人由起点沿跑道向首道横栏跑动,双脚交替落地;视频 2.4–2.55 s 下肢到达横栏位置,随后视频 2.6–2.9 s 身体翻倒,视频 3.0 s 仍躺在地面,未形成连续越栏跑。" + }, + { + "t": null, + "log_step": 76, + "text_zh": "观察:log step 76 的实时服务在重置后执行 episode 2,报告 control_steps=295、done=True、reason=failed;log step 102 再次读取仍为同一失败的 episode 2。评分记录 live_success=0、replay_success=0、deterministic=1,与视频首栏处倒地一致。" + }, + { + "t": null, + "log_step": null, + "text_zh": "推断:本任务以累计奖励达到700验收,实时模式达到门槛即结束;结合服务失败判决及视频末尾倒地,本节点未达到,不能把已跑向首栏算作完成。评分字段未报告 episode_return,不能用 success=0 代替累计奖励数值。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.3–2.35 s,机器人左右脚交替接触跑道地面承重并向首道栏前进;视频 2.9–3.0 s 翻倒后躯干落在地面。" + }, + { + "who": "腿/躯干", + "how": "其他", + "note_zh": "推断:视频 2.4–2.55 s,机器人下肢与首道横栏相交,腿未顺畅清过栏而身体继续前移,随后翻倒,支持腿部碰栏;单视角不能精确区分脚或小腿的接触。候选名称“碰撞”:这是意外撞上障碍,不属于抓握、推/扫、压/擦等主动接触方式。" + }, + { + "who": "腿/躯干", + "how": "避开", + "note_zh": "观察:视频 0.0–3.0 s,机器人沿两侧长墙之间的跑道前进并在首栏处倒地,身体与左右墙保持可见间隔;这里仅标避开侧墙,不把碰到的横栏标为避开。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "累计奖励", + "threshold": 700, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "hurdle_run_completed", + "t": 2.6, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频 2.4–2.55 s,机器人正在通过首道横栏;视频 2.6 s 躯干明显向前下方翻转,视频 2.65–2.85 s 双腿翻到身体上方,视频 2.9–3.0 s 倒地且未恢复。失败发生于跨栏跑本身,log step 76 的实时服务报告该次 episode 2 以 failed 结束。" + }, + "unmet": { + "hurdle_run_completed": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 2.6, + "log_step": null, + "evidence_zh": "观察:视频 2.55–2.85 s,机器人从直立跑动变为躯干向下翻转、双腿向上扬起,失去双脚的正常承重支撑;视频 2.9–3.0 s 躯干倒在跑道上,未再站起。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "推断:log step 75 的实际运行脚本采用 Gen 54 周期摆腿、髋膝踝目标和姿态反馈,log step 76 确认执行该脚本后失败;log step 69 的本地步态优化把横栏移到地下,log step 72 将横栏加入该步态测试后已出现首栏附近跌倒。结合视频 2.4–3.0 s 首栏处未能保持支撑并翻倒,支持实际跑步控制未能实现越栏与稳定落地;本地测试只用于解释控制缺陷,不计入评分进度。证据不足以进一步指定某个增益、逆解或反馈环节。", + "source_zh": "日志" + } + ], + "gap_zh": "没有状态日志或碰撞对、接触力记录;视频为256×256单一跟随视角,视频 2.4–2.55 s 的精确碰栏部位只能推断。评分字段和所检日志未报告 episode_return,原生累计奖励留空;无法拆分轨迹净空、落地反馈及具体参数各自的影响。视频时间直接取自帧,不将控制步、agent用时或本地仿真时间换算为视频时间。" + }, + "summary_zh": "机器人沿跑道跑到首道横栏,视频 2.6 s 左右在越栏过程中失去平衡并翻倒,视频 3.0 s 仍倒地,跨栏跑达标节点最终为0/1。log step 76 与评分记录确认第二次实时执行失败,后续本地越栏优化未再次执行到评分 episode,不能算作修复;实际运行的周期跑步控制未能完成稳定越栏。没有状态日志、精确接触对或累计奖励数值,碰栏部位及更细控制根因仍不能确定。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "视频 0–3.0 s 机器人跑到首栏,约 2.4–2.5 s 腿贴近横栏,2.6 s 起向前翻倒并留在地面;节点 0/1、断点在 hurdle_run_completed、平衡现象与控制根因均有帧和评分字段支持,未发现需要修改的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:07:31Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-standard-realtime--kidexu7/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-standard-realtime--kidexu7/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "d2c18024cc72d65691dc3fd87d3220a811e8d26ed64c7a9eb1f64c715ad950b5", + "video_sha256": "46e487813ca81d6d9464b30406135a9fbad31913114f93846b9f7d8d0ed44df3", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard-realtime--kwcvj28.json b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard-realtime--kwcvj28.json new file mode 100644 index 0000000000000000000000000000000000000000..1f0c6d6366a7bef95a348041d3f2b3b33181e042 --- /dev/null +++ b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard-realtime--kwcvj28.json @@ -0,0 +1,202 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-hurdle-i00-standard-realtime--kwcvj28", + "taskKey": "humanoidbench/loco_hurdle", + "dag_sha256": "ca09b7dc923221773af0078329063e8e11b1544a0f305c56d17055d1bf47910e", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 351, + "sim_steps": 21643, + "calls": 19144, + "rows": 19011, + "resets": 47, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1747.6, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 432.146, + "max_lag_s": 0.0315, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 3334, + "late_steps": 0, + "requests_refused": 28, + "request_busy_s": 8.809, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.307, + "peak_processes": 15, + "peak_disk_gb": 0.283, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-hurdle-i00-standard-realtime--kwcvj28/replay.mp4", + "duration_s": 3.6, + "fps": 20.0, + "frames": 72, + "sha256": "1503ece0abc17e5d762caa60ae1ff8aa001386567658e5bc08e243a1ff2dcb46" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "hurdle_run_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.0–2.8 s,机器人从直立起步,在两侧长墙之间交替抬腿向第一道横栏前进。" + }, + { + "t": 3.15, + "log_step": null, + "text_zh": "观察:视频 2.85–3.15 s,第一道横栏从机器人腿部附近移到身后,机器人继续向前但躯干明显倾斜;视频 3.35–3.55 s,机器人躯干贴地,未恢复站立。" + }, + { + "t": 3.55, + "log_step": 89, + "text_zh": "观察:log step 89 的机器人服务状态为 episode=48、control_steps=351、done=True、reason=failed;inputs/run.json 的评分字段 success=0、live_success=0、replay_success=0。推断:该末局未达到累计奖励700的验收门槛;越过首栏不等于跨栏跑达标。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–3.15 s,机器人的左右脚交替落在跑道地面承担支撑;视频 3.35–3.55 s,躯干倒在地面上。" + }, + { + "who": "腿/躯干", + "how": "避开", + "note_zh": "观察:视频 0.0–3.55 s,机器人在跑道内部前进及倒地,机体与左右长墙保持可见间隔。首栏处腿部与栏杆局部重叠,无法据此确认是否完全无擦碰。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "累计奖励", + "threshold": 700, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "hurdle_run_completed", + "t": 3.1, + "log_step": 89, + "where": "操作中", + "evidence_zh": "观察:视频 3.0–3.2 s,机器人在跨过第一道横栏并继续前进时躯干倾斜、骨盆下降,视频 3.35 s 已倒地,此后至视频 3.55 s 末帧未恢复;log step 89 确认最后第48局以 failed 结束。" + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 3.1, + "log_step": null, + "evidence_zh": "观察:视频 3.0–3.25 s,机器人躯干逐渐倾斜、双腿未能继续维持直立支撑,视频 3.35–3.55 s 躯干落在地上;这是失去支撑后的倒地。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "t": 3.1, + "log_step": 89, + "evidence_zh": "推断:agent 的相位步态及姿态修正控制未能维持越栏后的动态支撑稳定。log step 6、16、19、49 记录其用步态相位生成腿部目标,并按骨盆IMU读数修正髋、踝及航向;log step 81、84、89 记录最终控制器参数与末局启动,log step 89 的服务确认该局 failed;视频 3.0–3.55 s 显示该控制下机器人失衡倒地。现有记录不足以确定是哪一增益、模型误差或接触扰动所致。", + "source_zh": "日志" + } + ], + "gap_zh": "本目录没有状态日志,视频仅为256×256,无法确定首栏处有无细小擦碰、精确接触力或具体失稳参数;因此不另标碰撞,也不细分控制根因。评分字段未报告 episode_return,原生指标留null。本标注时间均取自已检查的视频帧,未将agent会话时间或其末局时长自述换算为视频时间。", + "unmet": { + "hurdle_run_completed": false + } + }, + "summary_zh": "最后第48局中,机器人向前跑并越过第一道横栏,但未达到“跨栏跑达标”,节点最终为0/1。视频约 3.1 s 机器人失衡,视频 3.35–3.55 s 已倒地;结合最终控制器记录,根因归为控制。缺少状态日志及累计奖励数值,无法确定细小栏杆擦碰或具体失稳参数。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "节点未达成、倒地发生在视频约3.1–3.35 s、断点与平衡/控制归因均与帧一致;256×256 画面无法判定首栏是否绊脚,标注已如实保留不确定性。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:05:34Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-standard-realtime--kwcvj28/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-standard-realtime--kwcvj28/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "0ed700af74a377ba9bf19b7c5a8b7f14eca263f167cb3e35d7d8780536133b7d", + "video_sha256": "1503ece0abc17e5d762caa60ae1ff8aa001386567658e5bc08e243a1ff2dcb46", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard-realtime--r3towv3.json b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard-realtime--r3towv3.json new file mode 100644 index 0000000000000000000000000000000000000000..118e85bdecb03b2224b8a0964153dc812a2fa335 --- /dev/null +++ b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard-realtime--r3towv3.json @@ -0,0 +1,201 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-hurdle-i00-standard-realtime--r3towv3", + "taskKey": "humanoidbench/loco_hurdle", + "dag_sha256": "ca09b7dc923221773af0078329063e8e11b1544a0f305c56d17055d1bf47910e", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 1000, + "sim_steps": 33172, + "calls": 33421, + "rows": 33280, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1836.0, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 662.555, + "max_lag_s": 0.0253, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 3285, + "late_steps": 0, + "requests_refused": 46, + "request_busy_s": 13.455, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.324, + "peak_processes": 10, + "peak_disk_gb": 0.836, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-hurdle-i00-standard-realtime--r3towv3/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "ca4e8be657413cb39c4ef76b8f3dd1e3cdf8bfcdcf3602fb833e19caa857fbc8" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "hurdle_run_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.5, + "log_step": null, + "text_zh": "观察:视频 0.0–0.3 s 机器人双脚站在起点附近;视频 0.4–1.2 s 机器人抬腿、屈膝并逐渐后倾塌下,前方第一道横栏始终没有移到机器人身后。" + }, + { + "t": 2.8, + "log_step": null, + "text_zh": "观察:视频 2.8–3.5 s 机器人已倒在跑道上,双腿抬起;视频 4.0–10.0 s 的抽帧及视频 9.0–10.0 s 的密集检查均显示其停在原处倒地,未恢复站立或越栏。" + }, + { + "t": null, + "log_step": 234, + "text_zh": "观察:log step 231 重置后,log step 233 的服务状态确认最后一次评分局为 episode 51、resets_left=0;log step 234 服务返回 control_steps=1000、reason=failed,评分字段 success=0、live_success=0、replay_success=0。" + }, + { + "t": 10.0, + "log_step": 234, + "text_zh": "推断:结合视频中未越过第一道横栏、倒地后没有恢复以及最后评分局的失败判决,唯一目标“跨栏跑达标”从未达到;评分记录未报告 episode_return,不能将二元 reward=0.0 当作累计奖励数值。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–0.8 s 机器人双脚在起点附近接触跑道地面并承重;视频 2.8–10.0 s 倒地后由躯干接触地面承重。未见机器人到达或接触横栏。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "累计奖励", + "threshold": 700, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "hurdle_run_completed", + "t": 0.8, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频 0.4–1.2 s 机器人在起点附近抬腿时躯干后倾、骨盆下降,未能形成连续向前跑动;视频 2.8 s 已倒地,直到视频 10.0 s 仍未恢复,前方横栏始终未被越过。" + }, + "unmet": { + "hurdle_run_completed": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 0.8, + "log_step": null, + "evidence_zh": "观察:视频 0.4–1.2 s 机器人屈膝后倾并失去直立支撑,视频 2.8–3.5 s 已倒地、双腿抬起;视频 9.0–10.0 s 仍保持倒地姿态。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "log_step": 233, + "evidence_zh": "推断:log step 230 的最终配置使用交替髋膝步态、IMU反馈及平衡姿态回退;log step 233 的最终局控制器输出中,倾斜量由控制步225的0.67增至控制步300的0.85,gate降至0,髋膝位置读数随后停在配置的回退姿态附近。结合视频 0.4–2.8 s 的塌下和此后持续倒地,控制器未能在起步换腿时维持或恢复支撑,归为控制失败;现有证据不足以确认某个反馈符号、增益或关节参数是直接原因。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "未提供状态日志,评分记录也未报告 episode_return,无法给出实际累计奖励或精确接触力;日志中的部分控制器修改只保留了调用摘要,不能进一步区分控制子类或证明某一参数错误。视频时间直接取自所检查帧,未将 agent 工作时间或控制步数换算为视频时间。" + }, + "summary_zh": "最后评分局在起点附近开始抬腿后即失衡,视频约 0.8 s 起塌下、约 2.8 s 已倒地,直到视频 10.0 s 未恢复,也未越过第一道横栏,最终节点为 0/1。log step 234 服务在第1000个控制步返回失败;最终步态和平衡回退未能维持支撑,根因归为控制。缺少状态日志、累计奖励数值及完整控制器修改内容,无法细分到具体参数或反馈环节。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "视频0.4–1.2 s塌下、2.8 s起倒地直到10 s与标签一致;节点false、breakpoint、平衡、控制根因均有证据支持。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:21:52Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-standard-realtime--r3towv3/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-standard-realtime--r3towv3/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "85193d6d7012b5accb7cc450860d39807395e672aeda5e5189c5fdb8f81fe510", + "video_sha256": "ca4e8be657413cb39c4ef76b8f3dd1e3cdf8bfcdcf3602fb833e19caa857fbc8", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard-realtime--ss2ipue.json b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard-realtime--ss2ipue.json new file mode 100644 index 0000000000000000000000000000000000000000..b7daf6354cd1fb6dba939946a98acb12bf6d0afb --- /dev/null +++ b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard-realtime--ss2ipue.json @@ -0,0 +1,200 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-hurdle-i00-standard-realtime--ss2ipue", + "taskKey": "humanoidbench/loco_hurdle", + "dag_sha256": "ca09b7dc923221773af0078329063e8e11b1544a0f305c56d17055d1bf47910e", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 164, + "sim_steps": 15514, + "calls": 22833, + "rows": 22757, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3567.8, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 309.638, + "max_lag_s": 0.0599, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 2302, + "late_steps": 0, + "requests_refused": 24, + "request_busy_s": 16.569, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.545, + "peak_processes": 19, + "peak_disk_gb": 1.06, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-hurdle-i00-standard-realtime--ss2ipue/replay.mp4", + "duration_s": 1.7, + "fps": 20.0, + "frames": 34, + "sha256": "432983d5f6d43a8c654b532285f63692a6401da14fa01f958b2b234dbe9146f0" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "hurdle_run_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00–1.35 s,机器人在跑道上交替迈腿、摆臂向前运动,双脚先后接触地面;首道横栏始终位于机器人前方,未见越栏。" + }, + { + "t": 1.5, + "log_step": null, + "text_zh": "观察:视频 1.40–1.65 s,机器人躯干逐渐偏斜,双腿向前伸而上身落后,末帧骨盆明显下沉;截至视频 1.65 s 未恢复直立,也未到达首栏。" + }, + { + "t": 1.65, + "log_step": null, + "text_zh": "推断:评分记录 success、live_success、replay_success 均为0,结合视频全段未越过首栏、末段失衡,判定跨栏跑达标事件未发生;评分记录未提供 episode_return,不能把任务失败值0当作累计奖励。" + }, + { + "t": null, + "log_step": 190, + "text_zh": "观察:log step 190 的服务状态为 episode 47、resets_left 4,所附试验输出涉及 episode 44–46;评分记录最终 resets=50。此前各局试验不能作为最后评分局的节点进度,日志未给出最后一局的完整终态读数。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–1.35 s,机器人左右脚交替踏在跑道地面上承重和推进;视频 1.40–1.65 s 支撑失稳。未见机器人实际越过或接触横栏,不追加典型方案中的避开。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "累计奖励", + "threshold": 700, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "hurdle_run_completed", + "t": 1.5, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频 1.40–1.65 s,机器人在向首栏前进时持续倾斜,脚步未能重新支撑上身,骨盆向地面下沉;视频结束前没有恢复。失败发生在跨栏跑步骤的起跑行进过程中,横栏仍在前方。" + }, + "unmet": { + "hurdle_run_completed": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 1.5, + "log_step": null, + "evidence_zh": "观察:视频 1.30–1.35 s 机器人尚能以双腿支承躯干;视频 1.40–1.65 s 躯干偏斜不断增加、双腿伸向身前、骨盆降低,机器人失去稳定支撑且未修复。画面没有显示其被横栏或侧墙挡住。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "推断:视频 0.00–1.35 s 机器人已执行连续迈步,视频 1.40–1.65 s 在尚未接触首栏或侧墙的空旷跑道上逐渐失稳,说明该次行进控制未能维持支撑和平衡。log step 188 的程序通过 gen.run 连续测试神经步态及关节修正方案,log step 190 仍记录试验中的明显俯仰、侧倾偏差;这些仅作控制能力的会话旁证,不把早先试验的碰栏或具体参数归给本次视频。", + "source_zh": "视频 / 日志" + } + ], + "gap_zh": "无状态日志;评分记录没有 episode_return,无法给出累计奖励数值或精确跌倒阈值触发时刻。log step 190 只读到 episode 47,log step 191 没有返回最后一局结果,最后评分局的具体控制器参数及完整传感轨迹无法核对;因此根因只标控制,不细分反馈、轨迹、估计误差或参数问题。视频末帧为视频 1.65 s,只主张可见的失衡和下沉,不另推断其后躯干触地。" + }, + "summary_zh": "评分视频中机器人已开始迈步前进,但没有越过首栏,跨栏跑达标节点最终为0/1。视频约 1.50 s 起机器人持续失衡、骨盆下沉,断点为跨栏跑达标的操作中,根因可归于控制未能维持平衡。日志含多次重置,末局完整传感记录和累计奖励均未提供,不能把早先试验的碰栏情况混入评分轨迹。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "视频仅 1.7 s,机器人在首栏前迈步后失衡前倾倒下,节点未达成、断点、平衡现象和控制根因均有画面支持;未发现需要修改的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:10:00Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-standard-realtime--ss2ipue/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-standard-realtime--ss2ipue/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "596c93ab13b591ecb4ee3ebd6dc87db2f9b6ead5b4c31b8ee87d7c6938e2054c", + "video_sha256": "432983d5f6d43a8c654b532285f63692a6401da14fa01f958b2b234dbe9146f0", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard-realtime--wmdx6oe.json b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard-realtime--wmdx6oe.json new file mode 100644 index 0000000000000000000000000000000000000000..38e8ce2eaca9abaf3057471b7772ad80c992ac92 --- /dev/null +++ b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard-realtime--wmdx6oe.json @@ -0,0 +1,217 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-hurdle-i00-standard-realtime--wmdx6oe", + "taskKey": "humanoidbench/loco_hurdle", + "dag_sha256": "ca09b7dc923221773af0078329063e8e11b1544a0f305c56d17055d1bf47910e", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 652, + "sim_steps": 12723, + "calls": 102097, + "rows": 101882, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1100.5, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 253.606, + "max_lag_s": 0.0369, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 2269, + "late_steps": 0, + "requests_refused": 36, + "request_busy_s": 28.654, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.175, + "peak_processes": 9, + "peak_disk_gb": 0.251, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-hurdle-i00-standard-realtime--wmdx6oe/replay.mp4", + "duration_s": 6.6, + "fps": 20.0, + "frames": 132, + "sha256": "7ad3b795851c836a65c22ddd02f08c6844d1e83f542a49f1db42b30f38d6add5" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "hurdle_run_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.65, + "log_step": null, + "text_zh": "观察:视频 0.0–1.7 s,机器人在起点附近屈膝、抬脚踏步并反复晃动,第一道横栏始终远在前方,没有横栏移到机器人身后。" + }, + { + "t": 2.0, + "log_step": null, + "text_zh": "观察:视频 1.75–2.0 s,机器人躯干迅速倾倒并落地;视频 2.0–6.55 s 始终躺在跑道上,未恢复行进或越栏。" + }, + { + "t": null, + "log_step": 56, + "text_zh": "观察:log step 56 中机器人服务 status 返回最后一局 episode=51、control_steps=652、done=True、reason=failed、resets_left=0;评分记录 live_success=0、replay_success=0。未报告累计奖励数值;该节点按本局服务及评分器的失败判决记为未达标,不把目录 reward=0.0 当作 episode_return。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–1.7 s,机器人的双脚接触跑道地面,抬脚踏步时由落地脚承重;视频约 2.0 s 后躯干落在跑道上。机器人未到达第一道横栏,未发生可见的越栏通过。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "累计奖励", + "threshold": 700, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "hurdle_run_completed", + "t": 1.75, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频 0.5–1.7 s,机器人已经抬脚踏步尝试行进;视频约 1.75 s 身体开始持续翻倒,视频约 2.0 s 躯干落地,此后至视频 6.55 s 未恢复,第一道横栏仍在前方。失败发生在跨栏跑的起步过程中。" + }, + "unmet": { + "hurdle_run_completed": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 1.75, + "log_step": null, + "evidence_zh": "观察:视频 1.70–2.05 s,机器人从屈膝站姿失去足部支撑,躯干旋转至近乎水平并倒在跑道上;视频 5.5–6.55 s 的逐帧画面仍显示其躺地,未恢复。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "推断:agent 的步态控制器未能稳定起步及支撑切换。log step 38 的脚本以周期足端目标、腿部逆解和 IMU 反馈生成关节目标,log step 53–55 继续修正预移重心相位并试验俯仰、横滚反馈;视频 0.55–1.7 s 机器人在抬脚时持续晃动,随后在视频约 1.75–2.0 s 倒地,倒地前没有接近横栏或可见的障碍阻挡。该证据支持控制失败,但不足以将最终失稳唯一归到某个参数、逆解或接触布局。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "未提供状态日志,评分字段及日志均未报告最后一局的 episode_return,故原生指标数值保留 null。日志包含多次重置和多组试验,不能把其他局的 fall_step 或 agent 自述的倒地时间换算为本视频时刻;最后一局缺少可逐帧核对的完整关节目标、受力和支撑状态,根因只标到控制。" + }, + "summary_zh": "机器人在起点附近抬脚踏步,但未到达或越过第一道横栏,跨栏跑达标节点始终未完成(最终 0/1)。视频约 1.75–2.0 s 机器人失衡倒地,直至视频 6.55 s 仍未恢复;结合控制脚本与试验日志,根因归为控制。最后一局累计奖励和完整支撑状态未报告,无法进一步确定失稳的具体控制子类。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 2, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "fix", + "conf": 8, + "summary_zh": "视频帧支持标注:机器人起点附近踏步,约1.75–2.0 s倒地并躺到6.55 s,节点未达成,断点、平衡现象、控制根因均成立。仅有一处log_step引用错位:服务status读数在step id 57,不在56。", + "n_issues": 1, + "n_major": 0 + }, + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 2, + "verdict": "agree", + "conf": 8, + "summary_zh": "视频帧支持标注:机器人起点附近踏步,约1.75–2.0 s倒地并躺到6.55 s,节点未达成,断点、平衡现象、控制根因均成立。上一轮log_step 56的争议已由标注者核实(step 56的调用含r.status()输出),不再重复。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:20:19Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-standard-realtime--wmdx6oe/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-standard-realtime--wmdx6oe/04_review_r1" + }, + { + "stage": "reconcile", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-standard-realtime--wmdx6oe/05_reconcile_r1" + }, + { + "stage": "traj_review", + "round": 2, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-standard-realtime--wmdx6oe/04_review_r2" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "b1f0bc691306861ec98bd3c373841822c01f367f98b47f10c4f2e22b01963494", + "video_sha256": "7ad3b795851c836a65c22ddd02f08c6844d1e83f542a49f1db42b30f38d6add5", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard-realtime--yyqbdh2.json b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard-realtime--yyqbdh2.json new file mode 100644 index 0000000000000000000000000000000000000000..21626708d8d0ce4b6c92136f84a542134b22c79c --- /dev/null +++ b/phase2/humanoidbench/loco_hurdle/trajectories/humanoidbench-loco-hurdle-i00-standard-realtime--yyqbdh2.json @@ -0,0 +1,195 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-loco-hurdle-i00-standard-realtime--yyqbdh2", + "taskKey": "humanoidbench/loco_hurdle", + "dag_sha256": "ca09b7dc923221773af0078329063e8e11b1544a0f305c56d17055d1bf47910e", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 288, + "sim_steps": 28129, + "calls": 273275, + "rows": 273220, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 2142.6, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 561.687, + "max_lag_s": 0.0045, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 27, + "late_steps": 0, + "requests_refused": 59, + "request_busy_s": 85.356, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.373, + "peak_processes": 10, + "peak_disk_gb": 1.049, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-loco-hurdle-i00-standard-realtime--yyqbdh2/replay.mp4", + "duration_s": 2.95, + "fps": 20.0, + "frames": 59, + "sha256": "f4e4ff00b87df80a31a3da5fde3db4da68b7df9492afc035f8e3b61e7bade267" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "hurdle_run_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 191, + "text_zh": "观察:视频 0.00–1.85 s,机器人在跑道起点附近屈膝踏步、躯干反复摇摆,前方第一根横栏始终未移到机器人身后;log step 191 启动标签为 FINAL 的最后一次 episode,log step 192 返回同一标签的失败记录。" + }, + { + "t": 2.1, + "log_step": 194, + "text_zh": "观察:视频 1.90–2.20 s,机器人持续向一侧倾斜并降低身体;视频 2.25–2.90 s,机器人倒地且没有恢复站立,第一栏仍在前方。log step 194 的服务状态为 episode=51、control_steps=288、done=true、reason=failed;评分字段 success、live_success、replay_success 均为0。" + }, + { + "t": 2.9, + "log_step": 194, + "text_zh": "推断:本次评分 episode 未形成越栏跑进展且以倒地失败结束,故跨栏跑达标节点全程未达到;不是仅根据零分认定未尝试。评分记录未报告 episode_return,不能将 outcome.reward=0.0 当作累计奖励数值。成功参考视频显示横栏会依次从机器人前方移至身后,仅用作视觉判据对照。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–2.10 s,机器人双脚在跑道地面承担站立与踏步支撑;视频 2.25–2.90 s,机器人身体塌低、倒向地面。未见腿或躯干到达前方横栏,不把任务典型方案中的越栏避开记为实际接触。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "累计奖励", + "threshold": 700, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "hurdle_run_completed", + "t": 2.1, + "log_step": 194, + "where": "操作中", + "evidence_zh": "观察:视频 1.90–2.20 s,正在踏步前进的机器人持续倾倒,视频 2.25–2.90 s 倒地,未到第一栏也未修复该步;log step 194 服务确认最后一次 episode=51 已失败。断点属于正在尝试的跨栏跑步骤。" + }, + "unmet": { + "hurdle_run_completed": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 2.1, + "log_step": 194, + "evidence_zh": "观察:视频 1.90–2.20 s,机器人身体相对双脚明显倾斜、支撑姿态塌下;视频 2.25–2.90 s,躯干降到地面附近并翻倒,直到末帧也未重新站立。log step 194 的服务状态为 done=true、reason=failed;这次是失去平衡倒地。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "反馈控制", + "source_zh": "日志", + "evidence_zh": "推断:log step 127 的控制代码以周期髋膝摆动叠加基于 IMU 的躯干俯仰、侧倾 PD 修正;log step 191 启动最终 walk2 控制器,log step 192 回显其 k_rr=1.5、k_rd=0.15、k_y=0.0,横向视觉位置反馈未启用。视频 0.70–1.55 s 已有反复倾摆,视频 1.90–2.90 s 则持续倾倒,说明该步态中的姿态反馈未能维持支撑平衡。log step 194 对 FINAL 本体感知记录的分析也报告明显姿态偏离;这些证据支持反馈稳定失败,尚不能证明某个增益或具体重心估计有误。" + } + ], + "gap_zh": "输入没有状态日志,评分字段也没有 episode_return 数值,无法恢复累计奖励曲线、接触力或精确重心位置。log step 194 的脚本把 FINAL 本体感知记录首次超过60度的时刻写为仿真 4.18 s,最终服务控制步数为288;视频仅长2.95 s,未独立核实控制步到视频帧的对应关系,因此未将这些仿真时刻换算为视频时间,断点锚点直接取自已检查的帧。反馈控制根因由代码、最终配置与失稳画面综合推断,具体参数或动力学原因不能进一步确定。" + }, + "summary_zh": "评分 episode 是 log step 191 启动、log step 194 确认为失败的第51次 episode;机器人只在第一栏前踏步和摇摆,跨栏跑达标节点未达到(最终0/1)。视频约 2.10 s 起机器人持续倾倒,至视频 2.90 s 末帧倒地,断点为跨栏跑操作中的平衡失败,代码与最终配置支持反馈控制未稳定住步态。没有状态日志和累计奖励数值,不能精确分析接触力、重心或具体失稳参数,也未混用视频与仿真时钟。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "节点未达到、断点(跨栏跑操作中,视频约2.1 s起倾倒、2.9 s倒地)、平衡现象和反馈控制根因均与视频及log一致;未换算agent/仿真时间,control_steps 288与视频2.95 s的对应(4.18 s仿真倾斜→约2.1 s)也与帧吻合。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:06:37Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-standard-realtime--yyqbdh2/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-loco-hurdle-i00-standard-realtime--yyqbdh2/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "da75fa9e41ed6c2643b314ada6108d9a075f721111042de5c2c3ae4766366a0b", + "video_sha256": "f4e4ff00b87df80a31a3da5fde3db4da68b7df9492afc035f8e3b61e7bade267", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-privileged--5wdq6n4.json b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-privileged--5wdq6n4.json new file mode 100644 index 0000000000000000000000000000000000000000..f04775be802390c8707a039c2e87d87694649885 --- /dev/null +++ b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-privileged--5wdq6n4.json @@ -0,0 +1,258 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-bookshelf-simple-i00-privileged--5wdq6n4", + "taskKey": "humanoidbench/manip_bookshelf_simple", + "dag_sha256": "2204f90476ae603026750aa49178a8c18a5f3ae6b8197cf0e7e8d3ecbe659563", + "outcome": { + "success": 0, + "reward": 0.7188156652393989, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 1000, + "final_reward": 0.7188156652393989, + "sim_steps": 1000, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.758, + "peak_processes": 10, + "peak_disk_gb": 0.239, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-bookshelf-simple-i00-privileged--5wdq6n4/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "de703b04a19965631a841357fe9969403b3c2e3822376ee520e52b9416bd10b6" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 5, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "object_1_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频0.0 s,最浅红色奖杯立在书架从上数第二层,顶部层格有其同色半透明目标副本,实体与目标明显分离。" + }, + { + "t": 5.0, + "log_step": null, + "text_zh": "观察:视频3.0–6.5 s,机器人前倾并抬起右手,但手停在奖杯所在层板下方,手指未包住奖杯;奖杯仍立在原层格。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频6.5–10.05 s及末秒密集帧中,奖杯没有离开原处,机器人两手没有携带奖杯,顶部目标仍未被实体占据;第1件到位事件未发生。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "压/擦", + "note_zh": "观察:视频约3.5–4.5 s,机器人前倾后两手接在中层格口的层板处;随后左掌仍贴低处层板,右手抬向奖杯所在层板下方并贴近板底。实际接触对象为书架层板,没有形成对奖杯的包夹抓握。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频0.0–10.05 s,机器人双脚接地、腿部承重,身体前倾并下沉,但未倒地。" + } + ] + }, + "object_2_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频0.0–10.05 s,机器人始终在奖杯附近伸手,未见其取走并搬运其余红色物体;可见红色实体保留初始位置,对应半透明目标未被搬入实体。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频全程第1件奖杯未到位,且未见后续搬放,故按DAG的顺序门控,第2件次浅红物体的到位事件未被记入;未把后续未启动步骤另列为失败现象。" + } + ] + }, + "object_3_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频0.0–10.05 s,机器人始终在奖杯附近伸手,未见其取走并搬运其余红色物体;可见红色实体保留初始位置,对应半透明目标未被搬入实体。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频全程第1件奖杯未到位,且未见后续搬放,故按DAG的顺序门控,第3件中红物体的到位事件未被记入;未把后续未启动步骤另列为失败现象。" + } + ] + }, + "object_4_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频0.0–10.05 s,机器人始终在奖杯附近伸手,未见其取走并搬运其余红色物体;可见红色实体保留初始位置,对应半透明目标未被搬入实体。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频全程第1件奖杯未到位,且未见后续搬放,故按DAG的顺序门控,第4件次深红物体的到位事件未被记入;未把后续未启动步骤另列为失败现象。" + } + ] + }, + "object_5_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频0.0–10.05 s,机器人始终在奖杯附近伸手,未见其取走并搬运其余红色物体;可见红色实体保留初始位置,对应半透明目标未被搬入实体。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频全程第1件奖杯未到位,且未见后续搬放,故按DAG的顺序门控,第5件最深红物体的到位事件未被记入;未把后续未启动步骤另列为失败现象。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 2000, + "direction": "higher_better" + }, + { + "name": "obj_goal_dist", + "value": null, + "unit": "m", + "threshold": 0.15, + "direction": "lower_better" + } + ], + "failure": { + "breakpoint": { + "ref": "object_1_at_slot", + "t": 5.0, + "log_step": 21, + "where": "操作中", + "evidence_zh": "观察:视频约5.0 s,机器人已向第1件最浅红奖杯伸右手,但手位于其承托层板下方,奖杯仍在原处;截至视频10.05 s没有抓起、移到顶部目标或修复这次取物失败。log step 21保存并重放的交付脚本确为朝奖杯伸手的1000步轨迹,重放哈希一致;此前会话内的跌倒测试不属于这段视频。" + }, + "unmet": { + "object_1_at_slot": false, + "object_2_at_slot": false, + "object_3_at_slot": false, + "object_4_at_slot": false, + "object_5_at_slot": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "没建立接触", + "t": 5.0, + "log_step": null, + "evidence_zh": "观察:视频约5.0–6.5 s及9.0–10.05 s,机器人右手伸在奖杯下方层格内,手指没有围住奖杯;奖杯一直立在上面的原层板,未随手移动。所需的可搬运抓取接触从未建立。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "抓取和接触设计", + "alt_feasible": "未知", + "source_zh": "日志", + "evidence_zh": "推断:log step 21的最终生成脚本以初始关节位置作为目标,只更新腿部平衡关节和右臂五个关节,手指目标保持初态;脚本只含伸手航点,没有闭指包夹奖杯及提取搬放的控制段。视频约5.0–10.05 s对应显示右手在奖杯下方伸展、奖杯不动,支持交付控制未设计出能携物的抓取接触。未验证可行替代握法,不归因为物体不可抓或任务无解。" + } + ], + "gap_zh": "无状态日志和成功参考视频;现有视频为256×256的单一房间视角,不能精确核实指尖与层板的接触点、接触力或物体距目标的数值。到位节点由可见奖杯未移动及后续顺序门控判为未达。评分器仅报告final_reward=0.7188156652393989等字段,未报告episode_return或obj_goal_dist,因此两项原生指标保留null;没有将末步奖励当作累计回报。" + }, + "summary_zh": "观察:视频0.0–10.05 s,机器人保持双脚支撑并前倾扶架,右手朝最浅红奖杯伸去,但奖杯未被抓起或移动,五个按序到位节点均未达成。推断:失败断在第1件奖杯的搬放操作,交付脚本只驱动右臂伸手而未加入闭指及提取动作,未形成搬运所需的抓取接触。无状态日志,接触细节与精确目标距离受单一低分辨率视角限制,累计回报也未报告。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "五个节点均未达成与视频一致:奖杯全程立在原层格,机器人未抓取;断点在object_1操作中、没建立接触、抓取和接触设计(交付脚本只含右臂伸手航点、无闭指/提取段)均有帧和log step 21支持。未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:53:42Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-privileged--5wdq6n4/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-privileged--5wdq6n4/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "2b90ac5560e9a36d1a2fc40b318def44e5c547aedef9e5b1abf8828f576eb478", + "video_sha256": "de703b04a19965631a841357fe9969403b3c2e3822376ee520e52b9416bd10b6", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-privileged--irfe4kr.json b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-privileged--irfe4kr.json new file mode 100644 index 0000000000000000000000000000000000000000..0b461db4984de7bd9a3f3f4f2cb1d82bcf714f89 --- /dev/null +++ b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-privileged--irfe4kr.json @@ -0,0 +1,237 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-bookshelf-simple-i00-privileged--irfe4kr", + "taskKey": "humanoidbench/manip_bookshelf_simple", + "dag_sha256": "2204f90476ae603026750aa49178a8c18a5f3ae6b8197cf0e7e8d3ecbe659563", + "outcome": { + "success": 0, + "reward": 0.6881264336081043, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 1000, + "final_reward": 0.6881264336081043, + "sim_steps": 1000, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 2.006, + "peak_processes": 9, + "peak_disk_gb": 0.253, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-bookshelf-simple-i00-privileged--irfe4kr/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "0804cd603d3f757171685e8e33845e53aa948f503356ea19a67fba87785056d8" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 5, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "object_1_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 11, + "text_zh": "观察:视频 0.0 s 最浅红色奖杯状实体位于一排竖书旁的层板上,其同色半透明目标位于上层;log step 11 的实例检查将该件记作 trophy、第1件,目标为 [0.75,-0.25,1.55]。" + }, + { + "t": 2.85, + "log_step": null, + "text_zh": "观察:视频 1.4–3.0 s 机器人左手在奖杯细杆附近收指、回撤并屈肘,但奖杯的球头和底座仍留在原层;上层半透明副本处未出现实体奖杯。" + }, + { + "t": 10.05, + "log_step": 119, + "text_zh": "观察:视频 3.0–10.05 s 奖杯始终未被搬到上层目标,末帧仍在原层、左手附近;log step 119 是最终交付生成及重放的输出,记录奖杯末位置 [0.911,0.048,1.25]、距目标 0.452 m,且重放 deterministic=true,与画面未到位相符。该日志值仅作视频判断的旁证。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "未知", + "note_zh": "观察:视频 1.4–3.0 s 机器人左侧灵巧手在最浅红奖杯的细杆附近合拢并回撤,随后停在奖杯旁;手与奖杯重叠且局部被遮挡,无法分辨实际为对向抓握、勾住还是擦碰,不将闭指命令当作抓握成立。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–10.05 s 机器人双脚接地、屈膝站立,双腿在左手操作奖杯时承重;未见摔倒。" + } + ] + }, + "object_2_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:已查看视频 0.0–10.05 s,机器人仅在最浅红奖杯附近伸手、收指和屈肘,未见搬运次浅红色竖书;视频中层靠近左侧的粉红色半透明书形目标没有出现从原位置搬来的对应实体。推断:第1件从未到位,且本件未发生搬运,按 DAG 的按序事件验收,本节点未达到。" + } + ] + }, + "object_3_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:已查看视频 0.0–10.05 s,机器人仅在最浅红奖杯附近伸手、收指和屈肘,未见搬运中红色细长物体;视频中下部层板上的红色细长半透明目标没有出现从原位置搬来的对应实体。推断:第1件从未到位,且本件未发生搬运,按 DAG 的按序事件验收,本节点未达到。" + } + ] + }, + "object_4_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:已查看视频 0.0–10.05 s,机器人仅在最浅红奖杯附近伸手、收指和屈肘,未见搬运次深红色竖书;视频下部层板上的对应半透明书形目标没有出现从原位置搬来的对应实体。推断:第1件从未到位,且本件未发生搬运,按 DAG 的按序事件验收,本节点未达到。" + } + ] + }, + "object_5_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:已查看视频 0.0–10.05 s,机器人仅在最浅红奖杯附近伸手、收指和屈肘,未见搬运最深红色竖书;视频下部层板上的对应半透明书形目标没有出现从原位置搬来的对应实体。推断:第1件从未到位,且本件未发生搬运,按 DAG 的按序事件验收,本节点未达到。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 2000, + "direction": "higher_better" + }, + { + "name": "obj_goal_dist", + "value": null, + "unit": "m", + "threshold": 0.15, + "direction": "lower_better" + } + ], + "failure": { + "breakpoint": { + "ref": "object_1_at_slot", + "t": 2.85, + "log_step": 112, + "where": "操作中", + "evidence_zh": "观察:视频约 2.85 s 机器人左臂结束伸入、回撤后屈肘停在奖杯旁,奖杯仍在原层;随后直到视频 10.05 s 也没有送到上层目标。log step 112 保存的最终生成器只对该奖杯实施进手、闭指及短距离回拉;log step 119 证实这份交付含1000步并已确定性重放。最早未修复的失败步骤是第1件到位。" + }, + "unmet": { + "object_1_at_slot": false, + "object_2_at_slot": false, + "object_3_at_slot": false, + "object_4_at_slot": false, + "object_5_at_slot": false + }, + "phenomena": [ + { + "cat": "运动", + "item": "位姿不准", + "t": 2.85, + "log_step": null, + "evidence_zh": "观察:视频 2.2–3.0 s 左手回撤、屈肘后,最浅红奖杯的球头与底座仍在竖书旁的原层,明显低于上层同色半透明目标;视频 3.0–10.05 s 未再向目标转移。这里仅标物体未被送至目标的可见位置偏差,不据遮挡推断已经提起、滑脱或松手。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志 / 视频", + "evidence_zh": "推断:agent未解决奖杯抓取后抽出并搬运的控制问题。视频 1.4–3.0 s 左手收指、回撤而奖杯未随手离开原层;log step 105 的会话测试显示手部参考点已撤至低处而奖杯仍留在层板,log step 108 的另一次抽出测试则记录奖杯脱离层板后下落,二者仅作为控制困难的会话证据,不算本次掉落。最终 log step 112 使用逆解、闭指和短回拉并保持,log step 119 生成结果仍距目标 0.452 m;未形成有效搬运。具体是接触几何还是腕臂跟踪问题,证据不足以细分。" + } + ], + "gap_zh": "输入有完整视频和运行日志,但没有状态日志、接触对或成功参考视频;256×256房间视角下左手与奖杯细杆局部重叠,无法核实是否形成对向抓握、短暂勾住或何时失去接触,故接触方式填未知,不标没保持住,也不细分控制根因。后四件除未开始搬运及前置未达外,没有独立操作过程。评分器只给 final_reward=0.6881264336081043,未报告 episode_return 或 obj_goal_dist;原生指标均留空,未把最终生成器自测的0.452 m冒充评分器指标。" + }, + "summary_zh": "评分轨迹只尝试最浅红奖杯,左手伸入、收指并回撤后停在原层旁;五个按序到位事件均未达到,最终为0/5。断点为最浅红物体到位,实体没有被搬到上层同色目标;视频与会话测试支持奖杯抓取搬运的控制困难,不能把调试中的掉落或摔倒算入本次轨迹。没有状态及接触日志,手部遮挡使实际握持方式和控制子原因无法确认。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "视频显示左手在奖杯旁收指回撤,奖杯始终留在原层,五节点均未达成、断点为第1件到位,与标注一致;未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:00:17Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-privileged--irfe4kr/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-privileged--irfe4kr/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "0faa4db23ac9ab6feea341beef58969c1061a4e639cf68baba96a005df8f9b4a", + "video_sha256": "0804cd603d3f757171685e8e33845e53aa948f503356ea19a67fba87785056d8", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-privileged--lciumdm.json b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-privileged--lciumdm.json new file mode 100644 index 0000000000000000000000000000000000000000..ce7fae5f4f0034eeb1dd8f8c0241a4fe953b9190 --- /dev/null +++ b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-privileged--lciumdm.json @@ -0,0 +1,253 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-bookshelf-simple-i00-privileged--lciumdm", + "taskKey": "humanoidbench/manip_bookshelf_simple", + "dag_sha256": "2204f90476ae603026750aa49178a8c18a5f3ae6b8197cf0e7e8d3ecbe659563", + "outcome": { + "success": 0, + "reward": 0.839549748804176, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 1000, + "final_reward": 0.839549748804176, + "sim_steps": 1000, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-bookshelf-simple-i00-privileged--lciumdm/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "aa4c2e64f53716f07ebf0a2c2e2d7de3ea01c4e18455e3e119b1208a391f1e79" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 5, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "object_1_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.0 s,浅色奖杯实体在直立书本旁的原层板上,同色半透明目标在更高一层;两者明显分离。" + }, + { + "t": 2.5, + "log_step": null, + "text_zh": "观察:视频 1.0–3.9 s,机器人转动躯干并抬起右臂,右手停在奖杯所在层板前下方;奖杯仍在书本旁,未见手指夹住它或将它移向上层目标。" + }, + { + "t": 10.05, + "log_step": 116, + "text_zh": "观察:视频 4.0–10.05 s,机器人维持右臂前伸,奖杯仍直立在原层板上,目标副本处未出现搬来的实体。log step 116 对已保存交付文件的会话重放检查也输出奖杯最大漂移约 1.6e-5 m,log step 122 再生成后状态哈希相同;这些日志仅作视频判断的佐证。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–10.05 s,机器人双脚着地、双腿承重,在右臂向奖杯伸出的过程中支撑身体;未见右手对奖杯形成夹持,故不记抓握接触。" + } + ] + }, + "object_2_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:对照视频 0.0 s、1.0–3.9 s、4.0–9.5 s及 9.1–10.05 s末段,次浅红色直立书本仍在原来的书本排中,未被机器人取走或送往中间层的同色副本。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频全段首件奖杯均未到位,本节点依赖的按序到位事件也未成立,因此本节点未被记入;不是仅凭评分失败推断未完成。" + } + ] + }, + "object_3_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:对照视频 0.0 s、1.0–3.9 s、4.0–9.5 s及 9.1–10.05 s末段,下方层板上的红色细长物体未被机器人拿起,未出现向其上方同色目标的搬运。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频全段首件奖杯均未到位,本节点依赖的按序到位事件也未成立,因此本节点未被记入;不是仅凭评分失败推断未完成。" + } + ] + }, + "object_4_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:对照视频 0.0 s、1.0–3.9 s、4.0–9.5 s及 9.1–10.05 s末段,红色直立书本排保持原位置,未见次深红书本被机器人取走并送至底部目标。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频全段首件奖杯均未到位,本节点依赖的按序到位事件也未成立,因此本节点未被记入;不是仅凭评分失败推断未完成。" + } + ] + }, + "object_5_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:对照视频 0.0 s、1.0–3.9 s、4.0–9.5 s及 9.1–10.05 s末段,红色直立书本排保持原位置,未见最深红书本被机器人取走并送至底部目标。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频全段首件奖杯均未到位,本节点依赖的按序到位事件也未成立,因此本节点未被记入;不是仅凭评分失败推断未完成。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 2000, + "direction": "higher_better" + }, + { + "name": "obj_goal_dist", + "value": null, + "unit": "m", + "threshold": 0.15, + "direction": "lower_better" + } + ], + "failure": { + "breakpoint": { + "ref": "object_1_at_slot", + "t": 2.5, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频约 1.0–2.5 s,机器人把右臂伸向首件浅红奖杯,但手停在其层板前下方,奖杯未随手移动;视频 3.0–10.05 s始终没有抓起或向上层目标搬运,首件到位步骤未修复。" + }, + "unmet": { + "object_1_at_slot": false, + "object_2_at_slot": false, + "object_3_at_slot": false, + "object_4_at_slot": false, + "object_5_at_slot": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "没建立接触", + "t": 2.5, + "log_step": null, + "evidence_zh": "观察:视频 1.0–3.9 s右手伸近后,未见手指在奖杯两侧形成夹持;视频 4.0–10.05 s奖杯仍在书本旁原处,右手保持前伸且未携带它。推断:未建立搬运所需的抓握接触;此处不把可能的瞬时擦碰等同于有效抓握。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "log_step": 111, + "source_zh": "日志", + "evidence_zh": "观察:log step 105 的会话伸手测试中,两组控制分别在控制步 98、145 结束,另两组也没有搬动奖杯;log step 109 中更大前倾的三组在控制步 119–124 结束。最终生成器(log step 111、114)仅用躯干及右臂逆解把掌心移近奖杯,手指目标保持执行器区间中点,没有随接近动作形成抓握和提起的控制阶段;log step 122 确认交付的是该生成器的重放结果。推断:会话未形成兼顾稳定支撑与有效抓取搬运的控制,最终交付停在伸手阶段;证据不足以细分为不可达、逆解失败或反馈失稳,故停在控制大类。" + } + ], + "gap_zh": "观察:inputs/ 中有视频和完整会话日志,但没有逐帧状态日志或成功参考视频;房间相机仅 256×256,手指与奖杯附近部分遮挡,不能排除瞬时擦碰,也不能量化接触力。评分器 reward 仅报告 final_reward=0.839549748804176、success=0、deterministic=1 和 sim_steps=1000,未报告 episode_return 或 obj_goal_dist,故两项原生指标留空;未把 final_reward 当作累计奖励。推断:有效抓握未形成可由物体始终留在原处及手的动作判断,但精确控制子类与可行替代方案仍不确定。" + }, + "summary_zh": "观察:视频 0.0–10.05 s中,机器人保持站立,右臂伸向首件浅红奖杯后停留,未见任何物体被搬到同色目标,五个按序到位节点最终均未完成。推断:断点在首件奖杯到位步骤,未建立搬运所需的抓握;会话测试及交付控制器表明稳定伸手尚未发展为有效抓取搬运,根因记控制。观察:缺少逐帧状态及接触对,评分器也未提供累计奖励与目标距离,无法进一步确定控制子类。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "视频 0–10 s 未见任何物体被搬到目标,五个节点均未到位,断点在首件到位、没建立接触、根因控制,与帧一致。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:01:52Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-privileged--lciumdm/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-privileged--lciumdm/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "72c9c0ff8feb073bfdd0ae1866d861f7aa26829c9cfa8545766384ac1fb86b2f", + "video_sha256": "aa4c2e64f53716f07ebf0a2c2e2d7de3ea01c4e18455e3e119b1208a391f1e79", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-privileged--qrkuwed.json b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-privileged--qrkuwed.json new file mode 100644 index 0000000000000000000000000000000000000000..5e5360e5590d58196680b2da0977f93343c83ad4 --- /dev/null +++ b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-privileged--qrkuwed.json @@ -0,0 +1,282 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-bookshelf-simple-i00-privileged--qrkuwed", + "taskKey": "humanoidbench/manip_bookshelf_simple", + "dag_sha256": "2204f90476ae603026750aa49178a8c18a5f3ae6b8197cf0e7e8d3ecbe659563", + "outcome": { + "success": 0, + "reward": 0.7914004920859983, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 1000, + "final_reward": 0.7914004920859983, + "sim_steps": 1000, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 1.963, + "peak_processes": 15, + "peak_disk_gb": 0.84, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-bookshelf-simple-i00-privileged--qrkuwed/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "7c0f98e034e7b8be915fc494dd601ec990092e287bc9c457b79a5b5df41e0b74" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 5, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "object_1_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 19, + "text_zh": "观察:视频 0.0 s 中,书架上层的同色半透明奖杯目标与实体奖杯所在的书本层分开;log step 19 的初态枚举将第一件确认为 trophy,目标为更高一层的位置。" + }, + { + "t": 2.0, + "log_step": null, + "text_zh": "观察:视频 0.0–2.4 s 的密集帧显示机器人调整站姿、抬起左手,手停在书本层下方附近;视频 2.5–10.0 s 及视频 10.05 s 末帧未见机器人握持奖杯、把奖杯提到上层或使上层奖杯实体与半透明目标重合。" + }, + { + "t": 10.05, + "log_step": 306, + "text_zh": "推断:视频 0.0–10.05 s 没有奖杯向上层目标的搬运,到位外观始终未出现;log step 306 所列最终生成脚本仅设定站姿并将左臂改为一个固定关节目标,随后保持,支持第一件未到位的视觉判断。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–10.05 s,机器人双脚接触地面,双腿承重,支撑其调整站姿和左臂接近第一件奖杯。" + }, + { + "who": "未知", + "how": "未知", + "note_zh": "观察:视频 2.0–10.05 s,左手靠近奖杯所在层的下方,躯干与层板遮挡手指和奖杯间的局部区域;未见形成抓握,无法确认是否有偶发的手指触碰,不将接近姿态记为抓握。" + } + ] + }, + "object_2_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频 0.0–10.05 s,机器人只调整站姿并抬近左手,未见将次浅红物体搬至对应半透明目标的动作;第一件奖杯未到位,按 DAG 的顺序事件规则,本节点的前置 object_1_at_slot 也未满足,本事件未被达到。" + } + ] + }, + "object_3_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频 0.0–10.05 s,机器人只调整站姿并抬近左手,未见将中红物体搬至对应半透明目标的动作;第一件奖杯未到位,按 DAG 的顺序事件规则,本节点的前置 object_2_at_slot 也未满足,本事件未被达到。" + } + ] + }, + "object_4_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频 0.0–10.05 s,机器人只调整站姿并抬近左手,未见将次深红物体搬至对应半透明目标的动作;第一件奖杯未到位,按 DAG 的顺序事件规则,本节点的前置 object_3_at_slot 也未满足,本事件未被达到。" + } + ] + }, + "object_5_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频 0.0–10.05 s,机器人只调整站姿并抬近左手,未见将最深红物体搬至对应半透明目标的动作;第一件奖杯未到位,按 DAG 的顺序事件规则,本节点的前置 object_4_at_slot 也未满足,本事件未被达到。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 2000, + "direction": "higher_better" + }, + { + "name": "obj_goal_dist", + "value": null, + "unit": "m", + "threshold": 0.15, + "direction": "lower_better" + } + ], + "failure": { + "breakpoint": { + "ref": "object_1_at_slot", + "t": 2.0, + "log_step": 306, + "where": "首步前", + "evidence_zh": "观察:log step 306 的最终 gen.py 只设固定站姿,并在动作循环索引150更新一次左臂关节目标,之后保持,没有抓取、抬升或搬运阶段;视频约2.0 s机器人左手抬近书架,视频2.5–10.05 s保持相近姿态,未见奖杯被握持或搬往上层目标。推断:交付的是站姿与手部接近基线,第一件奖杯的抓取搬运未启动,故位置为首步前;视频2.0 s仅作接近姿态的导航锚点。" + }, + "unmet": { + "object_1_at_slot": false, + "object_2_at_slot": false, + "object_3_at_slot": false, + "object_4_at_slot": false, + "object_5_at_slot": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 2.0, + "log_step": 306, + "evidence_zh": "观察:log step 277 写入、log step 306 打印的最终 gen.py 只有固定站姿和一次左臂关节目标更新,没有抓取、抬升或搬运阶段。视频0.0–2.5 s中机器人调整站姿并抬近左手,视频2.5–10.05 s保持接近姿态,未见奖杯随手离开原层或被送向上层目标。推断:评分局的交付基线未启动第一件的抓取搬运,按首步前规则标为没尝试。" + } + ], + "root_causes": [], + "gap_zh": "评分局未尝试抓取,最终脚本与视频只能确认站姿调整及左手接近保持,不能由没尝试反推推理/规划等根因。log step 277、306 的 NOTES.md 中,右臂逆解在层板处受阻、从前方放下手臂仍卡于层沿、保持时骨盆下沉,均是agent对会话尝试的自述;这些具体失败原因尚未独立核实,不能当作评分局事实,根因层留空。输入没有状态日志和成功oracle;视频为单一低分辨率视角,躯干及层板遮住奖杯与手指的局部关系,无法精确核实偶发触碰、接触力及距离。inputs/run.json的评分器仅报final_reward=0.7914004920859983,未直接报告episode_return或obj_goal_dist,故两个原生指标值保留null;不把该final_reward当作累计奖励。" + }, + "summary_zh": "评分轨迹只执行站姿调整和左手接近保持,五个按序到位事件均未达到,最终进度为0/5。第一件奖杯的抓取搬运未启动,断点为object_1_at_slot,位置为首步前,现象为没尝试。最终脚本及视频不足以确定未继续抓取的根因;会话中的受阻与骨盆下沉自述尚待核实,根因层留空,手指遮挡及缺少状态日志也限制了接触判断。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 3, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "fix", + "conf": 6, + "summary_zh": "进度 0/5、断点 object_1_at_slot、时钟与接触标注经核对无误;但根因“控制”缺乏支撑:其证据是会话里撤臂/抬臂的跟踪误差,与评分视频里没有抓取无因果关系,而 log step 277/306 显示 agent 有意交付“站姿+左手靠近、奖杯不动”的接近奖励策略,视频与之吻合,应按交付内容改判根因类别。", + "n_issues": 2, + "n_major": 1 + }, + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 2, + "verdict": "fix", + "conf": 7, + "summary_zh": "进度 0/5、各节点、时钟均无误;但交付脚本是不含任何抓取/搬运的“站姿+左手靠近”基线,按 03 规则应为 首步前 + 没尝试,且根因“推理/规划”是由 agent 决定只交付接近策略反推出来的(CASES.md FAIL-22),缺乏“方案本身错误”的证据。", + "n_issues": 2, + "n_major": 2 + }, + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 3, + "kind": "confirm", + "verdict": "agree", + "summary_zh": "两项 accept 均已在 prev/trajectory.json 中完整落实:断点改为首步前、现象改为没尝试,根因清空并在 gap_zh 说明缺口,未破坏其他字段。", + "n_issues": 0, + "n_checks": 2 + } + ], + "open_zh": [], + "confirm": { + "verdict": "confirmed", + "summary_zh": "两项 accept 均已在 prev/trajectory.json 中完整落实:断点改为首步前、现象改为没尝试,根因清空并在 gap_zh 说明缺口,未破坏其他字段。", + "checks_ok": "2/2", + "n_factual_errors": 0 + } + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:21:43Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-privileged--qrkuwed/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-privileged--qrkuwed/04_review_r1" + }, + { + "stage": "reconcile", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-privileged--qrkuwed/05_reconcile_r1" + }, + { + "stage": "traj_review", + "round": 2, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-privileged--qrkuwed/04_review_r2" + }, + { + "stage": "reconcile", + "round": 2, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-privileged--qrkuwed/05_reconcile_r2" + }, + { + "stage": "confirm", + "round": 3, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-privileged--qrkuwed/06_confirm_r3" + } + ], + "confirm_prompt": "prompts/06_confirm.md", + "confirm_prompt_sha256": "b9f56c095f7ea5e609e58bb21e076f565766c370333d71c7a3512cffb4d78267", + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "d930b5029b9246a4a8c8ea778fcad18864b6846be47a0b094f4dc5ebc9d118a5", + "video_sha256": "7c0f98e034e7b8be915fc494dd601ec990092e287bc9c457b79a5b5df41e0b74", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-privileged--vsutahw.json b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-privileged--vsutahw.json new file mode 100644 index 0000000000000000000000000000000000000000..d15e52094a7f39d06a7e681dcbbd2f5ecac133e4 --- /dev/null +++ b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-privileged--vsutahw.json @@ -0,0 +1,275 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-bookshelf-simple-i00-privileged--vsutahw", + "taskKey": "humanoidbench/manip_bookshelf_simple", + "dag_sha256": "2204f90476ae603026750aa49178a8c18a5f3ae6b8197cf0e7e8d3ecbe659563", + "outcome": { + "success": 0, + "reward": 0.5855822801938835, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 1000, + "final_reward": 0.5855822801938835, + "sim_steps": 1000, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.794, + "peak_processes": 11, + "peak_disk_gb": 0.24, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-bookshelf-simple-i00-privileged--vsutahw/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "dece4ae04bdae443d5707fe2f8e4f1e0ad3a7f14f0f12c6271ffc9c8fe3af115" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 5, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "object_1_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 18, + "text_zh": "观察:视频 0.0 s 的实体仍在原书架位置,顶层对应奖杯的半透明目标没有实体填入;log step 18 的初态模型检查将最浅红物体识别为 trophy(奖杯),不是顶层灰色耳机。" + }, + { + "t": 5.5, + "log_step": null, + "text_zh": "观察:视频 0.5–5.5 s 机器人双脚原地支撑,左手停在较低层前缘附近;视频 5.5–10.05 s 左手仍为空手,未见奖杯离开原处、随手抬升或移入顶层目标,最后一秒逐帧检查也未出现搬放。" + }, + { + "t": 10.05, + "log_step": 55, + "text_zh": "推断:视频 0.0–10.05 s 未发生将奖杯送至其较高目标位的运动,第一件到位事件未完成。log step 55 保存并回放的是最终 1000 步文件;log step 50 的大幅探身、抓取及摔倒属于较早试验,不计入本视频。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–10.05 s 机器人双脚接触地面承重,支撑原地伸左手尝试。左手没有围住或夹住第一件奖杯,不把空手伸出或屈指标成已经抓握;手与低层架板是否有轻微擦碰无法由此视角判定。" + } + ] + }, + "object_2_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 0.0–10.05 s 机器人仅在原地伸左手,书架可见红色书本和低层物体没有被携走或送入对应半透明目标;视频 9.0–10.05 s 的逐帧末段没有后续取放动作。" + }, + { + "t": 10.05, + "log_step": 18, + "text_zh": "推断:第2件为次浅红物体(book_c)(身份由 log step 18 的初态模型检查辅助辨认);视频全段没有先完成奖杯上移再依序处理该件的过程,因此按 DAG 的顺序门控,第2件到位事件未发生。" + } + ] + }, + "object_3_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 0.0–10.05 s 机器人仅在原地伸左手,书架可见红色书本和低层物体没有被携走或送入对应半透明目标;视频 9.0–10.05 s 的逐帧末段没有后续取放动作。" + }, + { + "t": 10.05, + "log_step": 18, + "text_zh": "推断:第3件为中红物体(chalk)(身份由 log step 18 的初态模型检查辅助辨认);视频全段没有先完成奖杯上移再依序处理该件的过程,因此按 DAG 的顺序门控,第3件到位事件未发生。" + } + ] + }, + "object_4_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 0.0–10.05 s 机器人仅在原地伸左手,书架可见红色书本和低层物体没有被携走或送入对应半透明目标;视频 9.0–10.05 s 的逐帧末段没有后续取放动作。" + }, + { + "t": 10.05, + "log_step": 18, + "text_zh": "推断:第4件为次深红物体(book_e)(身份由 log step 18 的初态模型检查辅助辨认);视频全段没有先完成奖杯上移再依序处理该件的过程,因此按 DAG 的顺序门控,第4件到位事件未发生。" + } + ] + }, + "object_5_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 0.0–10.05 s 机器人仅在原地伸左手,书架可见红色书本和低层物体没有被携走或送入对应半透明目标;视频 9.0–10.05 s 的逐帧末段没有后续取放动作。" + }, + { + "t": 10.05, + "log_step": 18, + "text_zh": "推断:第5件为最深红物体(book_b)(身份由 log step 18 的初态模型检查辅助辨认);视频全段没有先完成奖杯上移再依序处理该件的过程,因此按 DAG 的顺序门控,第5件到位事件未发生。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "unit": "reward", + "threshold": 2000, + "direction": "higher_better", + "value": null + }, + { + "name": "obj_goal_dist", + "unit": "m", + "threshold": 0.15, + "direction": "lower_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "object_1_at_slot", + "t": 5.5, + "log_step": 55, + "where": "操作中", + "evidence_zh": "观察:视频 5.5 s 机器人已在书架前伸左手,手仍停在奖杯所在层下方,既未持起奖杯,也未将它送向顶层半透明目标;这一状态持续至视频 10.05 s,第一件操作始终未修复。log step 55 记录最终文件已保存并确定性回放 1000 步,与视频中持续站立、没有早期试验摔倒的情形相符。" + }, + "unmet": { + "object_1_at_slot": false, + "object_2_at_slot": false, + "object_3_at_slot": false, + "object_4_at_slot": false, + "object_5_at_slot": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "没建立接触", + "t": 5.5, + "log_step": null, + "evidence_zh": "观察:视频 0.5–5.5 s 机器人左手在较低层附近伸出,视频 5.5–10.05 s 手仍未围住第一件奖杯,奖杯没有随手移动;机器人未建立搬放第一件所需的抓取接触。低层架板附近的手部投影重叠不能证明抓到了上层奖杯。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "log_step": 50, + "evidence_zh": "推断:控制根因来自 session 试验,而不是把交付基线解释成前移跟踪失败。log step 50 的探身加逆解/屈指试验没有带动奖杯,控制步 641 返回 done=True;agent 在 log step 56 NOTES 中称该次试验失稳摔倒,并说明随后改选站稳以获取站立奖励。log step 52 测得最小手—奖杯距离 0.425 m,但它测量的是 log step 51 生成的站立伸臂版本,不是探身试验或最终交付版本。log step 54–55 的最终代码只安排站立反馈、缓慢伸左臂及屈指,没有迈步前移或完整取放流程;log step 56–57 的自述与此一致,视频 0.5–10.05 s 也见左手停在低层、没有搬起奖杯。上述试验支持接近与抓取控制尚未实现,具体控制子类无法确认。", + "source_zh": "日志;agent 自述;视频" + } + ], + "gap_zh": "没有本次回放的状态日志或成功 oracle;256×256 单视角中躯干遮住部分奖杯和书本,无法精确测量物体原点距离、瞬时手指接触或确认手是否轻擦低层架板,但全段没有抬升/搬放动作且顺序第一件未完成,足以判五个事件均未到位。log step 52 的最小手距 0.425 m 来自随后被 log step 54–55 改写的较早文件,未用于本次回放的距离结论;agent 在 log step 51、56–57 声称必须迈步的几何判断也未获可达性求解验证,故不标“够不着”。评分器仅提供 final_reward=0.5855822801938835,没有 episode_return 或 obj_goal_dist,两个原生指标保留 null,不把该数当累计奖励。 agent 在 log step 51、56–57 决定停止推进并交付站立基线,本身不作为根因(docs/CASES.md FAIL-22);控制标签依据上述 session 试验。最终代码虽有伸臂和屈指指令,不能据此认定包含完整接近/抓取方案,也不能将原地站立写成试图前移却跟踪失败;迈步方案是否可行仍未验证。" + }, + "summary_zh": "评分轨迹中机器人维持站立并缓慢伸左手,未搬起第一件最浅红奖杯,五个按序到位事件均未完成,最终为0/5;断点在第一件放置的操作中,左手停在物体下方,没有建立抓取接触。交付脚本是 agent 在 session 接近与抓取控制试验失败后选定的站立基线,控制根因依据这些试验,不能把交付基线本身写成前移跟踪失败,也不把主动停止当成原因。没有状态日志或成功参考视频,部分物体受躯干遮挡,精确距离、轻微擦碰、更细控制根因及迈步可行性无法确认。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 2, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "fix", + "conf": 6, + "summary_zh": "节点值(0/5)、断点节点、时钟与 contact_actual 经帧核对均成立;唯一问题是根因证据把交付脚本写成“未能形成前移并送手”的控制失败,而 log 54 的 NOTES 与 log 52/57 显示这是 agent 明知够不到后主动选定的“站稳+慢伸左臂”基线。", + "n_issues": 1, + "n_major": 0 + }, + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 2, + "verdict": "agree", + "conf": 7, + "summary_zh": "经帧核对:视频 0–10.05 s 左手始终停在低层,奖杯未被搬起,五个节点 0/5、断点 object_1_at_slot、contact_actual(腿/躯干 支撑)成立;根因按 FAIL-22 由 session 试验支持、未把停止当原因,上轮意见已落实。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:21:04Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-privileged--vsutahw/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-privileged--vsutahw/04_review_r1" + }, + { + "stage": "reconcile", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-privileged--vsutahw/05_reconcile_r1" + }, + { + "stage": "traj_review", + "round": 2, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-privileged--vsutahw/04_review_r2" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "74afd2ce5aacfff9d90cb24413b30f41444be2489a747b5ebbd9fc480b912f33", + "video_sha256": "dece4ae04bdae443d5707fe2f8e4f1e0ad3a7f14f0f12c6271ffc9c8fe3af115", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-privileged--ylkcq6z.json b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-privileged--ylkcq6z.json new file mode 100644 index 0000000000000000000000000000000000000000..a5486511edfa59a9e5509bc68e1f4593013a2cae --- /dev/null +++ b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-privileged--ylkcq6z.json @@ -0,0 +1,257 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-bookshelf-simple-i00-privileged--ylkcq6z", + "taskKey": "humanoidbench/manip_bookshelf_simple", + "dag_sha256": "2204f90476ae603026750aa49178a8c18a5f3ae6b8197cf0e7e8d3ecbe659563", + "outcome": { + "success": 0, + "reward": 0.7876561228331058, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 1000, + "final_reward": 0.7876561228331058, + "sim_steps": 1000, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 6.112, + "peak_processes": 22, + "peak_disk_gb": 0.239, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-bookshelf-simple-i00-privileged--ylkcq6z/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "518ed7af8200ac9c5083a5be808cdeef8ed638cbd198bf430cceef252e4342c9" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 5, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "object_1_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.0 s,书架上层可见奖杯形半透明目标,未有实体奖杯与其重合;机器人站在书架前。" + }, + { + "t": 1.0, + "log_step": null, + "text_zh": "观察:视频 0.0–1.0 s,机器人调整双臂及站姿;视频 1.0–10.05 s 维持近似固定姿势,未见实体奖杯被带到上层半透明目标处,目标处始终没有实体奖杯到位。" + }, + { + "t": 10.05, + "log_step": 43, + "text_zh": "观察:视频 9.0–10.05 s 逐帧检查,机器人仍站立,未出现搬运到位;log step 43 最终生成脚本的复验输出 ti=0、replay steps 1000、sha match: True,与视频无首件到位事件相符。该 ti 是 agent 自编代理评分的阶段变量,仅作辅助证据。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–10.05 s,机器人双脚接地、腿部承重,维持首件操作前的站姿与接近准备。" + }, + { + "who": "未知", + "how": "未知", + "note_zh": "观察:视频 0.0–10.05 s,首件奖杯实体及靠近它的右手主要被机器人躯干遮挡;无法确认有无偶发接触或接触方式,未将静态屈指记为抓握。" + } + ] + }, + "object_2_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 0.0–10.05 s,机器人只在开头调整姿势,随后留在同一位置;未见次浅红物体被搬至相应半透明目标的动作。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频 0.0–10.05 s 首件奖杯未到位;本节点是要求此前各件依序被记入的 event,故第2件到位事件未成立,即使局部实体被遮挡也不能越过首件计入。" + } + ] + }, + "object_3_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 0.0–10.05 s,机器人只在开头调整姿势,随后留在同一位置;未见中红物体被搬至相应半透明目标的动作。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频 0.0–10.05 s 首件奖杯未到位;本节点是要求此前各件依序被记入的 event,故第3件到位事件未成立,即使局部实体被遮挡也不能越过首件计入。" + } + ] + }, + "object_4_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 0.0–10.05 s,机器人只在开头调整姿势,随后留在同一位置;未见次深红物体被搬至相应半透明目标的动作。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频 0.0–10.05 s 首件奖杯未到位;本节点是要求此前各件依序被记入的 event,故第4件到位事件未成立,即使局部实体被遮挡也不能越过首件计入。" + } + ] + }, + "object_5_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 0.0–10.05 s,机器人只在开头调整姿势,随后留在同一位置;未见最深红物体被搬至相应半透明目标的动作。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频 0.0–10.05 s 首件奖杯未到位;本节点是要求此前各件依序被记入的 event,故第5件到位事件未成立,即使局部实体被遮挡也不能越过首件计入。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 2000, + "direction": "higher_better" + }, + { + "name": "obj_goal_dist", + "value": null, + "unit": "m", + "threshold": 0.15, + "direction": "lower_better" + } + ], + "failure": { + "breakpoint": { + "ref": "object_1_at_slot", + "t": 0.0, + "log_step": 35, + "where": "首步前", + "evidence_zh": "观察:视频 0.0–1.0 s 机器人调整臂姿、站姿后,视频 1.0–10.05 s 保持站立,未启动将首件奖杯送往上层目标的搬运动作;log step 35 的最终候选控制器全程使用固定臂姿和手指弯曲量,仅腿部随质心反馈调整,log steps 41、43 保存并复验此轨迹。推断:这是首件搬放前的站立接近基线,断点取 object_1_at_slot,视频 0.0 s 为首步未启动的导航锚点。" + }, + "unmet": { + "object_1_at_slot": false, + "object_2_at_slot": false, + "object_3_at_slot": false, + "object_4_at_slot": false, + "object_5_at_slot": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 35, + "evidence_zh": "观察:视频 0.0–10.05 s,机器人调整站立和臂姿后停留,没有将首件奖杯提离原层并送至其上层目标;log step 35 的控制代码没有接近后抓取、抬升或送往目标的动作阶段,log step 44 的最终说明也明确搬放未实现。推断:交付轨迹未启动首件搬放,而非已经搬运后失手。" + } + ], + "root_causes": [ + { + "cat": "认知", + "item": "推理/规划", + "source_zh": "日志", + "evidence_zh": "观察:log step 29 明确改用最大化稠密奖励的静态姿态;log step 33 的自编代理回报等权相加站稳、躯干竖直、手近物及物近目标四项,未计入任务文本强调的逐件放置奖金;log step 35 的最终搜索仅优化固定臂姿和站立反馈,并用 300*max(0,moved-0.03) 惩罚奖杯位移。log steps 41、43 输出代理阶段 ti=0,复验后仍交付该轨迹;视频 1.0–10.05 s 也显示机器人静止维持该姿态。推断:最终策略把优化重点放在站立及手靠近物体,并抑制必须发生的物体移动,没有安排首件向目标搬运的阶段;该代理目标与依序搬放任务不等价,因而首件未启动。这一判断依据具体优化目标和动作结构,而非 agent 自述预算不足。" + } + ], + "gap_zh": "观察:inputs/ 提供视频和完整运行日志,没有状态日志或成功 oracle;视频分辨率为 256×256,机器人躯干遮挡首件奖杯及右手的局部区域,无法核实偶发接触和精确抓握状态。评分器只报告 final_reward=0.7876561228331058、success=0、deterministic=1 等结果,未报告 episode_return 或 obj_goal_dist;log step 20 的接口源码将 final_reward 定义为最后一步奖励,因此不把它或 agent 的代理总分 3518.2 填作累计回报。" + }, + "summary_zh": "机器人仅调整站姿和臂姿后停留,五个按序到位事件均未发生,最终为 0/5。断点在首件奖杯搬放启动前:最终脚本优化静态站立及手靠近首件,同时惩罚奖杯位移,没有安排送往目标的搬运阶段。首件奖杯与右手局部被躯干遮挡,且无状态日志,无法核实偶发接触及精确目标距离。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "视频全程只见机器人站立、手臂与手指近似静止,书架上层目标处始终没有实体物体到位,0/5、断点 object_1_at_slot(首步前·没尝试)与日志(CEM 静态姿态、移动惩罚 300*max(0,moved-0.03)、ti=0、sha match)一致;根因与证据基本站得住,未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:56:30Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-privileged--ylkcq6z/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-privileged--ylkcq6z/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "5cddd10f139cd63721650bc0f41c092b9bae84e81e3b7357e2ca997671d0bd5e", + "video_sha256": "518ed7af8200ac9c5083a5be808cdeef8ed638cbd198bf430cceef252e4342c9", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-privileged--zfv2sld.json b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-privileged--zfv2sld.json new file mode 100644 index 0000000000000000000000000000000000000000..4d0e4ba814b163d3cb6e293ac3ac5facdf6a5214 --- /dev/null +++ b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-privileged--zfv2sld.json @@ -0,0 +1,263 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-bookshelf-simple-i00-privileged--zfv2sld", + "taskKey": "humanoidbench/manip_bookshelf_simple", + "dag_sha256": "2204f90476ae603026750aa49178a8c18a5f3ae6b8197cf0e7e8d3ecbe659563", + "outcome": { + "success": 0, + "reward": 0.7462852771027741, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 1000, + "final_reward": 0.7462852771027741, + "sim_steps": 1000, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.954, + "peak_processes": 9, + "peak_disk_gb": 0.256, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-bookshelf-simple-i00-privileged--zfv2sld/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "f5c981f6a18acb13b9a20565d5610cc00af34fdded1f5a025131d589ecf358f2" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 5, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "object_1_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.0 s,浅色奖杯直立在红色书本旁,上一层可见对应的半透明目标;奖杯实体与目标分处两层,初态未到位。" + }, + { + "t": 4.9, + "log_step": null, + "text_zh": "观察:视频 0.0–4.8 s,机器人抬臂后停在书架前;视频 4.85–5.15 s,机器人前倾、双手靠近层板下方,奖杯仍直立在原层,没有随手抬起。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 5.5–10.05 s,机器人保持前倾伸手姿势,奖杯仍在书本旁;逐帧检查视频 9.0–10.05 s,未见奖杯被搬到上一层目标处。" + }, + { + "t": null, + "log_step": 17, + "text_zh": "观察:log step 17 的初态检查将首件识别为 trophy,初始原点为 [0.95, 0.02, 1.25],对应目标为 [0.75, -0.25, 1.55];此条用于识别物体和目标,不把会话仿真时间映射为视频时间。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察:视频 4.85–10.05 s,机器人双手在奖杯所在层板附近伸入;手掌内侧及部分指尖被身体和层板遮住,无法确定是否与奖杯有轻微接触。没有可确认的包握、夹持或奖杯随手移动,故不记抓握。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–10.05 s,机器人的双脚着地、双腿承重,伸手和前倾期间保持站立,未见摔倒。" + } + ] + }, + "object_2_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 0.0–10.05 s,机器人始终在首件奖杯附近伸手;红色书本仍成排留在原层,下面一层的物件也未被取走,未见针对次浅红书本(book_c)的独立搬放。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频全程首件奖杯未到位,按 DAG 的依序记入规则,本节点的前置事件未满足,因此本次评分回合未达到该节点。" + } + ] + }, + "object_3_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 0.0–10.05 s,机器人始终在首件奖杯附近伸手;红色书本仍成排留在原层,下面一层的物件也未被取走,未见针对中红物体(chalk)的独立搬放。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频全程首件奖杯未到位,按 DAG 的依序记入规则,本节点的前置事件未满足,因此本次评分回合未达到该节点。" + } + ] + }, + "object_4_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 0.0–10.05 s,机器人始终在首件奖杯附近伸手;红色书本仍成排留在原层,下面一层的物件也未被取走,未见针对次深红书本(book_e)的独立搬放。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频全程首件奖杯未到位,按 DAG 的依序记入规则,本节点的前置事件未满足,因此本次评分回合未达到该节点。" + } + ] + }, + "object_5_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 0.0–10.05 s,机器人始终在首件奖杯附近伸手;红色书本仍成排留在原层,下面一层的物件也未被取走,未见针对最深红书本(book_b)的独立搬放。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频全程首件奖杯未到位,按 DAG 的依序记入规则,本节点的前置事件未满足,因此本次评分回合未达到该节点。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "unit": "reward", + "threshold": 2000, + "direction": "higher_better", + "value": null + }, + { + "name": "obj_goal_dist", + "unit": "m", + "threshold": 0.15, + "direction": "lower_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "object_1_at_slot", + "t": 4.9, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频 4.85–5.15 s,机器人前倾向首件奖杯伸手后停在层板附近,奖杯没有离开原层;视频 5.5–10.05 s 持续保持这一姿势,未完成搬向上一层目标的动作。推断:首件放置中的取持阶段停滞且之后未修复,是最早未完成的步骤;视频 4.9 s 为靠近后停滞的粗略定位点。" + }, + "unmet": { + "object_1_at_slot": false, + "object_2_at_slot": false, + "object_3_at_slot": false, + "object_4_at_slot": false, + "object_5_at_slot": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "没建立接触", + "t": 4.9, + "log_step": null, + "evidence_zh": "观察:视频 4.85–5.5 s,机器人的手停在奖杯所在层的层板附近,奖杯始终直立在书本旁;视频 9.0–10.05 s,双手仍未带起奖杯,也没有向上一层搬运。推断:本次伸手未建立能够取持、搬运奖杯的有效接触;此处不排除被遮挡处有轻触,不宣称接触对数量为零。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "观察:log step 294 的最终生成脚本把双肩、双腕及腿部设为固定关节目标,手指目标保持初始化的零值,随后将同一 action 重复至多1000个控制步;脚本没有抓取后抬升、搬往目标的动作段。log steps 111–112 的会话测试中,双臂接近目标的控制也出现明显跟踪偏差;log step 179 的另一会话闭手测试未移动奖杯。推断:这些会话记录仅支持根因;结合视频 4.85–10.05 s 中伸手停滞、未取起奖杯,说明取持与搬运控制未实现。现有证据不足以进一步区分接触设计、轨迹或跟踪问题。", + "source_zh": "日志", + "log_step": 294 + } + ], + "gap_zh": "输入没有状态日志或成功参考视频;现有单路视频分辨率为256×256,身体和层板遮住部分手指,能判断奖杯未搬走,但不能确定手与奖杯是否曾轻触及精确接触方式。会话中的碰撞、闭手和摔倒测试不作为本次交付轨迹的接触或现象证据。评分器仅报告 final_reward=0.7462852771027741,未报告累计 episode_return、obj_goal_dist 或阶段计数,因此两项原生指标留 null;控制根因保留在大类,未进一步归因。" + }, + "summary_zh": "机器人在视频 0.0–10.05 s 保持站立、向首件浅色奖杯伸手,约视频 4.9 s 前倾后停在层板附近;奖杯始终留在原层,五个依序到位事件均未达到,最终为0/5。断点为首件到位的取持阶段:未建立可搬运奖杯的有效接触,最终脚本仅重复固定姿势,未实现取持及目标搬运控制。部分手指接触被遮挡,且无状态日志和累计回报字段,因此不细分接触方式或控制子因。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "视频全程奖杯未离开原层,五节点均未到达,final=0 与 final_reward 0.746(无阶段奖金)一致;断点为 object_1_at_slot 操作中,现象“没建立接触”、根因“控制”均有视频与 log step 294 脚本支撑,且已声明遮挡与会话测试的局限。未发现需要改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:59:27Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-privileged--zfv2sld/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-privileged--zfv2sld/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "6bcd962ede29836fadc8989abe2d842eea4fe7d16b870cea9f4adfaa1b41c7a4", + "video_sha256": "f5c981f6a18acb13b9a20565d5610cc00af34fdded1f5a025131d589ecf358f2", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard--8opf8kk.json b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard--8opf8kk.json new file mode 100644 index 0000000000000000000000000000000000000000..8d1c3f525678bcae58a89caee9bd89fa643a3d85 --- /dev/null +++ b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard--8opf8kk.json @@ -0,0 +1,219 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-bookshelf-simple-i00-standard--8opf8kk", + "taskKey": "humanoidbench/manip_bookshelf_simple", + "dag_sha256": "2204f90476ae603026750aa49178a8c18a5f3ae6b8197cf0e7e8d3ecbe659563", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 170, + "sim_steps": 725, + "calls": 68, + "rows": 725, + "resets": 5, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1185.9, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-bookshelf-simple-i00-standard--8opf8kk/replay.mp4", + "duration_s": 1.8, + "fps": 20.0, + "frames": 36, + "sha256": "528ef786823398e1905cd114d59f47bf38ec8c2176ebf341e7f247bb566d5ed2" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 5, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "object_1_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.75, + "log_step": null, + "text_zh": "观察:逐帧检查视频 0.00–1.75 s,机器人张臂、俯身再深蹲;最浅红色立柱状物体仍留在原来的层格,未被拿起送往上方同色半透明目标。视频 1.75 s 末帧仍未到位。" + } + ] + }, + "object_2_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.75, + "log_step": null, + "text_zh": "观察:视频 0.00–1.75 s 全部36帧中,机器人没有拿取或搬放次浅红物体,书架物体布局未发生可见变化。推断:首件最浅红物体始终未到位,此节点的按序前置条件未成立,故第2件到位事件未被达成。" + } + ] + }, + "object_3_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.75, + "log_step": null, + "text_zh": "观察:视频 0.00–1.75 s 全部36帧中,机器人没有拿取或搬放中红物体,书架物体布局未发生可见变化。推断:首件最浅红物体始终未到位,此节点的按序前置条件未成立,故第3件到位事件未被达成。" + } + ] + }, + "object_4_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.75, + "log_step": null, + "text_zh": "观察:视频 0.00–1.75 s 全部36帧中,机器人没有拿取或搬放次深红物体,书架物体布局未发生可见变化。推断:首件最浅红物体始终未到位,此节点的按序前置条件未成立,故第4件到位事件未被达成。" + } + ] + }, + "object_5_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.75, + "log_step": null, + "text_zh": "观察:视频 0.00–1.75 s 全部36帧中,机器人没有拿取或搬放最深红物体,书架物体布局未发生可见变化。推断:首件最浅红物体始终未到位,此节点的按序前置条件未成立,故第5件到位事件未被达成。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "unit": "reward", + "threshold": 2000, + "direction": "higher_better", + "value": null + }, + { + "name": "obj_goal_dist", + "unit": "m", + "threshold": 0.15, + "direction": "lower_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "object_1_at_slot", + "t": 0.0, + "log_step": 46, + "where": "首步前", + "evidence_zh": "观察:视频 0.00–1.75 s 只见机器人张臂、俯身和下蹲,未开始最浅红物体的拿取或向目标搬运。log step 38 是最后一次重置,log steps 38、40、42、44、46 的脚本只改变身体姿态以查看书架;log step 46 的机器人服务返回 done: True,此后没有新的机器人动作调用。推断:最后一局在首件搬放开始前已经结束;log step 84 的过滤器中断发生在该局结束之后。" + }, + "unmet": { + "object_1_at_slot": false, + "object_2_at_slot": false, + "object_3_at_slot": false, + "object_4_at_slot": false, + "object_5_at_slot": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 46, + "evidence_zh": "观察:视频 0.00–1.75 s 中机器人双手未握住目标物体,也没有搬放动作;最浅红物体及其上方目标保持分离。log step 46 已返回 done: True;log steps 71–83 后续仅做离线逆解与手指运动学检查,没有执行搬放。推断:评分局结束时,第一件物体的实际搬放尚未启动。" + } + ], + "root_causes": [], + "gap_zh": "没有状态日志或成功参考视频,无法核实手与书架的细小接触、物体精确距离和骨盆高度。log step 46 仅打印 done: True,没有打印 reason 或具体终止分量;视频末段显示深蹲但未显示明确摔倒,不能据此确定失衡或哪项高度规则触发,故不补写控制或认知根因。log step 84 是模型过滤器中断,但此前 log step 46 的在线回合已结束,按 D37 保留该回合的正常失败标注。评分字段没有报告 episode_return 或 obj_goal_dist,两个原生指标均留空;不把目录 reward=0.0 当作累计回报。" + }, + "summary_zh": "评分视频内机器人只张臂、俯身和深蹲观察书架,没有搬放物体,五个按序到位节点均未达成;断点为最浅红物体到位的首步前,现象为没尝试。log step 46 的在线回合已经结束,后来的 log step 84 过滤器中断不改变该局结果;缺少具体终止原因及状态日志,无法可靠归因,也不将深蹲直接标成摔倒。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "标注与证据一致:评分局视频(1.8 s,36帧)只见机器人张臂、俯身和深蹲,未触碰任何物体,5个节点均未达成;断点首件物体首步前、现象没尝试,log step 46 返回 done: True 且无终止原因,根因留空并写明 gap_zh,处理合理。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:00:03Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard--8opf8kk/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard--8opf8kk/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "f3b014423f591af090910e9a789a24ad5c4583ee5d93df83767e480881440619", + "video_sha256": "528ef786823398e1905cd114d59f47bf38ec8c2176ebf341e7f247bb566d5ed2", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard--bngpntm.json b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard--bngpntm.json new file mode 100644 index 0000000000000000000000000000000000000000..2d5365ed333f935a942c95aa53dbf8073ea9122a --- /dev/null +++ b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard--bngpntm.json @@ -0,0 +1,264 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-bookshelf-simple-i00-standard--bngpntm", + "taskKey": "humanoidbench/manip_bookshelf_simple", + "dag_sha256": "2204f90476ae603026750aa49178a8c18a5f3ae6b8197cf0e7e8d3ecbe659563", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 1000, + "sim_steps": 19192, + "calls": 19270, + "rows": 19192, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1824.6, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.303, + "peak_processes": 10, + "peak_disk_gb": 1.002, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-bookshelf-simple-i00-standard--bngpntm/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "45f8ac8b5bc01f3700225c227ec2e783cc57400d3fb69b188fe397f981bb6af1" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 5, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "object_1_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.0 s,最浅粉红色带底座立杆仍在书架上部、红色书本左侧;其同色半透明副本位于更高一层,实体与目标明显分离。" + }, + { + "t": 0.7, + "log_step": null, + "text_zh": "观察:视频 0.0–1.5 s,机器人右臂向书架伸出后弯肘,右手停在该立杆下方的层板附近,未包住立杆;视频 5.0 s、8.5–10.0 s 及末帧视频 10.05 s,立杆仍在原位置,没有提起或送到上层副本。" + }, + { + "t": null, + "log_step": 181, + "text_zh": "观察:log step 181 的第三次 trial 是最后一次重置后的动作:脚本用 G.REACH 设置右臂,插值70个控制步后保持700个控制步;log step 186 确认当前为 episode 51、control_steps=770、resets_left=0,log step 188 继续该局至1000步,服务返回 failed。先前各重置局不计入本轨迹。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–10.05 s,机器人双脚与腿在右臂伸向书架时承重;右手没有包住最浅粉红色立杆,未观察到对该物体的抓握接触。" + } + ] + }, + "object_2_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 0.0–10.05 s,机器人始终在同一书架前伸臂保持,没有搬运次浅红物体到对应副本的动作;视频 8.5–10.0 s 与末帧视频 10.05 s 未见新的上架结果。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频全段最浅粉红色立杆未离开原处、未达到其目标,第1件到位事件未发生;本节点是要求此前各件按序记入的第2件事件,故未达到,而非仅由总评分失败推定。" + } + ] + }, + "object_3_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 0.0–10.05 s,机器人始终在同一书架前伸臂保持,没有搬运中红物体到对应副本的动作;视频 8.5–10.0 s 与末帧视频 10.05 s 未见新的上架结果。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频全段最浅粉红色立杆未离开原处、未达到其目标,第1件到位事件未发生;本节点是要求此前各件按序记入的第3件事件,故未达到,而非仅由总评分失败推定。" + } + ] + }, + "object_4_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 0.0–10.05 s,机器人始终在同一书架前伸臂保持,没有搬运次深红物体到对应副本的动作;视频 8.5–10.0 s 与末帧视频 10.05 s 未见新的上架结果。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频全段最浅粉红色立杆未离开原处、未达到其目标,第1件到位事件未发生;本节点是要求此前各件按序记入的第4件事件,故未达到,而非仅由总评分失败推定。" + } + ] + }, + "object_5_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 0.0–10.05 s,机器人始终在同一书架前伸臂保持,没有搬运最深红物体到对应副本的动作;视频 8.5–10.0 s 与末帧视频 10.05 s 未见新的上架结果。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频全段最浅粉红色立杆未离开原处、未达到其目标,第1件到位事件未发生;本节点是要求此前各件按序记入的第5件事件,故未达到,而非仅由总评分失败推定。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "unit": "reward", + "threshold": 2000, + "direction": "higher_better", + "value": null + }, + { + "name": "obj_goal_dist", + "unit": "m", + "threshold": 0.15, + "direction": "lower_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "object_1_at_slot", + "t": 0.7, + "log_step": 181, + "where": "操作中", + "evidence_zh": "观察:视频约 0.7 s,机器人完成最初伸臂后,右手仍在最浅粉红色立杆下方,立杆没有被抓住或搬离;这一状态持续至视频 10.05 s,首件未到上层同色目标。log step 181 的最后一个 trial 启动该伸臂动作,log step 188 将同一 episode 51 继续至1000控制步,服务返回 failed,首件失败未修复。" + }, + "unmet": { + "object_1_at_slot": false, + "object_2_at_slot": false, + "object_3_at_slot": false, + "object_4_at_slot": false, + "object_5_at_slot": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "没建立接触", + "t": 0.7, + "log_step": null, + "evidence_zh": "观察:视频 0.0–1.5 s 机器人右手伸向书架后停在最浅粉红色立杆下方;视频 5.0 s、8.5–10.0 s 及末帧视频 10.05 s,手中没有该立杆,立杆始终留在红色书本左侧原处,没有建立搬运它所需的抓握接触。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志 / 视频", + "log_step": 132, + "evidence_zh": "推断:机器人未能把右手执行到可抓住首件的位置。会话测试中,log step 132 的脚本目标为手掌坐标[0.52,-0.1,0.6],由其本体读数计算的实际位置为[0.454,-0.276,0.076];log step 135 将肩俯仰指令从-0.50改到-2.00 rad,服务本体读数仅从-0.112变到-0.212 rad,显示指令与实际手臂姿态存在明显偏差。最后评分局在 log step 181 仍采用右臂姿态指令并保持,视频 0.7–10.05 s 的右手持续低于立杆,与该执行问题一致;这些测试不作为评分局已完成动作的证据,具体控制子类型尚不能确定。" + } + ], + "gap_zh": "没有状态日志、成功参考视频、物体与目标三维坐标或接触对,不能精确测量0.15米距离及层板附近手指的细微接触;但首件实体与上层目标明显分离,且没有搬运动作,足以判定事件未发生。log step 195 中 agent 自述物体超出手臂可达范围,相关定位和运动学计算均由其脚本生成,尚不足以证明所有合法姿态均不可达,也不足以判为任务或环境问题。评分器未报告原生累计回报及物体目标距离,两项 native_metrics 留空;视频时长10.1秒与该局1000×0.02秒控制时长不同,本标注时间直接取所看视频帧,不把控制步或agent时间换算为视频时间。" + }, + "summary_zh": "观察:评分视频中机器人伸出右臂并保持站立,但最浅粉红色立杆一直留在原处,五个按序到位节点均未达到(0/5)。推断:断点为首件到位,右手未建立抓握接触;会话的指令与实际关节姿态偏差支持控制问题,最后一局持续保持偏低的伸手姿态至失败。观察:输入没有状态日志,agent在 log step 195 所称物理不可达尚未独立核实。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "视频帧显示机器人右手始终停在最浅粉红立杆下方的层板处,立杆全程未动,五个节点均未达到,断点 object_1_at_slot、现象没建立接触、根因控制(item 控制,子类型不明)均有证据支持;时钟与 log step 均无误,会话测试已被标注为不计入评分局。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:01:48Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard--bngpntm/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard--bngpntm/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "169a506b42d03395c262a89642ce44d4bfcef7cd1c431792452c555659e0ea44", + "video_sha256": "45f8ac8b5bc01f3700225c227ec2e783cc57400d3fb69b188fe397f981bb6af1", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard--bv4rwq4.json b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard--bv4rwq4.json new file mode 100644 index 0000000000000000000000000000000000000000..cc30a466cb5fefe4aa4d52a269105b199643c145 --- /dev/null +++ b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard--bv4rwq4.json @@ -0,0 +1,263 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-bookshelf-simple-i00-standard--bv4rwq4", + "taskKey": "humanoidbench/manip_bookshelf_simple", + "dag_sha256": "2204f90476ae603026750aa49178a8c18a5f3ae6b8197cf0e7e8d3ecbe659563", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 1000, + "sim_steps": 6352, + "calls": 6208, + "rows": 6352, + "resets": 22, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1001.8, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-bookshelf-simple-i00-standard--bv4rwq4/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "c340041cd33fdffea6fa80ff3210e3478bb4ea8110e2a581d7b3b978f2f041df" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 5, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "object_1_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频0.0 s,最浅红色立柱状物体在自上而下第二层,顶层有其同色半透明目标;实体与目标分处两层。" + }, + { + "t": 2.4, + "log_step": null, + "text_zh": "观察:视频1.5–2.4 s,机器人右手向立柱状物体伸入并合指,手指与拇指围住柱身;物体仍在原层,未被提至顶层目标。" + }, + { + "t": 10.05, + "log_step": 70, + "text_zh": "观察:视频2.4–10.05 s,机器人右手和物体基本停留在第二层,顶层目标未被实体占据,机器人保持站立;视频9.0–10.05 s逐帧检查亦未见搬运。log step 70的服务状态为最后一局episode=23、control_steps=1000、done=true、reason=failed。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "抓握", + "note_zh": "观察:视频约2.0–2.4 s,机器人右手手指及拇指从两侧围住最浅红色立柱状物体的柱身,随后保持合指至视频10.05 s;物体未离开原层,不把合指等同于成功提起。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频0.0–10.05 s,机器人双脚接地、双腿屈曲承重,在右手接近及围握物体时支撑身体,未见倒地。" + } + ] + }, + "object_2_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频0.0–10.05 s全过程及视频9.0–10.05 s末秒,机器人只向最浅红色立柱状物体伸右手并保持,未见机器人另行搬运次浅红色物体到对应半透明目标。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频0.0–10.05 s最浅红物体从未进入顶层目标附近,按DAG逐件递进的验收顺序,第2件次浅红物体的到位事件未成立。" + } + ] + }, + "object_3_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频0.0–10.05 s全过程及视频9.0–10.05 s末秒,机器人只向最浅红色立柱状物体伸右手并保持,未见机器人另行搬运中红色物体到对应半透明目标。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频0.0–10.05 s最浅红物体从未进入顶层目标附近,按DAG逐件递进的验收顺序,第3件中红物体的到位事件未成立。" + } + ] + }, + "object_4_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频0.0–10.05 s全过程及视频9.0–10.05 s末秒,机器人只向最浅红色立柱状物体伸右手并保持,未见机器人另行搬运次深红色物体到对应半透明目标。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频0.0–10.05 s最浅红物体从未进入顶层目标附近,按DAG逐件递进的验收顺序,第4件次深红物体的到位事件未成立。" + } + ] + }, + "object_5_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频0.0–10.05 s全过程及视频9.0–10.05 s末秒,机器人只向最浅红色立柱状物体伸右手并保持,未见机器人另行搬运最深红色物体到对应半透明目标。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频0.0–10.05 s最浅红物体从未进入顶层目标附近,按DAG逐件递进的验收顺序,第5件最深红物体的到位事件未成立。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 2000, + "direction": "higher_better" + }, + { + "name": "obj_goal_dist", + "value": null, + "unit": "m", + "threshold": 0.15, + "direction": "lower_better" + } + ], + "failure": { + "breakpoint": { + "ref": "object_1_at_slot", + "t": 2.4, + "log_step": 69, + "where": "操作中", + "evidence_zh": "观察:视频约2.4 s机器人右手合指围住最浅红物体后,至视频10.05 s仍停留在原层,未将其搬到顶层目标。log step 69最终脚本只设接近、合指及保持,log step 70运行该脚本后的服务状态为episode=23、1000控制步、failed;第一件到位步骤已开始但一直未完成。" + }, + "phenomena": [], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志", + "evidence_zh": "推断:log step 55、59、62的测试脚本曾尝试抽回和抬高手中物体,log step 69最终脚本却只保留接近与合指保持;agent在log step 73自述其控制器不能后退,无法完成所需的取出和上移。结合视频2.4–10.05 s物体仍停在原层,将原因保守归为未实现取出及搬运的控制能力;测试动作仅作根因线索,不作为评分回合发生碰撞的证据,尚不能细分为接触设计、逆解或反馈控制。" + } + ], + "gap_zh": "评分视频存在,已检查全过程、右手合指附近的逐帧画面及最后一秒;没有状态日志和成功oracle。最终脚本在log step 69明确只接近、合指并保持,评分视频也没有向顶层目标执行抬升或搬运,故现象层留空:不能把静止未到位写成一次瞄准目标后的位置误差,也不能用测试局的卡挡自述替代评分局接触证据。房间视角仅256×256,无法确认握力或遮挡处是否存在阻断接触;agent在log step 73所称层板和自身躯干阻挡的具体机制及尺寸待核实。run.json未报告episode_return与obj_goal_dist,两个原生指标均填null;所有视频锚点直接取自帧,未将agent时间或控制步数换算成视频时间。", + "unmet": { + "object_1_at_slot": false, + "object_2_at_slot": false, + "object_3_at_slot": false, + "object_4_at_slot": false, + "object_5_at_slot": false + } + }, + "summary_zh": "评分回合中机器人保持站立,右手接近并围握最浅红色立柱状物体,随后停留在原层,五个按序到位事件均未完成,进度为0/5。断点为最浅红物体到位的操作中;测试记录和最终脚本支持将原因保守归为控制能力不足,但评分回合没有实际执行上移搬运,无法给出有证据的具体失败现象。无状态日志或成功oracle,握力与具体卡挡机制未能核实,原生回报和距离指标亦未报告。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "0/5、断点(第1件到位、操作中)与右手合指停留在第二层的描述与视频一致;现象层留空并给出 gap_zh,根因保守归为控制,已声明依据为日志线索而非评分局接触证据,未发现需要改动内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:59:02Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard--bv4rwq4/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard--bv4rwq4/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "7ac50656d48e7ed25b2e321306241e275073e46bbcaa7e325262012b29697475", + "video_sha256": "c340041cd33fdffea6fa80ff3210e3478bb4ea8110e2a581d7b3b978f2f041df", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard--hk5ursp.json b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard--hk5ursp.json new file mode 100644 index 0000000000000000000000000000000000000000..a1eca6185f0cdea70fb940bdc29d3c393e20a645 --- /dev/null +++ b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard--hk5ursp.json @@ -0,0 +1,253 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-bookshelf-simple-i00-standard--hk5ursp", + "taskKey": "humanoidbench/manip_bookshelf_simple", + "dag_sha256": "2204f90476ae603026750aa49178a8c18a5f3ae6b8197cf0e7e8d3ecbe659563", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 1000, + "sim_steps": 8868, + "calls": 2277, + "rows": 8643, + "resets": 40, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 2557.1, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-bookshelf-simple-i00-standard--hk5ursp/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "db6dd40434ff46e26d135b2c851ae882a093362855b1dde29d65505dcec05514" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 5, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "object_1_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频0.00 s最浅粉红色立柱状实体位于书架自上而下第二层左侧,对应细长半透明目标位于更高一层,两者明显分离。" + }, + { + "t": 1.5, + "log_step": null, + "text_zh": "观察:逐帧检查视频0.50–2.60 s,机器人抬臂朝书架伸手,但最浅粉红立柱仍留在原层板,没有随手离架或向上层目标移动;视频3.00–9.75 s的间隔帧及视频9.00–10.05 s的逐帧检查也未见搬运。" + }, + { + "t": 10.05, + "log_step": 376, + "text_zh": "观察:视频10.05 s最浅粉红立柱仍在原层板,上层对应目标未被该实体占据;log step 376记录最终第41回合在control_steps=1000时done=True、reason=failed。未将此前调试回合的结果计入。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察:视频0.50–2.60 s机器人向书架抬臂,操作侧手指受躯干和视角遮挡,无法确认是否发生轻微擦碰;最浅粉红立柱未随手移动,不能据脚本的抓取阶段名称标为抓握。log step 375–376显示脚本控制右手执行该尝试。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频0.00–10.05 s机器人双脚着地,屈膝承重支撑书架前的伸臂操作,未见跌倒。" + } + ] + }, + "object_2_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频0.00–10.05 s未见最浅粉红立柱离开第二层到达上层目标,第一件到位事件未发生;第2件次浅红物体的到位事件受DAG既定顺序约束,因此未达成。全段间隔帧及视频末秒逐帧检查未见机器人转而搬运该件。" + } + ] + }, + "object_3_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频0.00–10.05 s未见最浅粉红立柱离开第二层到达上层目标,第一件到位事件未发生;第3件中红物体的到位事件受DAG既定顺序约束,因此未达成。全段间隔帧及视频末秒逐帧检查未见机器人转而搬运该件。" + } + ] + }, + "object_4_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频0.00–10.05 s未见最浅粉红立柱离开第二层到达上层目标,第一件到位事件未发生;第4件次深红物体的到位事件受DAG既定顺序约束,因此未达成。全段间隔帧及视频末秒逐帧检查未见机器人转而搬运该件。" + } + ] + }, + "object_5_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频0.00–10.05 s未见最浅粉红立柱离开第二层到达上层目标,第一件到位事件未发生;第5件最深红物体的到位事件受DAG既定顺序约束,因此未达成。全段间隔帧及视频末秒逐帧检查未见机器人转而搬运该件。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "unit": "reward", + "threshold": 2000, + "direction": "higher_better", + "value": null + }, + { + "name": "obj_goal_dist", + "unit": "m", + "threshold": 0.15, + "direction": "lower_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "object_1_at_slot", + "t": 1.5, + "log_step": 376, + "where": "操作中", + "evidence_zh": "观察:视频0.50–1.50 s机器人抬臂伸向书架,但最浅粉红立柱没有随手移动;视频2.00 s及10.05 s该件仍留在原层板,与上层同色半透明目标分离。log step 376记录最后第41回合执行第一件的伸手、合手、抬举、松手程序后保持姿态,最终reason=failed;该步此后未完成。视频1.50 s仅为失败尝试的近似导航锚点。" + }, + "unmet": { + "object_1_at_slot": false, + "object_2_at_slot": false, + "object_3_at_slot": false, + "object_4_at_slot": false, + "object_5_at_slot": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "没建立接触", + "t": 1.5, + "log_step": 376, + "evidence_zh": "推断:视频0.50–2.60 s机器人完成一轮抬臂动作,最浅粉红立柱一直站在第二层原位;视频3.00–10.05 s该件也没有随手移动或到达上层目标。因此未建立能带走该件的有效抓取接触;本标签不表示已排除所有瞬时擦碰,也没有已抓住后滑落的证据。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "t": 1.5, + "log_step": 375, + "source_zh": "日志 / 视频", + "evidence_zh": "推断:log step 375的最终控制脚本按预设关节插值依次执行伸手、合手、抬举和松手,各阶段仅以done决定是否继续,随后保持既定姿态,未检查目标物是否随手运动并据此重试;log step 376确认这些阶段在最终第41回合执行。视频0.50–2.60 s及10.05 s显示第一件始终未被带走,说明该控制序列未实现有效取物且未修正;证据不足以细分为接触几何、逆解误差或跟踪问题。" + } + ], + "gap_zh": "没有状态日志或成功参考视频,256×256的房间相机画面部分遮挡操作侧手指,无法核实瞬时接触对、夹持力及精确三维距离,故实际手部接触标未知,控制根因不细分。评分器未报告episode_return或obj_goal_dist的原始数值,均保留null,不能把catalog reward=0.0当作累计回报。log step 376的1000控制步按脚本dt=0.02对应20 s,而所给视频仅10.1 s;没有核实两者映射,因此全部视频时间直接取自已检查帧,未把脚本阶段控制步换算为视频时刻。" + }, + "summary_zh": "最终第41回合尝试用右手操作最浅粉红立柱,但视频0.00–10.05 s该件始终留在原层板,未到上层目标,五个顺序到位事件均未达成(0/5);机器人保持双脚支撑,未见跌倒。断点为第一件到位的操作中:未建立有效抓取,预设控制序列继续抬举、松手并保持姿态,未修正取物失败。手指局部遮挡且无状态日志,无法细分控制根因或确认轻微擦碰,日志控制步与视频时间的对应关系也未核实。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "标签与视频一致:视频0.5–10.05 s最浅粉红立柱始终留在原层板,0/5,断点object_1、没建立接触、控制根因均有帧和log step 375–376支持,推断已标明,时钟未换算。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:54:47Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard--hk5ursp/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard--hk5ursp/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "d490aa6c9446e24da9e51a64aa3acc09acd11020ef8d43dd1e57f91e15284c23", + "video_sha256": "db6dd40434ff46e26d135b2c851ae882a093362855b1dde29d65505dcec05514", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard--huk5xts.json b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard--huk5xts.json new file mode 100644 index 0000000000000000000000000000000000000000..0a737cd886d19af3feb09c3530b792942ce02027 --- /dev/null +++ b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard--huk5xts.json @@ -0,0 +1,268 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-bookshelf-simple-i00-standard--huk5xts", + "taskKey": "humanoidbench/manip_bookshelf_simple", + "dag_sha256": "2204f90476ae603026750aa49178a8c18a5f3ae6b8197cf0e7e8d3ecbe659563", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 525, + "sim_steps": 14745, + "calls": 7439, + "rows": 14745, + "resets": 28, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3586.5, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-bookshelf-simple-i00-standard--huk5xts/replay.mp4", + "duration_s": 5.35, + "fps": 20.0, + "frames": 107, + "sha256": "eee17da127998f4208a76fb42a91265d11d4df20cc9069b1bdf7f8b404354aaa" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 5, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "object_1_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频0.00 s,最浅粉红色细杆、圆顶宽底物体立在书本所在层,匹配的半透明细杆目标位于上一层右侧,实体与目标明显分离。" + }, + { + "t": 3.7, + "log_step": null, + "text_zh": "观察:视频3.00–4.20 s,机器人右手在最浅红细杆物体旁收指并后撤,物体在原层架面上倾斜、挪动;视频5.00–5.30 s,右手继续抬至头旁,而该物体仍立在书本旁,没有随手升到上一层目标处。" + }, + { + "t": null, + "log_step": 207, + "text_zh": "观察:log step 206启动最后的episode 29,重放此前动作前缀后进行右手撤回、抬升;log step 207的服务返回episode=29、control_steps=525、done=False,评分器记录n_actions=525且success=0。该局末尾动作与视频相符,之前重置掉的测试不计作到位证据。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "推/扫", + "note_zh": "观察:视频3.65–4.20 s,机器人右手在最浅红细杆物体侧面接触并带动其倾斜、挪动;撤手抬高后物体仍留在架面。仅记录可见推拨,低分辨率画面不能确认手指形成对向包夹。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频0.00–5.30 s,机器人双脚着地、腿部承重,支撑右手接近和撤回的操作;末尾仍站立。" + } + ] + }, + "object_2_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 5.3, + "log_step": null, + "text_zh": "观察:视频0.00–5.30 s,机器人只对最浅红细杆物体进行取物动作,没有把次浅红物体搬至对应半透明目标的过程。" + }, + { + "t": 5.3, + "log_step": null, + "text_zh": "推断:视频全程最浅红物体未到上一层目标,第1件到位事件未成立,因此按DAG顺序门控,第2件到位事件也未成立;并非仅由总成功值为零推断。" + } + ] + }, + "object_3_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 5.3, + "log_step": null, + "text_zh": "观察:视频0.00–5.30 s,机器人只对最浅红细杆物体进行取物动作,没有把中红物体搬至对应半透明目标的过程。" + }, + { + "t": 5.3, + "log_step": null, + "text_zh": "推断:视频全程最浅红物体未到上一层目标,第1件到位事件未成立,因此按DAG顺序门控,第3件到位事件也未成立;并非仅由总成功值为零推断。" + } + ] + }, + "object_4_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 5.3, + "log_step": null, + "text_zh": "观察:视频0.00–5.30 s,机器人只对最浅红细杆物体进行取物动作,没有把次深红物体搬至对应半透明目标的过程。" + }, + { + "t": 5.3, + "log_step": null, + "text_zh": "推断:视频全程最浅红物体未到上一层目标,第1件到位事件未成立,因此按DAG顺序门控,第4件到位事件也未成立;并非仅由总成功值为零推断。" + } + ] + }, + "object_5_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 5.3, + "log_step": null, + "text_zh": "观察:视频0.00–5.30 s,机器人只对最浅红细杆物体进行取物动作,没有把最深红物体搬至对应半透明目标的过程。" + }, + { + "t": 5.3, + "log_step": null, + "text_zh": "推断:视频全程最浅红物体未到上一层目标,第1件到位事件未成立,因此按DAG顺序门控,第5件到位事件也未成立;并非仅由总成功值为零推断。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "unit": "reward", + "threshold": 2000, + "direction": "higher_better", + "value": null + }, + { + "name": "obj_goal_dist", + "unit": "m", + "threshold": 0.15, + "direction": "lower_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "object_1_at_slot", + "t": 4.2, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频3.65–4.20 s,机器人右手收指、后撤时,最浅红细杆物体仅在原层倾斜;视频5.00–5.30 s右手向上抬走,物体仍在书本旁。推断:第一件物体的取放尝试未形成可搬运的夹持,始终没有到达上层同色目标,之后没有修复。" + }, + "unmet": { + "object_1_at_slot": false, + "object_2_at_slot": false, + "object_3_at_slot": false, + "object_4_at_slot": false, + "object_5_at_slot": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "没建立接触", + "t": 4.2, + "log_step": null, + "evidence_zh": "观察:视频3.00–4.20 s,机器人右手虽触动最浅红物体使其倾斜,但未见物体被夹住并离开支撑面;视频5.00–5.30 s,右手抬高时物体仍在原层架面。推断:所需的搬运夹持未建立;可见触碰不等于成功抓取,亦无已提起后滑落的证据。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "视频 / 日志", + "evidence_zh": "推断:视频3.00–5.30 s机器人右手已接近正确物体并执行收指、后撤及上抬,却未能把物体带离原层,支持手部接触与搬运控制未奏效。观察:log step 206的脚本在最后一局加载拇指目标后直接撤手和上抬;log step 207返回的脚本计算显示手部抬升目标高度1.818米、实际估算1.788米,但视频末段物体仍留在架面。手部估算来自agent脚本,仅作辅助,不能据此证明物体被握住。" + } + ], + "gap_zh": "没有状态日志或成功oracle;256×256房间视角中,视频4.20–4.95 s物体与右手部分被躯干、头部遮挡,不能核实对向指尖接触、接触力或短暂夹持,故不声称已持物后滑落,控制根因保留到大类。评分器未报告episode_return或obj_goal_dist,两项原生指标留null;catalog reward=0.0不作为累计回报。log step 207的仿真10.5 s与视频5.35 s时长不同,未建立控制步到视频的映射,全部动作时间直接取自已检查的帧。" + }, + "summary_zh": "评分视频中机器人仅尝试搬放最浅红细杆物体:右手触动它使其在架面倾斜,随后抬手却没有带走物体,五个按序到位事件均未完成(0/5)。断点在第一件物体到位步骤,所需搬运夹持未建立,根因有证据支持到控制大类。没有状态日志,手指对向接触及遮挡期间的短暂夹持无法核实,未将此前重置测试的结果或仿真时钟直接作为视频进度。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "0/5、断点在第一件物体到位、没建立接触、控制根因均与视频一致;节点时间均取自视频帧,未转换agent时间,未发现需要修改的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:14:54Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard--huk5xts/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard--huk5xts/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "f96ef661fb360366a5ba48e577616cf0c07e347cbc52afc4c7af15d97ae109ec", + "video_sha256": "eee17da127998f4208a76fb42a91265d11d4df20cc9069b1bdf7f8b404354aaa", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard--lxm3boe.json b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard--lxm3boe.json new file mode 100644 index 0000000000000000000000000000000000000000..b026985836c79c7b8f9e50ffe984a02c1bab3316 --- /dev/null +++ b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard--lxm3boe.json @@ -0,0 +1,243 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-bookshelf-simple-i00-standard--lxm3boe", + "taskKey": "humanoidbench/manip_bookshelf_simple", + "dag_sha256": "2204f90476ae603026750aa49178a8c18a5f3ae6b8197cf0e7e8d3ecbe659563", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 995, + "sim_steps": 8401, + "calls": 172, + "rows": 8401, + "resets": 15, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 2141.0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-bookshelf-simple-i00-standard--lxm3boe/replay.mp4", + "duration_s": 10.05, + "fps": 20.0, + "frames": 201, + "sha256": "a201793e1cb11f32d902aa972a31b8709050ab1bfc4c4f8b4df5442bd38dd6f0" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 5, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "object_1_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.0 s,书本层的浅粉红实体与左侧同色半透明目标分开,尚未到位;机器人手位于更低的层板附近。" + }, + { + "t": 0.8, + "log_step": 47, + "text_zh": "观察:视频 0.0–1.5 s,机器人由屈膝姿态起身,两臂随后降至躯干两侧;手没有升到上部浅粉红书本的位置。视频 2.0、4.1、7.0、8.5、9.0–10.0 s,书本未被取走或移至对应副本;机器人保持站立。log step 47 的脚本针对最浅红书本计算右臂推扫目标,但画面不支持其已经推到书本。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–10.0 s,机器人双脚和双腿承重支撑起身及站立;未见右手与上部最浅红书本建立接触,因此不将脚本所称推书记为实际抓握或推扫接触。" + } + ] + }, + "object_2_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.0, + "log_step": 47, + "text_zh": "推断:视频 0.0–10.0 s 未见机器人把红色实体从原位搬到同色副本,末段视频 9.0–10.0 s 也没有放置动作;最浅红物体未到位。此第2件为按序锁存事件,前置 object_1_at_slot 未达成,故本节点全程未达到,而非仅凭 success=0 判否。log step 47 的最后一局脚本只尝试最浅红书本,未转入本件操作。" + } + ] + }, + "object_3_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.0, + "log_step": 47, + "text_zh": "推断:视频 0.0–10.0 s 未见机器人把红色实体从原位搬到同色副本,末段视频 9.0–10.0 s 也没有放置动作;最浅红物体未到位。此第3件为按序锁存事件,前置 object_2_at_slot 未达成,故本节点全程未达到,而非仅凭 success=0 判否。log step 47 的最后一局脚本只尝试最浅红书本,未转入本件操作。" + } + ] + }, + "object_4_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.0, + "log_step": 47, + "text_zh": "推断:视频 0.0–10.0 s 未见机器人把红色实体从原位搬到同色副本,末段视频 9.0–10.0 s 也没有放置动作;最浅红物体未到位。此第4件为按序锁存事件,前置 object_3_at_slot 未达成,故本节点全程未达到,而非仅凭 success=0 判否。log step 47 的最后一局脚本只尝试最浅红书本,未转入本件操作。" + } + ] + }, + "object_5_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.0, + "log_step": 47, + "text_zh": "推断:视频 0.0–10.0 s 未见机器人把红色实体从原位搬到同色副本,末段视频 9.0–10.0 s 也没有放置动作;最浅红物体未到位。此第5件为按序锁存事件,前置 object_4_at_slot 未达成,故本节点全程未达到,而非仅凭 success=0 判否。log step 47 的最后一局脚本只尝试最浅红书本,未转入本件操作。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "unit": "reward", + "threshold": 2000, + "direction": "higher_better", + "value": null + }, + { + "name": "obj_goal_dist", + "unit": "m", + "threshold": 0.15, + "direction": "lower_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "object_1_at_slot", + "t": 0.8, + "log_step": 47, + "where": "操作中", + "evidence_zh": "观察:log step 47 重置后执行第16局,以右臂逆解和关节位置指令尝试最浅红书本的连续推扫;log step 49 确认最后仍为第16局、control_steps=995。视频 0.8–1.5 s 手已降到书本层以下,此后视频 2.0–10.0 s 未见它触到该书本或使书本靠近副本,第一件始终未完成。此前各局的跌倒已被重置,视频中的本局没有跌倒。" + }, + "unmet": { + "object_1_at_slot": false, + "object_2_at_slot": false, + "object_3_at_slot": false, + "object_4_at_slot": false, + "object_5_at_slot": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "没建立接触", + "t": 0.8, + "log_step": 47, + "evidence_zh": "观察:视频 0.8–1.5 s 机器人两手停在上部书本层以下;视频 3.5–4.6 s、7.6–8.7 s 以及9.0–10.0 s 的密集画面中,手也没有到达最浅红书本,书本未被推动至左侧副本。没有出现握住后滑落或放置后失稳;log step 47 所称连续推书没有形成画面可见的目标接触。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "log_step": 31, + "source_zh": "日志", + "note_zh": "控制脚本关节索引映射错误。", + "evidence_zh": "推断:log step 17 的本体关节列表在左腕后先列完整左手关节,随后才列右肩至右腕;log step 7 将右肩、右肘、右腕列在关节索引40–44附近。log step 31 的 angles_to_u 却把 a69[0:21] 直接依次写给前21个身体执行器,没有按执行器关节名查 JIDX;右臂逆解写入的右肩至右腕角度因而没有传给相应身体执行器。log step 47 的最后一局仍调用这一转换函数;视频 0.8–10.0 s 右臂未抬到书本层,与此错误相符。该证据支持控制实现错误,不支持把 agent 自述的可达距离不足当作已验证的物理限制。" + } + ], + "gap_zh": "没有状态日志、逐件阶段计数、精确物体目标三维距离或实际累计回报,两个原生指标均保留null;局部物体会被机器人肩部遮挡,但已结合初态、手臂全程低于目标层及末段画面判断未发生到位事件。日志服务control_dt=0.02(log step 6),最终995控制步(log step 49)对应仿真19.90 s,而视频仅10.05 s;未核实二者采样映射,所有视频时间直接取自视频帧,不由控制步或agent工作时间换算。" + }, + "summary_zh": "评分对象是最后的第16局:机器人起身后一直站立,但未见任何红色物体移入同色目标,五个按序到位节点均未达成。断点是最浅红物体到位,现象为未建立目标接触;日志中的关节角到执行器索引映射错误可解释右臂未执行所计算的推书运动,早先试验中的跌倒不属于本局失败。没有状态日志、精确距离和累计回报,视频与控制步的时间映射也未核实。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "标签与视频一致:视频 0–10 s 机器人站立、双臂垂在身侧,红色书本及其余物体始终未移动,五个节点均未达成;断点在第1件、现象为没建立接触均成立。控制根因(angles_to_u 把 a69[0:21] 按序号直接写入执行器0–20,右臂逆解角度位于 proprio 索引约40以后,从未写入右臂执行器16–20)经 log step 29/31 的代码核实,与视频中右臂不抬起相符。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:17:40Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard--lxm3boe/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard--lxm3boe/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "ddfea5644ad6b867d342886984694d375c1b1f6c06bb80fedd0e6bdd849e054f", + "video_sha256": "a201793e1cb11f32d902aa972a31b8709050ab1bfc4c4f8b4df5442bd38dd6f0", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard--m56h4il.json b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard--m56h4il.json new file mode 100644 index 0000000000000000000000000000000000000000..f986be65d85da4507f818f04553fa223edb58ff9 --- /dev/null +++ b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard--m56h4il.json @@ -0,0 +1,271 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-bookshelf-simple-i00-standard--m56h4il", + "taskKey": "humanoidbench/manip_bookshelf_simple", + "dag_sha256": "2204f90476ae603026750aa49178a8c18a5f3ae6b8197cf0e7e8d3ecbe659563", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 880, + "sim_steps": 13044, + "calls": 406, + "rows": 12999, + "resets": 25, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3590.0, + "video_rendered": 1, + "first_success": 0, + "first_reason_code": 2, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.309, + "peak_processes": 9, + "peak_disk_gb": 0.845, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-bookshelf-simple-i00-standard--m56h4il/replay.mp4", + "duration_s": 8.9, + "fps": 20.0, + "frames": 178, + "sha256": "3b84cb25e4b3ea06319d96d5856051d197cd27a2b9881488a5ebc7741c327993" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 5, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "object_1_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 436, + "text_zh": "观察:视频0.0 s,最浅粉红色细长实体立在上层红书左侧,其同色半透明目标副本位于下一层板(左手所在层),与实体分离;左手停在目标层,没有伸到实体所在层。log step 436的服务返回episode=26、control_steps=0;该次reset后的最后一局是评分局,此前各局不计入本次进度。" + }, + { + "t": 1.0, + "log_step": 438, + "text_zh": "观察:视频0.6–2.4 s,机器人左手从较低层向上抬起并收拢手指,手停在粉红实体所在层板下方,未见手指包住实体;实体仍立在红书左侧。log step 438的脚本命令为left-reach-grip。" + }, + { + "t": 8.85, + "log_step": 445, + "text_zh": "观察:视频2.4–8.85 s,左臂继续上抬、躯干前倾,但最浅粉红实体始终未随手离开原层;视频7.8–8.85 s逐帧检查也未见实体移向下层同色目标。log step 440、442、444、445分别继续命令抬臂或保持,未产生可见搬运。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频0.0–8.85 s,机器人的双脚接地、双腿承重,支撑左臂伸向最浅粉红实体的尝试;未见倒地。左手在实体所在层板下方收指、抬升,未见其包夹实体,故不记抓握接触。" + } + ] + }, + "object_2_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 8.85, + "log_step": null, + "text_zh": "推断:视频0.0–8.85 s,机器人持续以左臂尝试取最浅粉红物体,未见搬运次浅红物体或将其送入同色目标;最浅红到位事件未发生,故按DAG的顺序前置条件,第2件到位事件也未触发。" + } + ] + }, + "object_3_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 8.85, + "log_step": null, + "text_zh": "推断:视频0.0–8.85 s,机器人持续以左臂尝试取最浅粉红物体,未见搬运中红物体或将其送入同色目标;最浅红到位事件未发生,故按DAG的顺序前置条件,第3件到位事件也未触发。" + } + ] + }, + "object_4_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 8.85, + "log_step": null, + "text_zh": "推断:视频0.0–8.85 s,机器人持续以左臂尝试取最浅粉红物体,未见搬运次深红物体或将其送入同色目标;最浅红到位事件未发生,故按DAG的顺序前置条件,第4件到位事件也未触发。" + } + ] + }, + "object_5_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 8.85, + "log_step": null, + "text_zh": "推断:视频0.0–8.85 s,机器人持续以左臂尝试取最浅粉红物体,未见搬运最深红物体或将其送入同色目标;最浅红到位事件未发生,故按DAG的顺序前置条件,第5件到位事件也未触发。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 2000, + "direction": "higher_better" + }, + { + "name": "obj_goal_dist", + "value": null, + "unit": "m", + "threshold": 0.15, + "direction": "lower_better" + } + ], + "failure": { + "breakpoint": { + "ref": "object_1_at_slot", + "t": 1.0, + "log_step": 438, + "where": "操作中", + "evidence_zh": "观察:视频约1.0 s,机器人左手抬起收指时仍在最浅粉红实体所在层板下方,没有包住实体;左手位于下一层的半透明目标所在层,而非实体所在层;随后视频2.4–8.85 s持续抬臂,实体仍留在红书左侧,未完成第一次搬运。log step 436明确切换为最后一局左手取物,log step 438、440、442、444、445执行该尝试,未再reset。" + }, + "phenomena": [ + { + "cat": "接触", + "item": "没建立接触", + "t": 1.0, + "log_step": 438, + "evidence_zh": "观察:视频0.6–2.4 s,机器人左手在目标实体下方收指,实体没有进入手指之间;视频2.4–8.85 s,手臂抬升而最浅粉红实体一直留在原位,未形成能带动物体的抓取接触。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "log_step": 440, + "source_zh": "日志 / 视频", + "evidence_zh": "推断:log step 438的脚本同时发出左手收指和伸臂命令,log step 440、442、444继续增大左肩抬升目标,并在log step 445保持;视频约1.0–8.85 s却显示左手始终没有把原层粉红实体包入手中,抬臂也没有带起实体。这支持取物动作的控制未建立有效抓取,不能仅凭该证据确定是接触几何、轨迹跟踪还是其他控制细项。" + } + ], + "gap_zh": "观察:inputs/中没有状态日志或成功oracle;视频仅有256×256的房间视角,能确认实体未搬动和未见包夹,不能分辨左指与层板是否有短暂轻触或测量接触力。评分器未报告episode_return、obj_goal_dist或逐件阶段;两项原生指标留null。日志末局累计880控制步,而视频长8.9 s,未验证其控制步到视频的映射,因此所有视频时刻均直接取自检查的帧,不由agent时间或控制步换算。", + "unmet": { + "object_1_at_slot": false, + "object_2_at_slot": false, + "object_3_at_slot": false, + "object_4_at_slot": false, + "object_5_at_slot": false + } + }, + "summary_zh": "评分的是log step 436重置后的第26局:视频中机器人左手尝试取最浅粉红物体,但该件始终留在原层,五个按序到位节点均未达成。断点为最浅红物体到位的操作中,左手未建立有效抓取接触;日志命令与画面支持控制失败,但缺少状态和力数据,无法细分控制原因或确认短暂层板接触。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 2, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "fix", + "conf": 7, + "summary_zh": "进度、断点(object_1_at_slot,操作中,没建立接触)与评分局界定(log step 436重置后的第26局)与视频和日志一致,五个节点均未达成。唯一需改处:证据文字把同色半透明目标写在实体“上层”,但视频首帧显示该半透明粉红副本在实体下方一层、正是左手停留的层板,实体在上一层、手始终没有够到它;这是对失败原因(手去了目标位而非实体)的错误描述。", + "n_issues": 1, + "n_major": 0 + }, + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 2, + "verdict": "agree", + "conf": 7, + "summary_zh": "第1轮修正已落实(目标副本在下一层、左手在目标层)。视频中五个节点均未达成,断点object_1_at_slot/操作中/没建立接触与画面一致,根因“控制”以推断标注且措辞保守,未发现需改内容。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:21:26Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard--m56h4il/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard--m56h4il/04_review_r1" + }, + { + "stage": "reconcile", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard--m56h4il/05_reconcile_r1" + }, + { + "stage": "traj_review", + "round": 2, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard--m56h4il/04_review_r2" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "6fc3ad446de3555644c2591b75338031f2d8e280c9ee7f6c89ae8378ea5116e9", + "video_sha256": "3b84cb25e4b3ea06319d96d5856051d197cd27a2b9881488a5ebc7741c327993", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard--zmpjuuc.json b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard--zmpjuuc.json new file mode 100644 index 0000000000000000000000000000000000000000..883f3ac2518abceec444a0e59135bcf0ae66ef89 --- /dev/null +++ b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard--zmpjuuc.json @@ -0,0 +1,247 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-bookshelf-simple-i00-standard--zmpjuuc", + "taskKey": "humanoidbench/manip_bookshelf_simple", + "dag_sha256": "2204f90476ae603026750aa49178a8c18a5f3ae6b8197cf0e7e8d3ecbe659563", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 1000, + "sim_steps": 19709, + "calls": 639, + "rows": 19709, + "resets": 42, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1938.4, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-bookshelf-simple-i00-standard--zmpjuuc/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "812c0a954a12114189d36af6fc10099d14476541c874c427101ae0f7da5c8aef" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 5, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "object_1_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.0 s,最浅红色瓶状实体位于自上而下第二层,顶部另一层可见其浅红色半透明目标副本;实体与目标明显分离。" + }, + { + "t": 3.0, + "log_step": null, + "text_zh": "观察:视频 2.4–3.5 s,机器人抬起右手向瓶子下方伸出;手没有包住瓶身,瓶子仍直立在原处,没有随手离架。" + }, + { + "t": 10.05, + "log_step": 159, + "text_zh": "观察:视频 5.2–6.7 s 的合指及调整后,直至视频 9.0–10.05 s,瓶子仍在原层,右手停在瓶子下方,没有把实体送到顶部目标;log step 159 的服务状态为 episode=43、control_steps=1000、reason=failed。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–10.05 s,机器人双脚接地、双腿承重支撑伸手操作;未见右手围住或夹住最浅红色瓶子,故不记目标抓握。手与书架边缘是否持续接触无法由该视角确定。" + } + ] + }, + "object_2_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频 0.0–10.05 s,机器人始终在最浅红色瓶状物附近伸手,没有完成第1件到位,也未见后续物体被搬往目标;第2件事件要求前序到位事件先记入,因此本局第2件到位未成立。" + } + ] + }, + "object_3_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频 0.0–10.05 s,机器人始终在最浅红色瓶状物附近伸手,没有完成第1件到位,也未见后续物体被搬往目标;第3件事件要求前序到位事件先记入,因此本局第3件到位未成立。" + } + ] + }, + "object_4_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频 0.0–10.05 s,机器人始终在最浅红色瓶状物附近伸手,没有完成第1件到位,也未见后续物体被搬往目标;第4件事件要求前序到位事件先记入,因此本局第4件到位未成立。" + } + ] + }, + "object_5_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频 0.0–10.05 s,机器人始终在最浅红色瓶状物附近伸手,没有完成第1件到位,也未见后续物体被搬往目标;第5件事件要求前序到位事件先记入,因此本局第5件到位未成立。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 2000, + "direction": "higher_better" + }, + { + "name": "obj_goal_dist", + "value": null, + "unit": "m", + "threshold": 0.15, + "direction": "lower_better" + } + ], + "failure": { + "breakpoint": { + "ref": "object_1_at_slot", + "t": 3.0, + "log_step": 159, + "where": "操作中", + "evidence_zh": "观察:视频 2.4–3.5 s,机器人右手抬至最浅红色瓶子下方后未抓到瓶身;视频 5.2–6.7 s 的合指和后续调整未使瓶子离开原位,视频 10.05 s 仍未修复。log step 159 记录最后一局 episode=43,服务在1000控制步后报告 failed;此前重置掉的测试不计入本段。" + }, + "unmet": { + "object_1_at_slot": false, + "object_2_at_slot": false, + "object_3_at_slot": false, + "object_4_at_slot": false, + "object_5_at_slot": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "没建立接触", + "t": 3.0, + "log_step": null, + "evidence_zh": "观察:视频 2.4–3.5 s 和视频 5.2–6.7 s,机器人右手在浅红瓶子的下方伸出并合指,始终未包握瓶身;瓶子没有随手运动,视频 9.0–10.05 s 仍留在原层。未建立搬运该瓶所需的抓持接触。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "参数调试", + "log_step": 121, + "source_zh": "日志", + "evidence_zh": "推断:任务文本明确动作±1映射到range_rad的两端;log step 48 给出右肩俯仰弧度范围[-2.87,2.87],log step 49 的q2a也按该弧度范围换算。可是log step 120 读到归一化lower/upper为[-1,1]后,log step 121 将它们直接作为弧度IK限值ARMLIM;log step 122 又将同一上下界用于弧度cmd的裁剪。log step 159 的最终脚本仍以QLO/QHI裁剪逆解和补偿指令。这把可用关节姿态人为限制到错误范围,是有代码证据的控制参数错误,解释了视频 3.0–10.05 s 伸手仍不到瓶身的失败;尚未验证纠正该参数便能完成抓取。" + } + ], + "gap_zh": "输入没有状态日志或成功参考视频,且房间视角仅256×256;无法确定手与书架边缘的细小接触、精确物体到目标距离或原生阶段数。log step 159 的0.259米是agent脚本基于固定脚位及运动学模型计算的掌心到估计瓶位距离,不是评分器obj_goal_dist,故不填作原生指标;评分器也未报告episode_return。任务文字称1000控制步为20秒机器人时间,但视频总长10.1秒,未验证其映射,所有导航时刻均直接取自视频,未把控制步换算成视频时间。参数错误由日志代码确认,其单独因果贡献及修正后可达性未经对照运行验证。" + }, + "summary_zh": "观察:评分视频对应最后一局,机器人双脚支撑着向最浅红色瓶状物伸手并合指,但瓶子没有被抓起或送到顶部目标,五个按序到位事件均未成立(0/5)。推断:断点为第1件放置步骤,现象为没建立接触;日志代码把归一化动作上下界误作弧度关节限值,支持控制·参数调试根因。缺少状态日志及原生距离、回报读数,无法细分手架接触或验证修正参数后的可达性。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "未发现需要改动的错误:视频帧显示右手始终在最浅红瓶下方,瓶子未离开原层,0/5、断点object_1_at_slot、没建立接触均有支持;控制·参数调试根因有日志代码证据(run.json id 120–122 将归一化±1当作弧度限值),并已标明为推断。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:01:11Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard--zmpjuuc/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard--zmpjuuc/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "f6940f855801702cb2c6b2fe5b3a46d6e170dc446bf824501bdb22df9c4348be", + "video_sha256": "812c0a954a12114189d36af6fc10099d14476541c874c427101ae0f7da5c8aef", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--bwyvw6f.json b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--bwyvw6f.json new file mode 100644 index 0000000000000000000000000000000000000000..933e3b700a123792a0bdc5168dedafe975031da8 --- /dev/null +++ b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--bwyvw6f.json @@ -0,0 +1,275 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-bookshelf-simple-i00-standard-realtime--bwyvw6f", + "taskKey": "humanoidbench/manip_bookshelf_simple", + "dag_sha256": "2204f90476ae603026750aa49178a8c18a5f3ae6b8197cf0e7e8d3ecbe659563", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 1000, + "sim_steps": 13246, + "calls": 18684, + "rows": 18624, + "resets": 24, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 850.5, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 264.608, + "max_lag_s": 0.0789, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 1935, + "late_steps": 0, + "requests_refused": 18, + "request_busy_s": 8.006, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.165, + "peak_processes": 8, + "peak_disk_gb": 0.246, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--bwyvw6f/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "88dbd0a0be5c32c67e7827cac11a66c648e0414fd055d271f6db96216f3cb15c" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 5, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "object_1_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频0.0 s中,最浅粉红色立杆状实体仍在从上往下第二层,配对的浅色半透明立杆目标在顶层;实体与目标明显分离,初始未到位。" + }, + { + "t": 3.5, + "log_step": null, + "text_zh": "观察:视频2.0–4.0 s逐帧检查可见机器人右臂向书架伸出,手停在书本右侧,未包住左侧最浅粉红色立杆;立杆仍留在原层。视频4.5–9.5 s及视频9.1–10.05 s末段逐帧检查均未见该实体离开原位或移向顶层目标。" + }, + { + "t": null, + "log_step": 51, + "text_zh": "观察:log step 51中,机器人服务记录最后一局episode=25、control_steps=1000、done=true、reason=failed;该局由最终版本脚本启动,之后没有重置。此前各局仅是会话试验,不作为本节点完成证据。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频0.0–10.05 s中,机器人双脚着地、双腿承重,在右臂伸向最浅粉红色立杆的过程中维持站立。右手未见与该实体建立接触或包夹,故不记抓握;左手保持在更低层前方,未见协同搬运。" + } + ] + }, + "object_2_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频0.0–9.5 s的全程抽帧及视频9.1–10.05 s末段逐帧画面中,机器人只伸出右臂并保持站立,书架上的红色实体未被取走或搬至同色半透明目标,未出现第2件的搬放。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频0.0、3.5和10.05 s均显示第1件最浅粉红色立杆仍在原层、未到顶层配对目标;第1件事件全程未达成,因此按DAG的顺序门控,第2件到位事件也没有成立。" + } + ] + }, + "object_3_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频0.0–9.5 s的全程抽帧及视频9.1–10.05 s末段逐帧画面中,机器人只伸出右臂并保持站立,书架上的红色实体未被取走或搬至同色半透明目标,未出现第3件的搬放。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频0.0、3.5和10.05 s均显示第1件最浅粉红色立杆仍在原层、未到顶层配对目标;第1件事件全程未达成,因此按DAG的顺序门控,第3件到位事件也没有成立。" + } + ] + }, + "object_4_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频0.0–9.5 s的全程抽帧及视频9.1–10.05 s末段逐帧画面中,机器人只伸出右臂并保持站立,书架上的红色实体未被取走或搬至同色半透明目标,未出现第4件的搬放。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频0.0、3.5和10.05 s均显示第1件最浅粉红色立杆仍在原层、未到顶层配对目标;第1件事件全程未达成,因此按DAG的顺序门控,第4件到位事件也没有成立。" + } + ] + }, + "object_5_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频0.0–9.5 s的全程抽帧及视频9.1–10.05 s末段逐帧画面中,机器人只伸出右臂并保持站立,书架上的红色实体未被取走或搬至同色半透明目标,未出现第5件的搬放。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频0.0、3.5和10.05 s均显示第1件最浅粉红色立杆仍在原层、未到顶层配对目标;第1件事件全程未达成,因此按DAG的顺序门控,第5件到位事件也没有成立。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "unit": "reward", + "threshold": 2000, + "direction": "higher_better", + "value": null + }, + { + "name": "obj_goal_dist", + "unit": "m", + "threshold": 0.15, + "direction": "lower_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "object_1_at_slot", + "t": 3.5, + "log_step": 51, + "where": "操作中", + "evidence_zh": "观察:视频约3.0–3.5 s中,机器人右臂完成向书架伸出的动作后空手停在书本右侧,最浅粉红色立杆仍在第二层,始终未被抓起或送到顶层同色目标;此状态持续至视频10.05 s,未修复。log step 51的最终局服务状态为episode=25、1000控制步后failed;断点锚点取可见的伸手后停滞,不取先前重置局的跌倒或碰撞。" + }, + "unmet": { + "object_1_at_slot": false, + "object_2_at_slot": false, + "object_3_at_slot": false, + "object_4_at_slot": false, + "object_5_at_slot": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "没建立接触", + "t": 3.5, + "log_step": null, + "evidence_zh": "观察:视频2.0–4.0 s逐帧画面显示机器人右手伸到书本右侧后停住,与左侧最浅粉红色立杆保持分离,没有手指包夹、提起或带动物体;视频9.1–10.05 s中立杆仍在原位、右手仍为空手。该件所需的操纵接触从未建立。" + } + ], + "root_causes": [ + { + "cat": "认知", + "item": "推理/规划", + "log_step": 50, + "source_zh": "日志", + "evidence_zh": "推断:log step 50列出的最终脚本把整局策略设为固定腿部站姿的平衡控制加三个右手接近航点,arm_at在最后航点后持续返回相同目标;hand默认为空字典,未加入合指抓取、持物移动至顶层配对目标的阶段。log step 51无参数启动该版本,视频3.5–10.05 s也只见机器人空手悬停。这个动作序列以手靠近物体为终点,缺少实现第1件到位的完整操作计划;归因依据脚本内容与可见动作,并非依据停止或零分。" + } + ], + "gap_zh": "观察:inputs/没有状态日志或成功参照视频;房间相机视频仅256×256像素,无法量化物体原点距离或排除细微擦碰,但可见目标实体未被抓起和搬放。评分字段没有报告episode_return或obj_goal_dist,两个原生指标均留null,不能把catalog的reward=0.0当作累计奖励。log step 51给出1000控制步,服务control_dt为0.02 s(log step 3),而视频长度为10.1 s,输入未给出同步映射;所有导航时间均直接取视频,不换算日志时间。agent关于必须行走才能触及目标的尺寸和可达性判断(log step 53)未获独立验证,故不标“够不着”或任务环境不可解。" + }, + "summary_zh": "最终评分局中机器人保持站立并伸出右臂,但未建立对最浅粉红色立杆的操纵接触,也未搬动任何目标物体,五个按序到位节点最终为0/5。断点是第1件到位:右手伸出后空停,最终脚本只有平衡和接近航点,缺少抓取及送往配对目标的完整操作序列。没有状态日志和成功参照,原生距离、累计奖励及日志与视频的精确时间映射均未提供。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "视频帧与日志均支持:最终局(episode 25)机器人站立、右臂伸向书架,最浅粉红立杆始终未被触及或移动,0/5、断点在第1件、没建立接触、推理/规划(脚本只有平衡+接近航点)均有依据;时钟未换算,未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:16:37Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--bwyvw6f/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--bwyvw6f/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "9e2c6a504cf9c69545f3508dc6e93a5dca78a10b19edc464937eaf0627c27622", + "video_sha256": "88dbd0a0be5c32c67e7827cac11a66c648e0414fd055d271f6db96216f3cb15c", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--ew5jrdt.json b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--ew5jrdt.json new file mode 100644 index 0000000000000000000000000000000000000000..7dcdd9492799feccc5f431283ff6f5d6b0822492 --- /dev/null +++ b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--ew5jrdt.json @@ -0,0 +1,255 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-bookshelf-simple-i00-standard-realtime--ew5jrdt", + "taskKey": "humanoidbench/manip_bookshelf_simple", + "dag_sha256": "2204f90476ae603026750aa49178a8c18a5f3ae6b8197cf0e7e8d3ecbe659563", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 1000, + "sim_steps": 6802, + "calls": 6177, + "rows": 6126, + "resets": 49, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 565.7, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 135.327, + "max_lag_s": 0.0029, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 1445, + "late_steps": 0, + "requests_refused": 34, + "request_busy_s": 3.112, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.179, + "peak_processes": 9, + "peak_disk_gb": 0.241, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--ew5jrdt/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "6e02a7cb78e0394c16f5299f024b86d1d9a3b85acad95d2920c5ce8f6d68efd9" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 5, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "object_1_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.0 s,浅红色实体仍在书架上部的实物排中,与上方同色半透明目标分开;视频 0.0–1.5 s 机器人只调整站姿、身体前倾,未见手包住或搬起该件。" + }, + { + "t": 10.05, + "log_step": 75, + "text_zh": "观察:全片按视频每 0.5 s 检查,并密查视频 9.0–10.05 s,机器人双臂一直停在书架中层附近,可见物体与目标的相对布局未出现搬放变化;log step 75 的最终控制器将双臂和手固定在初始关节目标,只调整站姿,没有抓取或送往目标的控制段。最浅红物体没有到位事件。" + } + ] + }, + "object_2_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": 75, + "text_zh": "观察:视频 0.0、0.5、1.0 s 至视频 9.5 s 的逐段画面及视频 9.0–10.05 s 的密集末段均未见机器人抓起或搬运次浅红物体;log step 75 的最终控制器只维持站姿、固定双臂与手。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频全程第1件未到位,此前置事件始终未成立,故按任务的原生依序记入规则,第2件的到位事件也未成立。" + } + ] + }, + "object_3_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": 75, + "text_zh": "观察:视频 0.0、0.5、1.0 s 至视频 9.5 s 的逐段画面及视频 9.0–10.05 s 的密集末段均未见机器人抓起或搬运中红物体;log step 75 的最终控制器只维持站姿、固定双臂与手。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频全程第1件未到位,此前置事件始终未成立,故按任务的原生依序记入规则,第3件的到位事件也未成立。" + } + ] + }, + "object_4_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": 75, + "text_zh": "观察:视频 0.0、0.5、1.0 s 至视频 9.5 s 的逐段画面及视频 9.0–10.05 s 的密集末段均未见机器人抓起或搬运次深红物体;log step 75 的最终控制器只维持站姿、固定双臂与手。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频全程第1件未到位,此前置事件始终未成立,故按任务的原生依序记入规则,第4件的到位事件也未成立。" + } + ] + }, + "object_5_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": 75, + "text_zh": "观察:视频 0.0、0.5、1.0 s 至视频 9.5 s 的逐段画面及视频 9.0–10.05 s 的密集末段均未见机器人抓起或搬运最深红物体;log step 75 的最终控制器只维持站姿、固定双臂与手。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频全程第1件未到位,此前置事件始终未成立,故按任务的原生依序记入规则,第5件的到位事件也未成立。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "unit": "reward", + "threshold": 2000, + "direction": "higher_better", + "value": null + }, + { + "name": "obj_goal_dist", + "unit": "m", + "threshold": 0.15, + "direction": "lower_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "object_1_at_slot", + "t": 0.0, + "log_step": 75, + "where": "首步前", + "evidence_zh": "观察:视频 0.0–10.05 s 机器人仅从初始站姿前倾并维持支撑,没有开始最浅红物体的抓取、搬运或送至半透明目标;视频 9.0–10.05 s 仍保持相同布局。log step 75 的最终控制器明确固定双臂和手,log step 80–82 确认其运行的是最后的第50局,服务返回 control_steps=1000、done=true、reason=failed;log step 83、90 表明 agent 保留此局作为评分局。" + }, + "unmet": { + "object_1_at_slot": false, + "object_2_at_slot": false, + "object_3_at_slot": false, + "object_4_at_slot": false, + "object_5_at_slot": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 75, + "evidence_zh": "观察:视频 0.0–1.5 s 的密集画面只有机器人站姿与前倾变化,视频 1.5–10.05 s 机器人持续停在书架前,没有对第1件最浅红物体执行抓取或搬放;log step 75 的最终脚本只调用站姿控制器,并将双臂和手保持在初始目标。" + } + ], + "root_causes": [], + "gap_zh": "没有状态日志、逐件阶段计数、物体目标坐标或接触对,视频中机器人遮挡了书架右侧,因此不能恢复精确三维距离或全部手部接触;本次未执行搬物,不给各节点虚构操作接触。log step 86、90 中 agent 把伸手描述为有失衡风险,但没有验证该搬物动作,也没有证据证明目标不可达或具体抓取、控制方案必定失败;此前站姿试验的失衡已在最终评分局修复,不能据此指定本次未启动搬物的根因。按 CASES.md FAIL-22,不把保留站姿局的停止决定本身当作推理/规划根因,根因层留空。log step 82 的1000控制步按任务规定对应仿真 20 s,而视频总长为 10.1 s,缺少经可见事件核对的时钟映射;所有标注时间直接取自视频,未将控制步或 agent 时间换算为视频时间。评分字段未提供 episode_return、obj_goal_dist,原生指标均留 null。" + }, + "summary_zh": "观察:评分视频中机器人前倾后一直维持站姿,未开始搬放最浅红物体,五个按序到位节点均未达成,最终为0/5。断点在首步前,现象为没尝试;log step 75 的最终程序只有站姿控制,log step 82 的服务状态确认最后一局结束且失败。缺少状态与阶段计数、精确距离及经验证的伸手尝试,不能把 agent 对失衡风险的自述进一步确定为控制或规划根因。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "视频 0–10.05 s 机器人只维持站姿、双臂停在书架中层,五件物体均无搬运,0/5、首步前/没尝试、根因层留空并附 gap_zh,均与画面和 log step 75/82 一致;未发现需修改之处。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:18:29Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--ew5jrdt/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--ew5jrdt/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "66cd67a3ef4dceac66ae869ba23c8b417929bd4ab37c3978b0f994c7c6e51dab", + "video_sha256": "6e02a7cb78e0394c16f5299f024b86d1d9a3b85acad95d2920c5ce8f6d68efd9", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--eysdqgy.json b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--eysdqgy.json new file mode 100644 index 0000000000000000000000000000000000000000..716e4a6ea79feb2c79852c7b473f06cd8a165e51 --- /dev/null +++ b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--eysdqgy.json @@ -0,0 +1,276 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-bookshelf-simple-i00-standard-realtime--eysdqgy", + "taskKey": "humanoidbench/manip_bookshelf_simple", + "dag_sha256": "2204f90476ae603026750aa49178a8c18a5f3ae6b8197cf0e7e8d3ecbe659563", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 1000, + "sim_steps": 8372, + "calls": 15308, + "rows": 15201, + "resets": 43, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3456.3, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 166.779, + "max_lag_s": 0.1093, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 6044, + "late_steps": 0, + "requests_refused": 20, + "request_busy_s": 8.924, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.31, + "peak_processes": 10, + "peak_disk_gb": 0.269, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--eysdqgy/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "d58f46c54cf6eb98c20c632f95ec76e711296dab7ce9484ddd91e8ddfbf03b8d" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 5, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "object_1_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00 s,最浅红色棋子状实体仍直立在从上数第二层左侧,与顶层的同色半透明目标分处不同层。" + }, + { + "t": 0.2, + "log_step": null, + "text_zh": "观察:逐帧查看视频 0.00–0.95 s,机器人右手从中层向下移动到灰色物体附近,未包住上方最浅红色棋子;该棋子没有离开原位。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 0.5 s 间隔的全程画面及视频 9.00–10.05 s 的逐帧画面中,最浅红色棋子仍在原层,顶层目标未被该实体占据,未见任何搬放成功。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–10.05 s,机器人的双脚接触地面、双腿承重站立,支撑右臂的首件搬放尝试;右手下降时未接触上方最浅红色棋子,未观察到对该件的抓握。" + } + ] + }, + "object_2_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 0.00–10.05 s 的全程抽帧及末秒逐帧画面未见机器人搬动次浅红物体,也未见首件离开原位或到达顶层目标。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:首件到位事件始终未成立,第2件的按序验收前置条件未达到,因此该事件未达到;并非由总成功分为零反推。" + } + ] + }, + "object_3_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 0.00–10.05 s 的全程抽帧及末秒逐帧画面未见机器人搬动中红物体,也未见首件离开原位或到达顶层目标。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:首件到位事件始终未成立,第3件的按序验收前置条件未达到,因此该事件未达到;并非由总成功分为零反推。" + } + ] + }, + "object_4_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 0.00–10.05 s 的全程抽帧及末秒逐帧画面未见机器人搬动次深红物体,也未见首件离开原位或到达顶层目标。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:首件到位事件始终未成立,第4件的按序验收前置条件未达到,因此该事件未达到;并非由总成功分为零反推。" + } + ] + }, + "object_5_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 0.00–10.05 s 的全程抽帧及末秒逐帧画面未见机器人搬动最深红物体,也未见首件离开原位或到达顶层目标。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:首件到位事件始终未成立,第5件的按序验收前置条件未达到,因此该事件未达到;并非由总成功分为零反推。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "unit": "reward", + "threshold": 2000, + "direction": "higher_better", + "value": null + }, + { + "name": "obj_goal_dist", + "unit": "m", + "threshold": 0.15, + "direction": "lower_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "object_1_at_slot", + "t": 0.2, + "log_step": 398, + "where": "操作中", + "evidence_zh": "观察:视频 0.00–0.30 s,机器人右臂已有伸手动作,但手下降到最浅红色棋子所在层的下方,棋子始终留在上方原位;视频 0.5–10.05 s 未再完成抓取或送达。log step 397 的最终脚本包含抓取、抬升、释放阶段,log step 398 将它用于最后的第44局;这次首件尝试此后没有完成。" + }, + "unmet": { + "object_1_at_slot": false, + "object_2_at_slot": false, + "object_3_at_slot": false, + "object_4_at_slot": false, + "object_5_at_slot": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "没建立接触", + "t": 0.2, + "log_step": null, + "evidence_zh": "观察:视频 0.00–0.95 s 的手部放大逐帧画面显示,机器人右手从中层下降并停在灰色物体附近,最浅红色棋子在更上一层保持直立;没有手指包围或夹住该棋子,也没有带起该棋子。视频 9.00–10.05 s 仍是空手停留、目标实体未动。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "log_step": 397, + "evidence_zh": "推断:最终控制脚本的实时指令调度未给抓取、抬升动作足够的执行时长。log step 397 用脚本循环变量 step 切换抬升、释放与回到站姿阶段,每次 r.submit 只提交一行且未指定 at;log step 10、359 的服务说明明确 submit 立即返回,默认从下一控制步替换计划。log step 398 的脚本在日志时钟 18:25:58–18:26:00 已打印到循环 Step 1000,但仍返回 done=False;结合视频 0.00–0.30 s 的短暂手臂下降及之后长时间停留,支持循环快速覆盖动作阶段、没有执行完整抓放轨迹。这里的循环 Step 不是服务 control_steps,未将它换算为视频时间。", + "source_zh": "日志", + "note_zh": "控制子类不足以细分:有实时提交语义与执行日志依据,但缺少最终局逐控制步的命令接收和执行记录,不能进一步断定具体伺服参数或接触几何错误。" + } + ], + "gap_zh": "输入有完整视频与会话日志,但没有状态日志、成功参考视频、逐物体目标距离或 task_index 读数;仅房间相机的256×256画面,无法确认手与下层灰色物体是否有轻微接触。最终脚本的中段在 run.json 中被截去;第44局循环计数不是服务控制步计数,agent 最终自述的20.0秒与本视频10.1秒不建立换算,按视频自身时间标注。评分字段未报告 episode_return 或 obj_goal_dist,二者保留 null。" + }, + "summary_zh": "五个按序到位事件均未达到,最终进度为0/5;视频 0.00–10.05 s 中最浅红色棋子始终留在原层,机器人右手短暂下降后停在下层灰色物体附近,未建立首件抓取接触。断点是最浅红物体到位的操作中,日志支持实时指令提交过快、循环阶段覆盖造成控制失败,不能把 agent 自述的完整抓放序列当作已执行。没有状态日志和逐件距离记录,原生奖励及精确接触不可补造。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "五个节点均未达到与视频一致:全程最浅红棋子留在原层,右手停在下层灰色物体附近,未包住棋子。断点、没建立接触、控制根因(脚本 1000 步在约 2 s 墙钟内提交完,log step 398)均有日志与画面支撑,未发现需要修改的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:18:40Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--eysdqgy/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--eysdqgy/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "cc5183bf77830a5f09c1fc415517781536b066c029fdd55743980532b3783fa5", + "video_sha256": "d58f46c54cf6eb98c20c632f95ec76e711296dab7ce9484ddd91e8ddfbf03b8d", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--s66z4kw.json b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--s66z4kw.json new file mode 100644 index 0000000000000000000000000000000000000000..5640cc4ade6ad8f1e404313b3b4432dcfd4dab0e --- /dev/null +++ b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--s66z4kw.json @@ -0,0 +1,262 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-bookshelf-simple-i00-standard-realtime--s66z4kw", + "taskKey": "humanoidbench/manip_bookshelf_simple", + "dag_sha256": "2204f90476ae603026750aa49178a8c18a5f3ae6b8197cf0e7e8d3ecbe659563", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 1000, + "sim_steps": 37300, + "calls": 27933, + "rows": 27800, + "resets": 47, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3465.0, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 745.412, + "max_lag_s": 0.049, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 15584, + "late_steps": 0, + "requests_refused": 45, + "request_busy_s": 46.987, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.324, + "peak_processes": 10, + "peak_disk_gb": 1.013, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--s66z4kw/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "0a655ace49af5aad213d0e800dab568ece8d70523743a2353d5eb81096d83435" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 5, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "object_1_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.0 s,最浅粉红的细柱状物体立在红色书本左侧,配对的浅色半透明细柱目标在上一层右侧,两者明显分离。" + }, + { + "t": 4.4, + "log_step": null, + "text_zh": "观察:视频 0.4–4.2 s,机器人左臂伸入该物体所在层;视频 4.4–5.5 s,左手回撤并抬到头顶附近,细柱状物体仍立在原层,未随手上移。" + }, + { + "t": 10.05, + "log_step": 172, + "text_zh": "观察:视频 9.1–10.05 s,机器人左手停在头顶上方,第一件物体未出现在上一层的配对目标处;log step 172 的服务状态为 episode 48、control_steps=1000、done=true、reason=failed。此处日志只核对最终评分局及结果,不把控制步或 agent 时间换算为视频时间。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察:视频 0.4–4.2 s,机器人左手伸向第一件浅粉红细柱状物体,手指与物体的接触区域被前臂及书本遮挡;不能确认是否发生轻触、推擦或具体接触方式,未见物体随手被带走,不据闭指动作标抓握。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–10.05 s,机器人的双脚与腿承重支撑伸臂操作;后段身体倾斜,但末秒仍由脚支撑,未见倒地。" + } + ] + }, + "object_2_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:已检查视频 0.0–10.05 s,机器人只围绕第一件浅粉红细柱状物体伸手、回撤和抬臂,没有接续搬放次浅红物体;第一件始终未送到上一层目标,故本节点所需的按序前置事件没有发生,不能记入第2件到位。" + } + ] + }, + "object_3_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:已检查视频 0.0–10.05 s,机器人只围绕第一件浅粉红细柱状物体伸手、回撤和抬臂,没有接续搬放中红物体;第一件始终未送到上一层目标,故本节点所需的按序前置事件没有发生,不能记入第3件到位。" + } + ] + }, + "object_4_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:已检查视频 0.0–10.05 s,机器人只围绕第一件浅粉红细柱状物体伸手、回撤和抬臂,没有接续搬放次深红物体;第一件始终未送到上一层目标,故本节点所需的按序前置事件没有发生,不能记入第4件到位。" + } + ] + }, + "object_5_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:已检查视频 0.0–10.05 s,机器人只围绕第一件浅粉红细柱状物体伸手、回撤和抬臂,没有接续搬放最深红物体;第一件始终未送到上一层目标,故本节点所需的按序前置事件没有发生,不能记入第5件到位。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "unit": "reward", + "threshold": 2000, + "direction": "higher_better", + "value": null + }, + { + "name": "obj_goal_dist", + "unit": "m", + "threshold": 0.15, + "direction": "lower_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "object_1_at_slot", + "t": 4.4, + "log_step": 168, + "where": "操作中", + "evidence_zh": "观察:视频 0.4–4.2 s,机器人左手在第一件浅粉红细柱状物体附近执行伸手动作;视频 4.4–5.5 s,手臂回撤上抬而该物体留在原层,直到视频 10.05 s 都未送至上一层目标。log step 168 的脚本安排最后两次固定关节伸臂试验,log step 170、172 确认最后评分局是 episode 48;未采用之前重置掉的试验作为本局进度。" + }, + "unmet": { + "object_1_at_slot": false, + "object_2_at_slot": false, + "object_3_at_slot": false, + "object_4_at_slot": false, + "object_5_at_slot": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "没建立接触", + "t": 4.4, + "log_step": null, + "evidence_zh": "推断:视频 4.2–5.5 s,机器人左手从物体所在层回撤并上抬,原先红色书本左侧的浅粉红细柱状物体仍留在架上;视频 9.1–10.05 s,手停在上方而物体仍未转移。因此本次伸手闭指未建立能将第一件物体带离原处的抓握;这不等于从画面断言指尖从未轻触物体。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志", + "evidence_zh": "推断:log step 168 的最终控制脚本关闭 vision、event_grasp 和 visual_pose,按固定 arm_schedule 伸臂、伸肘及回撤,并按 path 定时闭指,未等待本局抓稳确认再搬运。视频 4.4–5.5 s 对应的实际结果是左手上抬、第一件物体留在原层,说明伸手闭指控制没有实现有效夹持;现有证据不足以进一步区分接触几何、关节跟踪或具体参数问题。" + } + ], + "gap_zh": "视频仅有分辨率为256×256的房间视角,第一件物体附近的手指受前臂和书本遮挡,无法判清轻触及对向接触细节;没有状态日志或成功参考视频,也没有本局物体坐标、接触对、阶段编号及关节跟踪记录。评分字段未报告 episode_return 或 obj_goal_dist,两项原生指标均保留 null;日志中的早期视觉估计与调试读数不当作最后评分局的状态。" + }, + "summary_zh": "最后评分局只尝试第一件最浅红细柱状物体,五个按序到位节点均未达成(0/5)。视频 4.4–5.5 s 显示左手回撤上抬而物体留在原层,断点为第一件到位步骤,现象是未建立有效抓握,根因仅能归到控制。指尖接触受遮挡,且没有状态日志,无法细分控制原因或给出原生距离与累计回报。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "视频帧显示左臂伸入后上抬、物体留在原层,节点未达成、断点与现象与画面一致,未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:20:04Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--s66z4kw/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--s66z4kw/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "b79df86630d1858c41298c467ce1268c3f2c2d135df93d1b0f7922582d5ea9c7", + "video_sha256": "0a655ace49af5aad213d0e800dab568ece8d70523743a2353d5eb81096d83435", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--zmksjxb.json b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--zmksjxb.json new file mode 100644 index 0000000000000000000000000000000000000000..925ab3dcc117e04d722fdb130ac3613d237d383e --- /dev/null +++ b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--zmksjxb.json @@ -0,0 +1,279 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-bookshelf-simple-i00-standard-realtime--zmksjxb", + "taskKey": "humanoidbench/manip_bookshelf_simple", + "dag_sha256": "2204f90476ae603026750aa49178a8c18a5f3ae6b8197cf0e7e8d3ecbe659563", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 1000, + "sim_steps": 1439, + "calls": 1409, + "rows": 12974, + "resets": 5, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 2385.3, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 28.688, + "max_lag_s": 0.0452, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 381, + "late_steps": 0, + "requests_refused": 5, + "request_busy_s": 1.0, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.22, + "peak_processes": 6, + "peak_disk_gb": 0.255, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--zmksjxb/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "e004a77d7e49ccb622606fe5280e653736adc4d5254386db44cd4698591cc637" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 5, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "object_1_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.0 s,最浅红色立柱状物体在书架自上而下第二格左侧,其同形半透明目标在顶格;两者分处不同层。" + }, + { + "t": 2.0, + "log_step": null, + "text_zh": "观察:视频 1.4–2.6 s 的连续帧中,机器人左手在该实体下方一格伸入和合指,最浅红实体仍立在原处;视频 3.0–6.4 s 左手回撤、再次伸屈时也未携带该实体。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 9.1–10.05 s 末段,最浅红实体仍留在原层,未送至顶层同色目标,机器人保持前倾站姿;全段未见该件到位。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–10.05 s,机器人双脚接触地面,双腿承重,在左臂执行第1件的抓取搬放尝试时支撑身体;未见倒地。" + }, + { + "who": "未知", + "how": "未知", + "note_zh": "观察:视频 1.4–2.6 s,机器人左手伸到最浅红实体下方的书架格内,手与层板、侧板之间的接触点被侧板遮挡,无法确认是否有擦碰或按压。左手未包住最浅红实体,不记为实际抓握。" + } + ] + }, + "object_2_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.0 s,次浅红色书本位于自上而下第二格的书本组中,同色半透明书本目标在下一格;两者未重合。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 0.0–10.05 s,机器人没有从上方书本组取出次浅红书本送往下一格;末帧书本组仍在原层。第1件也始终未到位,故未出现按序完成第2件的事件。" + } + ] + }, + "object_3_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.0 s,中红色细长物体横放在白灰物体所在格的底部,同色细长目标位于上一格;起始不在目标位。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 0.0–10.05 s,中红色细长实体未被取出或送到上一格,末帧仍在原来的白灰物体旁;此前第1、2件的按序事件也未发生。" + } + ] + }, + "object_4_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.0 s,次深红书本在上方书本组中,下方靠近底部一格可见对应红色半透明目标,实体和目标分处不同层。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 0.0–10.05 s,机器人没有把上方次深红书本搬至下层目标,书本组未出现取出和向下搬运;此前各件按序到位事件均未发生。" + } + ] + }, + "object_5_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.0 s,最深红书本在上方书本组的右侧,下方靠近底部一格的深红半透明目标与之分层。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 0.0–10.05 s,机器人没有取出最深红书本并送到下层目标,末帧该书本仍在上方原处;此前各件未按序到位,第5件事件未发生。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "unit": "reward", + "threshold": 2000, + "direction": "higher_better", + "value": null + }, + { + "name": "obj_goal_dist", + "unit": "m", + "threshold": 0.15, + "direction": "lower_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "object_1_at_slot", + "t": 2.0, + "log_step": 80, + "where": "操作中", + "evidence_zh": "观察:视频 1.4–2.6 s,机器人左手在最浅红实体下一格伸入并合指,实体未离开原处;视频 3.0–6.4 s 的后续伸屈仍未带出物体,至视频 10.05 s 未修复。log step 80 启动的脚本以第1件抓取并送至顶层为目标;log step 88 的机器人服务状态已为 episode=6、control_steps=1000、done=true、reason=failed,之后直到 log step 115 的过滤中断未再重置或执行机器人动作。这是已结束的最后一局,不能用之前重置局的摔倒代替其失败现象;此处视频时间直接取自画面,未由控制步或 agent 时间换算。" + }, + "unmet": { + "object_1_at_slot": false, + "object_2_at_slot": false, + "object_3_at_slot": false, + "object_4_at_slot": false, + "object_5_at_slot": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "没建立接触", + "t": 2.0, + "log_step": null, + "evidence_zh": "观察:视频 1.4–2.6 s,机器人左手活动在最浅红立柱状实体下方一格,没有到达其杆身并形成相向夹持;视频 3.0–6.4 s 左手回撤和继续运动时没有带出该物体,最浅红实体一直留在上方原处,至视频 10.05 s 仍未抓起。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "推断:log step 80 的控制脚本按名义站姿预先求解第1件的抓取与搬放关节轨迹,执行时使用关节跟踪和腿部姿态反馈;视频 1.0–6.4 s 显示机器人前倾后,左手的实际运动始终落在目标实体下一格,未把手送到杆身处。log step 86 的 agent 离线运动学诊断也算出抓取阶段手部高度约1.05米,与其计划的约1.38米抓取高度明显不同;该诊断仅作会话辅助证据,不用作节点值或视频时钟。目标选择是第1件,但位置执行未实现抓取,归于控制;没有足够证据再细分逆解、姿态补偿或反馈控制。", + "source_zh": "日志" + } + ], + "gap_zh": "仅有256×256的房间视角视频,无状态日志或成功oracle;手与书架板件的接触局部被遮挡,无法确定擦碰方式和接触力,但实体未被搬出、未到位可由前后画面判断。评分器未报告逐件task_index、obj_goal_dist或episode_return数值,两个原生指标留null,不能把catalog reward=0.0当作累计回报。日志控制步与视频时长未建立可靠映射,全部视频锚点取自已检查画面。log step 115 确有模型过滤提示,但 log step 88 已记录最后一局结束,故不将此事后中断当作物理失败根因。" + }, + "summary_zh": "评分视频中机器人尝试以左手搬放最浅红物体,但手在实体下方一格伸屈,物体未被抓起,五个按序到位事件均未达成,最终为0/5。断点为最浅红物体到位,现象是没建立所需抓取接触,根因只归到控制:实际手部运动没有到达计划抓取位置;log step 88 已记录最后一局结束,log step 115 的模型过滤发生在其后。缺少状态与接触日志,无法细分控制误差来源及被遮挡的手与书架接触。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "逐帧核对:视频 0–10 s 中机器人左手始终在最浅红物体下一格伸屈,物体未被带出,五个节点均未达成,0/5、断点 object_1_at_slot、没建立接触、根因控制(推断已标注)均有画面和 log(step id 86 的 FK:抓取阶段手高约1.03–1.05 m)支持;视频时间均取自画面,未换算 agent 时间。未发现需要修改的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:20:03Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--zmksjxb/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--zmksjxb/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "50dae19455641d5180c811dbdab4b2ffee03cf4e08dee66c748b497150a7eb56", + "video_sha256": "e004a77d7e49ccb622606fe5280e653736adc4d5254386db44cd4698591cc637", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--zry4xj3.json b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--zry4xj3.json new file mode 100644 index 0000000000000000000000000000000000000000..9edab4932855c0c57e10e90dcc1429dbdd9d0c9c --- /dev/null +++ b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--zry4xj3.json @@ -0,0 +1,296 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-bookshelf-simple-i00-standard-realtime--zry4xj3", + "taskKey": "humanoidbench/manip_bookshelf_simple", + "dag_sha256": "2204f90476ae603026750aa49178a8c18a5f3ae6b8197cf0e7e8d3ecbe659563", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 1000, + "sim_steps": 28431, + "calls": 94080, + "rows": 94188, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1572.5, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 568.071, + "max_lag_s": 0.1318, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 6652, + "late_steps": 0, + "requests_refused": 36, + "request_busy_s": 41.863, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 1.927, + "peak_processes": 10, + "peak_disk_gb": 2.172, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--zry4xj3/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "49f2feb59c182669a69bc8403919820b590053777851cc2dbf568b395343f95c" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 5, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "object_1_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.0 s,最浅红色的细杆状实体立在上部红色物体旁的层板上;对应浅红半透明目标位于另一层,实体与目标分离。" + }, + { + "t": 2.5, + "log_step": null, + "text_zh": "观察:视频 0.0–2.5 s,机器人弯起左肘、抬起空手,手指停在该实体所在层板下方;视频 2.5–10.05 s,浅红实体一直留在原层板,未见被手包夹、提走或送入半透明目标。末秒视频 9.1–10.05 s 仍保持这一情况。" + }, + { + "t": null, + "log_step": 233, + "text_zh": "观察:log step 233 执行最后一次 reset 后的固定点接近及保持程序,机器人服务返回 episode=51、control_steps=1000、reason=failed、resets_left=0;log step 235 复查相同末局。此前多次重置的抓取、扫动测试不计入此视频的进度。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–1.0 s,机器人逐渐前倾屈膝,头部和躯干靠近书架层板;视频 1.0–10.05 s,机器人在抬起左手后继续维持屈膝前倾姿态,双脚仍与地面接触,末段右腿向后伸展。左手是本步的尝试作用端,但画面未见其包夹浅红实体,故不把接近动作记作已发生的抓握。" + } + ] + }, + "object_2_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 0.0–10.05 s,机器人仅抬起左手后保持,未见另取或搬送次浅红实体至对应半透明目标;视频 9.1–10.05 s 的末段也无该放置动作。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频中第1件最浅红实体始终未离开原位到达目标,第2件也无到位事件;按 DAG 的逐件顺序验收,本节点未达到。" + } + ] + }, + "object_3_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 0.0–10.05 s,机器人仅抬起左手后保持,未见另取或搬送中红实体至对应半透明目标;视频 9.1–10.05 s 的末段也无该放置动作。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频中第1件最浅红实体始终未离开原位到达目标,第3件也无到位事件;按 DAG 的逐件顺序验收,本节点未达到。" + } + ] + }, + "object_4_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 0.0–10.05 s,机器人仅抬起左手后保持,未见另取或搬送次深红实体至对应半透明目标;视频 9.1–10.05 s 的末段也无该放置动作。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频中第1件最浅红实体始终未离开原位到达目标,第4件也无到位事件;按 DAG 的逐件顺序验收,本节点未达到。" + } + ] + }, + "object_5_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 0.0–10.05 s,机器人仅抬起左手后保持,未见另取或搬送最深红实体至对应半透明目标;视频 9.1–10.05 s 的末段也无该放置动作。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频中第1件最浅红实体始终未离开原位到达目标,第5件也无到位事件;按 DAG 的逐件顺序验收,本节点未达到。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "unit": "reward", + "threshold": 2000, + "direction": "higher_better", + "value": null + }, + { + "name": "obj_goal_dist", + "unit": "m", + "threshold": 0.15, + "direction": "lower_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "object_1_at_slot", + "t": 2.5, + "log_step": 233, + "where": "操作中", + "evidence_zh": "观察:视频约 2.5 s,机器人左手抬起后停在最浅红实体所在层板下方,没有把该实体握起;视频 2.5–10.05 s 实体仍在原位,未发生到对应目标的搬送。log step 231 将最终固定点保持程序装为 gen.py,log step 233 运行最后一局并由机器人服务报告失败;早先测试中的摔倒与扫动不属于此评分局。" + }, + "unmet": { + "object_1_at_slot": false, + "object_2_at_slot": false, + "object_3_at_slot": false, + "object_4_at_slot": false, + "object_5_at_slot": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "没建立接触", + "t": 2.5, + "log_step": null, + "evidence_zh": "观察:视频 0.0–2.5 s,机器人左手从较低位置抬起,但未围住上层最浅红细杆状实体;视频 2.5、5.0、7.0、9.0、10.05 s 均见该实体立在同一层板上,手保持空着,始终未建立能带走该实体的抓持。" + } + ], + "root_causes": [ + { + "cat": "认知", + "item": "推理/规划", + "source_zh": "日志 / 视频", + "evidence_zh": "推断:log step 223 写入的最终程序只把左掌从 HOME_PALM 插值接近单个 target_rest,之后保持该点;log step 231 复制该程序,log step 233 以 TX=0.50、TY=0.03、TZ=0.49、HOLD_CURL=0.15 执行评分末局。这个已执行的动作方案没有后续抓持并搬往同色目标的阶段,视频 2.5–10.05 s 也只见空手保持、实体不动,因此固定点接近方案没有完成首件放置所需的后续动作。" + } + ], + "gap_zh": "没有状态日志,视频仅提供低分辨率房间视角,不能核实细微碰擦、物体原点距离、逐件 task_index 或累计回报数值;但实体没有离开原位置到目标的可见事实足以判定首件未到位。log step 233 的末局 control_steps=1000,任务文本控制周期为0.02秒,其名义仿真时长为20秒,而视频仅10.1秒;缺少可核对的逐帧映射,未把日志计时或控制步换成视频时间,所有视频锚点直接取自所检帧。" + }, + "summary_zh": "评分末局完成0/5个到位事件:机器人在书架前前倾屈膝,抬起左手后维持该姿态,最浅红实体仍留在原层板,其他物体也未被搬至目标。断点是首件到位,未建立有效抓持;最终执行方案仅接近并保持一个固定掌心点,没有后续搬往目标的阶段。没有状态日志或原生距离、累计回报读数,且末局服务计时与视频时长缺少已核对映射,时刻均按视频帧标注。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 2, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "fix", + "conf": 6, + "summary_zh": "进度 0/5、断点首件到位、无接触现象与视频一致(最浅红实体全程未被取走)。唯一需改处:证据把机器人描述为'保持站立、双脚接地',但视频 0–1 s 内机器人已向书架前倾下蹲,躯干/头部贴近层板,姿态与描述不符。", + "n_issues": 1, + "n_major": 0 + }, + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 2, + "verdict": "agree", + "conf": 7, + "summary_zh": "进度0/5、断点首件到位、没建立接触与规划根因均与视频一致;第1轮的姿态描述已修正,未发现新错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:21:52Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--zry4xj3/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--zry4xj3/04_review_r1" + }, + { + "stage": "reconcile", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--zry4xj3/05_reconcile_r1" + }, + { + "stage": "traj_review", + "round": 2, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--zry4xj3/04_review_r2" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "6add2aba9f2af51f8942c6ccef1e0c89f7ee4a62147d5e3c922a3f549c4001a9", + "video_sha256": "49f2feb59c182669a69bc8403919820b590053777851cc2dbf568b395343f95c", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--zufq7fk.json b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--zufq7fk.json new file mode 100644 index 0000000000000000000000000000000000000000..dcd29346db42f4ce9e2d945a0ae2b626bf1ceeb4 --- /dev/null +++ b/phase2/humanoidbench/manip_bookshelf_simple/trajectories/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--zufq7fk.json @@ -0,0 +1,259 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-bookshelf-simple-i00-standard-realtime--zufq7fk", + "taskKey": "humanoidbench/manip_bookshelf_simple", + "dag_sha256": "2204f90476ae603026750aa49178a8c18a5f3ae6b8197cf0e7e8d3ecbe659563", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 1000, + "sim_steps": 6148, + "calls": 2658, + "rows": 5771, + "resets": 20, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 2346.8, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 122.665, + "max_lag_s": 0.0172, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 2155, + "late_steps": 0, + "requests_refused": 3, + "request_busy_s": 1.796, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--zufq7fk/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "0fa14ca888b7158bc276946de11c1856364deb48724e05ff375c4296de2bbebe" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 5, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "object_1_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.5, + "log_step": null, + "text_zh": "观察:视频 0.0–7.0 s,机器人面向书架,左臂伸出、手停在最浅红色棋子状物体所在层板的下方;该实体仍在上方原位,与更高一层的同色半透明目标副本明显分离,未见夹住、提起或搬运。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 7.5–10.05 s,机器人躯干前倾、左肘弯曲,左手仍未带起最浅红实体;逐帧检查视频 9.05–10.05 s,未见该实体进入上层目标副本位置,机器人也未倒地。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–10.05 s,机器人的双脚与腿在左臂接近最浅红物体的过程中支撑身体;未观察到左手与该实体建立抓握,不记抓握接触。" + } + ] + }, + "object_2_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频 0.0–10.05 s,机器人始终只在最浅红物体下方伸手,未见将次浅红物体搬至对应半透明副本;第1件到位事件始终未发生,因此按本节点的顺序前置条件,第2件到位事件也未达到。" + } + ] + }, + "object_3_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频 0.0–10.05 s,机器人始终只在最浅红物体下方伸手,未见将中红物体搬至对应半透明副本;第1件到位事件始终未发生,因此按本节点的顺序前置条件,第3件到位事件也未达到。" + } + ] + }, + "object_4_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频 0.0–10.05 s,机器人始终只在最浅红物体下方伸手,未见将次深红物体搬至对应半透明副本;第1件到位事件始终未发生,因此按本节点的顺序前置条件,第4件到位事件也未达到。" + } + ] + }, + "object_5_at_slot": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:视频 0.0–10.05 s,机器人始终只在最浅红物体下方伸手,未见将最深红物体搬至对应半透明副本;第1件到位事件始终未发生,因此按本节点的顺序前置条件,第5件到位事件也未达到。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "unit": "reward", + "threshold": 2000, + "direction": "higher_better", + "value": null + }, + { + "name": "obj_goal_dist", + "unit": "m", + "threshold": 0.15, + "direction": "lower_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "object_1_at_slot", + "t": 0.5, + "log_step": null, + "where": "探索", + "evidence_zh": "观察:视频 0.0–7.0 s,机器人左臂伸向书架、手张开停在最浅红实体下方;视频 7.5–10.05 s 身体前倾、肘部弯曲,实体仍留在原层板,未见合手抓起或送往上层同色副本。推断:结合 log step 63–64 的固定臂目标实机探测代码,评分轨迹最可能停留在第1件操作前的伸臂探索,而非已执行的抓取或搬运;第1件到位判据始终未达到。" + }, + "unmet": { + "object_1_at_slot": false, + "object_2_at_slot": false, + "object_3_at_slot": false, + "object_4_at_slot": false, + "object_5_at_slot": false + }, + "phenomena": [], + "root_causes": [], + "gap_zh": "没有状态日志或接触传感记录,256×256房间视频的末段手部还被书架侧板部分遮挡。观察:log step 61 将控制程序改为不动臂、手张开的站立回合;log step 62 仅在 agent 自建模型内测试逆解;log step 63 的 probe5 在回合已结束或未运行时 reset,提交500行相同目标,左臂 qarm=[-1.15,0.03,0.04,0.23,0],手部目标也不随阶段变化,没有合手抓取或搬运阶段;log step 64 再次执行该探测程序,两步的 sim.reset 均为1。推断:视频中的伸臂姿态与 probe5 相容,评分局最可能是最后的实机探测,但没有回合编号或可靠的控制步—视频对应可唯一确认;两次500行提交与 reward.n_actions=1000 数值相合,不足以证明属于同一回合。按 docs/03_trajectory.md Step 3.1 的定义,将机器人实际伸臂探测记为探索;未见正式抓取动作,不能把探测停在实体下方直接标成没建立接触,也不能因没有合手指令就把已经发生的实机探索当作完全未启动。依 CASES.md FAIL-23,现象留空;日志与视频不足以确定没有转入抓取的原因,也不能由离线逆解收敛证明控制缺陷,或由 agent 的停止自述证明规划错误(FAIL-4、FAIL-22),故根因留空。log step 66 所称保留先前站立回合只是 agent 自述,不作为评分局身份依据。未将 agent 时间或控制步换算成视频时间;评分字段未报告 episode_return 或 obj_goal_dist,二者仍记 null,目录 reward=0 不等于累计回报。" + }, + "summary_zh": "评分视频中机器人在书架旁伸臂探测,最浅红实体始终未被抓起或送往上层同色目标,五个到位节点均未完成;断点仍为第1件到位步骤,位置改为探索。日志支持最后运行的是固定臂目标探测,但无法唯一确认评分局身份,亦不足以确定具体失败现象和根因,故两层均留空;置信度由7降至5,以反映回合对应及探索性质依赖推断。", + "conf": 5, + "review": { + "status": "agreed", + "rounds": 2, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "fix", + "conf": 6, + "summary_zh": "视频所见(最浅红物体始终未被抓起、五节点均未到位、断点为第1件)与标注一致。问题在根因:log 显示被评分回合最可能是 step 63/64 的 probe5 开环探测(固定左臂目标、手保持张开,两次各500步合计 n_actions=1000),而非一次真正的抓取尝试;文件把它当作“伸手控制失败”并据 agent 自建模型的逆解测试推断控制缺陷,证据不足。", + "n_issues": 2, + "n_major": 1 + }, + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 2, + "verdict": "agree", + "conf": 6, + "summary_zh": "视频所见与标注一致:五个到位节点均未达成,最浅红物体始终未被抓起;断点为第1件、位置探索,现象与根因留空且gap_zh说明了证据不足的理由,符合规则。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:23:14Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--zufq7fk/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--zufq7fk/04_review_r1" + }, + { + "stage": "reconcile", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--zufq7fk/05_reconcile_r1" + }, + { + "stage": "traj_review", + "round": 2, + "ok": true, + "space": "work/runs/humanoidbench-manip-bookshelf-simple-i00-standard-realtime--zufq7fk/04_review_r2" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "8dc2d579df2cb1c48fe66141bc55da36ac0d79a2d7b252e58a7099542dfac3cf", + "video_sha256": "0fa14ca888b7158bc276946de11c1856364deb48724e05ff375c4296de2bbebe", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-privileged--2bnpuhp.json b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-privileged--2bnpuhp.json new file mode 100644 index 0000000000000000000000000000000000000000..94429c405174a38ab248fa5ed7f08804e9841e9c --- /dev/null +++ b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-privileged--2bnpuhp.json @@ -0,0 +1,184 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-cube-i00-privileged--2bnpuhp", + "taskKey": "humanoidbench/manip_cube", + "dag_sha256": "ab5e9b50283b1821da97eeaa6a81c6bcbd6d347d014e9b3e7113b6656d951802", + "outcome": { + "success": 0, + "reward": 0.5773943782310549, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 500, + "final_reward": 0.5773943782310549, + "sim_steps": 500, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 2.028, + "peak_processes": 9, + "peak_disk_gb": 0.263, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-cube-i00-privileged--2bnpuhp/replay.mp4", + "duration_s": 5.1, + "fps": 20.0, + "frames": 102, + "sha256": "3acea75310e69f7dfea56bd8d0fe1449f7a36a1e713a3406f30e5e8e16510a9c" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "cubes_reoriented": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00 s 两个彩色方块分别位于机器人张开的双掌上方,悬浮半透明方块显示目标朝向;尚未进行调姿。" + }, + { + "t": 0.6, + "log_step": null, + "text_zh": "观察:视频 0.40–0.70 s 机器人两手手指逐渐合拢包围各自方块;视频 1.00–2.50 s 左臂抬起再向外下方转动,右手腕同时转动其方块,方块随手移动。" + }, + { + "t": 2.5, + "log_step": null, + "text_zh": "观察:视频 2.50–5.05 s 双手仍各持一个方块,身体维持双足站立;右手位于画面左侧,其方块的可见色面朝向仍与上方半透明目标不同,随后没有明显继续纠正。末段逐帧检查未见掉块或倒地。" + }, + { + "t": 5.05, + "log_step": 416, + "text_zh": "观察:inputs/run.json 的评分器字段 success=0、deterministic=1、n_actions=500、sim_steps=500;log step 416 对最终生成并交付的动作文件完成确定性回放,报告 scored=false、sim_steps=500。推断:按本节点累计奖励达到370的验收规则,评分结果确认该单一目标未达标;累计值单调增加,故不存在已达标的历史区间,不能仅凭持块及部分转向记为完成。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "观察:视频 0.00–0.30 s 双掌先从下方托住各自方块;视频约 0.50–0.70 s 两只灵巧手的手指合拢包围各自方块,随后双手各持一块并随腕臂转动,持续至视频 5.05 s。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–5.05 s 机器人双脚接触地面,腿与躯干在双手持块调姿期间持续承重。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 370, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "cubes_reoriented", + "t": 2.5, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频约 2.50 s 双臂的大幅转动结束,机器人仍抓持两个方块,但右手方块的可见色面朝向与悬浮目标不一致;视频 2.50–5.05 s 基本保持这一姿态,未继续完成纠正。评分器 success=0。推断:未修复的失败发生在双块调姿步骤;视频 2.50 s 仅作为转动后偏差持续的粗略导航锚点。" + }, + "unmet": { + "cubes_reoriented": false + }, + "phenomena": [ + { + "cat": "运动", + "item": "位姿不准", + "t": 2.5, + "log_step": null, + "evidence_zh": "观察:视频 1.00–2.50 s 机器人确实转动了持块的腕臂;视频 2.50–5.05 s 右手方块仍露出与半透明目标不一致的色面朝向,并保持到片尾。方块一直留在手边,机器人仍站立;本次可见问题是持物转向不足。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "log_step": 407, + "source_zh": "日志", + "evidence_zh": "观察:log step 407 的最终生成脚本将左臂插值到固定关节目标,右臂仅改变三个腕关节到固定值,运动结束后继续保持;脚本的反馈用于躯体平衡,没有读取方块朝向误差来继续修正。log step 394 的相近会话测试报告左右块四元数差范数分别为0.2083和0.9317;最终修改后 agent 仍自述右块误差0.9157(log step 418),不是评分器测量。推断:脚本能持块并站稳,但所生成的调姿控制仍留下较大右块朝向残差,与视频 2.50–5.05 s 的停止纠正一致;可归因到控制,现有证据不足以进一步区分抓持布局与腕臂轨迹求解。" + } + ], + "gap_zh": "没有状态日志。视频分辨率低且部分色面被手遮挡,不能精确重建左右块姿态、接触力或四元数符号;会话测试只用于根因支持,不代替交付视频。评分器只给 final_reward=0.5773943782310549 和 success=0,没有提供 episode_return,不能将该字段擅自当作累计奖励或推算累计值,故原生累计奖励指标填null;其未达标由成功判据与失败记录确认。" + }, + "summary_zh": "机器人完成了双手从托块到抓持、随腕臂部分转动,并在视频末段继续站立持块,但“转好两个方块”未达到累计奖励验收,最终节点为0/1。断点位于视频约 2.50 s 转动后仍残留朝向偏差且未再纠正,现象为位姿不准,结合最终固定关节插值脚本及会话测试推断根因为控制。缺少状态日志和原生累计奖励值,无法精确分解左右块误差及各奖励项贡献。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "节点未达成、断点(视频约2.5 s,agent 控制步80–230对应约0.8–2.3 s)、位姿不准、根因停在“控制”均有视频与 log 支持;未发现需要修改的内容。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:02:28Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-privileged--2bnpuhp/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-privileged--2bnpuhp/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "dd1897242ef127fe9a31dd55c7c5e91124b9d0852fe6203d76870823a1c0d2f8", + "video_sha256": "3acea75310e69f7dfea56bd8d0fe1449f7a36a1e713a3406f30e5e8e16510a9c", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-privileged--ahwdmjq.json b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-privileged--ahwdmjq.json new file mode 100644 index 0000000000000000000000000000000000000000..6f42b0b9f2497bd1b044365d6f31bc704a30faed --- /dev/null +++ b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-privileged--ahwdmjq.json @@ -0,0 +1,200 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-cube-i00-privileged--ahwdmjq", + "taskKey": "humanoidbench/manip_cube", + "dag_sha256": "ab5e9b50283b1821da97eeaa6a81c6bcbd6d347d014e9b3e7113b6656d951802", + "outcome": { + "success": 0, + "reward": 0.3705148112712424, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 480, + "final_reward": 0.3705148112712424, + "sim_steps": 480, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 6.733, + "peak_processes": 27, + "peak_disk_gb": 0.241, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-cube-i00-privileged--ahwdmjq/replay.mp4", + "duration_s": 4.9, + "fps": 20.0, + "frames": 98, + "sha256": "ed0cea629e91b6d5c37a789b5b33b6a8bfcd1b83d69e9822395a8d89bde2f630" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "cubes_reoriented": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.0–0.35 s,机器人两只平掌分别托着彩色方块,悬浮半透明方块显示目标朝向;起始托持不表示已完成调姿。" + }, + { + "t": 1.5, + "log_step": 141, + "text_zh": "观察:视频 0.45–1.0 s 两手收拢并下移,视频 1.1–2.0 s 画面右侧手臂持块转动到身体侧下方,画面左侧手臂恢复平掌托块;视频 2.0–4.3 s 后者方块的色面方向基本维持原样,未见两个方块均转向目标。log step 141 的最终生成脚本仅对 left 一侧执行朝向伺服,与画面右侧手臂的转动对应。" + }, + { + "t": 2.5, + "log_step": 146, + "text_zh": "观察:log step 146 明确重新读取最终交付动作文件验证,报告共 480 步、控制步 250 的左右方块几何朝向误差分别为 12.6°、177.2°,控制步 479 结束;这里只把该记录作为交付内容的佐证,不将其控制步或 agent 时间换算成视频时间。" + }, + { + "t": 4.8, + "log_step": null, + "text_zh": "观察:视频 4.45–4.8 s 机器人逐渐向后倾倒,躯干和骨盆下沉,低处方块仍在手边,画面未显示明确的离手下落;评分记录 success=0、deterministic=1、n_actions=480。" + }, + { + "t": 4.85, + "log_step": null, + "text_zh": "推断:该节点按整集累计奖励达到 370 验收,累计量一旦达标便不会撤销;评分器判本集未成功,因此事件节点从未达标,intervals=[]、final=false。左块局部转近目标不能单独计为该联合节点完成;reward.final_reward=0.3705148112712424 未标为 episode_return,不能充作原生累计分数。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "观察:视频 0.45–1.0 s,两只灵巧手的手指收拢包围各自方块;之后左手继续握着左块,随腕臂转动到身体侧下方(画面右侧)。未把腕臂带块旋转另标为手内调整。" + }, + { + "who": "双臂协作", + "how": "托/舀", + "note_zh": "观察:视频 0.0–0.35 s 两掌从下方分别承托方块;视频 1.1–4.3 s 右掌恢复向上托住右块(画面左侧),与左手握持转动同时发生。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–4.3 s 机器人双足接地、腿和躯干承重,支撑双手持块及左臂调姿;视频 4.45–4.8 s 支撑姿态逐渐失稳。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 370, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "cubes_reoriented", + "t": 1.1, + "log_step": 141, + "where": "操作中", + "evidence_zh": "观察:视频约 1.1 s,机器人右手恢复平掌托持,左手继续持块转动;视频 1.5–4.3 s 右块始终没有随之转向目标,视频 4.45–4.8 s 又出现明显后倾失衡,直到结束仍未完成双块调姿。log step 141 的交付脚本仅启用左手朝向伺服,评分器 success=0;断点锚定右手恢复托持、未再调姿的时刻,具体失效起点为近似判断。" + }, + "unmet": { + "cubes_reoriented": false + }, + "phenomena": [ + { + "cat": "运动", + "item": "位姿不准", + "t": 1.5, + "log_step": 146, + "evidence_zh": "观察:视频 1.5–4.3 s,机器人左手持块在身侧转动后保持,右手仍平掌托块,其可见色面方向没有转到悬浮目标所示方向;最终文件验证 log step 146 报告控制步 250 的右块朝向误差 177.2°。这是该次双块调姿的可见偏差,不把左块短暂接近目标等同于整步成功。" + }, + { + "cat": "全身", + "item": "平衡", + "t": 4.5, + "log_step": null, + "evidence_zh": "观察:视频 4.45 s 起机器人躯干持续向后倾,视频 4.7–4.8 s 身体明显倒向脚后方、骨盆和低处持块手一起下沉,没有恢复站稳;录像在倾倒途中结束,未显示身体落地。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志", + "evidence_zh": "推断:会话试验 log step 109 的右块误差从 66.2°降至 56.4°后终止、右块高度降至 0.577 m;log step 113 因右侧控制不稳改试仅左手伺服。最终脚本 log step 141 保留了只对左手求朝向控制量的实现,log step 146 对最终文件的验证仍得到右块误差 177.2°,与视频 1.5–4.3 s 的右块未调正一致;视频 4.45–4.8 s 又表明站立控制未能维持支撑。证据支持双手调姿及平衡控制未解决,但不足以把主因进一步唯一归到抓取几何、逆解或反馈参数。" + } + ], + "gap_zh": "无状态日志;视频分辨率仅 256×256,不能精确恢复接触力、方块与手的细微滑移或四元数符号。评分记录没有 episode_return 原生累计值,仅有 final_reward 与失败判定,因此原生指标 value=null,节点未达标依据其累计事件判据与评分器判定。agent 在 log step 157 自述末端“滑落”,但视频 4.45–4.85 s 只明确显示机器人后倾、左块随低处手下沉,未见方块与手分离,故未标没保持住;会话更早的右块掉落不作为本次现象。" + }, + "summary_zh": "机器人实际进行了双手收拢持块和左手调姿,右块随后恢复托持且未调正;联合累计验收节点未达到,最终进度 0/1。断在转好两个方块的操作中:右块朝向偏差一直未修复,视频约 4.5 s 又开始失衡后倾,日志支持双手调姿控制尚未解决。缺少原生累计奖励和状态日志,不能从低分或 agent 的“滑落”总结推定具体滑移、接触力或四元数符号错误。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "节点未达标(grader success=0)、断点、位姿不准与平衡现象与视频一致,未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:02:53Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-privileged--ahwdmjq/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-privileged--ahwdmjq/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "ed18727a094f9ea49e819971d83e9b6d44bdd6ef068f44a56207e18e67e50d88", + "video_sha256": "ed0cea629e91b6d5c37a789b5b33b6a8bfcd1b83d69e9822395a8d89bde2f630", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-privileged--anicu33.json b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-privileged--anicu33.json new file mode 100644 index 0000000000000000000000000000000000000000..5c248806855e1bb94160ce40bfdf8bea3efd95d8 --- /dev/null +++ b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-privileged--anicu33.json @@ -0,0 +1,185 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-cube-i00-privileged--anicu33", + "taskKey": "humanoidbench/manip_cube", + "dag_sha256": "ab5e9b50283b1821da97eeaa6a81c6bcbd6d347d014e9b3e7113b6656d951802", + "outcome": { + "success": 0, + "reward": 0.4819763936533149, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 500, + "final_reward": 0.4819763936533149, + "sim_steps": 500, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 1.669, + "peak_processes": 11, + "peak_disk_gb": 0.833, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-cube-i00-privileged--anicu33/replay.mp4", + "duration_s": 5.1, + "fps": 20.0, + "frames": 102, + "sha256": "dad3ec0cd48daf0d08774168d9db4e63e70a271bf0fd51fa5aee0d3821dff3bc" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "cubes_reoriented": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频0.00 s,两块彩色方块位于机器人的两只平掌上方,色面朝向与头旁悬浮半透明目标方块不同,初态尚未完成调姿。" + }, + { + "t": 0.4, + "log_step": null, + "text_zh": "观察:视频0.10–0.40 s,机器人两手收拢,分别包住各自方块并降到髋部附近;视频1.20–3.40 s,两块随手腕和手臂升降、翻转,说明本次确实进行了双手调姿。" + }, + { + "t": 2.8, + "log_step": null, + "text_zh": "观察:视频1.60–3.40 s,画面左侧手中方块的可见色面不断换向,手臂先抬起、再放低、又举向头顶,至少该块未形成稳定的目标朝向。视频4.05–5.05 s机器人仍站立,举高一侧的手块超出画面上缘,另一块部分被手与躯干遮挡;没有看见方块落地。" + }, + { + "t": 5.05, + "log_step": 163, + "text_zh": "推断:评分器字段success=0、all_success=0、deterministic=1、sim_steps=500;本节点按整段累计奖励达到370验收,累计量不会回退,因此本次全程未达到该验收节点。log step 163对最终交付文件的CLI重放报告500步、确定性状态哈希50462624cf53,但scored=false;不能把复现成功或瞬间近似对齐当作任务达标。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "观察:视频0.00–0.10 s机器人双掌先从下方承托各自方块;视频0.30–3.40 s左右灵巧手分别以弯曲手指包握一块,两个方块随各自手腕、手臂同时移动和转动。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频0.00–5.05 s,机器人双脚接地、双腿屈膝承重,在两手调姿过程中维持站立。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 370, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "cubes_reoriented", + "t": 2.8, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频1.60–3.40 s,机器人已包握两块并持续转动,画面左侧手块反复抬降、翻转,未稳定在目标朝向;视频2.80 s取作该调姿偏差的导航锚点。评分器最终success=0,整段联合调姿验收未达成;视频4.05–5.05 s未出现可见摔倒或掉块。" + }, + "unmet": { + "cubes_reoriented": false + }, + "phenomena": [ + { + "cat": "运动", + "item": "位姿不准", + "t": 2.8, + "log_step": null, + "evidence_zh": "观察:视频1.00 s两块仍露出与悬浮目标不同的色面;视频1.60–3.40 s至少画面左侧方块在握持中反复改变朝向,手臂在视频2.60 s抬高、视频2.80–3.00 s回落、视频3.40 s又举高,未形成稳定对齐。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "t": null, + "log_step": 152, + "evidence_zh": "推断:log step 152中的生成脚本以双臂雅可比逆解控制手掌位置与朝向,视频1.60–3.40 s显示机器人已握住方块却仍反复转动、偏离目标;log step 141同一最终候选的试验输出末态带符号目标四元数点积为0.933和0.813,agent在log step 162、168也自述最终交付具有这两个数值。结合log step 163最终文件500步确定性重放和可见调姿偏差,支持调姿执行精度不足的控制层归因,但不足以确定具体控制子类。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "无状态日志;录像仅256×256,手指遮挡色面,视频约3.60 s后举高手块超出画面,不能精确恢复两块朝向、接触力或手块相对运动。评分器未报告episode_return,final_reward=0.4819763936533149的字段不能直接当作累计奖励;仅由原生失败记录确定累计阈值未达到。日志中的其他候选试验不是本次评分片段,不能把其中掉块或跌倒记入本次;现有证据也不能进一步区分接触几何变化与逆解/反馈跟踪误差。" + }, + "summary_zh": "机器人双手从托块转为包握,并在维持站立的同时转动两块,但整段累计奖励验收未通过,节点最终为0/1。失败位于联合调姿操作中,可见至少一块在握持时反复改变朝向,支持位姿不准与控制层归因。缺少累计奖励及状态日志,且末段举高手块出画,无法精确判断终态误差或细分控制原因。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "节点 cubes_reoriented=false、双臂协作抓握与腿/躯干支撑的 contact_actual、操作中断点、位姿不准现象和停在“控制”的根因均有视频和日志支持;log_step 按 id 引用且都存在,未发现需要修改的内容。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:03:29Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-privileged--anicu33/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-privileged--anicu33/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "9f9bd3c31e751063132b5d73d581e88a0328902ecff98bf8bc1dd00040c398d3", + "video_sha256": "dad3ec0cd48daf0d08774168d9db4e63e70a271bf0fd51fa5aee0d3821dff3bc", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-privileged--j66gqku.json b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-privileged--j66gqku.json new file mode 100644 index 0000000000000000000000000000000000000000..1f95cec6a89eb3f876454c382cedb7a8d19c98fd --- /dev/null +++ b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-privileged--j66gqku.json @@ -0,0 +1,164 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-cube-i00-privileged--j66gqku", + "taskKey": "humanoidbench/manip_cube", + "dag_sha256": "ab5e9b50283b1821da97eeaa6a81c6bcbd6d347d014e9b3e7113b6656d951802", + "outcome": { + "success": 1, + "reward": 0.9962394985024506, + "source": "catalog", + "raw": { + "success": 1, + "all_success": 1, + "deterministic": 1, + "n_actions": 500, + "final_reward": 0.9962394985024506, + "sim_steps": 500, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 10.663, + "peak_processes": 37, + "peak_disk_gb": 0.838, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-cube-i00-privileged--j66gqku/replay.mp4", + "duration_s": 5.1, + "fps": 20.0, + "frames": 102, + "sha256": "bb9bf4d49a7032267ff21442b06785501ca5820e92004df25c97ac8db396c970" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "cubes_reoriented": { + "method": "visual", + "intervals": [ + [ + 5.0, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00 s,机器人双脚着地、双掌朝上,两个彩色方块分别位于掌面上方,尚未转成随后保持的姿态;悬浮半透明方块是目标朝向标记。" + }, + { + "t": 0.15, + "log_step": null, + "text_zh": "观察:逐帧查看视频 0.00–0.50 s,双手手指从张开转为围住各自方块,手臂和手腕快速转动;一手升至胸前,另一手降至髋部附近,彩色块面随手改变方向,随后大部分被手指遮挡。" + }, + { + "t": 5.0, + "log_step": null, + "text_zh": "观察:视频 0.50–5.05 s 的间隔帧及视频 4.00–5.05 s 的逐帧检查显示,机器人维持双脚支撑及两手姿势,未见方块从手边下坠或机器人倒地;遮挡和256像素画面不能给出精确朝向误差。" + }, + { + "t": null, + "log_step": 145, + "text_zh": "观察:log step 144 的 agent 脚本冻结已保存动作并执行最终审计;log step 145 输出500步、动作文件精确再生成及独立重放状态哈希相同,末步左右方块的有符号四元数距离分别约0.000142和0.000214。这是交付文件的会话审计,作为视频的辅助证据,不把该审计的控制步或agent用时换算成视频时刻。" + }, + { + "t": 5.0, + "log_step": null, + "text_zh": "推断:inputs/run.json 的评分器字段 success=1、all_success=1、deterministic=1、sim_steps=500,结合视频的转动和保持动作,确认唯一目标节点在本集结束时已完成。DAG按累计奖励达到370验收;评分器只报告 final_reward=0.9962394985024506,没有 episode_return 或逐步奖励,无法恢复首次过线时刻。区间起点视频 5.00 s 仅为已检查末段的保守确认锚点,不声称首次达标恰在视频 5.00 s,也不把视频早期的姿态稳定等同于累计达标。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "观察:视频 0.00 s 双掌从下方托着各自方块;视频 0.05–0.30 s 两只灵巧手的手指分别围住方块,并随腕臂转动,随后保持握持。两手各操作一个方块;不是两臂之间挤住同一物体。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–5.05 s 双脚接地,腿和躯干在双手调姿、持块期间承重,未见失去支撑。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 370, + "direction": "higher_better" + } + ], + "failure": null, + "summary_zh": "机器人由平掌托块转为双手分别握块并快速转动,随后保持至视频 5.05 s,未见掉块或跌倒;评分器 success=1、all_success=1,唯一节点“转好两个方块”最终完成,无失败断点。验收依据整集累计奖励达到370,评分器未报告原生 episode_return;没有状态日志,且方块后期多被手遮挡,因此不能精确恢复首次达标时刻,区间以视频 5.00 s 作为保守确认锚点,而非首次过线测量值。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "标签与视频和评分器字段一致:success=1,唯一目标节点完成,无失败断点;时间锚点说明诚实,未换算agent时间。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:02:08Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-privileged--j66gqku/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-privileged--j66gqku/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "bd510f060cdba6d52684df5859b35e4a67ce3a40ffd944a238168c217ee78360", + "video_sha256": "bb9bf4d49a7032267ff21442b06785501ca5820e92004df25c97ac8db396c970", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-privileged--lcrpnz8.json b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-privileged--lcrpnz8.json new file mode 100644 index 0000000000000000000000000000000000000000..a881990f54cefee26f2e1a1a70ae9bf94a7d3cca --- /dev/null +++ b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-privileged--lcrpnz8.json @@ -0,0 +1,185 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-cube-i00-privileged--lcrpnz8", + "taskKey": "humanoidbench/manip_cube", + "dag_sha256": "ab5e9b50283b1821da97eeaa6a81c6bcbd6d347d014e9b3e7113b6656d951802", + "outcome": { + "success": 0, + "reward": 0.2995868090780523, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 322, + "final_reward": 0.2995868090780523, + "sim_steps": 322, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 7.926, + "peak_processes": 27, + "peak_disk_gb": 0.239, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-cube-i00-privileged--lcrpnz8/replay.mp4", + "duration_s": 3.3, + "fps": 20.0, + "frames": 66, + "sha256": "3ed7f3e60bc266695cc191544a166363402774d6b47dbf9e5117392627523944" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "cubes_reoriented": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.4, + "log_step": null, + "text_zh": "观察:视频 0.00 s 两个彩色方块位于各自平掌上方;视频 0.05–0.60 s 机器人弯曲手指、用掌面兜托方块,视频 0.65–1.95 s 两臂先后抬起和转动,方块随手部移动;已开始联合调姿。" + }, + { + "t": 3.0, + "log_step": null, + "text_zh": "观察:视频 2.80–3.10 s 机器人身体逐渐前倾,视频 3.20–3.25 s 躯干翻至接近水平、骨盆明显下降,录像结束前未恢复站立。已检查全部66帧;方块部分被手掌及身体遮挡,不能从画面读出精确朝向误差。" + }, + { + "t": null, + "log_step": 60, + "text_zh": "观察:log step 59 保存最终322个动作的轨迹,log step 60 对该文件重放返回 deterministic=true、sim_steps=322;inputs/run.json 的评分器字段同为 n_actions=322、sim_steps=322、deterministic=1、success=0、grader_error=0。" + }, + { + "t": 3.25, + "log_step": null, + "text_zh": "推断:该节点以整集累计奖励达到370验收,评分器对这份确定性轨迹明确判定未成功,结合视频中的调姿尝试与末段翻倒,节点未达到;不能把局部转动或某一帧看似对齐计为完成。成功参考视频 0.20–5.05 s 显示机器人在改变双臂姿态后持续站立,仅用于比较外观,不作为本次进度证据。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "托/舀", + "note_zh": "观察:视频 0.05–0.60 s 两手分别以掌面和弯曲手指兜托各自方块,随后双臂先后抬升、转动;可见的是承托,未见足以确认拇指与其他手指对夹的画面,不将手指弯曲本身记作抓握。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–2.80 s 双脚接地、腿和躯干在双臂调姿期间承重;视频 2.85–3.25 s 支撑姿态失稳并翻倒。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 370, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "cubes_reoriented", + "t": 2.85, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:机器人已在视频 0.65–1.95 s 携带两块方块转动双臂,视频 2.85 s 起前倾加剧,视频 3.20–3.25 s 躯干翻至接近水平,未恢复即结束;评分器 success=0。失败发生在双块调姿步骤的操作中。" + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 2.85, + "log_step": null, + "evidence_zh": "观察:视频 2.80–3.25 s 机器人由直立逐渐前倾,头和躯干快速下落,腿部未能维持直立支撑,末帧躯干接近水平;调姿中的翻倒未修复。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "t": 2.85, + "log_step": 51, + "evidence_zh": "推断:log step 51 写出的最终生成器使用躯干倾角、角速度、质心位置和骨盆高度反馈调整踝、髋、膝,同时执行双臂插值动作;log step 59 修复生成器并保存322步轨迹,log step 60 验证该轨迹可确定性重放。视频 2.85–3.25 s 仍出现持续前倾和翻倒,说明交付控制器未能在双臂调姿中维持本体稳定;缺少力与关节跟踪记录,停留在控制层级,不指定某个增益、饱和或抓取几何为根因。", + "source_zh": "日志、视频" + } + ], + "gap_zh": "没有状态日志;录像为256像素视角,部分方块和手指相互遮挡,无法精确核实对向接触、方块朝向误差、四元数符号及具体终止高度。评分器未报告episode_return,final_reward=0.2995868090780523不能当作已报告的累计奖励;log step 53、56、57 的fitness是agent自建代理指标且对应较早候选,不用于本次原生指标。最终轨迹缺少力、关节跟踪和反馈诊断,无法进一步确定失稳机制。", + "unmet": { + "cubes_reoriented": false + } + }, + "summary_zh": "机器人已用双手兜托方块并先后转动双臂,但“转好两个方块”未通过整集累计奖励验收,最终节点为0/1。视频约 2.85 s 起机器人前倾加剧,至视频 3.20–3.25 s 翻倒,断点在调姿操作中,证据支持本体稳定控制失败。无状态日志或原生累计奖励记录,局部遮挡也限制了精确朝向与手指对夹判定。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "节点0/1、断点(cubes_reoriented,操作中,约视频2.85 s起前倾翻倒)、平衡现象与控制根因均与视频和log一致;视频时间均为视频钟,322步未被换算成视频时间。未发现需要修改的内容。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:04:15Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-privileged--lcrpnz8/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-privileged--lcrpnz8/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "1852a04b871ed794e37800f3f1182b1cd01210435bdf1d1dc61b985cd5a04320", + "video_sha256": "3ed7f3e60bc266695cc191544a166363402774d6b47dbf9e5117392627523944", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-privileged--oxckbzc.json b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-privileged--oxckbzc.json new file mode 100644 index 0000000000000000000000000000000000000000..5144055a180dcdbdde66549fb54d3483955d1f74 --- /dev/null +++ b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-privileged--oxckbzc.json @@ -0,0 +1,169 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-cube-i00-privileged--oxckbzc", + "taskKey": "humanoidbench/manip_cube", + "dag_sha256": "ab5e9b50283b1821da97eeaa6a81c6bcbd6d347d014e9b3e7113b6656d951802", + "outcome": { + "success": 1, + "reward": 0.995125418820322, + "source": "catalog", + "raw": { + "success": 1, + "all_success": 1, + "deterministic": 1, + "n_actions": 500, + "final_reward": 0.995125418820322, + "sim_steps": 500, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.888, + "peak_processes": 10, + "peak_disk_gb": 0.251, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-cube-i00-privileged--oxckbzc/replay.mp4", + "duration_s": 5.1, + "fps": 20.0, + "frames": 102, + "sha256": "8bc344905320cbde351920da96b8c3a41e10ddcfff49a693fa87d1a53bef7926" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "cubes_reoriented": { + "method": "visual", + "intervals": [ + [ + 5.0, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频0.0 s,机器人站立,两个彩色方块分别位于两只摊开的掌面上方;方块尚处初始朝向,画面左上有悬浮半透明目标方块。" + }, + { + "t": 0.15, + "log_step": null, + "text_zh": "观察:逐帧检查视频0.05–0.30 s,两只灵巧手的手指向各自方块弯曲包拢,彩色面随手部转动;视频0.30 s两手分向身体两侧,方块仍在各自手中。" + }, + { + "t": 0.7, + "log_step": null, + "text_zh": "观察:视频0.35–0.70 s,机器人一臂抬高弯折、另一臂降到髋旁,两手及所持方块随腕臂改变方向;视频0.70–5.05 s基本保持该姿态,双脚持续着地,未见方块下坠或身体倒下。低位手中的方块大部被手指和躯干遮挡,无法逐色面精确比较终态朝向。" + }, + { + "t": 5.0, + "log_step": null, + "text_zh": "观察:视频4.05–5.05 s逐帧检查,机器人持续站立并保持两手闭合的持物姿态;inputs/run.json的评分器字段success=1、all_success=1、deterministic=1,n_actions=500、sim_steps=500,final_reward=0.995125418820322。" + }, + { + "t": 5.0, + "log_step": null, + "text_zh": "推断:依据评分器成功记录及DAG规定的累计奖励验收规则,联合调姿节点终局为true;区间起点视频5.0 s取所见终局的确认锚点,不表示累计奖励恰在该帧首次达到370。输入没有逐步累计奖励或状态日志,不能将视频约0.70 s的稳定持块姿态直接当作累计达标,也不能从末步final_reward推算精确达标时刻。" + }, + { + "t": null, + "log_step": 133, + "text_zh": "观察:log step 133的最终生成脚本在单次recording内reset后记录最多500个动作,随后replay并保存该次动作;log step 134启动此脚本及重放核验,log step 138列出交付轨迹文件。log step 139中agent自述已验证确定性并报告小朝向误差;这些自述用于辨明交付,未替代视频中的持物观察或评分器成功记录。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "观察:视频0.0 s两手先以掌面分别托住各自彩色方块,视频0.05–0.30 s合指包拢形成各自抓握,随后两臂同时参与联合调姿;视频0.35–0.70 s一手抬高、一手下降,所持方块随腕臂转动。末段方块部分被手遮挡,但两手保持闭合,未见释放。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频0.0–5.05 s,人形机器人的双脚持续接触地面,双腿和躯干在双手持块、转动及保持期间承重。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 370, + "direction": "higher_better" + } + ], + "failure": null, + "summary_zh": "机器人在视频0.0–0.70 s由双掌托块转为两手各自抓握并随腕臂调姿,此后持续站立持块至视频5.05 s;评分器确认500动作确定性重放成功,联合调姿节点终局完成1/1,无失败断点。该节点按整段累计奖励达到370验收,区间用视频5.0 s作终局确认锚点,首次达标时刻未记录。没有状态日志或逐步累计奖励,末段部分方块受手指遮挡,无法精确核实朝向误差;episode_return未报告,保留null,末步final_reward不充当累计奖励。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "成功运行,评分器 success=1,节点终局 true 与视频一致;视频中机器人持续站立、两手持块,无失败断点,接触标注合理。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:59:38Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-privileged--oxckbzc/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-privileged--oxckbzc/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "6645c3197aafaeaf056405b1680cc2b0f2e84a7e0209442faa02e017ee30e6f0", + "video_sha256": "8bc344905320cbde351920da96b8c3a41e10ddcfff49a693fa87d1a53bef7926", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard--2rkbzz4.json b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard--2rkbzz4.json new file mode 100644 index 0000000000000000000000000000000000000000..413d838590972ab0df75d20eae099f4edd2785f5 --- /dev/null +++ b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard--2rkbzz4.json @@ -0,0 +1,192 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-cube-i00-standard--2rkbzz4", + "taskKey": "humanoidbench/manip_cube", + "dag_sha256": "ab5e9b50283b1821da97eeaa6a81c6bcbd6d347d014e9b3e7113b6656d951802", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 22, + "sim_steps": 630, + "calls": 37, + "rows": 630, + "resets": 11, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 2253.6, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-cube-i00-standard--2rkbzz4/replay.mp4", + "duration_s": 0.3, + "fps": 20.0, + "frames": 6, + "sha256": "c19b1c9ba1f0b42fd583666720f025f05badd2428254ea6152377c22d27fea1f" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "cubes_reoriented": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.05, + "log_step": 36, + "text_zh": "观察:视频 0.00–0.05 s,两只灵巧手从平掌姿态开始收指并移动,彩色方块仍在掌指上方或边缘,未见手指与拇指包围夹住方块;视频 0.10 s 两块已处于手部下方,视频 0.15–0.25 s 继续下坠。整段未见持块转至目标朝向。" + }, + { + "t": 0.25, + "log_step": 36, + "text_zh": "观察:log step 36 是最后一个实际机器人调用:脚本 reset 后执行接块段,服务返回 catch 22 True failed,脚本以 died catch 结束,后续 close 段未执行;评分器字段 live_success=0、finalized=1、infra_error=0。推断:结合视频中双块下落,该集未达到本节点所需的累计奖励验收;不能把 reward 中的 success=0 当作原生累计奖励为零。" + }, + { + "t": null, + "log_step": 37, + "text_zh": "观察:log step 37 的 calls 为空;其最终自述称执行了腕部接块、握紧、转动并保持,但没有相应实际调用或服务结果,且视频 0.10–0.25 s 显示双块下坠。推断:该自述不作为节点完成或抓握发生的依据。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "托/舀", + "note_zh": "观察:视频 0.00–0.05 s,两只灵巧手各以掌指从下方托接各自方块,随后边收指边移动;方块从掌指旁离开,未见合手夹持,故不标抓握。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.25 s,机器人双脚着地,双腿与躯干在双手接块动作期间承重,未见倒地。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 370, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "cubes_reoriented", + "t": 0.1, + "log_step": 36, + "where": "操作中", + "evidence_zh": "观察:视频 0.05 s 双手收指时方块仍未被包围夹住,视频 0.10 s 两个方块已在手部下方,随后持续下落至视频 0.25 s;机器人仍站立。log step 36 的服务结果为接块段 control_steps=22、done=True、reason=failed,握紧段未执行。推断:双块调姿在最初接块操作中失败,之后没有修复。" + }, + "unmet": { + "cubes_reoriented": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "没建立接触", + "t": 0.1, + "log_step": 36, + "evidence_zh": "观察:视频 0.00–0.05 s 方块位于张开或正在弯曲的掌指上方及边缘,未见拇指与其他手指围住方块;视频 0.10–0.25 s 两块从手旁下落,双手空留在上方。推断:短暂托接没有建立调姿所需的持块接触,不能算已抓住后滑脱。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志 / 视频", + "evidence_zh": "观察:log step 36 的脚本把双肩俯仰目标设为 −0.7 rad、双肘设为 0.8 rad,并同时给腕和手指固定目标,首段预定连续提交 60 个控制步,返回后才读取图像;服务在该段的第 22 个控制步结束。视频 0.05–0.10 s 显示双手改变姿态时没有留住方块。推断:接块动作的控制未能建立并维持有效持块接触;现有证据不足以把责任进一步分到接触几何、动作时序或伺服跟踪。" + } + ], + "gap_zh": "没有状态日志或接触力、接触对记录;视频虽能确认双块下落及未形成可见夹持,但不能确定短暂托接的精确接触点,也不足以区分接触几何、动作时序与跟踪误差。inputs/run.json 的评分器未报告 episode_return,因此该原生指标留空;节点终值由失败记录与录像共同确定。log step 37 的后续策略文字没有实际调用,不属于评分轨迹。" + }, + "summary_zh": "评分集未完成“转好两个方块”,最终进度为 0/1;视频 0.00–0.10 s 双手尝试托接、收指并移动,但未建立可见夹持,两个方块随后下落,机器人仍保持双脚支撑。最后实际调用在接块段失败(log step 36),根因归为控制,无法进一步确定具体控制机制;没有状态接触记录和原生累计奖励,最终自述中的转腕保持未获实际调用或录像支持。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "视频为末次评分集(22 控制步,0.3 s)。抽帧可见双手托接后方块下坠,机器人站立;节点未达成、断点 cubes_reoriented、没建立接触、控制根因均有证据,且 gap_zh 说明了不确定处。未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:18:54Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-standard--2rkbzz4/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-standard--2rkbzz4/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "918ee18608dd0b02a81f9e127468272e01a1f3f593c846951bffa83783928aae", + "video_sha256": "c19b1c9ba1f0b42fd583666720f025f05badd2428254ea6152377c22d27fea1f", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard--4zhjc4e.json b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard--4zhjc4e.json new file mode 100644 index 0000000000000000000000000000000000000000..730017030e54749cd7b194c5b62e2894b6731861 --- /dev/null +++ b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard--4zhjc4e.json @@ -0,0 +1,198 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-cube-i00-standard--4zhjc4e", + "taskKey": "humanoidbench/manip_cube", + "dag_sha256": "ab5e9b50283b1821da97eeaa6a81c6bcbd6d347d014e9b3e7113b6656d951802", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 116, + "sim_steps": 3841, + "calls": 488, + "rows": 3841, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1389.0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-cube-i00-standard--4zhjc4e/replay.mp4", + "duration_s": 1.25, + "fps": 20.0, + "frames": 25, + "sha256": "32e709b08e611fd1bf0c20b5a7dcea019b830c61feab49e6f889890445807257" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "cubes_reoriented": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:已检查视频 0.00–1.20 s 的全部25帧;机器人两手保持张开、掌心朝上,两个彩色方块各落在一只手掌上,未见合指包握或主动朝悬浮目标转块;视频 0.90–1.20 s 机器人持续后仰、骨盆降低。" + }, + { + "t": 1.2, + "log_step": 120, + "text_zh": "观察:log step 120 最后一次完成的试验为 nograsp,先保持初始姿态5个控制步,再令双膝目标为0.65 rad并保持,其余关节目标不变;服务返回 nograsp end 116 failed,下一次reset因预算耗尽被拒绝。log step 122 确认最后为episode 51、done=true、reason=failed;此前闭指试验不是受评集。" + }, + { + "t": 1.2, + "log_step": 122, + "text_zh": "推断:本任务节点按累计奖励达到370验收,服务达标即结束;最后受评集由服务判为failed,run.json评分字段live_success=0、replay_success=0且deterministic=1,结合全片无主动转块动作,节点未达到。评分字段未报告episode_return,不能将目录reward=0当作原生累计奖励。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "托/舀", + "note_zh": "观察:视频 0.05–1.20 s,机器人左右张开的手掌各从下方托住一个彩色方块,未见包夹;这是受评保持姿态时的实际承托,没有主动调姿。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.85 s,机器人双脚接地,双腿及躯干承重以维持托块姿态;视频约0.90 s起身体向后失衡。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 370, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "cubes_reoriented", + "t": 0.9, + "log_step": 120, + "where": "首步前", + "evidence_zh": "观察:视频 0.00–1.20 s 机器人只以双掌托块,未启动主动转块动作;视频约0.90 s起后仰,至视频1.20 s明显向后倾倒。log step 120 的受评末集代码是保持初始手臂和手指姿态、仅改变膝关节目标的基线,未含抓取或朝目标旋转;log step 132 中agent也自述最后受评集仅为保持姿态、没有尝试主动转块。" + }, + "unmet": { + "cubes_reoriented": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 120, + "evidence_zh": "观察:视频 0.00–1.20 s 两只手始终张开托块,没有启动抓住方块再转向目标的动作;log step 120 最后一次nograsp试验的控制目标保持手臂、手腕和手指初始角度,仅改变膝关节目标。log step 132 中agent自述受评集只是保持姿态,未尝试主动转块。" + }, + { + "cat": "全身", + "item": "平衡", + "t": 0.9, + "log_step": null, + "evidence_zh": "观察:视频 0.85–1.20 s,机器人躯干逐渐向后倾斜,骨盆降低,双腿伸向前方,至视频1.20 s已明显向后倾倒且没有恢复;此时两方块仍在手掌附近,未见方块先落地。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "t": 0.9, + "log_step": 120, + "source_zh": "日志", + "evidence_zh": "推断:log step 120 的受评末集脚本在双膝设为0.65 rad后反复提交同一组关节目标,虽读取结束状态,却不根据身体倾斜或关节偏差修正姿态;视频 0.90–1.20 s 在该保持控制下持续后仰而未恢复。会话中的站姿与反馈试验也未得到稳定控制:log step 107 各膝目标试验均failed,log step 111 两种自适应调节分别在控制步120、118返回failed;这些是根因的会话证据,不计入本集进度。证据支持站立控制未能稳住身体,但不足以进一步确定伺服力矩、反馈算法或特定参数哪一项是决定原因。" + } + ], + "gap_zh": "没有状态日志,原生episode_return未报告,视频不能给出精确朝向误差或原生终止高度;失败结局由服务和评分器确认。log step 122 报告116个控制步,按log step 6 的control_dt=0.02对应仿真2.32 s,而本视频只有1.25 s,缺少经可见事件核对的控制步到视频时间映射;所有标签时刻直接取视频帧,不将控制步或agent工作时间换算成视频时刻。控制根因可定位到未能稳定站立,缺少受评集逐步关节、力矩与接触数据以区分更细子类。" + }, + "summary_zh": "受评末集只用双掌托住两个方块并保持站姿,没有启动主动转块,转好两个方块节点最终为0/1。视频约0.90–1.20 s机器人后仰倾倒,日志显示恒定关节目标未能稳住身体,断点位于调姿首步前,现象为没尝试及平衡失败。原生累计奖励与状态日志缺失,且日志控制步时长与视频长度缺少可核实映射;节点结局依据服务及重放评分失败记录,时刻仅取视频。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "标签与视频和日志一致:受评末集(log id 120 的 nograsp 基线,服务 episode 51 failed)只以双掌托块,未启动转块;视频约0.9 s起后仰倾倒;节点未达、断点首步前/没尝试+平衡、根因控制(停在父类)均有证据支持,时钟未换算。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:04:51Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-standard--4zhjc4e/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-standard--4zhjc4e/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "a1bcd20bcd05b8a663174a7f758d799c32b65c095c26b7dabcd1b9e1c3b122b7", + "video_sha256": "32e709b08e611fd1bf0c20b5a7dcea019b830c61feab49e6f889890445807257", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard--5uz2mv2.json b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard--5uz2mv2.json new file mode 100644 index 0000000000000000000000000000000000000000..c54b5dc5e2e793f45e7d1ac92d2e7ab604e312dd --- /dev/null +++ b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard--5uz2mv2.json @@ -0,0 +1,173 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-cube-i00-standard--5uz2mv2", + "taskKey": "humanoidbench/manip_cube", + "dag_sha256": "ab5e9b50283b1821da97eeaa6a81c6bcbd6d347d014e9b3e7113b6656d951802", + "outcome": { + "success": 1, + "reward": 1.0, + "source": "catalog", + "raw": { + "success": 1, + "live_success": 1, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 477, + "sim_steps": 6150, + "calls": 3087, + "rows": 6150, + "resets": 20, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 2357.5, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.329, + "peak_processes": 13, + "peak_disk_gb": 0.911, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-cube-i00-standard--5uz2mv2/replay.mp4", + "duration_s": 4.85, + "fps": 20.0, + "frames": 97, + "sha256": "c861983d295eb6f5ab358dd944e751f45bee9106e459b15b5d19f22edebf39b0" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "cubes_reoriented": { + "method": "visual", + "intervals": [ + [ + 4.8, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00 s,机器人两手张开、掌心朝上,各有一个彩色方块位于掌上方;悬浮半透明方块为目标朝向,初始方块尚未转向目标。" + }, + { + "t": 0.5, + "log_step": null, + "text_zh": "观察:视频 0.10–0.60 s,两只灵巧手逐渐合拢包住各自的方块,方块随手掌倾转;视频 1.30–2.20 s,机器人两臂明显抬转、手腕改向,两个方块随各自的手移动,未见掉至地面;双脚持续支撑站立。" + }, + { + "t": 4.8, + "log_step": 159, + "text_zh": "观察:视频 3.85–4.80 s 逐帧可见机器人保持站立,上举的手及躯干旁的手保持合拢持块姿态,部分色面被手和身体遮挡,未见方块下坠;log step 159 的机器人服务返回 episode=21、control_steps=477、done=true、reason=success,log step 160 再次查询保持相同成功状态;run.json 的评分器字段 live_success=1、graded_by_live=1、finalized=1。" + }, + { + "t": 4.8, + "log_step": 159, + "text_zh": "推断:log step 145 重置并启动第21集,log step 146、153、154、159 连续记录该集,最终由服务判成功,与视频持块转动后保持的过程一致;此前20集试验不计入本次进度。节点按累计奖励达到370验收,标准模式达标即结束,故将达到事件锚定于已检查的末帧视频 4.80 s 附近;这是末尾近似定位,不将服务的仿真 9.54 s 或 agent 工作时间直接换算成视频时间。未报告累计奖励数值,不能从色面或agent自算朝向误差确定更早达标时刻。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "观察:视频 0.00 s 两手起初平掌承接各自方块,视频 0.10–0.60 s 手指合拢包住方块;随后左右手各握一块,配合腕臂转动和保持,视频 1.30–2.20 s 可见方块随手移动。实际是各手独立包握各自方块。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–4.80 s 机器人双脚着地,腿与躯干承重;双手持块转动时身体保持站立,未见跌倒。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 370, + "direction": "higher_better" + } + ], + "failure": null, + "summary_zh": "第21集为最终受评尝试:机器人双手分别握住方块,经腕臂转动后保持持块站立,服务在 log step 159 报告该集成功,唯一节点在视频末尾约 4.80 s 达到,终态为1/1。任务设计问题:任务文字要求两个方块转向目标,但DAG记录的实际验收是整段累计稠密奖励达到370,手块接近及站立等奖励也参与累计,成功不等于末帧两个方块精确对齐。没有状态日志或累计奖励数值,部分色面被遮挡,因此按服务成功记录确认达标,并仅以视频末帧作近似时间锚点。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "成功集,唯一节点在末帧约4.80 s(视频时钟)达到、终态1/1,与服务成功记录和帧一致;接触标注合理,任务设计问题说明有依据,未发现需修改之处。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:03:30Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-standard--5uz2mv2/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-standard--5uz2mv2/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "d58caee8a71bbd67abc09b23c53116618a8842a99a96df98890aef68005163e6", + "video_sha256": "c861983d295eb6f5ab358dd944e751f45bee9106e459b15b5d19f22edebf39b0", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard--c2mvlbm.json b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard--c2mvlbm.json new file mode 100644 index 0000000000000000000000000000000000000000..7a39ecd1707eb3230b26c0ec6131b5ec69530b3b --- /dev/null +++ b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard--c2mvlbm.json @@ -0,0 +1,202 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-cube-i00-standard--c2mvlbm", + "taskKey": "humanoidbench/manip_cube", + "dag_sha256": "ab5e9b50283b1821da97eeaa6a81c6bcbd6d347d014e9b3e7113b6656d951802", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 136, + "sim_steps": 4071, + "calls": 1892, + "rows": 4071, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 2376.0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-cube-i00-standard--c2mvlbm/replay.mp4", + "duration_s": 1.45, + "fps": 20.0, + "frames": 29, + "sha256": "23f9933d38f487b4ea4ac6a4a57a2d53c08b7092e96f48302a73bfd84a97e7c4" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "cubes_reoriented": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00 s 两个实体彩色方块位于机器人摊开的两掌上方,半透明目标方块悬在双手之间的高处;这是尚未进行调姿的初态。" + }, + { + "t": 0.6, + "log_step": null, + "text_zh": "观察:视频 0.20–0.60 s 机器人两手的手指逐渐围拢各自方块;视频 0.60–1.00 s 两个方块随手腕和手臂抬升、转动,确实进行了双块调姿尝试。" + }, + { + "t": 1.3, + "log_step": null, + "text_zh": "观察:视频 1.30 s,一个方块已脱离抬到头旁的手,出现在躯干侧边;视频 1.35–1.40 s 它继续降到髋部附近,另一方块仍在胸颈前的手边,未见重新接回掉落方块。" + }, + { + "t": null, + "log_step": 348, + "text_zh": "观察:log step 348 执行最后一次重置后的第51集,服务在该集 control_steps=136 返回 done=True、reason=failed;log step 349 的最终状态仍为第51集失败。run.json 的 reward.graded_by_live=1、live_success=0,未报告 episode_return 数值。" + }, + { + "t": null, + "log_step": 13, + "text_zh": "推断:log step 12–13 给出的服务规则是只评最后一集,且累计奖励达标会立即成功结束;最后一集以 failed 结束并出现视频 1.30–1.40 s 的掉块,因此联合累计验收事件未达到,节点 final=false、无达成区间。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "观察:视频 0.20–0.60 s 左右灵巧手分别合拢手指包围各自的彩色方块,随后两臂、手腕带着方块抬升转动;在此之前两手先以平掌从下方承接方块,视频 1.30 s 后其中一块从手边脱落。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–1.40 s 机器人双脚接触地面,腿与躯干承重支撑双手调姿;末段躯干明显前倾,录像结束前未见倒地。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 370, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "cubes_reoriented", + "t": 1.3, + "log_step": 348, + "where": "操作中", + "evidence_zh": "观察:机器人已在视频 0.60–1.00 s 持块抬臂转动,但视频 1.30 s 可见一个方块脱离头旁的手,至视频 1.40 s 继续下落且未被接回;log step 348–349 确认最后评分集失败,双块调姿步骤未修复。" + }, + "unmet": { + "cubes_reoriented": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "没保持住", + "t": 1.3, + "log_step": null, + "evidence_zh": "观察:视频 0.40–0.80 s 方块被围拢的手指保持并随抬手移动;视频 1.30–1.40 s 其中一块出现在远低于手掌的躯干侧边并持续下落,机器人没有重新接住。此前已建立持物接触,失败发生在保持接触的过程中。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志 / 视频", + "evidence_zh": "推断:log step 347 的实际执行脚本先按固定时序闭合手指,再插值旋转两臂和手腕;闭合阶段之后手指目标保持固定,读取的反馈只用于踝关节平衡修正。结合视频 0.40–0.80 s 的持块转动和视频 1.30–1.40 s 的掉块,该控制过程未能在调姿中维持方块;现有证据只能定位到控制层,不能区分接触几何、关节跟踪与身体倾斜各自的作用。" + } + ], + "gap_zh": "未提供状态日志;run.json 和最后一集日志未报告 episode_return、逐步方块姿态、接触力或具体失败终止分支,无法量化累计得分或进一步细分控制根因。头旁部分持物过程被手和头部遮挡,掉块时刻以视频 1.30 s 首次清楚可见的下落为锚点;末段身体倾斜,但录像截止前未见倒地。log step 347–348 的136个控制步按 dt=0.02 s 对应仿真 2.72 s,与视频总长 1.45 s 不同,未把控制步换算成视频时间,所有视频时刻均取自检查的帧。log step 351 在评分集结束后修改踝部控制,log step 352 只检查语法,后续没有再执行,不能作为本次评分轨迹的修复。" + }, + "summary_zh": "最后评分集完成了双手承接、合拢持块和抬臂转动的尝试,但视频约 1.30 s 一个方块脱手下落且未被接回,“转好两个方块”最终为0/1。断点在调姿操作中,现象为没保持住,日志与画面支持控制层归因。缺少累计奖励数值、接触力和状态日志,无法精确量化得分或细分掉块机制;结束后修改的脚本没有重新执行。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "逐帧核对:0.0 s 平掌承块,0.2–0.6 s 合拢持块,1.3 s 一块脱手下落且未接回,节点未达成、断点与『没保持住』『控制』归因均有帧和日志支持;时钟未把控制步换算成视频时间。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:03:21Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-standard--c2mvlbm/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-standard--c2mvlbm/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "589fe0de818b1c86d6ff0a24e01bb7bb9791fe44e135dc85257208c770cd1d90", + "video_sha256": "23f9933d38f487b4ea4ac6a4a57a2d53c08b7092e96f48302a73bfd84a97e7c4", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard--cgnqfnl.json b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard--cgnqfnl.json new file mode 100644 index 0000000000000000000000000000000000000000..182c3f8fd322090c344b4ece63f5f59fd13f3c70 --- /dev/null +++ b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard--cgnqfnl.json @@ -0,0 +1,197 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-cube-i00-standard--cgnqfnl", + "taskKey": "humanoidbench/manip_cube", + "dag_sha256": "ab5e9b50283b1821da97eeaa6a81c6bcbd6d347d014e9b3e7113b6656d951802", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 70, + "sim_steps": 70, + "calls": 5, + "rows": 70, + "resets": 0, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 2348.1, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-cube-i00-standard--cgnqfnl/replay.mp4", + "duration_s": 0.8, + "fps": 20.0, + "frames": 16, + "sha256": "cb606b0b665911adfba1d07dcdb8e511fb8b45e407e7d7195353074d3b5b5bfc" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "cubes_reoriented": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00–0.10 s,两只彩色实心方块分别位于机器人两手掌上方,尚未呈现悬浮半透明目标方块的朝向;视频 0.15–0.45 s,两手屈指包围各自方块,方块随接触转动。" + }, + { + "t": 0.55, + "log_step": 82, + "text_zh": "观察:视频 0.50 s 左手带块抬向胸前,视频 0.55 s 方块已悬在该手下方,视频 0.60–0.70 s 继续从胸前坠至骨盆、两腿之间;视频 0.75 s 末帧左手空着,右手方块仍在手边。log step 82 的脚本收紧左手并旋转左臂,服务返回 control_steps=70、done=False、reason=None;左右手身份由该动作与画面对应确认,视频时间直接取帧。" + }, + { + "t": 0.75, + "log_step": null, + "text_zh": "推断:评分器 success=0、live_success=0、replay_success=0,结合完整视频中左块离手且未恢复,联合节点“转好两个方块”未达到原生累计奖励验收;不把短暂转动算作完成。评分器未报告 episode_return,目录 reward=0.0 不能代替原生累计奖励。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "托/舀", + "note_zh": "观察:视频 0.00–0.10 s,机器人左右平掌分别从下方承接各自的彩色方块。" + }, + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "观察:视频 0.15–0.45 s,两只灵巧手的手指弯曲,分别包围各自方块;随后左手随腕臂转动并失去方块,右手继续托持其方块。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.75 s,机器人双脚与地面接触,腿和躯干承重;末段身体倾斜,未见倒地。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 370, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "cubes_reoriented", + "t": 0.55, + "log_step": 82, + "where": "操作中", + "evidence_zh": "观察:视频 0.45–0.55 s,机器人左手从腰前抬起并转动,原先手中的方块脱离手指,随后在视频 0.60–0.70 s 从胸前下坠到两腿之间,至视频 0.75 s 未重新抓回;右手仍在另一方块旁。log step 82 是本集左臂旋转和左手收紧的动作提交,评分器记录无重置、共70个动作。" + }, + "unmet": { + "cubes_reoriented": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "没保持住", + "t": 0.55, + "log_step": 82, + "evidence_zh": "观察:视频 0.40–0.50 s 左手手指包围方块并带其抬起,视频 0.55 s 方块从转动后的手下方离开,视频 0.60–0.70 s 自由下坠,至视频 0.75 s 左手空着;这是已有持物接触的丢失。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "抓取和接触设计", + "alt_feasible": "未知", + "t": 0.55, + "log_step": 82, + "source_zh": "日志 / 视频", + "evidence_zh": "推断:log step 58 给两手固定屈指目标,log step 82 在左臂旋转前先增加左手拇指与四指的闭合量,并在旋转过程中保持这些目标,没有释放左块的指令;但视频 0.50–0.60 s 左手转动后方块从手下方滑出。该握持布局未能在变换承载方向时约束方块,故归为抓取和接触设计;未在本次初态验证可行替代。" + } + ], + "gap_zh": "观察:inputs/ 无状态日志,评分器也未报告 episode_return;视频无法提供精确接触力、各指接触对及累计奖励。log step 82 服务仍返回 done=False、reason=None,log step 91 出现模型提供方过滤中断,不能据此确认方块已触发原生低高度终止。推断:按中断前已经可见且未修复的左块滑落标注物理失败,不把过滤中断归为机器人现象或根因;握持失效的具体力学机制仍待核实。视频锚点均直接取帧,未把控制步数或 agent 工作时间换算为视频时间。" + }, + "summary_zh": "机器人由双掌托块转为屈指握持,并开始左臂调姿,但“转好两个方块”最终为0/1。视频约 0.55 s 左块从转动的左手下方脱离,至视频 0.70 s 下坠到两腿之间;结合 log step 82 持续闭合手指的指令,推断握持布局未能在旋转时保持方块。无状态日志和原生累计奖励,且 log step 91 被模型提供方过滤中断;不将中断当作根因,也不声称已确认低高度终止。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "视频仅16帧(0.8 s),左块在约0.55–0.70 s从转动的左手滑落、右手仍托块,与标注一致;节点未达成、断点、现象与根因(带推断标记)均有证据支持,未发现需改动的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:01:45Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-standard--cgnqfnl/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-standard--cgnqfnl/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "2915f25937ab218ec1bc2b37f3f61bf9d522b52244df808068c32d88c59e7b0a", + "video_sha256": "cb606b0b665911adfba1d07dcdb8e511fb8b45e407e7d7195353074d3b5b5bfc", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard--prmkfab.json b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard--prmkfab.json new file mode 100644 index 0000000000000000000000000000000000000000..182b8c191ebe5653e8152fd135742d168c495fbb --- /dev/null +++ b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard--prmkfab.json @@ -0,0 +1,200 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-cube-i00-standard--prmkfab", + "taskKey": "humanoidbench/manip_cube", + "dag_sha256": "ab5e9b50283b1821da97eeaa6a81c6bcbd6d347d014e9b3e7113b6656d951802", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 269, + "sim_steps": 4360, + "calls": 4330, + "rows": 4360, + "resets": 27, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1466.8, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-cube-i00-standard--prmkfab/replay.mp4", + "duration_s": 2.75, + "fps": 20.0, + "frames": 55, + "sha256": "56ad2df181c62bcc5c8f5c6ae36dc97eb0f9b5f2f8ef4419553d369f3bb612c4" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "cubes_reoriented": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00 s,机器人站立,两只彩色方块分别位于张开的左右手掌上方,半透明目标方块悬浮在双手之间偏上方;初态尚未完成调姿。" + }, + { + "t": 0.4, + "log_step": null, + "text_zh": "观察:视频 0.20–0.55 s,机器人左右手的手指逐渐包围各自方块;视频 0.60–2.00 s,两臂带着握住的方块转动,方块离开原来平掌上方的位置,随后大部分被手遮挡。" + }, + { + "t": 2.65, + "log_step": 74, + "text_zh": "观察:视频 2.60 s,画面左侧的机器人右手指端下方露出彩色方块;视频 2.65–2.70 s,方块与该手分离并继续下坠,机器人仍站立。log step 74 最后执行返回 done 268 failed、steps 269 True failed,评分字段 live_success=0、graded_by_live=1、n_actions=269。" + }, + { + "t": 2.7, + "log_step": 74, + "text_zh": "推断:inputs/run.json 的 task_text 规定仅评分最后一集,成功会立即结束;log step 74 最后一集结束为 failed,且录像显示持块转动后掉块,因此该集从未达到 DAG 所绑定的累计奖励370验收条件,节点未达成。不能用被手遮住时可能出现的瞬时对齐替代累计验收。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "观察:视频 0.00 s 左右手起初张掌托在各自方块下方;视频 0.20–0.55 s 两只灵巧手分别合拢包住各自方块,此后两臂同时带块转动。视频 2.60–2.70 s,画面左侧的机器人右手未能继续握住方块,方块从指端下方滑出。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–2.70 s,机器人双脚接地,腿与躯干承重,双手调姿和方块掉落期间均未见摔倒。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 370, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "cubes_reoriented", + "t": 2.6, + "log_step": 74, + "where": "操作中", + "evidence_zh": "观察:机器人在视频 0.60–2.00 s 持块转动双臂,之后右手朝下;视频 2.60–2.70 s,右手(画面左侧)的方块从指端滑出并落向低处,片尾未接回。log step 74 最后执行返回 steps 269 True failed;失败发生在双块调姿操作中。" + }, + "unmet": { + "cubes_reoriented": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "没保持住", + "t": 2.6, + "log_step": 74, + "evidence_zh": "观察:视频 0.40–1.00 s,机器人右手包住并带动其方块转动;视频 1.50–2.55 s,该手转成指端朝下,方块被手遮挡;视频 2.60 s 方块从指端露出,视频 2.65–2.70 s 明显离手下坠。机器人没有执行目标处的计划松手,也没有接回方块。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "抓取和接触设计", + "alt_feasible": "未知", + "t": 2.6, + "log_step": 74, + "source_zh": "日志", + "evidence_zh": "推断:log step 41 的 hand 函数给定固定的各指弯曲目标;log step 45 的 grasp_keys 在合手后保持这些目标,执行部分只增加肩、肘、腕的调姿路径,log step 53 将合手幅度改为 GS 参数;log step 74 最后仍用 GS=3.0 执行同一流程。视频 1.50–2.55 s 显示右手转至指端朝下,视频 2.60–2.70 s 方块在手未主动张开的情况下滑出,支持该固定包握在翻转后的承载姿态下不足以保持方块。未验证具体可行的替代接触方案。" + } + ], + "gap_zh": "没有状态日志,录像中持块后大部分色面被手遮挡,无法精确判断两块瞬时朝向、四元数符号及滑动开始时刻;无接触力数据,不能进一步区分具体手指接触几何与伺服承载不足。inputs/run.json 的 reward 未报告 episode_return 原生数值,故该指标为 null;节点未达由最后一集失败记录及可见掉块共同确认。成功参考在其视频 0.00、0.10、0.30、0.60、5.00–5.05 s 可对照初态、持块转动和后续维持,但不能提供本次的奖励或接触力。" + }, + "summary_zh": "评分的最后一集先合拢双手持住两个方块并转动双臂,唯一节点“转好两个方块”未达到累计奖励验收,最终为0/1。视频约 2.60 s 右手(画面左侧)的方块滑出下坠,断点在调姿操作中;固定包握翻转后的保持不足,机器人本体仍站立。没有状态及接触力日志,持块后色面受遮挡,且未报告原生累计奖励,无法精确复核朝向误差与受力细节。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "节点未达成、断点(视频约2.6 s 右手方块滑出)、没保持住现象与抓取和接触设计根因均与帧和日志一致;时钟均为视频时间,无换算。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:03:38Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-standard--prmkfab/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-standard--prmkfab/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "97e75a7e1828ef160dcedaa630304a1fa4732426c34f0d8e3af448765f95d45a", + "video_sha256": "56ad2df181c62bcc5c8f5c6ae36dc97eb0f9b5f2f8ef4419553d369f3bb612c4", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard--tk3dj2e.json b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard--tk3dj2e.json new file mode 100644 index 0000000000000000000000000000000000000000..ac94580897eac579b38482e6cba7e522b996fbd2 --- /dev/null +++ b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard--tk3dj2e.json @@ -0,0 +1,199 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-cube-i00-standard--tk3dj2e", + "taskKey": "humanoidbench/manip_cube", + "dag_sha256": "ab5e9b50283b1821da97eeaa6a81c6bcbd6d347d014e9b3e7113b6656d951802", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 85, + "sim_steps": 2951, + "calls": 2682, + "rows": 2951, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1549.9, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-cube-i00-standard--tk3dj2e/replay.mp4", + "duration_s": 0.95, + "fps": 20.0, + "frames": 19, + "sha256": "f37ef6759705e9a8d9271da7375ff9781c7c6aacbc93e796b7b430b1c35e5719" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "cubes_reoriented": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00 s 两个彩色方块分别在机器人摊开的掌面上,朝向与上方半透明目标不同;逐帧检查视频 0.00–0.90 s,未见两手合拢抓取或主动转块到目标朝向。成功参考在参考视频 0.10–0.30 s 出现明显握持、抬臂和转腕,本集没有相应动作。" + }, + { + "t": 0.85, + "log_step": 102, + "text_zh": "观察:视频 0.70–0.85 s 机器人逐渐明显向后倾倒,两块仍在手边,未形成对准目标的操作结果。log step 102 的服务返回确认最后评分集为 episode 51,control_steps=85、done=True、reason=failed;run.json 的 live_success=0。" + }, + { + "t": null, + "log_step": 102, + "text_zh": "推断:DAG 以累计奖励达到370验收该唯一目标,最后评分集的服务失败结果确认该节点未达标;这一判断同时依据可见的未调姿与失衡过程,不以目录 reward=0 代替动作证据。原生 episode_return 未报告,不能把目录 reward=0 当作累计奖励。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "托/舀", + "note_zh": "观察:视频 0.00–0.85 s,机器人左右掌面分别从下方托住各自的彩色方块;本集保持初始托持,未见手指形成对向夹持或主动手内调整,后段双手随身体后仰。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.65 s,机器人双脚着地、屈膝承重;视频 0.70–0.85 s 躯干和骨盆明显向后下方偏移,支撑姿态失稳。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 370, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "cubes_reoriented", + "t": 0.7, + "log_step": 102, + "where": "首步前", + "evidence_zh": "观察:视频 0.00–0.65 s 机器人仅维持双掌托块的初始姿态,未开始将方块转向目标;视频 0.70–0.85 s 身体向后倾倒,之后没有恢复。log step 102 的最终控制循环只调用平衡控制器,服务在 episode 51 的控制步85返回 failed;方块调姿尚未启动即因支撑失稳告终。" + }, + "unmet": { + "cubes_reoriented": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 102, + "evidence_zh": "观察:视频 0.00–0.90 s 两手保持掌面托块,方块随手和身体移动,未见主动转向目标的动作;log step 102 的最终脚本只运行 Balance 控制循环。agent 在 log step 104 也自述未尝试手内旋转,与画面一致。" + }, + { + "cat": "全身", + "item": "平衡", + "t": 0.75, + "log_step": 102, + "evidence_zh": "观察:视频 0.70–0.85 s 机器人上身明显向后仰,骨盆后移下沉,双腿前伸,最后画面未恢复站姿;log step 102 的最终集轨迹打印俯仰角由控制步65的-0.114 rad扩大至控制步84的-0.592 rad,服务在控制步85结束该集。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "反馈控制", + "source_zh": "日志 / 视频", + "evidence_zh": "推断:最终平衡闭环未能抑制后仰,机器人在进入主动调姿前失稳。log step 96 将髋部修正下限夹为-0.09;log step 102 的最终集输出显示控制步71至84髋修正维持-0.09、踝修正从0.892增至3.421,而俯仰偏差从-0.166 rad继续扩大到-0.592 rad;该脚本对动作裁剪到[-1,1]。视频 0.70–0.85 s 的持续后仰与此一致,支持反馈修正不足并饱和的控制失稳;不能据此证明任何具体增益是唯一根因。" + } + ], + "gap_zh": "没有状态日志,原生累计奖励 episode_return 及精确终止高度未提供,不能判定具体是哪个低高度判据结束该集。视频共19帧、时长0.95 s,最后不同画面在视频 0.85 s;log step 102 的控制步85按脚本dt=0.02 s对应仿真1.70 s,与视频时长不符,因此不建立控制步到视频时间的映射,视频锚点全取实际帧。初始托持可见,但小尺寸和手部遮挡不足以恢复逐指接触或精确方块朝向;未把随身体后仰发生的方块转动算作主动调姿。" + }, + "summary_zh": "最后评分集 episode 51 仅维持双掌托块与站立,未启动主动方块调姿,唯一节点“转好两个方块”未完成。视频约 0.70–0.85 s 机器人明显向后失衡,最终集日志显示平衡反馈修正增大并受限却未阻止俯仰偏差扩大,服务在控制步85返回失败。原生累计奖励未报告,且控制步时长与视频时长不一致,故不换算两者时间,也不指定未经证实的低高度终止条件。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "标注与视频、日志一致:视频0.00–0.65 s双掌托块站立,0.70–0.90 s明显后倾;唯一节点未达成,断点首步前/没尝试+平衡,根因反馈控制有最终集脚本输出支撑,未把控制步换算成视频时间。未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:03:05Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-standard--tk3dj2e/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-standard--tk3dj2e/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "ef05b71889b96950899b7e2f95906000a5f67191afccbe6256d1aafd1e61d53a", + "video_sha256": "f37ef6759705e9a8d9271da7375ff9781c7c6aacbc93e796b7b430b1c35e5719", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard--vjpbwcr.json b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard--vjpbwcr.json new file mode 100644 index 0000000000000000000000000000000000000000..883024d5a6cabaededf7f78bfce72c8477b47afd --- /dev/null +++ b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard--vjpbwcr.json @@ -0,0 +1,194 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-cube-i00-standard--vjpbwcr", + "taskKey": "humanoidbench/manip_cube", + "dag_sha256": "ab5e9b50283b1821da97eeaa6a81c6bcbd6d347d014e9b3e7113b6656d951802", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 384, + "sim_steps": 6140, + "calls": 2511, + "rows": 6140, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1416.2, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-cube-i00-standard--vjpbwcr/replay.mp4", + "duration_s": 3.9, + "fps": 20.0, + "frames": 78, + "sha256": "de998201c5f5a59b30073a62b059f1131bd05c8c9be2b2379fc93e397e2d7a4a" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "cubes_reoriented": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00 s 两个彩色方块分别位于机器人张开的掌上,悬浮半透明目标方块位于双手上方;视频 0.05–0.60 s 两手手指收拢,手臂开始带块转动。" + }, + { + "t": 2.0, + "log_step": null, + "text_zh": "观察:视频 1.00–3.20 s 两个方块随各自手掌下移并改变可见色面;视频 3.40–3.85 s 机器人持续向后倾倒,两块仍在收拢的手指与掌部之间,未见从手边自由下坠。" + }, + { + "t": 3.85, + "log_step": 149, + "text_zh": "推断:log step 149 启动最后一次重置后的控制回合,log step 151、154 的 agent 自述该回合失败;inputs/run.json 的评分器字段 live_success=0、graded_by_live=1、n_actions=384 确认现场未成功。DAG 此事件以累计奖励达到370验收,标准模式达标即成功结束,因此该回合未到达此节点;不是仅以零奖励推断未操作。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "观察:视频 0.00 s 左右掌分别从下方托住各自方块;视频 0.05–0.30 s 两只灵巧手的手指向方块收拢,随后各自夹持方块随手腕、手臂转动,视频 1.00–3.85 s 仍见方块留在弯曲手指与掌部之间。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–3.20 s 机器人双脚接地、屈膝承重,支撑双手持块调姿;视频 3.40–3.85 s 身体向后倒,脚部支撑随后丧失。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 370, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "cubes_reoriented", + "t": 3.4, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频 0.05–3.20 s 机器人双手持块转动;视频 3.40 s 起躯干后仰加剧,视频 3.75–3.85 s 骨盆明显下降、双脚翻起,机器人向后倒,未恢复支撑。该时刻正在执行两个方块的联合调姿步骤。" + }, + "unmet": { + "cubes_reoriented": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 3.4, + "log_step": null, + "evidence_zh": "观察:视频 3.40–3.85 s 机器人在双手仍持块时向后倾倒,躯干由接近竖直变成大幅后仰,骨盆下降,脚掌翻起离开原先的站立支撑;视频结束前没有恢复。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "t": 3.4, + "log_step": 153, + "source_zh": "视频 / 日志", + "evidence_zh": "推断:视频 3.40–3.85 s 机器人持块调姿过程中后仰持续扩大,支撑控制未能抑制失稳。log step 153 读取最后评分回合的本体记录,末端 IMU 输出含 imu_gyro_y=-1.362、重力方向分量约(-0.81,-0.003,-0.587),与视频末尾的大幅后仰一致;log step 149 的 agent 自述最终脚本采用慢速转块与踝部补偿。综合判为控制失败,但日志中的最终修改仅保留截断的编辑调用,不能确定具体增益、接触布局或逆解问题。" + } + ], + "gap_zh": "没有状态日志,评分器未报告 episode_return,故原生累计奖励值留空;视频无法给出精确四元数误差、奖励分项及原生终止高度,不能确定终止条件的数值触发时刻。最终控制脚本的编辑正文未完整记录,无法细分控制根因或证明某个参数不当。所有时间锚点均直接取自视频;未将控制步数或 agent 工作时间换算成视频时间。" + }, + "summary_zh": "评分回合中机器人双手先托住再收拢夹持两个方块,并带块转动,但联合调姿节点未达到累计奖励370的原生验收,最终0/1。视频约3.40 s 起机器人持续向后失衡,视频3.85 s 末帧已大幅后倒,两个方块仍在手边;判为操作中的平衡失败,根因保守归为控制。没有状态日志、原生累计奖励值和完整最终控制脚本,无法精确分解朝向误差或确定失稳参数。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "视频显示双手持块调姿,约3.6–3.85 s 机器人向后倾倒,节点未达成、断点与平衡现象、保守的控制根因均有视频和 log step 151/153/154 支撑,未发现需要修改的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:16:41Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-standard--vjpbwcr/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-standard--vjpbwcr/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "8efcfaad71c36deb2b2dba06ef2b94a15d3a288558aa056019bba481ba855102", + "video_sha256": "de998201c5f5a59b30073a62b059f1131bd05c8c9be2b2379fc93e397e2d7a4a", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard-realtime--3zc2dgg.json b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard-realtime--3zc2dgg.json new file mode 100644 index 0000000000000000000000000000000000000000..cf14fdbd0db6a2517bd82208c313e259ddd55339 --- /dev/null +++ b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard-realtime--3zc2dgg.json @@ -0,0 +1,228 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-cube-i00-standard-realtime--3zc2dgg", + "taskKey": "humanoidbench/manip_cube", + "dag_sha256": "ab5e9b50283b1821da97eeaa6a81c6bcbd6d347d014e9b3e7113b6656d951802", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 113, + "sim_steps": 1209, + "calls": 4096, + "rows": 7916, + "resets": 20, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 2906.5, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 23.827, + "max_lag_s": 0.0027, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 16, + "late_steps": 0, + "requests_refused": 12, + "request_busy_s": 2.009, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-cube-i00-standard-realtime--3zc2dgg/replay.mp4", + "duration_s": 1.2, + "fps": 20.0, + "frames": 24, + "sha256": "abf14d762d16092b8991dde247b65160e87f8836f1f32b36f95bedb7979c328a" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "cubes_reoriented": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.3, + "log_step": null, + "text_zh": "观察:逐帧检查视频 0.00–1.15 s;两块彩色方块落到机器人各自张开的手掌上,视频 0.30 s 两手仍平掌承托,未见手指包握或主动转向悬浮目标;视频 0.90–1.15 s 机器人连同双手向后倾倒。成功参考视频 0.20–0.60 s 可见明显握持及腕臂转动,本次未见相应动作;参考仅用于比较可见线索。" + }, + { + "t": 1.15, + "log_step": 61, + "text_zh": "观察:log step 61 是最后一次重置后标为 graded 的启动,脚本只设置双腿髋、膝、踝目标并关闭反馈,服务状态输出 FINAL 113 True failed 30;inputs/run.json 的评分字段 live_success=0、graded_by_live=1、n_actions=113。该节点须按累计奖励达到370验收,现场判定未达标;没有报告 episode_return 数值,不能把目录 reward=0.0 当成累计奖励。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "托/舀", + "note_zh": "观察:视频 0.10–0.90 s 机器人两只张开的灵巧手分别从下方承接、托住各自方块,手指未形成对向夹持;末段手掌随身体后倒而倾斜。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.80 s 机器人双脚接地,屈膝支撑托块姿态;视频约 0.90 s 起向后失衡,视频 1.15 s 腿、躯干明显下倒。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 370, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "cubes_reoriented", + "t": 0.9, + "log_step": 61, + "where": "首步前", + "evidence_zh": "观察:逐帧检查视频 0.00–1.15 s,机器人双手始终保持张开承托姿态,未见主动抓握或转向目标朝向的调姿动作;视频约 0.90 s 起躯干明显后仰,末段双手随身体后倒而倾斜。log step 61 启动最后评分集,设置 closed=False 且路点仅涉及双腿髋、膝、踝,服务输出 FINAL 113 True failed 30。调姿步骤从未启动,断点位置为首步前;视频 0.90 s 作为后倒事件的导航锚点。" + }, + "unmet": { + "cubes_reoriented": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 0.9, + "log_step": 61, + "evidence_zh": "观察:视频 0.85–1.15 s 机器人在承托双块时持续向后失衡,头、躯干后仰,骨盆下移,双腿逐渐伸向前方;最后一帧仍处于下倒过程,未见躯干落地,不以触地或精确骨盆阈值作为已观察事实。" + }, + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.3, + "log_step": 61, + "evidence_zh": "观察:视频 0.00–1.15 s 机器人双手持续平掌承接、托住各自方块,未见主动抓握、腕臂转动或手内调姿,末段双手仅随躯干后倒而倾斜。log step 61 的最后评分集路点仅设置双腿髋、膝、踝目标,未启动方块调姿;log step 62 的 hands motionless 为 agent 自述,与所见一致,仅作旁证。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "反馈控制", + "log_step": 61, + "evidence_zh": "推断:log step 61 的最终脚本明确设置 closed=False,双腿仅按预设路点改变髋、膝、踝位置目标;log step 33 的控制代码显示该分支整段提交目标,不执行读取IMU并修正目标的反馈循环。结合视频 0.85–1.15 s 逐渐后仰而未纠正的表现,最后一集的开环支撑控制未能维持承托时的平衡;不将 agent 自述的零力矩或延迟抖动认定为已证实物理机制。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "未提供状态日志及原生累计奖励 episode_return,不能量化朝向误差、接触力或判定精确终止阈值。视频仅1.20 s(最后一帧为视频1.15 s),log step 61 报告113个控制步,按任务文本50 Hz对应仿真2.26 s,两者时长不一致且无已核实映射;本标签全部时间来自实际视频帧,不从控制步换算。末帧尚未见躯干触地,服务的 failed 未细分方块或骨盆低高度触发原因。" + }, + "summary_zh": "评分集双掌承接并托住两个方块,但从未启动主动调姿,唯一目标节点最终未达成(0/1),断点位置为首步前,现象为没尝试与平衡。视频约 0.90 s 机器人向后失衡并持续下倒;最终脚本关闭姿态反馈,支持平衡反馈控制不足的归因。原生累计奖励及状态日志缺失,日志控制步与视频时长不能直接对应,精确终止条件不明。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 2, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "fix", + "conf": 7, + "summary_zh": "节点未达、时间、接触均与视频一致;但评分集双手全程静止承托、从未开始调姿(log step 61 脚本只设腿路点,agent 总结写 hands motionless),断点应为首步前并补 没尝试,与 平衡 并列(参照 CASES.md FAIL-13 powerlift)。", + "n_issues": 2, + "n_major": 2 + }, + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 2, + "verdict": "agree", + "conf": 7, + "summary_zh": "节点未达、接触、首步前断点、没尝试与平衡现象均与视频一致(双手全程平掌承托,约0.9 s 起后倒);根因已标为推断,未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:22:21Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-standard-realtime--3zc2dgg/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-standard-realtime--3zc2dgg/04_review_r1" + }, + { + "stage": "reconcile", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-standard-realtime--3zc2dgg/05_reconcile_r1" + }, + { + "stage": "traj_review", + "round": 2, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-standard-realtime--3zc2dgg/04_review_r2" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "4a9b64806e00fd56415f0184c6d7329691fb194e296d97c1bc95058909aa7ac8", + "video_sha256": "abf14d762d16092b8991dde247b65160e87f8836f1f32b36f95bedb7979c328a", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard-realtime--7r8mzwl.json b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard-realtime--7r8mzwl.json new file mode 100644 index 0000000000000000000000000000000000000000..60e775fb3a3383e0edc2a3da990db6e09e6f86f9 --- /dev/null +++ b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard-realtime--7r8mzwl.json @@ -0,0 +1,207 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-cube-i00-standard-realtime--7r8mzwl", + "taskKey": "humanoidbench/manip_cube", + "dag_sha256": "ab5e9b50283b1821da97eeaa6a81c6bcbd6d347d014e9b3e7113b6656d951802", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 74, + "sim_steps": 686, + "calls": 616, + "rows": 7760, + "resets": 5, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3451.0, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 13.636, + "max_lag_s": 0.0113, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 0, + "late_steps": 598, + "requests_refused": 6, + "request_busy_s": 0.444, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.217, + "peak_processes": 9, + "peak_disk_gb": 0.257, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-cube-i00-standard-realtime--7r8mzwl/replay.mp4", + "duration_s": 0.8, + "fps": 20.0, + "frames": 16, + "sha256": "3d0b396eab77dded8a8c4bd82fc77db073a3c7d225d19a661e854cf9a60facff" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "cubes_reoriented": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00 s 两个彩色方块分别位于机器人两只朝上的手掌上方,手掌尚未收拢;视频 0.15–0.50 s 两手收拢并包住各自方块,开始转腕抬臂。" + }, + { + "t": 0.65, + "log_step": null, + "text_zh": "观察:视频 0.60 s 机器人右手(画面左侧)的方块移到手掌边缘;视频 0.65、0.70、0.75 s 方块已脱离右手,连续下坠至腿部附近,右手仍闭合,机器人保持站立。" + }, + { + "t": 0.75, + "log_step": 103, + "text_zh": "观察:log step 103 的机器人实时服务状态记录最后第6集 control_steps=74、done=true、reason=failed,calls=616、control_steps_total=686,与评分器字段一致;评分器 success=0、live_success=0。推断:结合视频 0.00–0.75 s 的持块后掉落过程,单一累计奖励验收节点未达到;不把前五集调试进度计入本集,也不将累计控制步数换算为视频时间。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "托/舀", + "note_zh": "观察:视频 0.00–0.10 s,机器人左右手掌分别从下方托住各自彩色方块,随后手指收拢。" + }, + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "观察:视频 0.20–0.50 s,机器人两只灵巧手分别用弯曲手指和手掌包住各自方块并转腕;视频约 0.65 s 右手(画面左侧)的方块脱出。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.75 s,机器人双脚接地,腿和躯干承重,在双手操作及方块掉落时仍保持站立。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 370, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "cubes_reoriented", + "t": 0.65, + "log_step": 103, + "where": "操作中", + "evidence_zh": "观察:视频 0.20–0.50 s 机器人已收拢双手并转腕抬臂,视频 0.60 s 右手方块处于掌缘,视频 0.65–0.75 s 该方块脱手下坠,未被接回;log step 103 服务记录最后第6集结束且 reason=failed。推断:失败发生在两个方块调姿过程中,首次明确失去方块的导航锚点取视频约 0.65 s。" + }, + "phenomena": [ + { + "cat": "接触", + "item": "没保持住", + "t": 0.65, + "log_step": null, + "evidence_zh": "观察:视频 0.25–0.50 s,机器人右手(画面左侧)收拢并把方块留在手边;视频 0.60 s 方块滑至掌缘,视频 0.65–0.75 s 手与方块分离,方块持续向下落,右手仍呈闭合姿态。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "t": 0.65, + "log_step": 102, + "source_zh": "日志 / 视频", + "evidence_zh": "观察:log step 102 的 agent 控制脚本按控制步数先收拢两手手指,再插值改变臂腕关节目标;手指在收拢后继续保持闭合目标,没有计划释放方块。log step 103 执行该修改后的最后第6集;视频 0.50–0.75 s 机器人转腕抬臂时右手仍闭合,方块却脱离下坠,身体没有翻倒。推断:这次调姿动作未能维持有效持块接触,支持控制层归因;现有证据不足以确定更细的控制子类。" + } + ], + "gap_zh": "观察:inputs 中有本次完整视频、运行日志和成功参考录像,但无状态日志;评分器未报告原生 episode_return,故该指标为 null。推断:手掌遮挡部分接触区域,且缺少接触对、力和物体姿态记录,不能确定失持来自抓取几何、夹持力还是臂腕跟踪;视频只用于可见过程与导航,不用于计算精确累计奖励或低高度终止阈值。", + "unmet": { + "cubes_reoriented": false + } + }, + "summary_zh": "机器人在评分片段中双手托块、收拢手指并开始转腕抬臂,联合调姿节点最终为0/1。视频约 0.65 s,机器人右手(画面左侧)的方块脱手并持续下坠;log step 103 确认最后第6集失败,现象为没保持住,根因支持控制层归因。无状态与接触力记录,原生累计奖励也未报告,无法细分控制原因或判定精确朝向误差。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "视频仅 0.8 s(16 帧),节点 cubes_reoriented 未达成、断点(操作中,约 0.65 s 右手方块脱手下坠)、现象“没保持住”和根因“控制”均与帧一致;log step 103 的服务读数(control_steps=74、reason=failed)与评分器字段相符,未换算 agent 时间。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:20:32Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-standard-realtime--7r8mzwl/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-standard-realtime--7r8mzwl/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "556dbea9560d42128def37628e25dc7597db250cf0ef3069990fba19d0cfb8d3", + "video_sha256": "3d0b396eab77dded8a8c4bd82fc77db073a3c7d225d19a661e854cf9a60facff", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard-realtime--cz4c9km.json b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard-realtime--cz4c9km.json new file mode 100644 index 0000000000000000000000000000000000000000..3c1adae8b9ed18ec618bc7b191798673e21802ef --- /dev/null +++ b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard-realtime--cz4c9km.json @@ -0,0 +1,206 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-cube-i00-standard-realtime--cz4c9km", + "taskKey": "humanoidbench/manip_cube", + "dag_sha256": "ab5e9b50283b1821da97eeaa6a81c6bcbd6d347d014e9b3e7113b6656d951802", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 86, + "sim_steps": 3498, + "calls": 3161, + "rows": 3359, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 939.3, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 69.291, + "max_lag_s": 0.0657, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 347, + "late_steps": 0, + "requests_refused": 45, + "request_busy_s": 2.665, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.389, + "peak_processes": 11, + "peak_disk_gb": 1.008, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-cube-i00-standard-realtime--cz4c9km/replay.mp4", + "duration_s": 0.95, + "fps": 20.0, + "frames": 19, + "sha256": "776bbcbe6b6a0e77447be8aa46296ff5ef7843549d141f78c91972da21f57ceb" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "cubes_reoriented": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00 s,机器人屈膝站立,两只张开的手掌各在一个彩色方块下方;上方有半透明目标方块。视频 0.05–0.45 s,两手手指合拢包住各自方块,未见成功参照视频 0.10–0.60 s 那样明显的双臂转向动作。" + }, + { + "t": 0.6, + "log_step": null, + "text_zh": "观察:逐帧检查视频 0.00–0.90 s(每帧间隔 0.05 s);视频约 0.55 s 起机器人躯干后仰逐渐加重,视频 0.75–0.90 s 身体向后倒低,双手仍在身体前上方,未见方块离手下落。" + }, + { + "t": null, + "log_step": 188, + "text_zh": "观察:log step 186 启动末次控制程序,服务返回 control_steps=86、reason=failed;log step 188 的服务状态确认 episode=51、done=true、resets_left=0。inputs/run.json 的评分字段 live_success=0、replay_success=0、deterministic=1,未报告 episode_return。" + }, + { + "t": null, + "log_step": 188, + "text_zh": "推断:该 DAG 唯一事件节点绑定整集累计奖励达到370;末集服务与评分器均判失败,结合视频中的持块后仰及未见目标调姿完成,节点未达到。此判断不把目录 reward=0.0 当作原生累计奖励。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "观察:视频 0.00 s 两个方块分别位于两掌上方;视频 0.05–0.45 s 左右灵巧手同时合指,各自包住一块方块,视频 0.65–0.85 s 后仰时仍随手保持,未见掉块;不是两臂夹同一物体。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.50 s 机器人双脚接地、双腿承重,在双手合指持块期间维持屈膝站立;视频 0.55–0.90 s 支撑姿态逐渐失稳并向后倒低。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 370, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "cubes_reoriented", + "t": 0.6, + "log_step": 186, + "where": "操作中", + "evidence_zh": "观察:视频约 0.55–0.65 s,机器人已合拢双手持块,躯干开始持续明显后仰;视频 0.75–0.90 s 继续向后倒低,未恢复站立或完成调姿。log step 186、188 确认这次末集以 failed 结束。" + }, + "unmet": { + "cubes_reoriented": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 0.6, + "log_step": null, + "evidence_zh": "观察:视频 0.50–0.90 s,机器人在双手持块时整体不断后仰,骨盆和躯干下沉,末帧已向后倒低;相邻各帧没有恢复支撑姿态。画面未显示方块先掉落,不能把失败写成没保持住。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "log_step": 186, + "source_zh": "日志 / 视频", + "evidence_zh": "推断:log step 186 的末集程序让机器人按控制步数从屈膝姿态渐变到较直腿姿态,并合拢双手;其中 ka_pitch=0、kd_pitch=0、hipstrateg=0、balance_from=9999,未启用这些身体倾斜反馈补偿。log step 188 的本体读数 imu_gravity_x 从末集控制步20的-0.116发展到控制步80的-0.788,视频 0.55–0.90 s 也显示持续后仰,说明这套实际执行的身体控制没有稳住持块姿态。证据支持控制层失败,不能据此独断为某个增益、关节限位或接触设计错误。" + } + ], + "gap_zh": "无状态日志,也未报告原生 episode_return,无法给出累计奖励或精确朝向误差。末集服务计数86步,按说明50 Hz对应仿真1.72 s,但提供的视频总长仅0.95 s(末帧视频0.90 s);未建立可靠映射,所有视频锚点均直接取自所看帧,不把控制步时间换成视频时间。视频末帧已明显后倒,但不能确定具体低高度终止阈值及触发原因。日志支持身体控制失效,尚不足以细分其机械或反馈失稳机制;log step 193–196 的程序修改发生在末集结束之后,不作为这次轨迹的动作。" + }, + "summary_zh": "评分末集中,机器人双手合指各持住一个方块,但“转好两个方块”未达到,最终为0/1。视频约0.60 s 起机器人在持块过程中明显后仰并继续倒低,断点在该步操作中,证据支持身体控制未稳住,不能归因于掉块。缺少原生累计奖励和状态日志,且末集86个控制步与0.95 s视频的时钟映射未确定,因此不标精确积分、朝向或终止阈值。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "标签与视频、日志一致:视频仅0.95 s,末集失败,机器人合指持块并后仰,节点未达成,断点、平衡现象与控制根因均有证据支持,时钟未做转换。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:19:33Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-standard-realtime--cz4c9km/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-standard-realtime--cz4c9km/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "eecc62d4ab765e0a9872cdaf686e80506a1f43847428e1a86550c9669b46d06e", + "video_sha256": "776bbcbe6b6a0e77447be8aa46296ff5ef7843549d141f78c91972da21f57ceb", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard-realtime--pwjzhq2.json b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard-realtime--pwjzhq2.json new file mode 100644 index 0000000000000000000000000000000000000000..c5e5a069e083eda3299e3cd5b335a4beac3a1f49 --- /dev/null +++ b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard-realtime--pwjzhq2.json @@ -0,0 +1,205 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-cube-i00-standard-realtime--pwjzhq2", + "taskKey": "humanoidbench/manip_cube", + "dag_sha256": "ab5e9b50283b1821da97eeaa6a81c6bcbd6d347d014e9b3e7113b6656d951802", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 158, + "sim_steps": 5044, + "calls": 4693, + "rows": 4578, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1820.4, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 100.074, + "max_lag_s": 0.0109, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 1024, + "late_steps": 0, + "requests_refused": 42, + "request_busy_s": 4.162, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.309, + "peak_processes": 12, + "peak_disk_gb": 0.832, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-cube-i00-standard-realtime--pwjzhq2/replay.mp4", + "duration_s": 1.65, + "fps": 20.0, + "frames": 33, + "sha256": "68b6942d709c6e4b32c55f414a2f6b078d3fed9ace9f4ef191bd2052d16757b9" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "cubes_reoriented": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00 s 两个彩色方块位于机器人摊开的双掌上方,悬浮半透明方块显示目标朝向;视频 0.05–1.35 s 双手合拢包住各自方块并转腕,方块随手改变朝向。" + }, + { + "t": 1.45, + "log_step": null, + "text_zh": "观察:视频 1.40–1.45 s 画面左侧手中的黄紫色方块移至手指下缘,视频 1.50–1.60 s 与手分离并下坠到膝部附近;另一只手仍在腰前持块,机器人双脚支撑且未倒下。" + }, + { + "t": null, + "log_step": 195, + "text_zh": "观察:log step 194 启动第 51 集,log step 195 的服务 FINAL_STATUS 报告 episode=51、control_steps=158、done=true、reason=failed、resets_left=0;评分器 success、live_success、replay_success 均为 0,未报告 episode_return。" + }, + { + "t": 1.6, + "log_step": 195, + "text_zh": "推断:结合视频末段的方块脱手下坠与最后一集服务判负,且任务达标会立即结束,该联合累计奖励事件未达到;不把转腕或一帧近似对齐当作节点完成。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "观察:视频 0.00 s 双掌先从下方承托各自的彩色方块,视频 0.05–1.35 s 两只灵巧手手指分别合拢包住一块并随腕转动;视频约 1.45 s 画面左侧方块从该手下缘脱离。承托到抓握为依次发生的接触。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–1.60 s 机器人双脚与地面接触,屈膝站立,在双手转块期间承重;末帧仍未倒下。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 370, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "cubes_reoriented", + "t": 1.45, + "log_step": 195, + "where": "操作中", + "evidence_zh": "观察:机器人正在双手转块,视频约 1.45 s 画面左侧方块从合拢手指下缘脱出,视频 1.50–1.60 s 持续下坠而未被接回,另一块仍在腰前手中;log step 195 确认最后一集以 failed 结束。" + }, + "unmet": { + "cubes_reoriented": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "没保持住", + "t": 1.45, + "log_step": 195, + "evidence_zh": "观察:视频 0.10–1.35 s 画面左侧方块处于合拢手指间并随手转动,视频 1.45–1.60 s 从手指下方滑出、与手分离并下坠;机器人仍站立,未见计划放置或张手释放。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "推断:log step 194 的最后一集控制器启动参数采用 CUBE_HAND_FLEX=5.0、双腕偏置均为 1.0、CUBE_WRIST_RAMP=80;视频 0.10–1.35 s 可见其合手与转腕执行,但视频 1.45–1.60 s 转动中的一块仍滑出且未被控制住,说明该手指与转腕控制未维持住持块接触。现有证据不能进一步区分接触布局、实际夹持力或反馈跟踪问题,归因止于控制。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "没有状态日志、接触力或方块位姿记录,且给定日志未呈现控制器完整源码,无法确认具体失稳机制;评分器未报告原生 episode_return,不能将 outcome.reward=0 当作累计奖励。log step 195 的最后一集有 158 个控制步,任务文字给出的控制周期为 0.02 s,而所供视频全长仅 1.65 s,二者未建立可靠映射;所有视频锚点均直接取自逐帧查看,未将控制步或 agent 工作时间换算为视频时间。" + }, + "summary_zh": "最后评分的第 51 集中,机器人双手合拢持块并转腕,但“转好两个方块”未通过,最终为 0/1 节点。视频约 1.45 s 画面左侧方块从手中滑出,视频 1.60 s 已下坠到膝部附近,机器人仍站立;结合最后一集控制参数,失败归于持块转动控制未保持住接触。原生累计奖励、精确接触状态及视频与控制步的时钟映射均不足,未进一步细分控制根因。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "节点未达成、断点(cubes_reoriented 操作中,约视频1.45 s左块脱手)、接触与“没保持住”现象均与视频和log一致;控制根因保守且已声明证据边界,时钟未被换算。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:20:50Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-standard-realtime--pwjzhq2/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-standard-realtime--pwjzhq2/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "670284233d8e73d91a5cdb79ac3b0e801a12cb4ca4c3eeee044ed46e3dd2dd95", + "video_sha256": "68b6942d709c6e4b32c55f414a2f6b078d3fed9ace9f4ef191bd2052d16757b9", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard-realtime--sjrypsn.json b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard-realtime--sjrypsn.json new file mode 100644 index 0000000000000000000000000000000000000000..dae03d02ee0896845c1c94d4388078e438487f2a --- /dev/null +++ b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard-realtime--sjrypsn.json @@ -0,0 +1,200 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-cube-i00-standard-realtime--sjrypsn", + "taskKey": "humanoidbench/manip_cube", + "dag_sha256": "ab5e9b50283b1821da97eeaa6a81c6bcbd6d347d014e9b3e7113b6656d951802", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 500, + "sim_steps": 3852, + "calls": 7042, + "rows": 7012, + "resets": 22, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 456.6, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 76.659, + "max_lag_s": 0.0089, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 151, + "late_steps": 0, + "requests_refused": 23, + "request_busy_s": 3.372, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.199, + "peak_processes": 9, + "peak_disk_gb": 0.527, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-cube-i00-standard-realtime--sjrypsn/replay.mp4", + "duration_s": 5.1, + "fps": 20.0, + "frames": 102, + "sha256": "bdedf98f637fb62f517f267dc08985f7cb42bbca055c412a7e10ebca6ef511fd" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "cubes_reoriented": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 36, + "text_zh": "观察:视频0.00、0.05、0.10、0.20、0.40、0.60 s显示机器人双掌向上,两个彩色方块落在各自掌面附近;随后全片的四分之一秒采样中,机器人主要俯仰调整站姿,双手没有包住方块并将其转向悬浮目标。log step 36以TAG=final运行控制器,服务返回episode=23、control_steps=500、done=True、reason=failed;此前各次重置的试验不属于本次评分集。" + }, + { + "t": 5.05, + "log_step": 38, + "text_zh": "观察:逐帧检查视频4.00–5.05 s,两个方块仍在摊开的掌面上,可见色面方向与开头近似,机器人没有倒地,也没有方块下坠离开手边。agent在log step 38自述最后一集只保持站立,未将方块转向目标;该说法与本集画面及log step 36的运行命令相符。" + }, + { + "t": 5.05, + "log_step": 36, + "text_zh": "推断:评分字段success=0、live_success=0及log step 36最终服务判定failed,结合全片没有目标转向动作,联合调姿节点未达成;不是仅以目录reward=0推断未操作。原生episode_return没有报告,不能从图像精确恢复累计奖励或四元数误差。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "托/舀", + "note_zh": "观察:视频0.10–5.05 s中,机器人左、右摊开的手掌分别从下方托住一个彩色方块,随身体的平衡调整小幅运动;未见手指合拢构成对向夹持或手内重定向。此处记录实际承托,目标调姿本身未启动。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频0.00–5.05 s中,机器人的双脚接地,屈膝的腿与躯干承重;身体前后摆动后仍站立,双掌同时承托方块。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 370, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "cubes_reoriented", + "t": 0.0, + "log_step": 36, + "where": "首步前", + "evidence_zh": "观察:视频0.00–5.05 s中,机器人只用双掌托块并调整站姿,没有启动使两块方块朝目标转向的握持、腕臂旋转或手内调整;log step 36的最后一集调用未设置ARM_转腕参数,最终服务判定failed。视频0.00 s是未启动目标步骤的导航锚点,并非掉块或跌倒时刻。" + }, + "unmet": { + "cubes_reoriented": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 36, + "evidence_zh": "观察:视频0.00、0.10、0.20、0.40、0.60 s以及全片采样和视频4.00–5.05 s的逐帧检查中,机器人双掌一直摊开托块,没有朝目标调姿的动作;log step 38的agent自述也明确说本次评分集未转方块。此前log step 34的转腕测试已被重置,不能把该测试当成本集尝试。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志 / agent 自述", + "evidence_zh": "推断:log step 26的控制器以初始关节目标为基准,只用重心及骨盆IMU修正腿部;log step 34才加入可选ARM_腕部轨迹,并以左右腕相反方向各转0.8 rad测试,服务在该测试的control_steps=145返回failed,agent在log step 38自述该次测试掉了左块,且尚无法实现所需重抓、物体跟踪与腕臂转向控制。log step 36最终运行未启用该腕部轨迹;结合视频0.00–5.05 s只托块保平衡,根因可归于未实现可执行的持块调姿控制,不能细分为某个已证实的逆解、接触几何或反馈参数问题。" + } + ], + "gap_zh": "没有本次状态日志,也没有原生累计episode_return或逐步朝向、接触力记录;本集低分可由最终失败判定确认,但不能量化累计分数或朝向误差。控制根因结合脚本和agent对已重置测试的自述推断,不能核实该测试掉块的具体接触机制,故只标一般控制;测试掉块不作为本集现象。视频时长5.10 s与任务所述500个50 Hz控制步的时长不同,未获得并核对控制步到视频的映射,所有视频锚点直接取帧时间,不由控制步或agent时间换算。" + }, + "summary_zh": "评分采用log step 36的最后一集:视频0.00–5.05 s中机器人保持站立并用双掌承托两个方块,但没有启动目标转向,联合调姿节点最终为0/1。断点为转好两个方块的首步前,现象为没尝试;脚本及此前转腕测试支持一般控制能力不足,不能把重置测试中的掉块算入本集。缺少状态日志和原生累计奖励,无法细化朝向误差、接触机制及控制子类。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "评分集为log step 36的episode 23(500步,failed),视频画面显示双掌一直托块、没有转向动作,节点未达成、断点首步前/没尝试、根因控制(按FAIL-22右侧写法由会话证据推断)均有证据支持,未发现需要改动的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:20:21Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-standard-realtime--sjrypsn/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-standard-realtime--sjrypsn/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "6def72ecc8aa86bf21169f4402090759fd2b80b158205ce9b943c34558cdd1bf", + "video_sha256": "bdedf98f637fb62f517f267dc08985f7cb42bbca055c412a7e10ebca6ef511fd", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard-realtime--v5hkhxy.json b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard-realtime--v5hkhxy.json new file mode 100644 index 0000000000000000000000000000000000000000..f47c067343d4dcc1a487869a22a46b93420ce541 --- /dev/null +++ b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard-realtime--v5hkhxy.json @@ -0,0 +1,208 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-cube-i00-standard-realtime--v5hkhxy", + "taskKey": "humanoidbench/manip_cube", + "dag_sha256": "ab5e9b50283b1821da97eeaa6a81c6bcbd6d347d014e9b3e7113b6656d951802", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 179, + "sim_steps": 4135, + "calls": 31308, + "rows": 31194, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 414.7, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 81.958, + "max_lag_s": 0.0031, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 45, + "late_steps": 0, + "requests_refused": 46, + "request_busy_s": 10.722, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.198, + "peak_processes": 9, + "peak_disk_gb": 0.275, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-cube-i00-standard-realtime--v5hkhxy/replay.mp4", + "duration_s": 1.85, + "fps": 20.0, + "frames": 37, + "sha256": "c4f156951e7b35b43e9c9e22ecaa0b2258d7ace09d42e56c47b8e3939c9770f0" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "cubes_reoriented": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.1, + "log_step": null, + "text_zh": "观察:逐帧检查视频 0.00–1.80 s,两个方块在起始下落、转动后停在各自朝上的平掌上;双手没有合拢包握或执行面向悬浮目标的调姿,视频 1.50–1.80 s 随身体倾倒一起下降。成功参考视频 0.25–5.00 s 可见不同的腕臂转动与持块姿态,仅用作外观对照。" + }, + { + "t": 1.8, + "log_step": 57, + "text_zh": "观察:log step 55–57 将最终控制器用于第51集,服务最终状态为 control_steps=179、done=True、reason=failed、resets_left=0;inputs/run.json 的 live_success 与 replay_success 均为0。控制代码的 balance 分支仅改变踝、髋及膝目标,没有方块调姿动作;agent 在 log step 61 也自述调姿尚未开始。" + }, + { + "t": 1.8, + "log_step": null, + "text_zh": "推断:结合全段可见的平掌托块基线、没有目标调姿动作,以及本集失败记录,累计验收节点未达到;不从目录奖励0推定具体累计奖励数值。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "托/舀", + "note_zh": "观察:视频 0.10–1.50 s,左右灵巧手各以朝上平掌从下方承托一个彩色方块,未见手指对向夹持;这是平衡基线中的实际托块接触,尚未开始目标调姿。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–1.50 s,机器人双脚接地,腿与躯干承重维持屈膝站姿;视频 1.50–1.80 s 支撑失稳,身体俯倒下降。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 370, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "cubes_reoriented", + "t": 1.6, + "log_step": 56, + "where": "首步前", + "evidence_zh": "观察:视频 0.00–1.50 s 机器人仅平掌托着两个方块维持站姿,没有启动目标调姿;视频约 1.60 s 起身体倾斜迅速增大,到视频 1.80 s 已明显俯倒、双手下降。log step 55–57 确认最后第51集执行 balance 基线后以 failed 结束;调姿首步尚未开始。" + }, + "unmet": { + "cubes_reoriented": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0, + "log_step": 55, + "evidence_zh": "观察:视频 0.00–1.80 s 双手保持平掌托块,没有将两个方块主动转向目标。log step 30、32、38–42、55 给出的最终代码沿用 balance 分支,只对站立关节做反馈修正,腕指与臂目标沿用初态;log step 56 启动的也是 balance 模式。" + }, + { + "cat": "全身", + "item": "平衡", + "t": 1.6, + "log_step": 56, + "evidence_zh": "观察:视频 1.50–1.80 s 机器人上身向前俯倾加剧、骨盆与双手快速降低、双脚支撑姿态被破坏,片尾仍未恢复。log step 56 的最终集输出中 gravity_x 从 control step 162 的0.297增至 control step 177 的0.619,随后服务在 control step 179 报 failed;这些是服务步数,未换算成视频时刻。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "反馈控制", + "log_step": 56, + "source_zh": "日志 / 视频", + "evidence_zh": "推断:站立反馈未能稳定机器人,阻止了后续调姿。log step 30 的 balance 代码依据骨盆重力方向与角速度修正关节目标,log step 55–56 的最终集使用 kp=0.2、kd=0.5、kk=1.5;最终集服务反馈在 control step 162–177 显示前倾量及俯仰角速度持续增大,最后报 failed,视频 1.50–1.80 s 也显示俯倒未获纠正。不据此断言某个增益数值或执行器能力是唯一原因。" + } + ], + "gap_zh": "没有状态日志,也未报告 episode_return 数值,无法恢复精确累计奖励或判定触发终止的具体高度条件。最终集服务记录179个控制步,按说明每步0.02 s应为仿真3.58 s,而视频只有1.85 s,缺少两者逐帧对应关系;全部视频时间直接取所检查帧,不将服务步数换算为视频时间。agent 在 log step 61 自述“向后倒”,但本次视频末段显示向前俯倒,且 log step 56 的最终集 gravity_x 为正;未采用其倒地方向自述。" + }, + "summary_zh": "受评分第51集仅以双掌分别托块并运行站立平衡控制,未启动目标调姿,节点最终0/1。视频约1.60 s起明显向前俯倒且未恢复,断点在调姿首步前,根因依据最终集控制代码与反馈归为反馈控制未能稳定身体。没有状态日志或原生累计奖励数值,且服务步数与短视频的时间映射未明。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "标注与视频和日志一致:节点未达成、断点在调姿首步前、视频约1.6 s起向前俯倒(log step 56 的 gx 为正),agent 自述的“向后倒”未被采用;agent 步数未换算成视频时间。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:20:47Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-standard-realtime--v5hkhxy/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-standard-realtime--v5hkhxy/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "52a544b235005392e9ea4b6e92640643b45ef5281e0140d7d62cd2529d109116", + "video_sha256": "c4f156951e7b35b43e9c9e22ecaa0b2258d7ace09d42e56c47b8e3939c9770f0", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard-realtime--yamw2rc.json b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard-realtime--yamw2rc.json new file mode 100644 index 0000000000000000000000000000000000000000..2ff3a563e665abf456bc70853ad18bb52e4b6e58 --- /dev/null +++ b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard-realtime--yamw2rc.json @@ -0,0 +1,208 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-cube-i00-standard-realtime--yamw2rc", + "taskKey": "humanoidbench/manip_cube", + "dag_sha256": "ab5e9b50283b1821da97eeaa6a81c6bcbd6d347d014e9b3e7113b6656d951802", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 39, + "sim_steps": 4486, + "calls": 3007, + "rows": 6402, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 2928.3, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 88.928, + "max_lag_s": 0.0086, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 334, + "late_steps": 0, + "requests_refused": 37, + "request_busy_s": 2.916, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.346, + "peak_processes": 9, + "peak_disk_gb": 0.306, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-cube-i00-standard-realtime--yamw2rc/replay.mp4", + "duration_s": 0.45, + "fps": 20.0, + "frames": 9, + "sha256": "a43ee0857adb24c00823a820ef1d8921f54dba63a0865d0c2507542bd7130b88" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "cubes_reoriented": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.05, + "log_step": null, + "text_zh": "观察:逐帧检查视频 0.00–0.40 s 的全部9帧;机器人两手张开,彩色方块先位于掌面上方并由掌面承托,视频 0.10–0.25 s 随掌面倾斜发生翻滚,未见手指包覆方块。" + }, + { + "t": 0.35, + "log_step": null, + "text_zh": "观察:视频 0.30 s 两个方块滑至掌缘,视频 0.35–0.40 s 均与手部分离并下坠至膝部附近或以下,两手已空,机器人仍站立;未见完成双块目标调姿后的保持。" + }, + { + "t": null, + "log_step": 485, + "text_zh": "观察:机器人服务在 log step 485 报告最后第51集 control_steps=39、done=True、reason=failed、resets_left=0;评分器字段 live_success=0、replay_success=0、deterministic=1。" + }, + { + "t": null, + "log_step": 485, + "text_zh": "推断:该节点按整集累计奖励达到370验收,最后一集的服务失败结果及方块下坠画面共同支持未达标;此前重置前的会话试验、log step 487 的关节运动学误差与“verified”抓取自述均不能给该集计入完成。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "托/舀", + "note_zh": "观察:视频 0.05–0.25 s,机器人左、右灵巧手分别用张开的掌面从下方承托各自彩色方块;之后方块滚向掌缘并掉落,未见形成对向包覆抓握。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.40 s,机器人双脚接地、双腿承重,在双手承托及失去方块期间维持站立。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 370, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "cubes_reoriented", + "t": 0.3, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频 0.15–0.25 s 机器人以倾斜的张开掌面承托两个方块,视频约 0.30 s 方块滑至掌缘,视频 0.35–0.40 s 两块均脱离手部下坠,随后录像结束,未恢复承托或完成调姿;log step 485 的机器人服务确认最后第51集失败。" + }, + "phenomena": [ + { + "cat": "接触", + "item": "没保持住", + "t": 0.3, + "log_step": null, + "evidence_zh": "观察:视频 0.05–0.25 s 方块与机器人掌面接触并受承托,视频 0.30 s 滚至掌缘,视频 0.35–0.40 s 两块均与手部分离下坠;这是已有承托接触后丢失物体,未见计划中的目标处松手。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "抓取和接触设计", + "alt_feasible": "已验证有", + "t": 0.3, + "log_step": 476, + "source_zh": "日志 / 视频", + "evidence_zh": "推断:最终控制脚本在 log step 476 将手指初始目标设为零角度,控制计数小于25时维持张手,25至60时才渐变闭合,没有先确认包覆接触再继续的反馈条件;视频 0.10–0.30 s 掌面已倾斜而手指仍未包住方块,方块沿掌缘滚落,支持承载接触未能约束滚滑这一根因。成功参考 inputs/oracle/replay.mp4 的参考视频 0.05–0.30 s 可见手指先收拢包覆、随后持块转动,参考视频至 5.05 s 未见同样下坠,oracle.json success=1,验证同任务存在包覆持块的可行替代;参考仅用于可行性比较,不作为本集进度证据。" + } + ], + "gap_zh": "没有状态日志或接触力数据,无法精确恢复手指接触对、摩擦及方块终止高度;评分器未报告原生 episode_return,不能给出累计奖励数值。log step 485 的39个控制步乘 log step 8 的 control_dt=0.02 得仿真时长0.78 s,与视频时长0.45 s不一致,缺少可核对映射,所有事件时刻均直接取视频帧,不换算控制步或agent时间;失败原因仅到可见承载接触设计层,不推测具体增益或力参数。", + "unmet": { + "cubes_reoriented": false + } + }, + "summary_zh": "最后第51集只短暂以双掌承托两个方块,未完成“转好两个方块”,最终节点0/1。视频约0.30–0.35 s两块从倾斜且未形成包覆抓握的掌面滑落,机器人仍站立;结合最终脚本延后闭手的安排,归因为抓取和接触设计。没有状态日志及原生累计奖励数值,控制步与视频的时间映射也未能核实。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "标签与视频、日志一致:9帧视频中双掌承托方块后两块滑落,节点未达成,断点、现象(没保持住)与根因(抓取和接触设计,参考视频支持可行替代)有证据支持,时钟处理得当。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:20:04Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-standard-realtime--yamw2rc/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-standard-realtime--yamw2rc/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "7e21b0095960e519d4926b6223f53b9ff82e7d424e08a6e010b854214ac74055", + "video_sha256": "a43ee0857adb24c00823a820ef1d8921f54dba63a0865d0c2507542bd7130b88", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard-realtime--zj3kkq2.json b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard-realtime--zj3kkq2.json new file mode 100644 index 0000000000000000000000000000000000000000..1d37093c5a107d0db15763e54f59c8c973467024 --- /dev/null +++ b/phase2/humanoidbench/manip_cube/trajectories/humanoidbench-manip-cube-i00-standard-realtime--zj3kkq2.json @@ -0,0 +1,189 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-cube-i00-standard-realtime--zj3kkq2", + "taskKey": "humanoidbench/manip_cube", + "dag_sha256": "ab5e9b50283b1821da97eeaa6a81c6bcbd6d347d014e9b3e7113b6656d951802", + "outcome": { + "success": 1, + "reward": 1.0, + "source": "catalog", + "raw": { + "success": 1, + "live_success": 1, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 491, + "sim_steps": 3638, + "calls": 3144, + "rows": 3109, + "resets": 16, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1277.6, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 72.506, + "max_lag_s": 0.032, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 529, + "late_steps": 0, + "requests_refused": 16, + "request_busy_s": 7.308, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.253, + "peak_processes": 11, + "peak_disk_gb": 0.9, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-cube-i00-standard-realtime--zj3kkq2/replay.mp4", + "duration_s": 5.0, + "fps": 20.0, + "frames": 100, + "sha256": "a532a58a2fb3f7e9e3029bb32491feda3380f83e8357d2903e42bd27fae7bc1a" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "cubes_reoriented": { + "method": "visual", + "intervals": [ + [ + 4.9, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00 s 两个彩色方块分别位于机器人张开的左右手掌上方,尚未经历调姿;悬浮半透明方块给出目标朝向。" + }, + { + "t": 0.4, + "log_step": null, + "text_zh": "观察:视频 0.15–0.40 s 机器人两只灵巧手的手指逐渐从方块周围合拢;视频 0.60–1.40 s 两个方块随各自手腕和手臂一起转动,近侧左手方块的可见色面由黄绿转为红粉,远侧右手方块随后大部分被合拢的手指遮挡。" + }, + { + "t": 4.95, + "log_step": null, + "text_zh": "观察:逐帧检查视频 0.00–4.95 s;视频 1.45–4.95 s 机器人保持站立及双手握持姿态,近侧左手方块保持在髋前,远侧右手方块被手指遮挡,仅有少量色面可见,未见方块下坠或机器人倒下;视频 4.00–4.95 s 逐帧检查结果相同。" + }, + { + "t": null, + "log_step": 85, + "text_zh": "观察:log step 84 启动最终控制器;log step 85 读取最终 episode=17 的服务状态为 done=true、reason=success、control_steps=491、sim_time_s=9.82(仿真 9.82 s),与评分器 reward.live_success=1、graded_by_live=1、n_actions=491 一致;此前各次 reset 的试验不作为本段视频的进度。" + }, + { + "t": 4.9, + "log_step": 85, + "text_zh": "推断:依据最终集服务成功记录和 DAG 的原生累计奖励验收规则,联合调姿节点最终达标;实时模式达标即终止,故以已检查的视频末端约 4.9 s 作为首次达标的近似锚点,而非把视频 1.4 s 附近完成主要转动当作达标。缺少逐帧累计奖励及明确的仿真到视频时钟映射,不能精确定位越过370的帧,也未将仿真 9.82 s 直接当作视频时间。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "观察:视频 0.15–0.40 s 机器人左右手的手指分别包住各自方块,视频 0.60–1.40 s 两块随各自腕臂转动;其后保持握持姿态至视频 4.95 s。初段从平掌托块过渡到合手抓握,后段远侧右手的方块受手指遮挡。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–4.95 s 机器人双足着地,腿和躯干承重,在双手抓持和转动方块期间维持站立。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 370, + "direction": "higher_better" + } + ], + "failure": null, + "summary_zh": "最终评分集为 episode=17:机器人双手合拢抓住两个方块,随腕臂转动后维持站立与握持,服务在 log step 85 确认成功,联合调姿节点最终完成1/1;累计达标以视频约 4.9 s 作末端近似锚点。任务设计问题:DAG 记载原生验收为朝向、手块接近及站立控制等奖励累计达到370,不能由成功记录证明终态两块均精确对齐。没有状态日志或逐帧累计奖励,评分器未报告 episode_return 数值,且远侧方块后段受手遮挡,因此保留原生指标数值为空,达标帧与精确朝向仍有证据缺口。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "成功集标注与视频一致:联合调姿节点最终达成,4.9 s 为近似锚点且证据明确说明未把仿真 9.82 s 当视频时间;接触(双臂协作抓握+腿/躯干支撑)符合画面,无失败层。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:20:45Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-standard-realtime--zj3kkq2/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-cube-i00-standard-realtime--zj3kkq2/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "2afa170e34cbd85beab7f07b6b28c79d07fed2e4057313d4d35ded53f2a8c023", + "video_sha256": "a532a58a2fb3f7e9e3029bb32491feda3380f83e8357d2903e42bd27fae7bc1a", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-privileged--2glojxy.json b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-privileged--2glojxy.json new file mode 100644 index 0000000000000000000000000000000000000000..369310284dc9f671b51bd6f95ea08056b0ea9e24 --- /dev/null +++ b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-privileged--2glojxy.json @@ -0,0 +1,243 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-door-i00-privileged--2glojxy", + "taskKey": "humanoidbench/manip_door", + "dag_sha256": "0e2b91c83745073b5462f5a378fa464b0e90842fe7004cf6de2891ede1995c32", + "outcome": { + "success": 0, + "reward": 0.2069931504119996, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 1000, + "final_reward": 0.2069931504119996, + "sim_steps": 1000, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.961, + "peak_processes": 7, + "peak_disk_gb": 0.253, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-door-i00-privileged--2glojxy/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "3cf02461321b845e9c1b58f25fb438ffb57ddcee7481d938432c227daf8f971d" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 3, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "handle_pressed": { + "method": "visual", + "intervals": [ + [ + 0.65, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.65, + "log_step": null, + "text_zh": "观察:视频 0.35–0.65 s,机器人左前臂伸向门把手;视频 0.65–0.85 s,把手由接近水平转为明显向下,机器人双脚仍支撑在地面。约视频 0.65 s 记为压下事件;后续把手姿态变化不撤销该事件。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "压/擦", + "note_zh": "观察:视频 0.5–0.85 s,机器人左前臂及无主动手指的手部末端压动门把手,没有合手夹持。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.5–0.85 s,机器人双脚接触地面,支撑伸臂压把手的身体。" + } + ] + }, + "door_opened": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.2, + "log_step": null, + "text_zh": "观察:视频 0.9–1.9 s,把手已经向下,机器人左臂仍在把手附近调整,但门扇未明显转离初始闭合位置,未出现开口。视频 2.0–6.8 s 以及视频 9.1–10.05 s,门扇与门框的相对位置仍基本不变,未见开门事件。" + }, + { + "t": null, + "log_step": 114, + "text_zh": "观察:log step 113 写入最终生成脚本,log step 114 生成并回放最终交付,两个回放均为 1000 个控制步、相同末态哈希,且 scored=false;这是确定性检查,不是开门完成报告。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "压/擦", + "note_zh": "观察:视频 0.9–1.9 s,机器人左前臂及手部末端抵在向下转动的把手附近并调整,门扇未随之打开;只标可见的按抵,不把脚本计划的勾拉当作已建立的勾接触。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.9–6.3 s,机器人屈膝站在门前,双脚在地面支撑左臂的开门操作。" + } + ] + }, + "doorway_crossed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 6.8, + "log_step": null, + "text_zh": "观察:视频 0–6.8 s,机器人躯干始终处于门前一侧,随后逐渐被门扇遮挡;视频 7.0–10.05 s 门扇仍封住门口,未出现供躯干穿过的开口,也未见机器人到达门的另一侧。打开门这一前置事件始终未达到,因此穿门事件未达到。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "原生奖励分", + "threshold": 600, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "door_opened", + "t": 1.2, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频约 1.2 s,机器人已压下把手并在把手处继续操作,门扇却没有被带开;视频 2.0–6.8 s 的持续操作及视频 9.1–10.05 s 的末段均未修复开门失败。机器人尚未开始穿门,故断点为打开门。" + }, + "unmet": { + "door_opened": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "接触力异常", + "t": 1.2, + "log_step": null, + "evidence_zh": "观察:视频 0.65–0.85 s,左前臂接触使把手明显向下转动;视频 1.1–1.9 s,左臂仍在把手处按抵和调整,门扇却未产生可见的打开运动。视频 2.0–5.8 s 仍维持这一停滞结果,表现为有机构接触但没有把门带开;画面不能给出具体力值。" + }, + { + "cat": "全身", + "item": "平衡", + "t": 6.8, + "log_step": null, + "evidence_zh": "观察:视频 6.4 s 机器人仍屈膝站在门前;视频 6.8–7.2 s 可见下肢倒向地面,从站立支撑变为双腿近水平伸在地面。视频 7.2–10.05 s 双腿姿态基本不变,未恢复站立;躯干被门扇遮挡,具体倒地方式和触地部位不能判定。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "推断:视频 0.65–1.9 s 显示机器人能够到达并压动把手,但接下来的左臂操作未转化为门扇打开。log step 113 的最终脚本以左手及左肘位置误差做雅可比逆解,压下后持续给定相对把手的手部目标,没有按门角进展切换后续动作;log step 114 确认该脚本生成的交付已回放。结合视频 2.0–5.8 s 的停滞,支持开门操作的控制失败;缺乏实际接触力和末端跟踪记录,不能进一步确定是接触几何、施力还是逆解跟踪问题。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "没有逐帧状态日志或接触力记录;视频分辨率为 256×256,把手附近有局部遮挡,视频约 6.8 s 后躯干和左臂大部被门扇挡住,无法确认后段具体接触或是否依靠门扇支承。下肢由站立变为倒地的过程可见,足以记录平衡现象,但不能确定倒地机制、具体触地部位或骨盆是否低于任务终止阈值,控制根因仍不细分。极小门角无法量化,但全片未见 DAG 判据所述门扇转离门框形成开口。评分字段仅报告 final_reward=0.2069931504119996,未报告以原生奖励分计的 episode_return,故该原生指标留空;不把会话测试中的门角或接触力当作评分回合的状态。" + }, + "summary_zh": "机器人在视频约 0.65 s 压下门把手,完成 1/3 个节点,但随后未能把门扇带开,也未穿过门口。失败断点为打开门:视频 1.1–5.8 s 左臂持续操作把手而门扇停滞,支持接触力异常现象及控制层根因,视频约 6.8–7.2 s 又出现可见倒地,末段未恢复站立;log step 114 仅证实最终交付可确定性回放,评分 success=0。缺少评分回合的状态、接触力和跟踪记录,后段躯干被门扇遮挡,无法可靠细分控制原因或倒地机制。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 2, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "fix", + "conf": 7, + "summary_zh": "节点值(压下把手 0.65 s 达到、开门/穿门未达到)与断点 door_opened 经帧核对无误;但视频约 6.4–7.2 s 机器人由站立变为腿部平躺在地面,属于可见的摔倒/失去支撑,label 却称仅被门扇遮挡并明确不追加失衡现象,漏记了平衡现象。", + "n_issues": 1, + "n_major": 1 + }, + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 2, + "verdict": "agree", + "conf": 7, + "summary_zh": "节点值(压下把手约 0.65 s 达到,开门、穿门未达到)、断点 door_opened、倒地的平衡现象与控制根因均与帧一致;未发现需要修改的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:15:39Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-privileged--2glojxy/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-privileged--2glojxy/04_review_r1" + }, + { + "stage": "reconcile", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-privileged--2glojxy/05_reconcile_r1" + }, + { + "stage": "traj_review", + "round": 2, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-privileged--2glojxy/04_review_r2" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "d379346bc6c971e4e57805cd3b1f9ae0b04c59aa124a4d177f606590ecb09f7a", + "video_sha256": "3cf02461321b845e9c1b58f25fb438ffb57ddcee7481d938432c227daf8f971d", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-privileged--3seifdr.json b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-privileged--3seifdr.json new file mode 100644 index 0000000000000000000000000000000000000000..0a7184182f113f183ac5a01283dd2cd2b57fec39 --- /dev/null +++ b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-privileged--3seifdr.json @@ -0,0 +1,190 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-door-i00-privileged--3seifdr", + "taskKey": "humanoidbench/manip_door", + "dag_sha256": "0e2b91c83745073b5462f5a378fa464b0e90842fe7004cf6de2891ede1995c32", + "outcome": { + "success": 0, + "reward": 0.17128802538718313, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 1000, + "final_reward": 0.17128802538718313, + "sim_steps": 1000, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-door-i00-privileged--3seifdr/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "36f4dc54ee545fca5c6599732a1930aa2eae6253a4b6b059e8706b9b647654e0" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 3, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "handle_pressed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00–1.00 s 逐帧及视频 1.5–10.0 s 每约 0.5 s 的画面中,机器人由屈膝站姿略向门前移动,未出现抬臂压把手动作;可见把手没有向下转动。" + }, + { + "t": 10.05, + "log_step": 82, + "text_zh": "观察:视频 9.00–10.05 s 逐帧中机器人维持门前站姿,把手仍无下压变化;log step 72 的生成代码只调整双侧髋、膝、踝目标,逐步重复同一个动作,log step 82 仅修改髋、膝偏置后重新生成最终交付,并重放 1000 个控制步。" + } + ] + }, + "door_opened": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00 s 门扇位于门框内;视频 0.00–1.00 s 逐帧、视频 1.5–10.0 s 每约 0.5 s 及视频 9.00–10.05 s 逐帧均未见门扇向机器人侧转出并形成开口。机器人始终留在门前。" + }, + { + "t": 10.05, + "log_step": 91, + "text_zh": "观察:视频 10.05 s 门扇仍在原位置,机器人未做拉门动作;agent 最终报告(log step 91)明确称交付为恒定站姿轨迹,并承认没有打开门。报告与画面相符,会话中的拉门测试不计入该轨迹。" + } + ] + }, + "doorway_crossed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": 82, + "text_zh": "观察:视频 0.00–10.05 s 中机器人仅在门前略向前移并稳定站立,没有交替迈步穿门,躯干始终处于门的原来一侧;视频 9.00–10.05 s 逐帧确认结束前也未穿过。log step 82 交付的是重复恒定关节目标的站姿轨迹。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "原生奖励分", + "threshold": 600, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "handle_pressed", + "t": 0.0, + "log_step": 82, + "where": "首步前", + "evidence_zh": "观察:视频 0.00–10.05 s 只有站姿调整和小幅前移,没有启动压把手动作,三个语义步骤均未完成。log step 72 的基线代码重复恒定动作,log step 82 修改腿部偏置后保存、重放的最终文件仍是这一基线,log step 85 确认为 1000×19 的动作数组;会话中的伸臂、拉门测试未被交付。" + }, + "unmet": { + "door_opened": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 82, + "evidence_zh": "观察:视频 0.00–1.00 s 逐帧及视频 1.5–10.05 s 检查中,机器人保持站立并略向门前漂移,没有实施压把手、拉门动作;log step 82 交付恒定站姿动作,log step 91 的最终报告也承认这只是站姿基线。评分轨迹中的压把手步骤没有启动。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "推断:agent 未能在会话调试中得到有效的开门控制动作,最后仅交付维持站姿的基线。依据是 log step 65 的伸臂回拉测试中门角随后仍约为零,log step 78 的三次勾拉测试均输出 door≈0、max≈0.010,未使门持续打开;log step 82 随后生成并核验恒定站姿文件,agent 在 log step 91 承认勾住把手和实现行走均未解决。这些测试只支持为何退回基线的根因,不作为评分轨迹里的接触现象;证据不足以细分控制子类。", + "source_zh": "日志" + } + ], + "gap_zh": "视频仅有 256×256 的单一斜侧视角,门扇遮挡部分上肢,且无状态日志,不能确定接触对、微小把手角或门角;但可见把手、门扇和躯干均未出现 DAG 所要求的完成变化,最终交付代码也未包含机构操作。会话测试只用于支持控制根因,无法进一步确认是接触几何、轨迹跟踪还是其他控制细节导致失败。评分器仅报告 final_reward=0.17128802538718313,未提供 episode_return,故原生累计奖励值留空,未将其擅自换算为累计分。" + }, + "summary_zh": "评分轨迹中机器人仅略向门前移动并保持站立,压把手、开门和穿门均未完成,最终节点为 0/3。断点为压把手首步前,现象是没尝试;会话中的伸臂和勾拉测试未形成有效开门动作,之后交付恒定站姿基线,支持控制能力不足这一根因。单一低分辨率视角及缺少状态日志限制了细微接触和具体控制子类的判断,评分器也未报告累计原生奖励。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "三个节点均未达成与画面一致;断点 handle_pressed/首步前/没尝试,与 FAIL-22 先例一致;根因停在「控制」并明确依据会话证据,未把停止当原因。未发现需要修改的内容。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:00:19Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-privileged--3seifdr/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-privileged--3seifdr/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "bd19c2b25d5129e020e17f255376f1e60fd5536da171242b3f37c4f8921d831a", + "video_sha256": "36f4dc54ee545fca5c6599732a1930aa2eae6253a4b6b059e8706b9b647654e0", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-privileged--m7n4qra.json b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-privileged--m7n4qra.json new file mode 100644 index 0000000000000000000000000000000000000000..4fe12e1efdd0f8e31269be993e28180d5291b1e2 --- /dev/null +++ b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-privileged--m7n4qra.json @@ -0,0 +1,237 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-door-i00-privileged--m7n4qra", + "taskKey": "humanoidbench/manip_door", + "dag_sha256": "0e2b91c83745073b5462f5a378fa464b0e90842fe7004cf6de2891ede1995c32", + "outcome": { + "success": 0, + "reward": 0.13096873439398238, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 1000, + "final_reward": 0.13096873439398238, + "sim_steps": 1000, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 1.959, + "peak_processes": 10, + "peak_disk_gb": 0.834, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-door-i00-privileged--m7n4qra/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "3997513a9f29a9aec1bdba9348b59ace5cee34ae9e76b274fe8afd887d6225dc" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 3, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "handle_pressed": { + "method": "visual", + "intervals": [ + [ + 5.2, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 5.2, + "log_step": null, + "text_zh": "观察:视频 2.5、3.5、4.5 s 把手仍接近初始水平方向;视频约 5.0–5.5 s 才开始可辨地向下转动,视频 6.5、8.0 s 下倾逐步增大,视频 9.0–10.05 s 接近竖直向下。首次微小转动受低分辨率影响,取视频约 5.2 s,误差约半秒;这是压下事件,不要求持续满分角度。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "压/擦", + "note_zh": "推断:视频约 5.0 s 起机器人抬起的一侧前臂在门内侧把手处施压,使把手下转;结合 log step 193 的左臂控制代码,将作用端识别为左前臂/肘附近。视频 2.6–4.5 s 只能观察到前臂靠近把手区域,不能据此确认已压动把手。门扇遮住精确接触点,未按抓握标注。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频约 5.0–6.4 s 机器人双脚接触地面、屈膝支撑身体进行把手操作。" + } + ] + }, + "door_opened": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 5.6, + "log_step": null, + "text_zh": "观察:视频 5.0–6.4 s 机器人髋部向后移动、上身俯向门把手区域,但门扇没有随身体向机器人侧转离门框;视频 6.4–10.05 s 把手继续下转,门扇仍停在闭合位置,未出现开口。" + }, + { + "t": 10.05, + "log_step": 206, + "text_zh": "观察:视频 10.05 s 门仍关闭;agent 最终自述“把手约 1.59 rad,但主门铰链仍关闭”(log step 206),与视频所见一致。该自述不用于给视频补写精确门角。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "压/擦", + "note_zh": "推断:视频 5.0–10.05 s 机器人后移时,操作臂仍在门内侧把手处施力并使把手继续下转;结合 log step 193 识别为左前臂。可确认压动把手,无法确认形成了可靠的勾拉接触。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 5.0–6.4 s 双脚在地面支撑,机器人通过屈髋、髋部后移尝试带开门;视频 9.0–10.05 s 双脚仍支撑身体。" + } + ] + }, + "doorway_crossed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 0–10.05 s 机器人始终在门的同一侧,前段靠近门、后段髋部后移并俯身操作把手;门一直未打开,躯干没有穿过门口,亦未启动穿门行走。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "原生奖励分", + "threshold": 600, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "door_opened", + "t": 5.6, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频约 5.0 s 起把手下转与机器人髋部后移几乎同时发生,视频 5.0–6.4 s 上身仍贴近门;把手转动而门扇没有跟随打开,直到视频 10.05 s 仍无开口。机器人在压把手的同时尝试带开门扇,打开门的步骤始终未完成。" + }, + "phenomena": [ + { + "cat": "接触", + "item": "接触力异常", + "t": 5.6, + "log_step": null, + "evidence_zh": "观察:视频 5.4–6.4 s 机器人后移并在把手处继续施力,视频 6.4–10.05 s 把手进一步向下转到近竖直,而门扇始终没有被带开;可见接触只产生了把手相对转动,未产生所需的开门运动。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "抓取和接触设计", + "alt_feasible": "未知", + "evidence_zh": "推断:视频 5.0–10.05 s 的后移只继续压转把手,没有带开门扇,说明所用前臂接触与身体承载布局未建立有效的开门作用。log step 193 的最终候选脚本用固定左手位置目标 [0.659,0.30,1.28] 配合髋关节后移,log step 194 和 log step 206 将此策略说明为最终交付;此前相近策略的会话诊断在 log step 163 记录 door–torso_link 及 door_hatch–left_elbow_link 接触,支持接触布局问题这一推断。这些诊断不是评分回放状态,不据此断言最终轨迹的具体力值;也没有验证可行的替代接触方案。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "无状态日志、无成功示范;视频仅 256×256,门扇和门框遮住内侧前臂的精确接触点,故把手首次下转时间约有半秒误差,开门失效的接触布局归因为结合视频及会话代码的推断,最终回放的接触力和转矩无法直接核实。评分字段只有 final_reward=0.13096873439398238 等结果,未明确报告累计原生 episode_return,故该指标留 null,不擅自将其换算为累计分。", + "unmet": { + "door_opened": false + } + }, + "summary_zh": "观察:评分视频中机器人靠近门并操作把手,约视频 5.2 s 起出现可辨的向下转动,完成压把手事件;门扇未打开,机器人未穿门,节点最终为 1/3。推断:视频约 5.0 s 起压把手与髋部后移几乎同时发生,前臂接触和身体姿态只继续带动把手,未有效带开门扇,断点为打开门;具体接触布局归因参考会话代码与诊断,评分回放没有状态或力日志可核实。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 2, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "fix", + "conf": 7, + "summary_zh": "门未开、断点 door_opened 合理;但 handle_pressed 的起始时间标为视频 3.5 s 与帧不符:把手在 2.5–4.5 s 仍为水平,约 5–5.5 s 才开始下转,偏差约 1.5 s。", + "n_issues": 2, + "n_major": 1 + }, + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 2, + "verdict": "agree", + "conf": 6, + "summary_zh": "第二轮:把手下转起点已改为视频约 5.2 s,与帧一致;门全程关闭、断点 door_opened、接触力异常(门没开到位)及抓取和接触设计根因(带推断标注、alt_feasible 未知)均有依据,未发现新错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:13:56Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-privileged--m7n4qra/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-privileged--m7n4qra/04_review_r1" + }, + { + "stage": "reconcile", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-privileged--m7n4qra/05_reconcile_r1" + }, + { + "stage": "traj_review", + "round": 2, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-privileged--m7n4qra/04_review_r2" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "07f2c45c9d1031b28094fa4ee8293adfcc1936bd632f6cb5a131b01acd913ff6", + "video_sha256": "3997513a9f29a9aec1bdba9348b59ace5cee34ae9e76b274fe8afd887d6225dc", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-privileged--mexyr7z.json b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-privileged--mexyr7z.json new file mode 100644 index 0000000000000000000000000000000000000000..22466484f984c469e379500e242c45653f701558 --- /dev/null +++ b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-privileged--mexyr7z.json @@ -0,0 +1,237 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-door-i00-privileged--mexyr7z", + "taskKey": "humanoidbench/manip_door", + "dag_sha256": "0e2b91c83745073b5462f5a378fa464b0e90842fe7004cf6de2891ede1995c32", + "outcome": { + "success": 0, + "reward": 0.1915368826354997, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 1000, + "final_reward": 0.1915368826354997, + "sim_steps": 1000, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 14.71, + "peak_processes": 22, + "peak_disk_gb": 0.257, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-door-i00-privileged--mexyr7z/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "cbdd561a464facbe05a4d9fb53243e2edf5d08207a813592190fea5216ee0f09" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 3, + "base": 0, + "peak": 2, + "final": 2, + "nodes": { + "handle_pressed": { + "method": "visual", + "intervals": [ + [ + 0.9, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.9, + "log_step": null, + "text_zh": "观察:视频 0.6–0.8 s 门外侧把手近水平,视频 0.9–1.1 s 把手相对门扇明显向下转动;机器人上肢在把手另一侧操作。首次转动时刻约为视频 0.9 s,误差约 0.2 s。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "压/擦", + "note_zh": "推断:左前臂/手部末端从门内侧压动把手;依据上肢下降与外侧把手同步向下转动。H1 无闭合手指,不记抓握。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:机器人双脚接触地面,支撑压把手时的身体。" + } + ] + }, + "door_opened": { + "method": "visual", + "intervals": [ + [ + 2.1, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 2.1, + "log_step": null, + "text_zh": "观察:视频 2.0–2.4 s 机器人屈膝后撤,门扇开始绕铰链转离门框;视频 2.5–3.1 s 开口进一步增大。视频 3.4–10.05 s 门虽回摆至较小开度,仍曾打开,按 event 规则保留达成;时间锚点约为视频 2.1 s。" + }, + { + "t": 2.5, + "log_step": 265, + "text_zh": "推断:视频 2.1–2.6 s 机器人上肢随身体回撤并带开门扇,结合 log step 265 交付脚本的左臂回撤及躯干转动阶段,判断左前臂/末端在把手内侧形成勾带接触;门扇遮挡了接触面的细节。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "勾", + "note_zh": "推断:左前臂/手部末端在把手内侧勾带并回撤,拉动门扇;无手指合拢抓握。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:双脚支撑屈膝、后撤和重新站稳的身体。" + }, + { + "who": "手指/夹爪", + "how": "压/擦", + "note_zh": "推断:后段右肩/上臂贴住门边撑门;这是拉开之后的支撑接触,与左臂勾带依次发生。上肢按手指/夹爪归类。" + } + ] + }, + "doorway_crossed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 3.5, + "log_step": 265, + "text_zh": "观察:视频 3.1–4.5 s 机器人从后仰调整为靠近门边的站姿,双腿移位但躯干仍在原来一侧;视频 4.5–10.05 s 基本停留在门边,未穿过门口。log step 265 交付脚本的第三阶段明确含前移与撑门动作,故按已尝试但未完成标注。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:机器人双脚在原侧地面支撑并调整站位,未走到门另一侧。" + }, + { + "who": "手指/夹爪", + "how": "压/擦", + "note_zh": "推断:前移调整后右肩/上臂与门边相贴,保持门缝;没有主动抓握。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "原生奖励分", + "threshold": 600, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "doorway_crossed", + "t": 3.5, + "log_step": 265, + "where": "操作中", + "evidence_zh": "观察:视频 3.1–4.5 s 机器人开门后调整双腿和身体向门边靠近,随后停留在门前;包括视频 9.0–10.05 s 在内始终没有躯干穿到另一侧。log step 265 交付了含前移阶段的完整轨迹,log step 267 验证其 1000 个控制步可确定性重放;评分器 success=0。" + }, + "unmet": { + "door_opened": true + }, + "phenomena": [ + { + "cat": "全身", + "item": "步态", + "t": 3.5, + "log_step": null, + "evidence_zh": "观察:视频 3.1–4.5 s 机器人调整站位并恢复直立,但前进距离不足以穿门;视频 4.5–10.05 s 双脚与躯干保持在门的原侧,没有持续穿门行进,也没有摔倒。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "log_step": 265, + "evidence_zh": "推断:log step 202 的前移控制搜索用三组腿部/躯干关节路点尝试向门口推进;log step 265 最终交付脚本沿用该路点结构,第三阶段完成三段插值后一直输出固定站姿目标,没有继续推进的步行段。视频 3.1–4.5 s 只有有限站位调整,此后直到视频 10.05 s 都停在原侧,说明交付的腿部/全身动作控制未形成足够的穿门位移;不足以进一步确定轨迹或反馈控制子类。", + "source_zh": "日志" + } + ], + "gap_zh": "观察:inputs/ 内有完整视频和日志,但没有状态日志或成功示范;单一低分辨率视角中门扇遮挡部分上肢,勾带把手及肩部撑门的具体接触面为结合动作前后与交付脚本的推断。评分器仅报告 final_reward=0.1915368826354997,未报告 episode_return,故原生累计奖励填 null;不将该字段擅自换算为累计分,也无法细分穿门控制失败的子类。" + }, + "summary_zh": "观察:机器人在视频约 0.9 s 压下把手、约 2.1 s 拉开门,完成 2/3 个事件节点,随后调整站姿但始终未穿过门口,评分器 success=0。推断:断点为穿过门口,交付的腿部/全身路点控制只产生有限前移并转为固定撑门站姿,穿门步态未完成。观察:缺少状态日志及累计原生奖励,上肢接触细节受门扇遮挡,相关接触方式为上下文推断。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "把手约0.9 s转动、门约2.1 s开启(log 里 step 210 门角0.029 对应约2.1 s)、全程未穿门,节点值、断点(doorway_crossed)与视频一致;log_step 为 steps[].id(265 gen.py、267 回放)均存在,未发现需要修改的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:53:01Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-privileged--mexyr7z/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-privileged--mexyr7z/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "c298dfff7852d2f9fe533370cfe9912a62e7335ef6bab76057c8eb8eb38c4ad7", + "video_sha256": "cbdd561a464facbe05a4d9fb53243e2edf5d08207a813592190fea5216ee0f09", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-privileged--mq5dddp.json b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-privileged--mq5dddp.json new file mode 100644 index 0000000000000000000000000000000000000000..47d3ee2e51a965287e41ee70c5aa2be835ff0df5 --- /dev/null +++ b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-privileged--mq5dddp.json @@ -0,0 +1,224 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-door-i00-privileged--mq5dddp", + "taskKey": "humanoidbench/manip_door", + "dag_sha256": "0e2b91c83745073b5462f5a378fa464b0e90842fe7004cf6de2891ede1995c32", + "outcome": { + "success": 0, + "reward": 0.15547608307574418, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 1000, + "final_reward": 0.15547608307574418, + "sim_steps": 1000, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 6.828, + "peak_processes": 34, + "peak_disk_gb": 0.839, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-door-i00-privileged--mq5dddp/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "87f1c2df8b071e8cce8a32abf2e76fe09b2637fd2a41ec5d012a32879ceba35e" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 3, + "base": 0, + "peak": 2, + "final": 2, + "nodes": { + "handle_pressed": { + "method": "visual", + "intervals": [ + [ + 0.9, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.9, + "log_step": null, + "text_zh": "观察:视频 0.0–0.5 s 把手外露杆接近横向;视频 0.8–1.15 s,机器人靠门侧前臂下移并带动把手杆相对门扇向下转动,随后门扇开始转开。压下事件约在视频 0.9 s 已可辨认,时间精度受低分辨率限制。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "压/擦", + "note_zh": "观察:视频 0.6–1.15 s,机器人靠门侧前臂或手部末端贴靠把手并下压;没有可见的合指夹持。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–1.15 s,机器人双脚接触地面,屈膝支撑身体完成上肢操作。" + } + ] + }, + "door_opened": { + "method": "visual", + "intervals": [ + [ + 0.9, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.9, + "log_step": null, + "text_zh": "观察:视频 0.9–2.8 s,门扇绕铰链向机器人所在一侧转离门框,门扇自由边与门框的相对位置持续改变,形成开口;因此开门事件已发生。" + }, + { + "t": 9.3, + "log_step": 168, + "text_zh": "观察:视频 8.7–9.3 s 门扇明显回摆,至视频 10.05 s 仍有小角度开口。log step 167–168 对最终交付动作的重放检查报告最大门角 0.888 rad、末门角约 0.275 rad,且两次末状态哈希一致;与视频的先开大再回摆一致。此节点是锁存事件,回摆不撤销完成。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "勾", + "note_zh": "推断:视频 0.9–1.5 s,机器人靠门侧弯曲的前臂/手部末端贴着把手向身体侧带动门扇,符合前臂勾拉,未见主动手指包围抓握;门扇随后遮挡接触部位。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.9–1.5 s 可见双脚承重,机器人通过屈膝和身体位移配合拉门。" + } + ] + }, + "doorway_crossed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": 168, + "text_zh": "观察:视频 0.0–1.5 s 机器人在门前拉门,视频 2.1–8.5 s 门扇遮挡大部分身体且没有机器人从门口另一侧出现;视频 9.3–10.05 s 外露的手臂和脚仍位于原来一侧。log step 167–168 明确检查最终交付文件,报告机器人根部世界 x 最大仅 0.272 m、末值约 0.093 m;log step 166 的最终生成代码在前 280 个动作后仅过渡到固定关节目标,未加入穿门行走段。" + }, + { + "t": null, + "log_step": 166, + "text_zh": "推断:结合完整视频中没有躯干通过门口的运动、最终交付动作的构成,以及 log step 168 对同一交付文件的重放检查,穿门事件未发生;未把会话中另行试验的行走动作计入评分轨迹。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "原生奖励分", + "threshold": 600, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "doorway_crossed", + "t": null, + "log_step": 166, + "where": "首步前", + "evidence_zh": "推断:视频 0.0–10.05 s 的评分轨迹完成压把手和开门后,机器人仍留在门前,没有启动躯干穿过门口的行走段。log step 166 的最终生成代码只把前 280 个动作接上固定关节目标,log step 167–168 核对该文件并明确未穿门;断点取穿门步骤启动前,没有可单独定位的失败瞬间。" + }, + "unmet": { + "door_opened": true + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": null, + "log_step": 166, + "evidence_zh": "推断:视频 0.0–10.05 s 仅见机器人在门前操作门和保持姿态,没有穿门步态。log step 166–168 显示最终交付的后段为固定关节目标,交付文件中的根部 x 最大仅 0.272 m;会话中的穿门试验没有编入该轨迹,因此是穿门步骤未启动。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "推断:未交付穿门动作有行走控制失败的会话证据:log step 143 的脚本确实尝试交替抬脚并将身体移向门后,但 log step 146 的同次试验输出显示根部高度由约 1.02 m 下降到 0.573 m,前向位置随后退至 -0.420 m;log step 149 另一次绕门试验也出现身体降低而门角回落。这些是会话试验,不作为评分轨迹的摔倒现象。最终 log step 166 仅保留拉门前缀和固定目标后段,说明尚未形成能稳定穿门的交付动作;证据不足以细分控制子类。", + "source_zh": "日志" + } + ], + "gap_zh": "视频只有低分辨率固定视角,门扇在视频约 1.5–9.0 s 遮挡了大部分身体和把手接触区;无逐帧状态日志,不能精确确认后段接触脱离时刻或受力细节。穿门未发生由前后画面及 log step 166–168 明确针对最终交付文件的检查共同支持;控制根因仅有会话试验支持,无法细分接触设计、轨迹求解或反馈跟踪。评分字段只报告 final_reward=0.15547608307574418 和 success=0,没有直接报告 episode_return,故累计原生奖励值留空,未将归一化含义不明的分数当作累计奖励。" + }, + "summary_zh": "评分轨迹压下了把手并将门扇拉开,两个事件节点完成;门扇在视频 8.7–9.3 s 回摆不撤销开门事件,机器人始终没有穿过门口。断点为穿门步骤启动前:最终交付仅保留拉门和固定关节目标,会话中的行走试验未能形成稳定穿门动作,根因保守标为控制。门扇遮挡了后段接触细节,且无逐帧状态日志或累计原生奖励读数。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "把手压下、开门两节点与视频一致(约 1 s 把手转动,1.4–2.8 s 门扇转开,8.7–9.3 s 回摆),穿门未发生,交付文件 log 166–168 显示无行走段;断点穿门 / 首步前 / 没尝试与根因控制(来自会话试验证据,已标推断并写明缺口)符合 FAIL-22 的处理,未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:31:08Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work_rerun1/runs/humanoidbench-manip-door-i00-privileged--mq5dddp/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work_rerun1/runs/humanoidbench-manip-door-i00-privileged--mq5dddp/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "0517737998350fd11762dcde137320abad703e14a24788adfe28ed9662fa7b2c", + "video_sha256": "87f1c2df8b071e8cce8a32abf2e76fe09b2637fd2a41ec5d012a32879ceba35e", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-privileged--ngfake3.json b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-privileged--ngfake3.json new file mode 100644 index 0000000000000000000000000000000000000000..d2753979b93da44ffc67a0978bc7bd6822e6f15d --- /dev/null +++ b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-privileged--ngfake3.json @@ -0,0 +1,208 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-door-i00-privileged--ngfake3", + "taskKey": "humanoidbench/manip_door", + "dag_sha256": "0e2b91c83745073b5462f5a378fa464b0e90842fe7004cf6de2891ede1995c32", + "outcome": { + "success": 0, + "reward": 0.18753487802183605, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 1000, + "final_reward": 0.18753487802183605, + "sim_steps": 1000, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 7.05, + "peak_processes": 26, + "peak_disk_gb": 0.505, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-door-i00-privileged--ngfake3/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "b9d740e3043e9dbb67c75bb00bf22508027b00a45082c7ba4d9f8b8f8703e447" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 3, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "handle_pressed": { + "method": "visual", + "intervals": [ + [ + 2.95, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 2.95, + "log_step": null, + "text_zh": "观察:视频 2.90–3.00 s 的逐帧裁剪显示,门外侧可见的把手从近水平位置突然向下转动,视频 3.10 s 左右回弹;视频 6.0–10.05 s 又持续下转。压把手为事件节点,回弹不撤销达成;首次时刻约为视频 2.95 s,误差约 0.1 s。" + }, + { + "t": 7.0, + "log_step": 81, + "text_zh": "推断:视频 2.5–3.1 s 及视频 6.0–10.05 s,机器人左前臂在门内侧把手处运动,门外侧联动把手随之下转,说明左上肢实际压动了把手;内侧受力点被门扇部分遮挡。log step 81 的最终生成脚本与 log step 85 的参数更新均以左手末端跟踪 door_hatch_inside,把手动作与视频相符。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "压/擦", + "note_zh": "机器人左前臂/手部末端在门内侧把手处施压,使联动把手下转;H1 未使用闭合手指抓握。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "机器人双脚接触地面支撑前倾身体,操作期间维持站立。" + } + ] + }, + "door_opened": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 3.0, + "log_step": null, + "text_zh": "观察:视频 0–10.05 s 的全程采样及视频 2.5–3.3 s、视频 4.0–6.0 s、视频 7.0–10.05 s 的密集帧中,把手虽转动,门扇仍贴在闭合门框位置,没有可见的向机器人侧转开或通行开口。" + }, + { + "t": null, + "log_step": 85, + "text_zh": "观察:log step 81 的最终生成脚本将 t_pull、t_unbow、t_pnorm 都设为 10**9,1000 步期间只执行站立前倾和左臂跟踪把手;log step 85 只将 t_press 改为 300、hand_gain 改为 1.0,仍禁用拉门。log step 86 保存该版本,log step 97 再生成并重放得到相同 state_sha;这些交付记录与视频中没有拉门阶段相符。" + } + ] + }, + "doorway_crossed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 0–10.05 s 机器人躯干一直在门前同一侧,双脚仅随站立姿态调整,没有经门口走到另一侧;视频 9.0–10.05 s 的密集末段帧仍然如此。门未打开,穿门前置事件也未达到。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "原生奖励分", + "threshold": 600, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "door_opened", + "t": null, + "log_step": 85, + "where": "首步前", + "evidence_zh": "观察:视频 2.95 s 已压动把手,此后直到视频 10.05 s,机器人仍在门前操作把手,未启动将门朝身体拉开的动作。log step 81、85 的最终脚本把拉门阶段置于第 10**9 个控制步,log step 86、97 实际保存并重放的只有 1000 步,因此断点为未启动的打开门步骤,不能把会话中其他版本的拉门测试当成交付动作。" + }, + "unmet": { + "door_opened": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": null, + "log_step": 85, + "evidence_zh": "观察:视频 0–10.05 s 只有站立、前倾、左臂接近和压动把手,没有拉开门扇的阶段;log step 81、85 将最终控制器的 t_pull 固定为 10**9,log step 86 保存 1000 步动作,开门步骤未开始。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "抓取和接触设计", + "alt_feasible": "未知", + "source_zh": "日志", + "evidence_zh": "推断:会话中 log step 79 的低速摩擦拉门测试记录,左手末端从 [0.689,0.348,1.108] 退至 [0.357,0.262,1.182] m,门把手杆仍在约 [0.659,0.300,1.200] m,门角始终接近 0;结合 log step 73 的圆柱把手几何,说明这一前臂贴靠方案没有形成有效带门运动的接触。log step 80 随即因拉门无效改为保留压把手版本,log step 81、85 的交付脚本禁用拉门。这支持接触设计困难导致最终轨迹缺少开门动作;没有已验证成功的替代接触方案。" + } + ], + "gap_zh": "没有状态日志或成功示范;视频分辨率为 256×256,门内侧前臂与把手的受力点部分被遮挡,不能从交付视频确定接触力或诊断拉门滑脱。根因依据会话测试及其与最终脚本选择的关联推断,会话中的摔倒、滑脱未标作交付现象;未验证可行替代方案。评分器只报告 final_reward=0.18753487802183605,未报告 DAG 指标 episode_return 的原生累计分,故该 native_metrics.value 留 null。" + }, + "summary_zh": "交付视频中机器人保持站立,并在视频约 2.95 s 首次压动门把手,事件进度最终为 1/3;门扇未打开,躯干未穿过门口。断点为打开门步骤的首步前:会话中的前臂贴靠拉门方案未能带动门扇,最终交付脚本禁用拉门,只保留把手操作。无状态日志,接触受力点部分遮挡,接触设计根因来自会话测试,原生累计奖励未报告。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "视频中门始终闭合、把手在约 2.95 s 后下转,节点进度 1/3、断点 door_opened(首步前/没尝试)与交付脚本(log 81/85 禁用拉门,log 86 把手 0.906 rad、门 -1e-4)一致;根因为会话测试推断且已标注为推断,未发现需要修改的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:59:13Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-privileged--ngfake3/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-privileged--ngfake3/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "15eba84357718405ccd96ed9a33aa51ac13102cf5a258ad92e4fcf226689bd33", + "video_sha256": "b9d740e3043e9dbb67c75bb00bf22508027b00a45082c7ba4d9f8b8f8703e447", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard--93yxyts.json b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard--93yxyts.json new file mode 100644 index 0000000000000000000000000000000000000000..94e63c491247e52c4735574554ef4c9acceb2d95 --- /dev/null +++ b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard--93yxyts.json @@ -0,0 +1,222 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-door-i00-standard--93yxyts", + "taskKey": "humanoidbench/manip_door", + "dag_sha256": "0e2b91c83745073b5462f5a378fa464b0e90842fe7004cf6de2891ede1995c32", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 1000, + "sim_steps": 15737, + "calls": 14654, + "rows": 15737, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3399.8, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.364, + "peak_processes": 8, + "peak_disk_gb": 0.304, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-door-i00-standard--93yxyts/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "c635ac880106fd6b5c0e3e11fbace9c11267550113bb174cb22d2e9f2a0f6a34" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 3, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "handle_pressed": { + "method": "visual", + "intervals": [ + [ + 1.1, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 1.1, + "log_step": null, + "text_zh": "观察:视频 0.9–1.25 s 的逐帧放大图中,机器人靠门的前臂向下运动,银色把手的自由端相对门扇由斜向上转到近水平、略向下;视频约 1.1 s 已出现明确下转,时间误差约 0.1 s。此事件随后即使回弹仍保留。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "压/擦", + "note_zh": "推断:视频 1.05–1.25 s,靠门侧前臂或手部末端的下移伴随把手下转,形成按压;接触点被竖直门框部分遮挡,无可闭合手指,不记抓握。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 1.05–1.25 s,机器人双脚接地支撑压把手时的身体。" + } + ] + }, + "door_opened": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.7, + "log_step": null, + "text_zh": "观察:视频 1.25–2.5 s,机器人留在门侧,前臂停留在把手附近并调整身体;把手虽已下转,深色门扇仍覆盖门框内的开口,未见绕铰链向机器人侧摆开。视频 3.0–10.05 s,机器人逐渐被门扇及门框遮住,但门扇仍未出现明确打开的变化。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察:视频 1.25–2.5 s,机器人靠门侧前臂在门框后操作;门框遮住末端与把手的交界,不能确认拉门阶段是否实际勾住把手、压住门扇或接触门框。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 1.25–2.5 s,机器人双脚接地并调整站姿,支撑门前的操作。" + } + ] + }, + "doorway_crossed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 3.5, + "log_step": null, + "text_zh": "观察:视频 2.5–4.0 s,机器人迈腿朝门扇方向移动,躯干逐渐被门扇和竖直门框遮住;未见门扇打开或躯干穿到门后。视频 4.0–10.05 s,门旁仍露出同侧腿部,位置基本停留在门前,没有从门口另一侧出现;其前置开门事件也未达到。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 2.5–4.0 s,双脚交替接地并向门前移动;视频 4.0–10.05 s,腿部继续支撑在门前,未完成通过。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "原生奖励分", + "threshold": 600, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "door_opened", + "t": 1.7, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频约 1.1 s 把手被压下后,机器人在视频 1.25–2.5 s 继续留在把手附近操作并调整站姿,门扇却没有向机器人侧摆开;视频 2.5 s 后机器人转入迈腿前进,直到视频 10.05 s 仍未修复开门失败。最早未修复的步骤是打开门。" + }, + "unmet": { + "door_opened": false + }, + "phenomena": [], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "推断:log step 615 给出的最终控制脚本以预设关节目标和固定步数切换阶段,反馈只对髋、踝作惯性姿态修正;代码中可见的行进阶段没有门扇开口反馈。视频 1.25–2.5 s 把手操作未带开门扇,视频 2.5–4.0 s 机器人仍继续向未打开的门方向迈步,说明操纵执行没有实现所需门扇运动且未作有效修正;归入控制,现有证据不足以细分接触几何、轨迹或力控问题。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "输入只有固定房间视角视频,没有状态日志或成功参考视频;门框遮挡了拉门阶段前臂与把手的接触点。可判定把手下转而门扇未打开,但无法可靠区分未建立拉门接触、接触丢失或机构受阻,因此现象层留空,不凭门未开推定具体接触异常。根因仅作控制层面的推断;log step 615 的脚本中段在原始日志中被截断,log step 622 的完整阶段叙述属于 agent 自述,不能补作动作完成证据。评分器未报告 episode_return,不能把目录 reward=0.0 当成累计原生奖励;未建立控制步与视频时钟映射,所有视频时间均来自检查过的画面。" + }, + "summary_zh": "机器人在视频约 1.1 s 压下门把手,完成 1/3 个节点;随后未能带开门扇,继续迈腿也未穿过门口,断点为打开门。视频 10.05 s 机器人仍站在门前;预设关节控制未使门扇产生所需运动,控制根因属推断,拉门接触被门框遮挡使具体失败现象无法确认。评分器判定失败,但未提供累计原生奖励值。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "视频中把手约1.1 s转平、门扇全程未开、机器人未过门,节点与断点(door_opened,操作中)与画面一致;现象层留空且有gap_zh,根因标为推断,可接受。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:57:08Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard--93yxyts/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard--93yxyts/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "d1804675aa9cc88ee9ee5e9ce155b92fe232088112d0fc019c642e369b381bd9", + "video_sha256": "c635ac880106fd6b5c0e3e11fbace9c11267550113bb174cb22d2e9f2a0f6a34", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard--cfxzsqf.json b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard--cfxzsqf.json new file mode 100644 index 0000000000000000000000000000000000000000..47ba6b45568447cc014d8663d0759bb7e79fe64f --- /dev/null +++ b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard--cfxzsqf.json @@ -0,0 +1,212 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-door-i00-standard--cfxzsqf", + "taskKey": "humanoidbench/manip_door", + "dag_sha256": "0e2b91c83745073b5462f5a378fa464b0e90842fe7004cf6de2891ede1995c32", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 1000, + "sim_steps": 16223, + "calls": 16123, + "rows": 16223, + "resets": 45, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1612.5, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.187, + "peak_processes": 10, + "peak_disk_gb": 0.245, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-door-i00-standard--cfxzsqf/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "d69635a09ecb4d678d707c8c7c408ae5f7ea9acdd79e5ec8ddf95805762b6158" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 3, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "handle_pressed": { + "method": "visual", + "intervals": [ + [ + 1.5, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 1.5, + "log_step": 80, + "text_zh": "观察:视频 1.3–1.7 s,机器人左侧上肢末端向下运动,门把手从近水平转为明显向下倾斜;视频 2.0–9.6 s 仍可见下压姿态,视频 9.7–10.05 s 把手回弹。按压事件已发生,回弹不撤销该 event 节点;时间取自视频帧,约有 0.2 s 误差。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "压/擦", + "note_zh": "观察:视频 1.3–1.7 s,机器人左侧前臂/球形手部末端压在门把手处并将其压下,没有合手夹持;左右侧依据 log step 60 的最终脚本和 log step 80 的执行记录确认。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–10.05 s,机器人双脚接触地面支撑身体,屈膝前倾以完成左上肢按压;未见倒地。" + } + ] + }, + "door_opened": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.8, + "log_step": 80, + "text_zh": "观察:视频 0.0–10.05 s,门扇相对门框保持原位,没有出现绕铰链打开的开口;视频约 1.8 s 后机器人保持门前按压姿态,没有向后拉开门扇。log step 80 最终调用的是 log step 60 编写的 final.py:下压后反复 c.step(q) 保持关节目标,没有拉门段;log step 77–78 的拉门脚本属于被重置的先前回合。" + } + ] + }, + "doorway_crossed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": 80, + "text_zh": "观察:视频 0.0–10.05 s,机器人始终在门的同一侧,双脚留在门前,躯干没有穿过门口;已逐帧检查视频 9.0–10.05 s,末尾也未穿门或倒地。log step 80 最终脚本只有平衡、前倾、按压和保持,没有穿门动作。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "原生奖励分", + "threshold": 600, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "door_opened", + "t": 1.8, + "log_step": 80, + "where": "首步前", + "evidence_zh": "观察:视频约 1.5 s 压下把手后,机器人在视频约 1.8–10.05 s 保持门前姿态,门扇未转开,也没有启动拉门动作。log step 80 明确执行按压后保持至回合结束的 fallback,服务返回 DONE failed 1000;log step 81 中 agent 自述最终评分局只按住把手。断点为尚未启动的打开门步骤,而非先前调试局的拉门失败。" + }, + "unmet": { + "door_opened": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 1.8, + "log_step": 80, + "evidence_zh": "观察:视频约 1.8–10.05 s,机器人按压把手后没有实施拉开门扇的动作。log step 60 的 final.py 下压后进入保持同一关节目标的循环,log step 80 将它作为最终评分回合运行;开门动作没有开始。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "log_step": 78, + "evidence_zh": "推断:会话中未能形成有效的开门控制,是最终退回仅按压策略的依据。log step 77–78 的脚本实际尝试从把手下方勾住、向后拉并调整身体前倾,均返回 DONE failed 1000;log step 78 的 agent 自算末端位置在拉动段与目标持续相差约 0.10 m。log step 80 的脚本注释将此前勾拉和双手实验未能使门转开列为采用按压保持方案的背景,log step 81 中 agent 也自述这些拉门尝试未成功。以上仅用于解释为何最终未启动拉门,不作为本次视频的接触现象;证据不足以细分控制子类。", + "source_zh": "日志 / agent 自述" + } + ], + "gap_zh": "没有状态日志、门角/把手角和接触力读数;episode_return 未在评分器字段中报告,不能把目录 reward=0.0 当成原生累计奖励。先前会话拉门只有脚本、agent 自算运动学和自述,不能判定究竟是接触几何、施力还是门锁联动导致无效,因此根因仅标控制;这些被重置回合的接触不计入本次评分轨迹。视频只有房间视角,按压动作和门扇不动可见,精确接触点不可辨;所有时间直接取视频帧,未将 agent 时间或控制步数换算为视频时间。" + }, + "summary_zh": "评分回合在视频约 1.5 s 压下门把手,完成 1/3 个节点,门扇未打开,机器人也未穿门且始终站立。最终脚本按压后只保持姿态,断点为打开门步骤启动前,现象为没尝试;会话中反复勾拉控制失败后采用该保底方案,根因仅归入控制。缺少门关节、接触力及原生累计奖励记录,不能进一步确定先前拉门控制失效的机制。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "节点(压把手 1.5 s 达成,门未开、未穿门)、断点(door_opened / 首步前 / 没尝试)与根因(控制,依据会话证据)与视频和日志一致,符合 FAIL-22 的处理方式。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:57:27Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard--cfxzsqf/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard--cfxzsqf/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "6a00ad11760e0df62199e7c554a1bf38ab8963c0768b04573deff44a66b5818d", + "video_sha256": "d69635a09ecb4d678d707c8c7c408ae5f7ea9acdd79e5ec8ddf95805762b6158", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard--dq2avt4.json b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard--dq2avt4.json new file mode 100644 index 0000000000000000000000000000000000000000..b4bec4d34193cf1d250ea4cd01c097e5e2a6064c --- /dev/null +++ b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard--dq2avt4.json @@ -0,0 +1,229 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-door-i00-standard--dq2avt4", + "taskKey": "humanoidbench/manip_door", + "dag_sha256": "0e2b91c83745073b5462f5a378fa464b0e90842fe7004cf6de2891ede1995c32", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 250, + "sim_steps": 17748, + "calls": 362, + "rows": 17673, + "resets": 32, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3500.1, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.331, + "peak_processes": 9, + "peak_disk_gb": 0.858, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-door-i00-standard--dq2avt4/replay.mp4", + "duration_s": 2.6, + "fps": 20.0, + "frames": 52, + "sha256": "c88de3a68002a00bea35973fc163695e65e2d913dcbd281f6b76f46c4845d926" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 3, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "handle_pressed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.45, + "log_step": null, + "text_zh": "观察:视频 0.00–0.65 s,机器人一侧上肢向门边伸出后抬高,末端出现在把手上方;可见的门把手未相对门扇向下转动。" + }, + { + "t": 2.55, + "log_step": null, + "text_zh": "观察:已检查视频全部 52 帧,包括视频 1.60–2.55 s 的末段;把手保持原有朝向,未出现下压事件。门后侧上肢逐渐被门扇遮挡。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.65 s,机器人双脚接触地面,支撑伸臂及身体前倾;末段仍可见脚部位于同侧地面。" + }, + { + "who": "未知", + "how": "未知", + "note_zh": "观察:视频 0.00–0.65 s,上肢末端在门边、把手上方运动,此后逐渐被门扇遮挡;单一房间视角无法确定末端是否触及门后侧把手或门扇,也不能确认具体施力方式。" + } + ] + }, + "door_opened": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.65, + "log_step": 383, + "text_zh": "观察:log step 383 中 agent 重置后提交伸臂、回拉的关节位置序列;视频 0.00–2.55 s 显示机器人向门边前倾并逐渐被遮挡,但门扇相对门框未出现可辨认的转开或新开口。" + }, + { + "t": 2.55, + "log_step": null, + "text_zh": "观察:视频 1.60–2.55 s,门扇仍占据同一门框开口,机器人脚部仍在原侧;未见门扇被拉开。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.65 s,机器人双脚接触地面,支撑伸臂及身体前倾;末段仍可见脚部位于同侧地面。" + }, + { + "who": "未知", + "how": "未知", + "note_zh": "观察:视频 0.00–0.65 s,上肢末端在门边、把手上方运动,此后逐渐被门扇遮挡;单一房间视角无法确定末端是否触及门后侧把手或门扇,也不能确认具体施力方式。" + } + ] + }, + "doorway_crossed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 2.55, + "log_step": null, + "text_zh": "观察:视频 0.00–2.55 s,机器人由门旁前倾到门扇遮挡区,末段脚部仍露在同侧地面,未见躯干通过打开的门口到达另一侧;开门前置事件也未发生。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "原生奖励分", + "threshold": 600, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "handle_pressed", + "t": 0.45, + "log_step": 383, + "where": "操作中", + "evidence_zh": "观察:视频 0.35–0.65 s,机器人伸向门边的末端抬至把手上方,门把手没有下转;后续直至视频 2.55 s 也未见下压完成。log step 383 是最后一次重置后的伸臂回拉序列,先前回合不计入此次评分。推断:首先未修复的是压下门把手这一步。" + }, + "unmet": { + "door_opened": false + }, + "phenomena": [ + { + "cat": "运动", + "item": "位姿不准", + "t": 0.45, + "log_step": 383, + "evidence_zh": "观察:视频 0.35–0.65 s,机器人末端抬到门把手上方,而可见把手仍保持原朝向;随后身体前倾到门扇遮挡区,视频 2.55 s 前未见有效的下压。推断:本次伸臂未形成压下把手所需的末端位置和运动方向。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "log_step": 383, + "evidence_zh": "观察:log step 383 的 agent 脚本从初始关节读数生成固定的 250 行位置目标,右臂先到 [-1.5,-0.43,0.68,0.4],再到 [0.8,-0.4,0.68,1.7],整段一次提交,中途没有观测或接触反馈修正。视频 0.35–0.65 s 可见末端抬到把手上方,把手未被压下。推断:该预设关节运动未实现所需的末端操作,归为控制;缺少实际接触与跟踪数据,不能细分控制子类。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "观察:inputs/ 只有单一房间相机视频,没有状态日志或成功示范;机器人靠近门后,上肢与躯干被门扇遮挡,无法确定实际接触及其受力。log step 383 提交 250 个控制步,log step 7 的服务规格为每步 0.02 s,而所提供视频仅 2.60 s,缺少已验证的逐步视频映射,因此未将控制步换算为视频时间,也不从 agent 自述补全隐藏运动。评分字段未报告原生 episode_return;没有证据把失败进一步归于某个控制子类。" + }, + "summary_zh": "评分回合中,机器人伸臂并前倾到门边,但未见把手下压、门扇打开或躯干穿过门口,三个节点均未完成。断点为压下门把手:可见末端抬至把手上方,预设关节序列未实现有效下压,根因仅归为控制。门扇遮挡了后段上肢,且视频与控制步缺少可靠时间映射,实际接触和更细根因无法确定。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "三个节点均未达成与视频一致:把手始终未见相对门扇下转,门扇未开,躯干未过门口;断点 handle_pressed、位姿不准与控制根因有证据且已诚实标注接触与时间映射的不确定性。未发现需要修改的内容。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:14:00Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard--dq2avt4/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard--dq2avt4/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "3d260fcae96210e69de167e47a102008965b41eb41eefd9fa5548721a5b466e2", + "video_sha256": "c88de3a68002a00bea35973fc163695e65e2d913dcbd281f6b76f46c4845d926", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard--jrwcmjt.json b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard--jrwcmjt.json new file mode 100644 index 0000000000000000000000000000000000000000..b1b9d82fd7fa1fe44e103cca558df8a4476b4801 --- /dev/null +++ b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard--jrwcmjt.json @@ -0,0 +1,229 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-door-i00-standard--jrwcmjt", + "taskKey": "humanoidbench/manip_door", + "dag_sha256": "0e2b91c83745073b5462f5a378fa464b0e90842fe7004cf6de2891ede1995c32", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 127, + "sim_steps": 3441, + "calls": 582, + "rows": 3441, + "resets": 25, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3245.2, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-door-i00-standard--jrwcmjt/replay.mp4", + "duration_s": 1.35, + "fps": 20.0, + "frames": 27, + "sha256": "91b845eb2b16df92884847a32e90683c09944e1ac3b029b19c02588ad2d1019f" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 3, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "handle_pressed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.55, + "log_step": null, + "text_zh": "观察:已逐帧检查视频 0.00–1.30 s;机器人前臂在视频约 0.40–0.85 s 向门把手一侧伸出、抬起,但门扇上外露把手的朝向未见相对门扇向下转动;随后手臂回收,未观察到压下把手事件。" + }, + { + "t": null, + "log_step": 95, + "text_zh": "观察:log step 95 调用最终版本的 attempt(cycles=1),服务返回 ep 26、control_steps=127、done=True、reason=failed;log step 88 定义该版本的站立、伸手、按压、回拉序列。此处只用日志确定评分回合与动作意图,不把脚本中的 press 命名当成把手已经压下。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察:视频约 0.40–0.85 s 机器人单侧前臂伸向把手所在区域;结合 log step 88、95 的右臂命令判断为右臂尝试。门框立柱遮挡末端受力点,不能确认右前臂或手部末端是否接触把手、门扇或门框,也不能确认接触方式。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.90 s 机器人双脚接触地面、屈膝支撑伸臂操作;视频约 0.95–1.30 s 身体持续后仰下坠,支撑失稳。" + } + ] + }, + "door_opened": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.05, + "log_step": null, + "text_zh": "观察:视频 0.00–1.30 s 门扇始终与门框保持原有闭合相对位置,未见绕铰链转开形成通路;视频约 0.85–1.10 s 机器人弯肘回收前臂、身体远离门,门扇没有随之向机器人侧摆开。" + }, + { + "t": null, + "log_step": 95, + "text_zh": "观察:log step 95 的最后回合执行 log step 88 所定义的伸手、按压、回拉程序后以 failed 结束;log step 96 中 agent 自述门从未打开。节点未达由视频中的门扇决定,未把先前重置回合的尝试当成评分回合进展。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察:视频约 0.85–1.10 s 机器人前臂向身体回收;结合 log step 88、95 的开门回拉程序,可判断其尝试带开门扇,但门框遮住先前末端接触点,未能看清与把手或门扇形成勾挂,不能标为已建立勾或抓握。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频约 0.85–1.10 s 机器人双脚仍接触地面支撑回臂动作,随后全身向后倾倒;未见迈步穿门。" + } + ] + }, + "doorway_crossed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.3, + "log_step": null, + "text_zh": "观察:视频 0.00–1.30 s 机器人躯干一直位于门前同一侧,未穿过门口;末段视频 1.10–1.30 s 机器人向远离门的方向后仰下坠,门仍闭合,穿门事件未发生,也未见启动穿门行走。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "原生奖励分", + "threshold": 600, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "handle_pressed", + "t": 0.95, + "log_step": 95, + "where": "操作中", + "evidence_zh": "观察:视频约 0.40–0.85 s 机器人在门把手一侧伸臂操作,未见把手向下转动;视频约 0.95 s 起身体明显后仰,视频 1.10–1.30 s 躯干和骨盆快速下坠,之后没有修复压把手这一步。log step 95 确认最后的第 26 回合在 control_steps=127 返回 failed。推断:最先尝试而未完成、亦未修复的步骤是压下门把手,断点归该步骤的操作中。" + }, + "unmet": { + "door_opened": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 1.0, + "log_step": 95, + "evidence_zh": "观察:视频约 0.95–1.30 s 机器人从屈膝站立变为持续向后倾倒,躯干倾角逐帧增大,骨盆明显降低,末帧身体已大幅斜倒;机器人未在视频结束前恢复稳定支撑。log step 95 的服务状态为 done=True、reason=failed。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志", + "log_step": 82, + "evidence_zh": "推断:log step 82 的 Balance/run_chunk 代码以骨盆重力方向的低通比例反馈调整踝、髋目标,每批执行固定关节目标;log step 88 的固定右臂操作序列由 log step 95 调用。在视频约 0.95–1.30 s,机器人执行该序列时持续后仰下坠,说明该控制实现未能在操作过程中维持全身稳定。没有足够证据进一步区分接触几何、轨迹跟踪或具体控制参数的责任,根因停在控制。" + } + ], + "gap_zh": "输入没有状态日志或成功示范;视频只有单一 256×256 房间视角,门框遮挡末端与门的具体接触点,故不能证实接触反力、压把手的精确角度或勾挂关系。把手外露部分与门扇可见,未出现 DAG 要求的可见转动和开口,因此相应节点按视觉判为未达。log step 95 的最终回合为 127 个控制步,log step 4 的 control_dt=0.02 对应仿真 2.54 s,但视频时长为 1.35 s,未建立精确映射;所有 t 均直接使用所检视频帧时间,没有将 agent 时间或控制步换成视频时间。评分字段没有报告 episode_return,原生指标值留空。" + }, + "summary_zh": "评分对象是 log step 95 的最后第 26 回合;视频逐帧显示机器人在门前伸臂、回臂,未见把手压下或门扇打开,也未穿门,三个节点均未达。机器人在压把手的未完成操作中向后失衡下坠,断点为 handle_pressed,控制实现未能维持操作所需的全身稳定。门框遮挡具体接触点,且无状态日志、原生累计奖励或成功示范,无法进一步确认接触力和控制子类。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "逐帧核对 27 帧:门始终闭合,把手未见下压,机器人约 0.95 s 起后仰倒下,三个节点未达、断点 handle_pressed、平衡/控制标注与证据一致;时钟未换算,validator 通过。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:16:22Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard--jrwcmjt/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard--jrwcmjt/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "89d20ef66023692e0d08d6167baac28c501631a2e7c70ba63688d968ebe2b095", + "video_sha256": "91b845eb2b16df92884847a32e90683c09944e1ac3b029b19c02588ad2d1019f", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard--lfilchg.json b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard--lfilchg.json new file mode 100644 index 0000000000000000000000000000000000000000..0e0d1177693eed71f1cdf103aad8407316a3fcb8 --- /dev/null +++ b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard--lfilchg.json @@ -0,0 +1,238 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-door-i00-standard--lfilchg", + "taskKey": "humanoidbench/manip_door", + "dag_sha256": "0e2b91c83745073b5462f5a378fa464b0e90842fe7004cf6de2891ede1995c32", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 490, + "sim_steps": 16290, + "calls": 11357, + "rows": 16290, + "resets": 33, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3596.1, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-door-i00-standard--lfilchg/replay.mp4", + "duration_s": 5.0, + "fps": 20.0, + "frames": 100, + "sha256": "e9ead43f75255dbfbc43dcd60261db10789286bbbd852a90c480f114f95418e2" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 3, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "handle_pressed": { + "method": "visual", + "intervals": [ + [ + 3.0, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 3.0, + "log_step": null, + "text_zh": "观察:视频 2.80–3.25 s,机器人右前臂末端向把手下压,门扇外侧可见的灰色把手由近水平转为明显向下,视频约 3.0 s 首次确认压转;时间误差约 0.1 s。" + }, + { + "t": 3.6, + "log_step": null, + "text_zh": "观察:视频 3.45–3.75 s,把手回弹至近水平;此节点为已发生即保留的事件,回弹不撤销压下事件。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "压/擦", + "note_zh": "机器人右前臂及无主动手指的手部末端压在门把手处,使把手向下转动;没有手指合拢或夹持。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "机器人双脚接触地面,在下蹲姿态支撑上肢压把手。" + } + ] + }, + "door_opened": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 3.45, + "log_step": null, + "text_zh": "观察:视频 3.0–3.75 s,机器人压转把手后右前臂向自身一侧移动,把手在视频约 3.45 s 回弹,前臂逐渐离开把手;门扇未随动出现可见开口。" + }, + { + "t": 4.95, + "log_step": null, + "text_zh": "观察:已查看视频 0–4.95 s 全段及视频 4.0–4.95 s 的逐帧画面,门扇一直处于闭合门框内,末段机器人抬起右前臂,仍未把门带开。" + }, + { + "t": null, + "log_step": 160, + "text_zh": "观察:log step 157 重置并执行 fastfeedback,log step 158 返回该动作共 390 个控制步;log step 159 又执行 zclear 至 490 个控制步,log step 160 服务状态为 episode=34、control_steps=490、done=False。评分字段 n_actions=490、resets=33 与之吻合;此前回合的开门自述不计入本回合。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "压/擦", + "note_zh": "右前臂末端在把手处施压并向自身一侧移动,随后离开,把手回弹;未见抓握或可靠勾住并带动门扇。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚始终在门前地面支撑,下蹲并调整身体姿态配合上肢操作。" + } + ] + }, + "doorway_crossed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 4.95, + "log_step": null, + "text_zh": "观察:视频 0–4.95 s 机器人躯干始终位于门前同一侧;视频 4.0–4.95 s 双脚仍支撑在原侧地面,没有躯干通过门口的动作,开门前置事件也未发生。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "原生奖励分", + "threshold": 600, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "door_opened", + "t": 3.45, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频约 3.0 s 机器人已压转把手,但视频 3.45–3.75 s 前臂回移、把手回弹,门扇没有被带开;直到视频 4.95 s 仍未修复开门步骤,躯干未穿门。" + }, + "unmet": { + "door_opened": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "没保持住", + "t": 3.45, + "log_step": null, + "evidence_zh": "观察:视频 3.0–3.25 s 机器人右前臂与把手作用使其向下转动,说明接触已建立;视频 3.45–3.75 s 前臂与把手分离、把手回弹,而门扇仍关闭。失去的是已建立的把手操作接触,并非物体曾被手指抓握。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "推断:视频 3.0–3.75 s,机器人能以右前臂压转把手,但随后的末端移动没有维持把手接触并带动门扇,接触在开门完成前丢失;故归为操作接触的控制问题。视频只支持这一层,不能确定是接触几何、目标轨迹还是跟踪误差所致。", + "source_zh": "视频" + } + ], + "gap_zh": "未提供状态日志或成功示范;侧面低分辨率画面中门框遮住部分右前臂接触区,虽可从把手转动及回弹确认接触建立与丢失,仍无法判定接触力、勾挂几何或控制跟踪的具体问题,故根因止于“控制”。评分字段未报告 episode_return,不能把 success=0 或归一化 reward=0 当作原生累计奖励值。所有时间均直接取视频,不将控制步计数或 agent 工作时间换算为视频时间。" + }, + "summary_zh": "评分回合在视频约 3.0 s 压转门把手,完成 1/3 个节点,但门扇未打开,机器人未穿过门口。断点为打开门:视频约 3.45 s 右前臂未保持住把手接触、把手回弹,末段机器人仍在门前下蹲支撑,未见摔倒;控制根因只能确认到大类。缺少状态及接触力日志,也未报告原生累计奖励。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "节点值与时间(把手约 3.0 s 压下、约 3.5 s 回弹,门全程闭合,未穿门)与视频一致;断点 door_opened、现象“没保持住”、根因“控制”(已标明为推断且只到大类)有证据且无时钟换算问题。未发现需要改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:00:40Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard--lfilchg/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard--lfilchg/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "4e11dd42283fb0389efdff6b6c588a41e4dd6de470f75981713c98dcba753d68", + "video_sha256": "e9ead43f75255dbfbc43dcd60261db10789286bbbd852a90c480f114f95418e2", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard--lun3tpt.json b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard--lun3tpt.json new file mode 100644 index 0000000000000000000000000000000000000000..e5cab478ae6b7296309e3a7777d150381d9f6790 --- /dev/null +++ b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard--lun3tpt.json @@ -0,0 +1,214 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-door-i00-standard--lun3tpt", + "taskKey": "humanoidbench/manip_door", + "dag_sha256": "0e2b91c83745073b5462f5a378fa464b0e90842fe7004cf6de2891ede1995c32", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 1000, + "sim_steps": 7008, + "calls": 6972, + "rows": 7008, + "resets": 18, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1506.0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-door-i00-standard--lun3tpt/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "7ba6ce2e4bc5ca6bbb163b5b70d2dd41742d6aa4adfebbaa591ea2f955c55ab9" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 3, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "handle_pressed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.5, + "log_step": null, + "text_zh": "观察:视频 0.0–2.0 s 的逐帧画面中,机器人由屈膝姿态站起并抬动上肢;门把手可见部分始终保持近水平,没有相对门扇向下转动。" + }, + { + "t": 10.05, + "log_step": 116, + "text_zh": "观察:视频 2.0–8.8 s 的密集抽帧及视频 9.0–10.05 s 的逐帧画面中,机器人保持站立、手臂小幅摆动,把手仍未见下压。log step 116 的最终脚本将左臂朝把手侧轻扫;该回合为 episode 19,服务返回 control_steps=1000、reason=failed。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 1.5–10.05 s,机器人的双脚接触地面,支撑其站立和伸臂尝试。" + }, + { + "who": "未知", + "how": "未知", + "note_zh": "观察:视频 1.4–2.0 s、视频 9.0–10.05 s,朝把手方向伸出的上肢末端被门扇和门框遮挡;能确认尝试操作把手,但不能确定末端有无轻触及其接触方式,不据此标抓握、勾或按压。" + } + ] + }, + "door_opened": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.0 s,深色门扇位于门框内,处于闭合位置。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频全程的粗看、视频 0.0–2.0 s 与视频 9.0–10.05 s 的逐帧画面,以及视频 2.0–8.8 s 的密集抽帧均未见门扇相对门框转开或出现通行开口;末帧仍闭合。" + } + ] + }, + "doorway_crossed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 0.0–10.05 s,机器人站在门的同一侧,由屈膝变为站立并摆臂,没有躯干穿过门口的运动;视频 9.0–10.05 s 逐帧仍见双脚留在门前原区域、门扇闭合。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "原生奖励分", + "threshold": 600, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "handle_pressed", + "t": 1.5, + "log_step": 116, + "where": "操作中", + "evidence_zh": "观察:视频约 1.4–2.0 s,机器人已经朝把手一侧抬臂,但把手没有向下转动;之后持续同类小幅摆臂至视频 10.05 s,未修复此步。log step 116 的最终回合脚本保留朝把手侧轻扫的动作,移除了此前回合的回拉阶段;因此断点取正在尝试的压把手步骤,不把先前重置回合中的拉门或摔倒算入本次。" + }, + "unmet": { + "door_opened": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "没建立接触", + "t": 1.5, + "log_step": null, + "evidence_zh": "推断:压把手所需的有效下压接触未建立。视频 1.4–2.0 s 及视频 2.0–8.8 s 的密集画面显示机器人伸臂后仍只在门前小幅摆动;视频 9.0–10.05 s 逐帧检查中,把手仍保持初始近水平姿态、门扇闭合。这里判的是未建立可使把手下转的接触,并不据遮挡画面断言从未有过轻微擦碰。" + } + ], + "root_causes": [ + { + "cat": "认知", + "item": "推理/规划", + "log_step": 116, + "source_zh": "日志", + "evidence_zh": "推断:最终方案没有解决先走近门口这一阶段,直接依赖原地站立前倾和摆臂来操作把手。log step 112 的脚本说明已估计站位与门有间距,但其主循环只设置固定站姿及上肢角度;log step 116 的最终修改进一步采用固定前倾参考和周期性小幅摆臂,没有加入向门迈步的阶段,视频 1.5–10.05 s 也未见接近门口的连续步行。agent 在 log step 119 自述需要接近步行却未构建有效步态;该自述支持接近阶段未解决,不将其声称的臂长和门距当作已验证尺寸。" + } + ], + "gap_zh": "观察:inputs/ 仅有本次视频和会话记录等文件,没有状态日志或成功示范;门框与门扇遮住上肢末端,无法确认微小擦碰、接触力或把手精确关节角,所需有效接触未建立属于结合动作和把手无响应的推断。评分器未报告 episode_return 数值,目录 reward=0.0 不作为原生累计分。log step 116 的回合计数不直接换算到视频时钟,全部视频锚点来自已检查画面。" + }, + "summary_zh": "受评的是 log step 116 启动并结束的 episode 19;视频中机器人保持站立并朝把手侧摆臂,压把手、开门和穿门均未达到,终态为 0/3。断点在压把手:推断其未建立有效下压接触,最终方案没有解决先走近门口的阶段便持续原地摆臂。末端受门框遮挡且无状态日志,不能确定微小接触与精确关节角;原生累计奖励也未报告。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "三个节点均未达成、终态 0/3 与视频一致;断点取压把手、现象为没建立接触(已标推断)、根因推理/规划有 log 支持。未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:59:06Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard--lun3tpt/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard--lun3tpt/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "1455dcf9effd4a7031b2f86543f36a00f739b33516f1b3e4597297de28bdf0a7", + "video_sha256": "7ba6ce2e4bc5ca6bbb163b5b70d2dd41742d6aa4adfebbaa591ea2f955c55ab9", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard--ucjya8k.json b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard--ucjya8k.json new file mode 100644 index 0000000000000000000000000000000000000000..10e1d8d4aadda067c877f50fcd5f65eba7ebed41 --- /dev/null +++ b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard--ucjya8k.json @@ -0,0 +1,221 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-door-i00-standard--ucjya8k", + "taskKey": "humanoidbench/manip_door", + "dag_sha256": "0e2b91c83745073b5462f5a378fa464b0e90842fe7004cf6de2891ede1995c32", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 142, + "sim_steps": 1396, + "calls": 27, + "rows": 1396, + "resets": 10, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3534.7, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-door-i00-standard--ucjya8k/replay.mp4", + "duration_s": 1.5, + "fps": 20.0, + "frames": 30, + "sha256": "a62c7b36082410ea6caf3801634c2be2f61ea5136c0b267d8190245cc9ab3b4e" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 3, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "handle_pressed": { + "method": "visual", + "intervals": [ + [ + 0.55, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.55, + "log_step": null, + "text_zh": "观察:视频 0.00–0.45 s 把手杆朝斜上方;视频 0.50–0.65 s 机器人一侧前臂下移并使把手相对门扇向下转动,视频 0.65–0.90 s 把手杆明显朝下。约视频 1.00 s 把手回弹,不撤销压下事件。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "压/擦", + "note_zh": "观察:视频 0.50–0.90 s,机器人一侧前臂或手部末端抵压门把手使其转下;未见包夹式抓握。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.90 s,机器人双脚接触地面,屈膝支撑上肢操作。" + } + ] + }, + "door_opened": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.9, + "log_step": null, + "text_zh": "观察:逐帧检查视频 0.00–1.45 s,把手曾转下,但门扇与门框之间未出现随开门扩大的开口;视频 0.95–1.15 s 前臂回撤,门扇未被带开。视频 1.20–1.45 s 机器人继续向后倾倒,门仍留在闭合位置;画面整体随相机移动,不能把门在图像中的平移算作开门。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "压/擦", + "note_zh": "观察:视频 0.55–0.90 s,机器人前臂抵在把手区域后回撤;可确认抵压,未见可靠勾住把手并带动门扇,故不记抓握或勾。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.55–1.15 s,机器人双脚着地支撑回撤动作;视频 1.20–1.45 s 身体明显后倒、支撑失稳。" + } + ] + }, + "doorway_crossed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.45, + "log_step": null, + "text_zh": "观察:视频 0.00–1.45 s 机器人躯干一直留在门前同一侧;视频 1.20–1.45 s 身体向远离门口的方向后倒,未见躯干通过门口到达另一侧。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "原生奖励分", + "threshold": 600, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "door_opened", + "t": 1.15, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频 0.55–0.90 s 机器人压动把手后回撤前臂,门扇未被带开;视频约 1.15 s 起躯干持续后倾,至视频 1.45 s 已明显向后失衡,始终没有恢复并打开门。失败发生在拉门动作中,尚未进入穿门动作。" + }, + "unmet": { + "door_opened": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 1.25, + "log_step": null, + "evidence_zh": "观察:视频 1.15–1.45 s,机器人在门前回撤时骨盆后移、躯干越来越向后倾斜;视频 1.35–1.45 s 上身倒向脚后方,双脚未通过迈步重新建立稳定支撑。视频结束前未见恢复站稳。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "log_step": 104, + "evidence_zh": "推断:视频 1.15–1.45 s 回撤过程中机器人持续后倾且未恢复支撑,支持全身动作控制失稳。日志末次实际提交(log step 104)把压把手、回拉和双腿姿态变化预先插值成 142 个关节位置动作一次提交,代码没有在该动作序列中依据新观测修正平衡;log step 104 输出末端姿态估计也显示骨盆后移及降低。此日志仅作为控制方案的会话证据,不把其采样时刻换算为视频时间;不足以进一步区分接触载荷、轨迹或反馈控制的具体责任。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "没有状态日志或成功示范,无法核实把手及门关节角、接触力、骨盆终止高度;评分字段也没有报告 episode_return,故原生累计奖励留空。日志有 10 次重置,末次实际提交为 log step 104 的 episode 11(142 个控制步),其后至 log step 119 均为检查或离线计算,所有 say 为空、没有最终自述;提供的视频仅 1.50 s,未建立与该回合完整采样序列的时间映射,节点和失衡时刻均只取视频观察,不合并其他调试回合的进展。单一房间视角中上肢接触处部分被门框遮挡,可见抵压和回撤,但不能确定回拉是否曾形成可靠的勾持,因此不额外标注接触丢失现象。" + }, + "summary_zh": "观察:视频约 0.55 s 机器人压下门把手,完成 1/3 个节点;门扇未被打开,机器人也未穿过门口。观察:视频约 1.15–1.45 s 机器人在拉门回撤中向后失衡,断点为打开门;推断:动作控制失稳是有证据支持的控制层面原因。缺少状态日志和日志到视频的精确对应,无法判断具体接触载荷、关节角及原生累计奖励。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "逐帧检查视频 0–1.45 s:把手压下、门未开、身体后倾失衡与标注一致;断点、现象与根因证据充分,无需修改。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:57:51Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard--ucjya8k/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard--ucjya8k/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "fbc6e01fee92899c818a41929a51c9a4ab1ff692b5063c20244642d72266c02e", + "video_sha256": "a62c7b36082410ea6caf3801634c2be2f61ea5136c0b267d8190245cc9ab3b4e", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard--xuztkgj.json b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard--xuztkgj.json new file mode 100644 index 0000000000000000000000000000000000000000..2643397beb6d0c757a05ec48ff0f7c25d5b45a5c --- /dev/null +++ b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard--xuztkgj.json @@ -0,0 +1,245 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-door-i00-standard--xuztkgj", + "taskKey": "humanoidbench/manip_door", + "dag_sha256": "0e2b91c83745073b5462f5a378fa464b0e90842fe7004cf6de2891ede1995c32", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 535, + "sim_steps": 7479, + "calls": 7334, + "rows": 7479, + "resets": 30, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 2583.9, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-door-i00-standard--xuztkgj/replay.mp4", + "duration_s": 5.45, + "fps": 20.0, + "frames": 109, + "sha256": "554c406876015ff22b9599f850828d250a50a6c57f4fde062c354a20d7d25743" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 3, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "handle_pressed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.65, + "log_step": null, + "text_zh": "观察:视频 0.00–0.75 s,机器人在门前站立并向把手所在高度伸臂;可见把手保持近水平,没有相对门扇向下转动。" + }, + { + "t": 1.8, + "log_step": null, + "text_zh": "观察:逐帧检查视频 0.80–2.70 s 及视频 2.75–5.40 s,可见把手未出现下压转动;门扇遮挡机器人操作侧的手和把手接触区域,不能确认是否曾短暂触碰。" + }, + { + "t": null, + "log_step": 142, + "text_zh": "观察:log step 142 更新最终驱动的把手目标为 (0.84, 0.42, 1.20) m,并运行伸手、下压、回拉序列;log step 145 重置后重新运行该驱动,log step 148 显示运行结束,log step 149 确认留档驱动一致。这些记录用于识别尝试的步骤,不代替视频中的把手运动。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–5.40 s,机器人双脚接触地面支撑身体进行伸臂操作,未见倒地。" + }, + { + "who": "未知", + "how": "未知", + "note_zh": "观察:视频 0.80–2.70 s,门扇遮挡操作侧上肢与把手之间的区域;无法判定左前臂或手部末端是否碰到把手及具体接触方式,不能记为抓握或压/擦。" + } + ] + }, + "door_opened": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 2.75, + "log_step": null, + "text_zh": "观察:视频 0.00–5.40 s,门扇与门框的相对位置保持不变,没有绕铰链转开形成开口;视频 2.75–5.40 s 逐帧检查也未见门扇被带动。" + }, + { + "t": 4.8, + "log_step": null, + "text_zh": "观察:视频 3.75–5.40 s,机器人仍在门后可见的一侧活动手臂,手臂一度抬高后再放下,门扇仍关闭。" + }, + { + "t": null, + "log_step": 148, + "text_zh": "观察:log step 148 的最终驱动输出处于 pull 阶段;其中 phi 是脚本生成的期望门弧参数,而非服务提供的门角读数,不能作为已经开门的证据。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 2.75–5.40 s,机器人以双脚在地面支撑身体,躯干及手臂继续活动。" + }, + { + "who": "未知", + "how": "未知", + "note_zh": "观察:视频 2.75–5.40 s,门扇遮住操作侧手臂与门把手或门边之间的接触区域;无法确认是否形成勾拉、推压或其他上肢接触。" + } + ] + }, + "doorway_crossed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 5.4, + "log_step": null, + "text_zh": "观察:视频 0.00–5.40 s,门扇一直关闭,机器人躯干没有穿过门框到达另一侧;最后一秒逐帧检查仍见机器人留在原侧,没有穿门事件。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "原生奖励分", + "threshold": 600, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "handle_pressed", + "t": 1.8, + "log_step": 142, + "where": "操作中", + "evidence_zh": "观察:视频 0.00–0.75 s,机器人向门把手所在区域伸臂;视频 0.80–2.70 s 可见把手始终未下转,直至视频 5.40 s 仍没有完成该事件。log step 142、145 的最终驱动包含伸手、下压和回拉,随后 log step 148 的回拉阶段及视频后段也未修复压把手这一步;视频 1.80 s 仅作为失败操作阶段的粗略导航锚点。" + }, + "phenomena": [], + "root_causes": [], + "gap_zh": "视频只有侧面房间视角,门扇持续遮挡操作侧手与把手的接触区域;没有状态日志或接触、力读数,无法区分从未接触、触碰但未正确施力等现象,故现象层留空,上肢接触记未知。inputs/ 中没有 oracle,评分器字段只给成功与否,未报告 episode_return,原生累计奖励值留空。日志含 30 次重置,前期跌倒、定位及步态试验只属会话证据;最终驱动的 phi 不是门角测量,脚本手部世界坐标估计也未经独立状态验证。未建立控制索引与视频时钟的可靠映射,未用其换算视频时间。核对 log step 148:控制索引 k=300–360 与 k=380–400 均标为 pull,不能作为伸手或下压阶段的直接记录;其中 k=360 的 hand.z=0.968 m、des.z=1.200 m,亦非全段仅约 0.04–0.08 m 的跟踪偏差。回拉阶段的手部估计偏差及肩关节限位不能解释更早未完成的压把手步骤,也不与视频后段抬臂建立对应;缺少压把手阶段可验证的手部位置、接触和施力记录,无法区分定位、接触方案与控制执行等根因,因此根因层留空。", + "unmet": { + "door_opened": false + } + }, + "summary_zh": "观察:评分视频中机器人伸臂尝试操作门把手,随后继续活动手臂,但把手未见下转、门扇未打开、躯干未穿过门口,三个节点均未完成,断点为压下门把手。门扇遮挡了关键接触区域,log step 148 的回拉阶段偏差不足以解释压把手失败,故现象层和根因层均留空;评分器未报告原生累计奖励数值。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 2, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "fix", + "conf": 6, + "summary_zh": "节点(三者均未达成)、断点 handle_pressed、现象留空及 gap_zh 与视频一致;问题在于根因「控制」的证据取自回拉阶段(k=380–400),而断点是压把手,且把脚本控制索引与视频后段抬臂隐含对应,但文件自己承认没有映射。", + "n_issues": 1, + "n_major": 1 + }, + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 2, + "verdict": "agree", + "conf": 6, + "summary_zh": "三个节点均未达成、断点 handle_pressed、现象/根因留空并附 gap_zh 与视频一致;门扇遮挡接触区域,上肢接触记未知合理。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:14:17Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard--xuztkgj/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard--xuztkgj/04_review_r1" + }, + { + "stage": "reconcile", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard--xuztkgj/05_reconcile_r1" + }, + { + "stage": "traj_review", + "round": 2, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard--xuztkgj/04_review_r2" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "49bcdab6d739155b424630d4177716dcec142b39338367a0f9cc85df40c34671", + "video_sha256": "554c406876015ff22b9599f850828d250a50a6c57f4fde062c354a20d7d25743", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard-realtime--be9r5ht.json b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard-realtime--be9r5ht.json new file mode 100644 index 0000000000000000000000000000000000000000..d4c9a42c364b0996b91f5d04d928038fa17548f6 --- /dev/null +++ b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard-realtime--be9r5ht.json @@ -0,0 +1,266 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-door-i00-standard-realtime--be9r5ht", + "taskKey": "humanoidbench/manip_door", + "dag_sha256": "0e2b91c83745073b5462f5a378fa464b0e90842fe7004cf6de2891ede1995c32", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 531, + "sim_steps": 38257, + "calls": 122419, + "rows": 37466, + "resets": 46, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3492.7, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 764.303, + "max_lag_s": 0.0922, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 1921, + "late_steps": 0, + "requests_refused": 4, + "request_busy_s": 69.461, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.315, + "peak_processes": 10, + "peak_disk_gb": 0.943, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-door-i00-standard-realtime--be9r5ht/replay.mp4", + "duration_s": 5.4, + "fps": 20.0, + "frames": 108, + "sha256": "ab2713d7cb9a9bdf578ef36c3a3757441b72f113adcb1f5bc3c1c6a183a828d7" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 3, + "base": 0, + "peak": 2, + "final": 2, + "nodes": { + "handle_pressed": { + "method": "visual", + "intervals": [ + [ + 2.65, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 2.65, + "log_step": null, + "text_zh": "观察:视频 2.30–2.55 s 门把手仍近于横向;视频 2.60–2.75 s,机器人左臂末端下移时,把手相对门扇转成向下倾斜、接近竖直,确认发生压下事件。时间为逐帧视觉估计,不代表把手达到某个数值角度。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "压/擦", + "note_zh": "观察:视频 2.60–2.75 s,机器人操作侧前臂及手部末端在把手处向下压,把手随之转动;结合 log step 106、117 的左臂控制代码识别为左臂,未见手指夹握。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 2.30–2.75 s,机器人双脚接触门前地面,屈膝支撑压把手动作。" + } + ] + }, + "door_opened": { + "method": "visual", + "intervals": [ + [ + 2.85, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 2.85, + "log_step": null, + "text_zh": "观察:视频 2.75–3.05 s,机器人在把手处回收操作臂,门扇自由边向机器人一侧转动,与固定门框分离;视频 3.60–4.75 s 门扇已明显偏离初始位置。按 DAG 的开门事件判据计为达到,不要求开到 1 rad,也不因后来失衡撤销。" + }, + { + "t": 5.35, + "log_step": 162, + "text_zh": "观察:视频 5.00–5.35 s 门扇仍在转动,机器人却向后倒下;log step 162 的机器人服务记录最后一局 episode=47、control_steps=531、done=true、reason=failed,与评分字段 n_actions=531 一致;服务没有给出累计原生奖励。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "勾", + "note_zh": "推断:视频 2.75–4.95 s,机器人左臂弯曲的末端在把手附近随身体后移带动门扇向机器人侧打开,属于勾带接触;局部受门扇遮挡,不能确定具体接触面。没有可闭合手指的夹握证据。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 2.85–4.95 s,机器人双脚在门前地面支撑并随拉门调整身体姿势;视频 5.00–5.35 s 身体后倾并失去稳定支撑。" + } + ] + }, + "doorway_crossed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 5.35, + "log_step": null, + "text_zh": "观察:已检查视频 0.00–5.35 s 全段及末秒各帧,机器人躯干一直位于门前原来一侧;门扇打开后机器人向后倾倒,没有躯干越过门口到达另一侧的画面。最终局没有开始穿门的前进行走。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "原生奖励分", + "threshold": 600, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "door_opened", + "t": 5.0, + "log_step": 160, + "where": "操作中", + "evidence_zh": "观察:视频 4.75–5.35 s,机器人仍在把手附近回拉并向后移身,随后躯干持续后仰、骨盆下落,未恢复。log step 160 的最终局脚本采用更强的向后重心目标;log step 161 将此局描述为继续扩大开门,log step 162 确认该局失败。推断:开门事件虽然已经发生,但倒下时机器人仍在继续开门,并未开始穿门,故断点归于 door_opened。" + }, + "unmet": { + "door_opened": true + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 5.0, + "log_step": null, + "evidence_zh": "观察:视频 4.95–5.35 s,机器人从屈膝支撑转为明显后仰,躯干与骨盆向后下方坠落、双腿前伸,末帧仍在倒下,未见恢复稳定支撑;此时仍处于门前拉门动作。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "推断:log step 160 显示最终控制脚本沿用拉门动作并将 com_path 改为较强的向后重心偏移;log step 162 的该局反馈中,control step 450、500 的 imu_gravity_x 从约 -0.025 变为 -0.179,最终服务判该局 failed。视频 4.75–5.35 s 同样显示后倾持续扩大直至倒下,说明开门过程中全身姿态与支撑未被控制住;没有足够证据进一步锁定增益、逆解或接触几何中的某一项。上述 control step 仅作日志依据,不换算成视频时间。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "没有状态日志、把手及门关节角、接触力和累计原生奖励读数,也没有成功示范。视频为单一 256×256 房间视角,操作侧末端与近侧把手部分被门扇遮挡,勾带接触由前后帧联动推断,不能确定接触点或失衡的更细控制原因。log step 162 的最后一局仿真时长为 10.62 s,视频时长为 5.4 s;所有标签时间均直接取自视频帧,没有将仿真秒数或 agent 工作时间当成视频秒数。" + }, + "summary_zh": "观察:最终评分局在视频约 2.65 s 压下把手、约 2.85 s 将门扇向机器人侧带开,达到 2/3 个事件节点,但躯干始终没有穿过门口。推断:机器人继续拉门时的全身控制未维持支撑,视频约 5.0 s 开始向后倒下,断点归于打开门,评分成功值为 0;局部接触和更细控制原因受单一低分辨率视角及缺少状态、力和累计奖励日志限制。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 2, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "fix", + "conf": 6, + "summary_zh": "节点时间与接触(把手约2.65 s、门约2.85 s、未穿门)经帧核对基本成立;但断点 ref 落在已达成且未丢失的 door_opened 上,与 progress 自洽性冲突,应为首个未达成步骤 doorway_crossed。", + "n_issues": 1, + "n_major": 1 + }, + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 2, + "verdict": "agree", + "conf": 6, + "summary_zh": "节点时间(把手约2.65 s、开门约2.85 s、未穿门)、接触和 全身·平衡/控制 经帧核对成立。断点取 door_opened 有争议:该节点已达成,但机器人倒下时仍在拉门、从未开始穿门,按 S14“正在进行的步骤”读法可接受,第1轮已被带证据驳回,无新证据,不再坚持。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:21:49Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard-realtime--be9r5ht/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard-realtime--be9r5ht/04_review_r1" + }, + { + "stage": "reconcile", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard-realtime--be9r5ht/05_reconcile_r1" + }, + { + "stage": "traj_review", + "round": 2, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard-realtime--be9r5ht/04_review_r2" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "cb7a7783e6eff589fbdffdc4379b2177d269b53ba789a85a875e57f80ce9df0e", + "video_sha256": "ab2713d7cb9a9bdf578ef36c3a3757441b72f113adcb1f5bc3c1c6a183a828d7", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard-realtime--fgzzms2.json b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard-realtime--fgzzms2.json new file mode 100644 index 0000000000000000000000000000000000000000..23a94f3a3827ade3cedeb192cff6fe2059af2154 --- /dev/null +++ b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard-realtime--fgzzms2.json @@ -0,0 +1,219 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-door-i00-standard-realtime--fgzzms2", + "taskKey": "humanoidbench/manip_door", + "dag_sha256": "0e2b91c83745073b5462f5a378fa464b0e90842fe7004cf6de2891ede1995c32", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 574, + "sim_steps": 25262, + "calls": 20615, + "rows": 20880, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1669.3, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 504.361, + "max_lag_s": 0.0624, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 5073, + "late_steps": 0, + "requests_refused": 37, + "request_busy_s": 22.924, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.314, + "peak_processes": 10, + "peak_disk_gb": 1.014, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-door-i00-standard-realtime--fgzzms2/replay.mp4", + "duration_s": 5.8, + "fps": 20.0, + "frames": 116, + "sha256": "c6c8ea8c111f584e942b0b36b7a1e21f30b9644282a321f866ab970c0b62292a" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 3, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "handle_pressed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.0, + "log_step": null, + "text_zh": "观察:视频 0.00–5.25 s,机器人双臂朝门把手附近伸出并改变臂姿;可见把手相对门扇保持原朝向,未见向下转动。已检查视频 0.40–1.20 s 的逐帧动作及视频 4.70–5.75 s 的逐帧末段,手端局部被竖直门框遮挡。" + }, + { + "t": 5.5, + "log_step": 239, + "text_zh": "观察:视频 5.40–5.75 s,机器人向后倾倒,双臂离开门框附近,过程中可见把手没有下转。log step 239 的最后回合脚本先伸臂、逐步增加踝部偏置,再执行下压臂姿;服务返回 episode=51、control_steps=574、done=True、reason=failed。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–5.25 s,机器人双脚在门前地面承重,支撑双臂朝把手区域操作;视频 5.40–5.75 s 身体向后失衡。" + }, + { + "who": "未知", + "how": "未知", + "note_zh": "观察:视频 0.75–5.40 s,双臂伸向门框后方,但手端与把手或门扇的交界处被门框遮挡;无法确认哪只手是否接触及其接触方式,不据伸臂动作认定压擦或抓握。" + } + ] + }, + "door_opened": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 5.75, + "log_step": null, + "text_zh": "观察:视频 0.00、1.00、3.00、5.00 s 及视频 5.40–5.75 s 末段,黑色门扇一直保持相对门框的初始闭合位置,未出现绕铰链转开的开口。末段机器人向后倒退,画面取景随之改变,但门扇与门框没有可见相对转动。" + } + ] + }, + "doorway_crossed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 5.75, + "log_step": null, + "text_zh": "观察:视频 0.00–5.25 s,机器人躯干始终在门前原来一侧,双脚未迈步穿门;视频 5.40–5.75 s,躯干朝远离门的方向后仰下落,没有到达门的另一侧。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "原生奖励分", + "threshold": 600, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "handle_pressed", + "t": 5.4, + "log_step": 239, + "where": "操作中", + "evidence_zh": "观察:视频 0.75–5.25 s,机器人正在伸双臂操作把手附近,未见把手压下;视频约 5.40 s 开始明显后仰,至视频 5.75 s 骨盆下落、身体大幅向后倾倒,未恢复操作。log step 239 为最后的第 51 回合,服务返回失败;log step 241 确认该回合已经结束且不能再重置。" + }, + "unmet": { + "door_opened": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 5.5, + "log_step": 239, + "evidence_zh": "观察:视频 5.40–5.75 s,机器人上身持续向后倾斜,双臂远离门,膝部弯曲且骨盆快速降低,末帧仍处于向后倒下的过程;没有恢复稳定站姿。log step 239 的服务状态显示最后回合在 574 个控制步结束并判失败。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志", + "evidence_zh": "推断:log step 239 的最后回合控制程序在双臂保持 PRESS 姿态时,将平衡控制的踝部偏置 ab 从 0 逐步增至 0.9,再在该偏置下改变双臂下压姿态;该程序对应的最后回合中,机器人在视频 5.40–5.75 s 向后失衡并结束。控制程序没有在操作过程中维持住全身稳定;缺少逐步关节与接触力记录,不能进一步确定是反馈律、具体参数还是接触承载布局导致。" + } + ], + "gap_zh": "观察:inputs/ 中有完整运行日志和单视角视频,但没有状态日志或成功示范。门框遮挡手端与把手的接触点,无法确定上肢实际接触方式,也不足以另标没建立接触或没保持住;可见把手及门扇均无完成动作。log step 239 的 574 个控制步按任务规定 50 Hz 对应仿真约 11.48 s,而视频长度为 5.80 s,未验证逐步对应关系,所有标签时间均直接取所检视频帧,不将日志动作时刻换算为视频时刻。评分器未报告 episode_return 数值,原生指标保留 null。" + }, + "summary_zh": "受评的是最后第 51 回合,机器人在门前伸双臂操作把手区域,但未见把手下压、门扇打开或躯干穿门,三个节点均未完成。视频约 5.40–5.75 s,机器人在操作中向后失衡,断点为压下门把手,日志与视频支持归因于操作过程中全身控制未能保持稳定。手端接触点被门框遮挡,且无状态日志,具体接触方式及控制失稳的子类无法确定。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "三个节点均未达成、断点为 handle_pressed(操作中)、末段向后倾倒的平衡现象与停在“控制”的根因均有视频帧和 log step 239 支持;时钟未把 agent/控制步换算成视频时间。未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:16:11Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard-realtime--fgzzms2/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard-realtime--fgzzms2/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "8484d8e6252ca453059b22fd8457e5cce5201ae160b916cc28ad0d3d8dd22225", + "video_sha256": "c6c8ea8c111f584e942b0b36b7a1e21f30b9644282a321f866ab970c0b62292a", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard-realtime--mfmtphq.json b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard-realtime--mfmtphq.json new file mode 100644 index 0000000000000000000000000000000000000000..1bff4ad840bbbb266a1a504bb85f1fca93d8e209 --- /dev/null +++ b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard-realtime--mfmtphq.json @@ -0,0 +1,243 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-door-i00-standard-realtime--mfmtphq", + "taskKey": "humanoidbench/manip_door", + "dag_sha256": "0e2b91c83745073b5462f5a378fa464b0e90842fe7004cf6de2891ede1995c32", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 399, + "sim_steps": 4032, + "calls": 4979, + "rows": 19836, + "resets": 9, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3510.0, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 80.461, + "max_lag_s": 0.0302, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 1240, + "late_steps": 0, + "requests_refused": 4, + "request_busy_s": 2.44, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.379, + "peak_processes": 8, + "peak_disk_gb": 0.264, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-door-i00-standard-realtime--mfmtphq/replay.mp4", + "duration_s": 4.05, + "fps": 20.0, + "frames": 81, + "sha256": "a6277d37a161c8eff48776424c11943ebc85720d5d27bcc3f05bb1d6f06ef6f2" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 3, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "handle_pressed": { + "method": "visual", + "intervals": [ + [ + 0.9, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.9, + "log_step": null, + "text_zh": "观察:视频 0.75–1.25 s,机器人操作侧前臂向下经过门把手,把手由近水平转为明显向下;视频 1.65–1.85 s 把手回弹。按事件节点规则,回弹不撤销已完成的压下事件,首次可辨时刻约为视频 0.9 s。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "压/擦", + "note_zh": "观察:视频 0.85–1.5 s,机器人操作侧前臂/手部末端下压门把手,使把手向下转动;没有手指闭合包夹。日志控制器操作左臂(log step 127)。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0–1.5 s,机器人双脚接触地面,屈膝支撑压把手动作。" + } + ] + }, + "door_opened": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.75, + "log_step": null, + "text_zh": "观察:视频 1.25–2.5 s,机器人操作臂屈肘撤回,把手回到近水平,但门扇仍处在门框内,未出现向机器人侧打开的可见门缝;逐帧查看视频 0–4.0 s,未见门扇转离闭合位置。" + }, + { + "t": 3.75, + "log_step": null, + "text_zh": "观察:视频 3.65–4.0 s,机器人离开把手并明显向后倾倒,门扇仍未打开,开门步骤未补救。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "压/擦", + "note_zh": "观察:视频 1.25–1.5 s,机器人操作侧前臂仍在已下压的把手附近接触,随后屈肘撤回,把手回弹;画面未显示包夹抓握或持续勾住把手并带动门扇。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 1.5–3.6 s,机器人双脚在门前地面支撑上身,随后在视频 3.7–4.0 s 向后失衡。" + } + ] + }, + "doorway_crossed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 4.0, + "log_step": null, + "text_zh": "观察:视频 0–4.0 s,机器人躯干一直留在门前同一侧;末帧向远离门的方向后仰,未穿过门框。门未打开,亦未见启动穿门行走。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "原生奖励分", + "threshold": 600, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "door_opened", + "t": 1.7, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频约 0.9 s 机器人已压下把手;视频 1.5–1.85 s 操作臂开始撤回,把手回弹,门扇未随手臂打开,此后直至视频 4.0 s 都未补救。机器人在视频 3.7–4.0 s 又向后失衡;第一个未完成且未修复的操作步骤是打开门。" + }, + "unmet": { + "door_opened": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "没保持住", + "t": 1.7, + "log_step": null, + "evidence_zh": "推断:视频 0.9–1.5 s 把手已被机器人前臂压下;视频 1.65–1.85 s 操作臂撤离,把手回弹,而门扇没有跟随向外转动,表明压下后未能继续保持带开门扇所需的接触。这里只判断接触脱离,不推定曾有手指抓握。" + }, + { + "cat": "全身", + "item": "平衡", + "t": 3.7, + "log_step": null, + "evidence_zh": "观察:视频 3.65 s 后机器人上身开始向后偏离双脚,视频 3.8–4.0 s 后仰持续加剧、骨盆下降,末帧躯干已接近横斜姿态,未能恢复站立支撑。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "视频", + "evidence_zh": "推断:视频 0.9–1.85 s,机器人能够触及并转动把手,却在操作臂撤回时未能保持开门接触;视频 3.65–4.0 s 又未能稳定门前站姿而后仰塌落。操作执行中的接触维持及身体稳定均失败,支持归入控制;无足够证据细分接触设计、轨迹逆解或反馈参数。" + } + ], + "gap_zh": "输入无状态日志、无成功示范,视频只有低分辨率侧面视角;门边部分遮挡操作手部,无法确定接触脱离的精确几何、是否曾短暂勾住及接触力大小。日志包含十次回合(log step 128 重置到 episode 10),不能将早先回合的诊断或仿真秒直接用作视频时间;所有节点时间均由视频帧判断。run.json 未报告 episode_return 数值,原生累计奖励留空。" + }, + "summary_zh": "观察:机器人在视频约 0.9 s 压下门把手,完成 1/3 个节点;随后操作臂撤回、把手回弹,未打开门,也未穿门。推断:断点为打开门,接触未保持且末段出现向后失衡,根因保留为控制;无状态日志且操作手部部分遮挡,不能细分控制原因,评分器也未报告原生累计奖励。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "视频帧确认:约 0.8–1.4 s 前臂压下把手,随后撤回、把手回弹,门扇始终未开;3.7 s 后向后倾倒。节点值、断点(door_opened,操作中)、现象与根因均与帧一致,无需修改。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:16:53Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard-realtime--mfmtphq/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard-realtime--mfmtphq/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "aefbba98cfed39124286b4897a9233a29f7cd4685e8bc834b3787865da4bb666", + "video_sha256": "a6277d37a161c8eff48776424c11943ebc85720d5d27bcc3f05bb1d6f06ef6f2", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard-realtime--sthgqmg.json b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard-realtime--sthgqmg.json new file mode 100644 index 0000000000000000000000000000000000000000..4657671ce369025113c2aa5e85fed91ccf9085a1 --- /dev/null +++ b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard-realtime--sthgqmg.json @@ -0,0 +1,235 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-door-i00-standard-realtime--sthgqmg", + "taskKey": "humanoidbench/manip_door", + "dag_sha256": "0e2b91c83745073b5462f5a378fa464b0e90842fe7004cf6de2891ede1995c32", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 1000, + "sim_steps": 31403, + "calls": 20286, + "rows": 20086, + "resets": 49, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1366.4, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 627.228, + "max_lag_s": 0.0834, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 12647, + "late_steps": 0, + "requests_refused": 15, + "request_busy_s": 11.784, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.224, + "peak_processes": 9, + "peak_disk_gb": 0.241, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-door-i00-standard-realtime--sthgqmg/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "f8afaa7150f2b44ba165973f8091e6f4a1ce13dbd612c3e60c1cb3cbe613d7ed" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 3, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "handle_pressed": { + "method": "visual", + "intervals": [ + [ + 3.5, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.0–3.3 s,门把手伸杆大致水平;机器人在门前站立并将左臂伸向把手区域。" + }, + { + "t": 3.5, + "log_step": null, + "text_zh": "观察:视频 3.5–4.1 s,门把手伸杆相对门扇明显向下转动;视频 4.5–8.2 s 仍可见下垂把手。按事件节点记首次压下约为视频 3.5 s,时间误差约 0.2 s;视频末段把手回弹不撤销该事件。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "压/擦", + "note_zh": "推断:视频 3.3–4.1 s,机器人左臂末端在把手区域向下运动,同时可见另一侧相连把手下转,支持左手末端压下把手;接触点部分被门框遮挡,没有围合抓握的证据。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–5.0 s,机器人双脚接触地面,支撑站姿完成伸臂和压把手。" + } + ] + }, + "door_opened": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 5.7, + "log_step": null, + "text_zh": "观察:视频 3.5–5.7 s,把手已下转;机器人随后收回操作臂并调整躯干,但门扇仍与门框闭合,没有出现门扇绕铰链转开及开口。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 6.0–10.05 s,机器人继续在门前调整上肢和躯干,门扇始终保持闭合;加密检查视频 8.0–10.05 s 后仍未见门扇脱离闭合门框,不能把画面投影变化算作开门。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "压/擦", + "note_zh": "推断:开门尝试包含视频 3.5–5.7 s 对门把手的下压及随后收臂,实际可由把手转动确认的是左手末端的按压;收臂阶段是否勾住把手背面不可见,不标为勾或抓握。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 4.5–10.05 s,机器人双脚在门前地面支撑身体,配合收臂、前倾及躯干转动,未见摔倒。" + } + ] + }, + "doorway_crossed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 0.0–10.05 s,机器人躯干一直在门前同一侧;双脚没有形成穿过门口的行走,门扇也未打开。视频 9.1–10.05 s 的末段仍未见躯干到达另一侧。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "原生奖励分", + "threshold": 600, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "door_opened", + "t": 5.7, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频约 5.7 s,机器人在压下把手后回收操作臂,门扇没有被带开;视频 6.0–10.05 s 的再次伸臂、前倾与转动躯干仍未打开门。失败停留在拉门操作,尚未开始穿门;评分器记录 success=0、finalized=1、infra_error=0。" + }, + "unmet": { + "door_opened": false + }, + "phenomena": [], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "t": 5.7, + "log_step": 89, + "source_zh": "日志 / 视频 / agent 自述", + "evidence_zh": "推断:log step 89 的末次机器人控制调用使用左臂从把手上方向下压、再向身体方向退回的关节轨迹,行走计划为空;视频 3.5–5.7 s 对应可见把手下转及收臂,却未产生门扇转开。log step 91 中 agent 自述尝试勾住把手再拉动但从未确认勾挂成立,与评分视频中门始终闭合一致,支持将瓶颈归于门机构接触操控,证据不足以细分接触设计、逆解或反馈控制。" + } + ], + "gap_zh": "评分视频只有 256×256 的单路房间视角,拉门阶段机器人手末端与把手背面被门扇、门框遮挡;没有状态日志、接触对、力或门关节角读数。压下把手可由机构转动确认,但拉门失败无法可靠区分未建立勾挂、短暂勾挂后滑脱或施力方向/大小不当,因此现象层留空,根因只到控制大类。日志含 49 次重置,先前测试不计入本次节点;log step 90 修改脚本后未见再次运行调用,不能以最终脚本文本或 agent 对其运行的自述替换评分视频。评分器未报告 episode_return 的数值,成功标志 0 不是累计奖励 0;未将日志控制步数或 agent 时间换算成视频时间。" + }, + "summary_zh": "评分回合在视频约 3.5 s 压下门把手,完成 1/3 个节点;随后收臂并调整身体,至视频 10.05 s 门仍闭合,机器人未穿门。断点为打开门,日志与视频支持门机构接触操控未奏效;拉门接触点被遮挡,无法可靠选定具体失败现象,累计原生奖励数值也未报告。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "把手约 3.5 s 压下、门全程闭合、未穿门,节点值与断点(door_opened,操作中)与帧一致;根因停在“控制”大类、现象留空并写明 gap_zh,符合分类顺序,未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:17:48Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard-realtime--sthgqmg/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard-realtime--sthgqmg/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "bc342fe3e6a5b3537fe62f697a3099b45dadde019f56f9724fc7008aad06b61c", + "video_sha256": "f8afaa7150f2b44ba165973f8091e6f4a1ce13dbd612c3e60c1cb3cbe613d7ed", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard-realtime--xa8pvhw.json b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard-realtime--xa8pvhw.json new file mode 100644 index 0000000000000000000000000000000000000000..543cc4d32f3114d8fb265d6953e6dfdfe599856a --- /dev/null +++ b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard-realtime--xa8pvhw.json @@ -0,0 +1,282 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-door-i00-standard-realtime--xa8pvhw", + "taskKey": "humanoidbench/manip_door", + "dag_sha256": "0e2b91c83745073b5462f5a378fa464b0e90842fe7004cf6de2891ede1995c32", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 787, + "sim_steps": 9428, + "calls": 536, + "rows": 10436, + "resets": 17, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3459.8, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 188.268, + "max_lag_s": 0.0152, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 34, + "late_steps": 0, + "requests_refused": 17, + "request_busy_s": 5.699, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-door-i00-standard-realtime--xa8pvhw/replay.mp4", + "duration_s": 7.95, + "fps": 20.0, + "frames": 159, + "sha256": "40d98c3cd110d61c13e10b125bbe14a2b272e9ecdbbd5b9c2072628f3472ac3f" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 3, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "handle_pressed": { + "method": "visual", + "intervals": [ + [ + 6.5, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.0 s 门把手近乎水平,机器人一侧前臂伸向门把手附近;门扇仍在门框内。" + }, + { + "t": 6.5, + "log_step": null, + "text_zh": "观察:逐帧检查视频 6.0–7.9 s,把手的自由端从近水平位置向下转动,视频约 6.5 s 已能辨认下转,视频 7.0–7.9 s 更明显;这不是只有手臂靠近。按门把手下转的事件判据计为达到,首次时刻有约 0.3 s 的视觉误差。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "压/擦", + "note_zh": "推断:机器人一侧前臂或手部末端在把手处下压;视频 6.5–7.9 s 上肢靠在该区域且把手向下转动。门扇遮住精确触点及左右侧,未见包围式抓握,不标抓握或勾。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–7.9 s 机器人双脚与地面接触并调整站位;后段屈膝、身体下沉,未见明确翻倒。" + } + ] + }, + "door_opened": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 4.0, + "log_step": null, + "text_zh": "观察:检查视频 0.0–7.9 s,门扇一直留在门框内,未见绕铰链向机器人侧转出或形成通行开口;视频 4.0 s 机器人靠近门且上半身部分被门扇遮挡。" + }, + { + "t": 7.0, + "log_step": null, + "text_zh": "观察:视频 6.5–7.9 s 把手虽下转,门扇仍未打开;机器人屈膝下沉、腿向外伸,未见勾住门把手后向身体方向拉门的动作。" + }, + { + "t": null, + "log_step": 82, + "text_zh": "观察:log step 82 才启动最终回合,机器人服务返回 episode=18、control_steps=297、done=False;此前 log step 77 的长回合和 log step 79 关于接触的自述属于已被重置的测试,不计入本视频开门进度。" + } + ] + }, + "doorway_crossed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 7.9, + "log_step": null, + "text_zh": "观察:检查视频 0.0–7.9 s,机器人躯干始终在门的原来一侧,末段仍在闭门前下沉,没有躯干穿过打开门口到另一侧的过程;未开始穿门。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "原生奖励分", + "threshold": 600, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "door_opened", + "t": 7.0, + "log_step": 82, + "where": "首步前", + "evidence_zh": "观察:视频约 6.5 s 把手已发生下转,但视频 7.0–7.9 s 机器人仍屈膝留在闭门前,尚未启动可辨认的勾拉开门动作;压把手事件已达到,打开门是其后未启动且未完成的步骤。log step 82 才启动评分所用的最终回合,先前测试里的开门动作计划不能当作本视频已经执行。" + }, + "unmet": { + "door_opened": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 7.0, + "log_step": 82, + "evidence_zh": "观察:视频 0.0–7.9 s 只有门前站位调整、身体下沉及末段把手下转,没有向机器人侧勾拉门扇的动作,门始终未形成开口。log step 82 启动的是最后评分回合;不能把 log step 77 的前一测试回合当作已开始本回合的开门。" + } + ], + "root_causes": [], + "gap_zh": "观察:inputs/ 中只有低分辨率房间视角视频和会话日志,没有状态日志或成功示范;门扇遮住上肢精确触点,无法确定接触的是哪一侧手臂,也无法恢复把手角度或骨盆高度。视频 4.0–7.9 s 可见机器人屈膝下沉,但未见明确失去支撑或摔倒,不能仅据此确定平衡失败,依赖该判断的控制根因也缺少支持。log step 80 的 time_left=19.6 s、episode=17、done=True;log step 81 返回 bc: command not found;log step 82 补启动 episode=18,服务返回 control_steps=297、done=False、running=True。最终视频止于视频 7.95 s,未显示清楚的勾拉;log step 70 的程序计划在仿真 8.5 s 伸手、仿真 11.0 s 勾住、仿真 12.5 s 回拉,计划不代表已执行。推断:结合 inputs/run.json 的 wall_s=3603 和 time_limit_s=3600,最终回合可能因会话预算结束而未执行到计划勾拉阶段;按 docs/03_trajectory.md Step 3.2 和 CASES.md FAIL-4/FAIL-22,不把预算耗尽或晚启动当作失败根因。缺少最终回合的终止原因和完整控制反馈,根因留空,也不判断拉门接触几何或力是否足够。评分字段报告 success=0,但未报告 episode_return,故原生累计奖励留空。" + }, + "summary_zh": "评分视频中机器人在门前调整站位并屈膝下沉,视频约 6.5 s 起可见门把手下转,完成压把手事件;门扇一直未打开,也未穿门,节点最终为 1/3,断点是尚未启动勾拉的打开门步骤。屈膝下沉不足以证明失去支撑,启动器报错及最终回合时长不足也不足以支持交付根因,因此只保留“没尝试”现象,根因留空。置信度降至 5:没有状态日志、成功示范或原生累计奖励读数,上肢精确触点及最终回合的终止原因仍无法核实。", + "conf": 5, + "review": { + "status": "agreed", + "rounds": 3, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "fix", + "conf": 6, + "summary_zh": "节点 1/3 与断点基本可接受,但根因「交付·交付操作出错」实质是把最终回合因预算所剩无几而只跑了约 8 s 当作原因(FAIL-4/22:超时/预算用尽不是原因),平衡现象也缺少帧证据(视频中机器人始终站立、蹲姿靠门,没有失去支撑或摔倒)。", + "n_issues": 3, + "n_major": 2 + }, + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 2, + "verdict": "fix", + "conf": 6, + "summary_zh": "节点 1/3 与断点基本可接受,但「没尝试」现象把一次被会话时限截断的最终回合(视频仅 7.95 s,程序计划的勾拉在仿真 11 s 才开始)当成机器人未尝试;这属于外因截断,应清空现象层并用 gap_zh 说明。", + "n_issues": 1, + "n_major": 1 + }, + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 3, + "kind": "confirm", + "verdict": "agree", + "summary_zh": "I1 的 reject 成立:最终回合由 agent 自己的启动器(缺 bc)延误后在预算将尽时启动,会话预算耗尽属 agent 自身责任而非外因,且该回合确为受评回合;保留没尝试与断点无误。", + "n_issues": 0, + "n_checks": 1 + } + ], + "open_zh": [], + "confirm": { + "verdict": "confirmed", + "summary_zh": "I1 的 reject 成立:最终回合由 agent 自己的启动器(缺 bc)延误后在预算将尽时启动,会话预算耗尽属 agent 自身责任而非外因,且该回合确为受评回合;保留没尝试与断点无误。", + "checks_ok": "1/1", + "n_factual_errors": 0 + } + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:26:18Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard-realtime--xa8pvhw/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard-realtime--xa8pvhw/04_review_r1" + }, + { + "stage": "reconcile", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard-realtime--xa8pvhw/05_reconcile_r1" + }, + { + "stage": "traj_review", + "round": 2, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard-realtime--xa8pvhw/04_review_r2" + }, + { + "stage": "reconcile", + "round": 2, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard-realtime--xa8pvhw/05_reconcile_r2" + }, + { + "stage": "confirm", + "round": 3, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard-realtime--xa8pvhw/06_confirm_r3" + } + ], + "confirm_prompt": "prompts/06_confirm.md", + "confirm_prompt_sha256": "b9f56c095f7ea5e609e58bb21e076f565766c370333d71c7a3512cffb4d78267", + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "7376f440028edd70bb02c06b88b6777d4cf971bc783c9af16d5b67e283d168b7", + "video_sha256": "40d98c3cd110d61c13e10b125bbe14a2b272e9ecdbbd5b9c2072628f3472ac3f", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard-realtime--xrpue3w.json b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard-realtime--xrpue3w.json new file mode 100644 index 0000000000000000000000000000000000000000..0febc4d2dac7a8491fc774640e7b7a1689b8e00d --- /dev/null +++ b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard-realtime--xrpue3w.json @@ -0,0 +1,235 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-door-i00-standard-realtime--xrpue3w", + "taskKey": "humanoidbench/manip_door", + "dag_sha256": "0e2b91c83745073b5462f5a378fa464b0e90842fe7004cf6de2891ede1995c32", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 1000, + "sim_steps": 42147, + "calls": 17694, + "rows": 17589, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 2957.4, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 842.039, + "max_lag_s": 0.0759, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 25418, + "late_steps": 0, + "requests_refused": 26, + "request_busy_s": 8.422, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-door-i00-standard-realtime--xrpue3w/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "a92bab8587ed747ca0eefaedcf72b44ab8dd49017b50a0a1ad2d5915e8cb2a2a" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 3, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "handle_pressed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.6, + "log_step": null, + "text_zh": "观察:视频 0.00–1.30 s 的逐帧近看中,机器人在门把手高度附近活动前臂,随后可见前臂末端降到把手下方;门上可见的银色把手未相对门扇向下转动。视频 1.4–10.05 s 也未出现压下把手的变化。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "视频 0.00–1.30 s 可见机器人上肢向门把手区域运动,但手部末端与把手的交界被门扇和门框遮挡,无法确认实际触碰对象与受力方式;不据脚本宣称标为抓握或按压。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "视频 0.00–1.30 s,机器人双脚接触地面,支撑身体进行门把手区域的伸臂尝试。" + } + ] + }, + "door_opened": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.4, + "log_step": null, + "text_zh": "观察:视频 0.00、0.50、1.00、1.40 s 及其间密集帧中,门扇相对门框保持闭合位置,未见门扇转离门框形成开口;视野整体平移时门框也一起移动,不能把该画面移动当作开门。视频 2.0–10.05 s 门扇仍未打开。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "视频约 0.5–2.0 s,机器人前臂在门旁回收、身体靠近门扇,手部实际是否勾住或压到门的区域被遮挡,无法确认手与门的接触方式。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "视频约 0.5–2.0 s,机器人双脚在地面支撑身体,进行门前上肢动作与身体位置调整。" + } + ] + }, + "doorway_crossed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 1.0–10.05 s,机器人身体持续位于关闭门扇的原来一侧,双腿在门前小幅运动,末段身体前倾并降低;躯干始终没有穿过门口到达另一侧,视频 9.0–10.05 s 的密集帧中亦未发生穿门。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "视频 2.0–10.05 s,机器人双脚在门前地面支撑并作小幅迈步调整,身体未通过门口;末段仍可见脚与地面接触。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "原生奖励分", + "threshold": 600, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "handle_pressed", + "t": 0.6, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频 0.00–0.80 s,机器人已在门把手区域进行伸臂动作,但可见前臂末端随后降到把手下方,把手没有随之向下转动;到视频 10.05 s 都未补做出这一变化。log step 522 启动最后一次控制器,log step 525 的机器人服务状态确认最后回合为 episode 51、control_steps=1000、done=True、reason=failed、resets_left=0;服务状态用于确认评分回合结束,不用于推算视频时间。" + }, + "unmet": { + "door_opened": false + }, + "phenomena": [ + { + "cat": "运动", + "item": "位姿不准", + "t": 0.6, + "log_step": null, + "evidence_zh": "观察:视频 0.30–0.80 s,机器人可见前臂末端向下、向身体侧收回,降到门把手高度以下,未形成使把手下转的操作姿态;视频 0.90–1.30 s 身体继续靠近门而把手仍保持原朝向。这里标注可见的上肢位置偏差,遮挡区是否曾短暂碰触把手无法判断。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "t": 0.6, + "log_step": 520, + "source_zh": "日志", + "evidence_zh": "推断:log step 520 编写、log step 522 执行的最终控制器按 control_steps 推进预设右臂 traj 关节目标,可见代码中的反馈仅用骨盆 IMU 重力方向和角速度修正踝关节;没有以实际把手转动结果修正上肢动作的记录。结合视频 0.30–1.30 s 前臂下移而把手未转、随后继续后续运动,支持上肢操作控制未实现压把手;证据不足以细分为接触设计、逆解、参数或反馈控制中的某一项。" + } + ], + "gap_zh": "输入有完整视频及会话日志,但无状态日志、成功示范或本回合的其他相机视频;门扇与门框遮住手部和把手的交界,无法确定是否短暂碰触以及力、关节跟踪误差。可见把手和门扇的机构变化均未出现,因此节点判为未达到;根因仅保留宽泛的控制。日志控制步长为 0.02 s、末回合为 1000 控制步,与视频 10.1 s 的长度不一致,未建立控制步到视频的映射,全部视频锚点直接取自已查看帧。评分字段未报告累计原生奖励,episode_return 留空;log step 527 关于压把手、开门及穿门的完成叙述无画面支持,未作为节点证据。" + }, + "summary_zh": "评分回合中,机器人在门前活动上肢并随后小幅迈步,但未观察到把手下转、门扇打开或躯干穿门,三个节点均未达到。最早未修复的断点为压下门把手:前臂末端下移到把手下方而机构未动,预设上肢控制未实现该操作;log step 525 的服务状态确认最后回合失败。手部接触区被遮挡且无状态日志,无法进一步确定接触方式与控制失效的具体环节。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "视频 0–10 s 中把手形态与门扇始终未变,三节点未达成、断点在压把手、接触标注保守,均与帧一致;根因仅保留宽泛的控制,证据已声明不足以细分,可接受。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:17:37Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard-realtime--xrpue3w/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard-realtime--xrpue3w/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "21854e9619e4d13dc109919a7c6e797f4458683f4e1071fc42d7562f0ac0363d", + "video_sha256": "a92bab8587ed747ca0eefaedcf72b44ab8dd49017b50a0a1ad2d5915e8cb2a2a", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard-realtime--zjhusvr.json b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard-realtime--zjhusvr.json new file mode 100644 index 0000000000000000000000000000000000000000..379aefbb946a58a2432c150335095be2ba092918 --- /dev/null +++ b/phase2/humanoidbench/manip_door/trajectories/humanoidbench-manip-door-i00-standard-realtime--zjhusvr.json @@ -0,0 +1,247 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-door-i00-standard-realtime--zjhusvr", + "taskKey": "humanoidbench/manip_door", + "dag_sha256": "0e2b91c83745073b5462f5a378fa464b0e90842fe7004cf6de2891ede1995c32", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 1000, + "sim_steps": 15406, + "calls": 55288, + "rows": 54643, + "resets": 26, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 2904.8, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 307.662, + "max_lag_s": 0.1008, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 5842, + "late_steps": 0, + "requests_refused": 26, + "request_busy_s": 24.013, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.195, + "peak_processes": 9, + "peak_disk_gb": 0.241, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-door-i00-standard-realtime--zjhusvr/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "08aab73ee79acc7e0097be42cb1de30c905199e4b0c4279f353b026a874436a9" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 3, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "handle_pressed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.0 s 门扇上的浅色把手可见;视频 0.0–10.05 s 的全段抽帧与视频 9.0–10.05 s 的逐帧检查中,把手均未出现相对门扇向下转动。" + }, + { + "t": 3.0, + "log_step": 169, + "text_zh": "观察:视频 1.5–3.5 s 机器人一侧手臂向画面右上方伸展,视频 3.5–8.5 s 可见手端持续高于浅色把手,未见压下动作使把手运动;log step 169 启动最终回合,log step 170–171 的最终脚本记录只出现 stand、reach、press 阶段。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 1.5–10.05 s 机器人双脚接触地面,支撑伸臂和抬臂尝试,未见跌倒。" + }, + { + "who": "未知", + "how": "未知", + "note_zh": "观察:视频 1.5–10.05 s 靠门的上肢局部被门框遮挡;可见手端偏离把手,无法确认被遮挡部位是否有短暂接触及其方式。" + } + ] + }, + "door_opened": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 5.0, + "log_step": null, + "text_zh": "观察:视频 0.0、2.0、5.0、8.5 s 及视频 9.0–10.05 s 的密集帧中,深色门扇与门框的相对位置不变,未见门扇向机器人一侧转开或形成新开口。" + }, + { + "t": 10.05, + "log_step": 171, + "text_zh": "观察:视频 10.05 s 门扇仍处于初始闭合位置;log step 171 的服务返回 reason=failed、control_steps=1000。该失败字段仅作为整体结果,开门未达由画面判断。" + } + ] + }, + "doorway_crossed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 0.0–10.05 s 机器人躯干始终留在门前同一侧,双脚在门前支撑,末尾视频 9.0–10.05 s 未见躯干穿过门口;门扇始终未打开。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "原生奖励分", + "threshold": 600, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "handle_pressed", + "t": 3.0, + "log_step": 169, + "where": "操作中", + "evidence_zh": "观察:视频 1.5–3.5 s 机器人已开始伸臂,但可见手端伸向画面右上方而非压在把手上;视频 3.5–8.5 s 保持明显偏高的伸臂姿态,随后视频 9.0–10.05 s 继续调整仍未使把手下转。log step 169 启动最终评分回合,log step 171 确认该回合结束;首次压把手尝试未在本回合修复。" + }, + "unmet": { + "door_opened": false + }, + "phenomena": [ + { + "cat": "运动", + "item": "位姿不准", + "t": 3.0, + "log_step": null, + "evidence_zh": "观察:视频 3.0–8.5 s 机器人可见手端举在把手的右上方,手端与门扇上的浅色把手有明显空间偏离;视频 9.0–10.05 s 手臂虽继续改变姿态,把手仍未下转。" + } + ], + "root_causes": [ + { + "cat": "认知", + "item": "感知", + "log_step": 169, + "source_zh": "日志 / 视频 / agent 自述", + "evidence_zh": "推断:机器人可能依据偏位的把手位置估计伸臂。log step 159 的脚本对所选把手像素 (98.9,103.8) 反投影,在假设 z=1.12 m 时输出位置 [0.866,0.707,1.12],假设 z=1.08–1.16 m 时输出 y=0.807–0.607 m;这是 agent 自编相机模型的计算结果,并非实测三维坐标。log step 165 据此将 HANDLE.y 设为0.75,而 log step 169 启动最终回合前又把 HANDLE 恢复为 [0.79,0.45,1.12]、按压目标恢复为 y=0.44+ys,并将 HINGE_Y 恢复为-0.5;这些是脚本中的实际目标配置。视频 3.0–8.5 s 可见手端持续在把手右上方,视频 9.0–10.05 s 调整后仍未压下把手,与目标定位偏差相符,但画面不独立证明世界 y 坐标。agent 在 log step 172 自述门口位于 y≈+0.5–0.9、此前可能瞄准门旁,仅作为其诊断推断。因缺少独立三维真值及末端跟踪记录,不能断言目标估计是唯一原因。" + } + ], + "gap_zh": "无状态日志和成功示范;单个侧面视角及门框遮挡使上肢局部接触无法确定,但把手未下转、门扇未打开和躯干未穿门均可从可见画面判断。log step 171 服务报告最终回合为 1000 控制步,任务文本给定每步 0.02 s,agent 在 log step 172 自述回合长 20 s;所提供视频仅 10.1 s,缺少已验证的控制步到视频时钟映射,所有视频锚点直接取自画面。原生累计奖励未报告,不能把 outcome.reward=0.0 当作 episode_return。目标定位错误为日志计算与画面偏位共同支持的推断:log step 159、162 的反投影依赖 agent 自编模型、像素选择及假定高度,不能当作独立三维真值。log step 101 的 FK 未使用 sy、log step 102 修正前臂向量,均可从代码核对;sy 缺项仅是可能的控制误差来源,缺少最终回合实际末端位姿与目标的同步对照,无法判定其贡献,也不能确认运动本身准确跟踪目标;不另立控制根因。可行接触替代方案未验证。" + }, + "summary_zh": "评分视频中机器人完成了门前支撑和伸臂尝试,压下把手、打开门、穿过门口均未达成,最终进度 0/3,断点为压下门把手。视频 3.0–8.5 s 手端持续偏离把手,log step 159 的反投影计算与 log step 169 恢复的目标配置存在偏差,支持目标位置估计错误的感知根因推断;不再将 FK 未使用 sy 认定为单一原因。三维目标真值及实际末端跟踪未独立核实,根因判别仍有不确定性,故置信度由7降至6;门框遮挡、两种时钟缺少可靠映射及未报告原生累计奖励的缺口仍在。", + "conf": 6, + "review": { + "status": "agreed", + "rounds": 2, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "fix", + "conf": 6, + "summary_zh": "视频与节点标注一致:0–10.05 s 门始终闭合、把手未下转、躯干未穿门,进度 0/3,断点 handle_pressed 合理,位姿不准现象有画面支持。问题在根因:把末端失准归因于 FK 缺少 sy 项,只是对日志代码的单一推断,且忽略了日志中更直接的目标定位证据。", + "n_issues": 1, + "n_major": 1 + }, + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 2, + "verdict": "agree", + "conf": 7, + "summary_zh": "视频 0–10 s 门始终闭合、把手未下转、躯干未穿门,0/3 与断点 handle_pressed 正确;位姿不准有画面支持,感知根因已明确标为推断并写出局限,第 1 轮意见已妥善处理,未发现需改动之处。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:23:07Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard-realtime--zjhusvr/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard-realtime--zjhusvr/04_review_r1" + }, + { + "stage": "reconcile", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard-realtime--zjhusvr/05_reconcile_r1" + }, + { + "stage": "traj_review", + "round": 2, + "ok": true, + "space": "work/runs/humanoidbench-manip-door-i00-standard-realtime--zjhusvr/04_review_r2" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "feae9660c93570db475c66b41b2e57587daf0323c7cfb908b27bc0ba50e80ce6", + "video_sha256": "08aab73ee79acc7e0097be42cb1de30c905199e4b0c4279f353b026a874436a9", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-privileged--6vjp8rn.json b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-privileged--6vjp8rn.json new file mode 100644 index 0000000000000000000000000000000000000000..43dfa60bf6c40e4e1544ea4a7c24df6656bc4d64 --- /dev/null +++ b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-privileged--6vjp8rn.json @@ -0,0 +1,172 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-powerlift-i00-privileged--6vjp8rn", + "taskKey": "humanoidbench/manip_powerlift", + "dag_sha256": "388cfca325afca3d4eb2a3cee3aacbb2213d1a0ad552b6c823429491cd5d2e19", + "outcome": { + "success": 0, + "reward": 0.3405565805962514, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 1000, + "final_reward": 0.3405565805962514, + "sim_steps": 1000, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-powerlift-i00-privileged--6vjp8rn/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "32bd81cdf46b3630a2625eac9df7df0c0547f218aef5dbbaf414e424bd4a66ab" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "dumbbell_raised_and_held": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00–1.00 s 逐帧检查可见机器人从屈膝初姿调整为较直的站姿,双手停在身前、明显高于哑铃杆,没有下探或合手夹住哑铃;哑铃两端配重仍落在地面。" + }, + { + "t": 5.0, + "log_step": null, + "text_zh": "观察:视频 1.0–9.0 s 每隔约0.5 s检查,机器人持续站立,手与哑铃分离,哑铃保持原来的地面位置,没有随手臂上升。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 9.00–10.05 s 逐帧检查,机器人没有摔倒,也没有伸手取哑铃;最后一帧哑铃仍由地面承托。整个评分轨迹未启动抓取、举高或持举,因此不记录该节点的操作接触。" + }, + { + "t": null, + "log_step": 71, + "text_zh": "观察:log step 71 的生成脚本保存1000步站立基线,使用固定站姿目标、重力补偿与质心反馈,未设置取铃或抬举阶段;log step 81 对该文件重复回放得到 deterministic=true、sim_steps=1000,与视频只站立的动作相符。评分器字段 success=0、final_reward=0.3405565805962514。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "dumbbell_raised_and_held", + "t": 0.0, + "log_step": 71, + "where": "首步前", + "evidence_zh": "观察:视频 0.00–10.05 s 中机器人仅调整站姿并站立,双手始终没有下探抓取地面哑铃,哑铃始终未被抬离地面;log step 71 保存的正是站立基线,log step 81 验证了这个文件。举高保持步骤在评分轨迹内未启动。" + }, + "unmet": { + "dumbbell_raised_and_held": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 71, + "evidence_zh": "观察:视频 0.00–1.00 s 的初始站姿调整及视频 1.0–10.05 s 的后续站立都没有取铃动作,手未靠近哑铃杆,哑铃一直在地面。log step 71 保存的脚本只维持站立;会话中的其他抬举测试未写入最终交付轨迹。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "log_step": 79, + "source_zh": "日志", + "evidence_zh": "推断:未能实现稳定的取铃控制,是最终仅交付站立基线的能力缺口。依据为会话中 log step 73 的逆解抬举测试输出 FELL 51,log step 75、79 的改版均输出 FELL 65,三次记录的哑铃最高位置均为初始约0.209米;这些测试未得到能举起哑铃的控制结果。agent 在 log step 81、84 也明确说明接近哑铃阶段使本体失稳,所以保留站立交付文件;归因为控制而不细分子类,测试中的摔倒不作为评分视频的现象。" + } + ], + "gap_zh": "没有状态日志或成功示范,无法独立复核接触力、关节跟踪及所有可行举重方案;根因只由会话测试与最终交付记录推断,不能据 agent 的力矩估算认定任务不可解。评分器未报告 episode_return,final_reward 的归一化口径未给出,故原生累计奖励保留 null;节点未完成同时有视频中始终未举铃及 success=0 支持。" + }, + "summary_zh": "评分视频中机器人仅调整站姿并持续站立,哑铃始终在地面,“哑铃举高保持”未完成,节点最终为0/1。断点为该步骤的首步前,现象是没尝试;会话测试反复在接近阶段失稳,支持尚未实现稳定取铃控制这一根因,最终交付的是已验证回放的站立基线。没有状态日志及原生累计奖励记录,控制子类与任务物理可解性无法确定。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "视频全程机器人站立、哑铃在地面,节点未达、断点首步前/没尝试与交付站立基线(log 71、81)一致;根因控制来自会话测试(FELL 51/65,哑铃未离地),按FAIL-22范式可接受,gap_zh已说明局限。未发现需改动处。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:48:20Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-privileged--6vjp8rn/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-privileged--6vjp8rn/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "8c38aeb2447dd82a355def210e39857938ab2d22d6acc4537fe918add3bdec59", + "video_sha256": "32bd81cdf46b3630a2625eac9df7df0c0547f218aef5dbbaf414e424bd4a66ab", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-privileged--7qrxmpn.json b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-privileged--7qrxmpn.json new file mode 100644 index 0000000000000000000000000000000000000000..ca72a9d79f2110d6b272c67920cec01a50debd73 --- /dev/null +++ b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-privileged--7qrxmpn.json @@ -0,0 +1,173 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-powerlift-i00-privileged--7qrxmpn", + "taskKey": "humanoidbench/manip_powerlift", + "dag_sha256": "388cfca325afca3d4eb2a3cee3aacbb2213d1a0ad552b6c823429491cd5d2e19", + "outcome": { + "success": 0, + "reward": 0.15152150342000587, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 411, + "final_reward": 0.15152150342000587, + "sim_steps": 411, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-powerlift-i00-privileged--7qrxmpn/replay.mp4", + "duration_s": 4.2, + "fps": 20.0, + "frames": 84, + "sha256": "480595e401eb69a2e399e1929756045c51ee997d75bae4e343c273504e8e9047" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "dumbbell_raised_and_held": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 11, + "text_zh": "观察:视频 0.0–0.9 s,机器人双手悬在身前,哑铃两端配重仍落在地面,没有下蹲取物、合手抓杆或托举动作。log step 11 的交付生成脚本仅将初始关节位置归一化后反复执行同一动作向量,并保存录制轨迹;log step 12 启动该基线的生成与重放。" + }, + { + "t": 1.3, + "log_step": null, + "text_zh": "观察:视频 1.0–1.3 s,机器人向前倾倒,双手落向地面,随后躯干贴地、双腿抬起;哑铃没有随手臂被举离地面。视频 1.4–4.15 s 机器人保持倒地,最后一秒逐帧检查也没有恢复或举重。" + }, + { + "t": 4.15, + "log_step": null, + "text_zh": "推断:评分视频全程没有启动举重,本节点从未达成;评分器 success=0 与此一致。初始双足站地和随后倒地支承属于基线背景接触,不是对本节点的取物操作,因此不填写 contact_actual;不能把会话中另外测试的高处承载姿态算入本次评分轨迹。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "dumbbell_raised_and_held", + "t": 1.0, + "log_step": 11, + "where": "首步前", + "evidence_zh": "观察:视频 0.0–0.9 s 机器人只维持原站姿,未向地面哑铃开展抓取或托举;视频约 1.0 s 开始明显前倾,至视频 1.3 s 已倒地,直到视频 4.15 s 均未恢复。log step 11 保存的生成脚本只是重复初始关节目标的站立基线,没有举重动作序列。" + }, + "unmet": { + "dumbbell_raised_and_held": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 11, + "evidence_zh": "观察:log step 11 的录制循环始终调用 s.step(norm_a),norm_a 来自初始关节位置,没有接近哑铃、闭合手指、抬臂举重的阶段;视频 0.0–4.15 s 也未见这些动作,评分交付仅为站立基线。" + }, + { + "cat": "全身", + "item": "平衡", + "t": 1.1, + "log_step": null, + "evidence_zh": "观察:视频 0.95–1.20 s,机器人躯干快速向前倾,双手下降触地;视频 1.25–1.35 s 躯干落地、双腿抬起,随后到视频 4.15 s 一直俯卧,未能恢复双足站立。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "推断:log step 11 的交付控制器只重复初始关节位置对应的固定目标,不根据返回的 obs 调整目标;视频 0.95–1.35 s 的前倾、落地表明这套基线控制未能在实际动力学中维持全身支撑。可归因为控制不足,但现有证据不足以进一步确定具体增益、力矩或关节跟踪问题。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "inputs/ 没有状态日志,评分器也未报告 episode_return 原始累计值,因此该原生指标留空,不能把 final_reward=0.15152150342000587 当作累计奖励。日志可确认保存了站立基线且后续有独立控制测试,但未明确记录为何最终没有将举重测试整合进交付;不据此增标规划或交付操作出错。" + }, + "summary_zh": "评分交付仅为固定初始关节目标的站立基线,未启动哑铃抓取或举高保持,节点完成 0/1。机器人在视频约 1.0–1.3 s 向前倒地并保持至结束,现象为没尝试及平衡,固定目标基线未能稳住本体支持控制根因。无状态日志及原始累计奖励,无法细分失稳机制或还原累计分数。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "节点未达成、断点首步前、没尝试+平衡的判断与视频一致:0–0.9 s 站立不动,约1.0–1.3 s 前倒,之后俯卧到 4.15 s,哑铃未被举起。交付为固定初始关节目标的基线(log step 11),日志末段仍在做会话内测试,未见交付被替换。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:44:05Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-privileged--7qrxmpn/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-privileged--7qrxmpn/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "e0287be0002e6de96626c36a270ff7cd9a865199d0874ab4be9b09b9114c9cdf", + "video_sha256": "480595e401eb69a2e399e1929756045c51ee997d75bae4e343c273504e8e9047", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-privileged--a2z2efz.json b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-privileged--a2z2efz.json new file mode 100644 index 0000000000000000000000000000000000000000..df5a665b800d8f39c74266035f51e578e9c275d1 --- /dev/null +++ b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-privileged--a2z2efz.json @@ -0,0 +1,178 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-powerlift-i00-privileged--a2z2efz", + "taskKey": "humanoidbench/manip_powerlift", + "dag_sha256": "388cfca325afca3d4eb2a3cee3aacbb2213d1a0ad552b6c823429491cd5d2e19", + "outcome": { + "success": 0, + "reward": 0.15152992040481741, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 1000, + "final_reward": 0.15152992040481741, + "sim_steps": 1000, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 7.224, + "peak_processes": 23, + "peak_disk_gb": 0.239, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-powerlift-i00-privileged--a2z2efz/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "319866f6edc4178f903a5574be0da0c9b39344d304c1a371e34b6cce5a23e67a" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "dumbbell_raised_and_held": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.0–0.5 s,哑铃两端重块落在地面,机器人双手抬向空中,没有握住哑铃杆。" + }, + { + "t": 1.3, + "log_step": null, + "text_zh": "观察:视频 0.6–1.4 s,机器人躯干逐渐前折并塌倒,双臂伸向地面;哑铃没有被提离地面。视频 1.5–10.05 s,机器人持续低伏,哑铃仍在地面,未出现举高保持。" + }, + { + "t": null, + "log_step": 47, + "text_zh": "观察:log step 47 的交付生成脚本只设置腿部伸展、固定肩肘目标及质心反馈,没有抓取或举升哑铃的动作阶段;log step 48–49 修正属性访问后成功保存1000步动作,log step 51 对交付文件重放得到 deterministic=true。举重步骤从未启动,故省略 contact_actual,不把静态足地支撑记成举重接触。" + }, + { + "t": null, + "log_step": 52, + "text_zh": "观察:log step 52,agent 明确自述交付轨迹没有举起哑铃;评分器 reward.success=0,final_reward=0.15152992040481741。该结果与视频中未举起的表现一致,但 final_reward 未明确标为累计 episode_return,不能将其作为累计分数。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "dumbbell_raised_and_held", + "t": 0.6, + "log_step": 47, + "where": "首步前", + "evidence_zh": "观察:视频 0.0–0.5 s,机器人没有向哑铃建立抓取,随后在视频约 0.6 s 开始明显前倾,视频约 1.3 s 塌倒,直至视频 10.05 s 未恢复,哑铃始终在地面。log step 47–49 保存的是姿态稳定控制序列,log step 52 的 agent 自述也承认未举起;失败发生在举重步骤启动之前。" + }, + "unmet": { + "dumbbell_raised_and_held": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 47, + "evidence_zh": "观察:视频 0.0–10.05 s,机器人没有握住、托起或举升哑铃,开始时双手上抬,随后塌倒。log step 47 的交付脚本没有抓取或举升阶段,log step 52 的 agent 明确自述所交付轨迹不举哑铃。" + }, + { + "cat": "全身", + "item": "平衡", + "t": 0.6, + "log_step": null, + "evidence_zh": "观察:视频约 0.6 s,机器人上身明显前倾,视频 0.8–1.1 s 躯干接近水平,视频 1.2–1.4 s 本体塌倒、双臂落至地面附近;视频 1.5–10.05 s 一直低伏,未重新站起。这是评分视频中的失衡,不能由log step 52 的 agent 自述“survives the whole episode”视为保持站立。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "推断:log step 47 的交付控制器以腿部关节目标缓慢伸展,并用机器人质心横向位置及前向速度反馈调整踝和髋;log step 49、51 确认该序列已保存且可确定性重放。视频 0.6–1.4 s 仍出现前折塌倒,随后直至视频 10.05 s 未恢复,说明所交付的本体稳定控制未能维持站姿,阻断了后续举重。log step 44 的会话测试也记录了髋部偏离目标、骨盆下沉,但它仅作控制困难的会话佐证,不用于定位本次视频时刻。现有证据不足以锁定某个增益或关节力矩为唯一原因,保留在“控制”。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "没有状态日志或成功参考视频,无法精确核验手、足与哑铃的接触对、关节受力或累计奖励。评分器只给出 final_reward,未给出明确的 episode_return,故该原生指标留空;不把 agent 自述的负载分析视为任务物理不可解的证明。视频中的塌倒可以直接判断,但不能从“未提前终止”或末态画面确定骨盆终止阈值是否触发;不新增人工约束标签。" + }, + "summary_zh": "评分视频中哑铃始终在地面,“哑铃举高保持”未完成,节点终值为0/1。机器人在视频约 0.6 s 开始失衡、约 1.3 s 塌倒并持续至末尾;交付内容仅为姿态稳定控制,举重未启动,断点为首步前,根因证据支持本体控制失败。无状态日志,累计 episode_return 未报告,无法进一步确认具体受力或控制参数原因。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "哑铃始终在地面、节点未达成、断点首步前、没尝试+平衡、根因控制,均与视频和log一致;交付脚本确无举重阶段,未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:47:02Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-privileged--a2z2efz/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-privileged--a2z2efz/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "c5203ef5ba5b33e6bb7a1630e3781345f3297173ce42cd35f0e3f2ec9e189729", + "video_sha256": "319866f6edc4178f903a5574be0da0c9b39344d304c1a371e34b6cce5a23e67a", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-privileged--m2efkww.json b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-privileged--m2efkww.json new file mode 100644 index 0000000000000000000000000000000000000000..b9c9827d76924cf06b5dbe832b771d7e655f3a13 --- /dev/null +++ b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-privileged--m2efkww.json @@ -0,0 +1,169 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-powerlift-i00-privileged--m2efkww", + "taskKey": "humanoidbench/manip_powerlift", + "dag_sha256": "388cfca325afca3d4eb2a3cee3aacbb2213d1a0ad552b6c823429491cd5d2e19", + "outcome": { + "success": 0, + "reward": 0.3283699313677372, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 1000, + "final_reward": 0.3283699313677372, + "sim_steps": 1000, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 1.853, + "peak_processes": 10, + "peak_disk_gb": 0.834, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-powerlift-i00-privileged--m2efkww/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "7cc15b0050e7cef9e072c4265408bf3194736abc2e3909f484ba0c2909ab8a5a" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "dumbbell_raised_and_held": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.0 s 哑铃横置地面,机器人屈膝站在后方,双手停在腰前且明显高于铃杆;视频 0.0–10.0 s 的连续抽帧未见伸手下探、握杆或举起动作。该步未启动,因此不添加 contact_actual;站立的背景支撑不作为举重接触。" + }, + { + "t": 10.0, + "log_step": 156, + "text_zh": "观察:视频 9.0–10.0 s 的密集末段画面仍见机器人站立、双手空置、哑铃两端落地,没有高处持举;评分器 success=0、deterministic=1、n_actions=1000。agent 在 log step 156 也自述交付只保持站立,哑铃仍在地面。" + }, + { + "t": null, + "log_step": 150, + "text_zh": "观察:log step 35 写入的生成脚本仅以髋、膝、踝反馈维持初始站姿;log step 94 再次打印相同脚本。log step 150 明确选择该站立基线并运行生成脚本保存交付;log step 148、152 的触杆、横移和倒下是另行测试,不是视频中的评分动作。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "dumbbell_raised_and_held", + "t": 0.0, + "log_step": 150, + "where": "首步前", + "evidence_zh": "观察:视频 0.0–10.0 s 机器人仅调整并维持站姿,双手始终没有下探铃杆,哑铃一直落地;log step 150 选择并重新生成站立基线作为交付,log step 156 明确承认未完成举起。评分轨迹没有启动哑铃举高保持,断点以视频 0.0 s 为导航锚点。" + }, + "unmet": { + "dumbbell_raised_and_held": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 150, + "evidence_zh": "观察:视频 0.0–10.0 s(含视频 9.0–10.0 s 密集末段)双手维持腰前空手姿态,机器人没有向地面哑铃发起抓取或抬升;log step 150 交付的是站立基线。会话里的抓取试验没有进入该评分轨迹。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "t": null, + "log_step": 148, + "source_zh": "日志", + "evidence_zh": "推断:控制实现未能形成可用的下蹲抓取与起举动作,因而最终保留仅站立的基线。依据为会话测试 log step 141:脚本检测到手与哑铃接触后,铃中心高度仍约0.2米且机器人骨盆继续降低;log step 148 改为充分屈指后仍打印 maxz≈0.209米,随后骨盆降至约0.2米并终止;log step 150 因这些举起测试失败而重新生成站立交付。这些是控制根因的会话证据,不据此给评分视频添加接触或跌倒现象。", + "note_zh": "下蹲稳定性与抓取承载均未解决,现有证据不足以唯一细分为接触设计、轨迹或反馈控制;未由“没尝试”本身反推根因。" + } + ], + "gap_zh": "没有状态日志或成功示范,无法逐帧核验精确接触和累计奖励;run.json 的评分器字段仅报告 final_reward=0.3283699313677372,未明确报告 episode_return,因此该原生指标留空,不将单步/归一化字段擅自换算为累计值。控制根因由会话测试与最终交付选择推断,尚不能确定具体控制子类。视频可见哑铃全程落地,节点未完成及首步前断点不受这些缺口影响。" + }, + "summary_zh": "评分轨迹仅维持站姿,哑铃在视频 0.0–10.0 s 始终落地,“哑铃举高保持”未启动,最终完成0/1节点。断点为该步首步前,现象为没尝试;会话抓取测试反复无法抬升且失稳,支持控制实现未解决后交付站立基线的归因,但这些测试动作不计入评分轨迹。无状态日志,精确累计奖励未报告,控制子类仍有不确定性。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "评分轨迹(视频 0–10 s)机器人仅站立、哑铃全程落地,节点 false、断点首步前/没尝试、根因控制(会话证据,符合 FAIL-22 的处理)均有据;未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:45:19Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-privileged--m2efkww/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-privileged--m2efkww/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "e3f81ce5eb24105bed68177cbdc94e17252e5cc1e3b0becf758395532d6348de", + "video_sha256": "7cc15b0050e7cef9e072c4265408bf3194736abc2e3909f484ba0c2909ab8a5a", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-privileged--pok5bqz.json b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-privileged--pok5bqz.json new file mode 100644 index 0000000000000000000000000000000000000000..09ee9979fd821e51efa9323d120ad9f8c7adb195 --- /dev/null +++ b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-privileged--pok5bqz.json @@ -0,0 +1,168 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-powerlift-i00-privileged--pok5bqz", + "taskKey": "humanoidbench/manip_powerlift", + "dag_sha256": "388cfca325afca3d4eb2a3cee3aacbb2213d1a0ad552b6c823429491cd5d2e19", + "outcome": { + "success": 0, + "reward": 0.15152991948500694, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 1000, + "final_reward": 0.15152991948500694, + "sim_steps": 1000, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 1.008, + "peak_processes": 11, + "peak_disk_gb": 0.253, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-powerlift-i00-privileged--pok5bqz/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "6bb0f36bb8e0e21c6103362498e187ba9b544d7c09a7af052cdf1bc769f8e740" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "dumbbell_raised_and_held": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 159, + "text_zh": "观察:视频 0.00–0.65 s,机器人双手停在身前并随身体后仰,未向地面哑铃建立抓取或托举;视频 0.70 s 已倒地,哑铃仍在原地。log step 159 的最终生成脚本只在初始关节目标上调节双侧髋、踝俯仰,没有抓取或举升动作。" + }, + { + "t": 10.05, + "log_step": 212, + "text_zh": "观察:全片粗查并逐帧检查视频 0.00–2.05 s、9.00–10.05 s,哑铃始终在地面,机器人倒地后未恢复站立,也未举起哑铃。log step 205、212 对最终文件的重放均报告 deterministic=true、scored=false、sim_steps=1000;评分器 success=0。该步骤从未启动,故不记录该步骤的 contact_actual;会话中另行测试的手物或踝物接触不属于交付轨迹。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "dumbbell_raised_and_held", + "t": 0.7, + "log_step": 159, + "where": "首步前", + "evidence_zh": "观察:视频 0.00–0.65 s 机器人仅在原地失稳后仰,没有向哑铃伸手抓取;视频约 0.70 s 躯干已落地,之后直到视频 10.05 s 都未起身,哑铃未离地。log step 159、205、212 确认交付的是仅调节髋、踝俯仰的1000步平衡轨迹,重放未达标;举高保持步骤未开始。" + }, + "unmet": { + "dumbbell_raised_and_held": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 159, + "evidence_zh": "观察:视频 0.00–10.05 s 未见机器人尝试抓住或托起哑铃;机器人双手先随躯干后仰,倒地后保持朝上,哑铃始终留在地面。log step 159 的交付脚本没有抓取、下蹲接近或抬升哑铃的动作段。" + }, + { + "cat": "全身", + "item": "平衡", + "t": 0.7, + "log_step": null, + "evidence_zh": "观察:逐帧检查视频 0.00–0.70 s,机器人由站立逐渐后仰,视频 0.55–0.65 s 躯干接近水平,视频约 0.70 s 背部/上躯干落地;视频 0.70–10.05 s 始终保持倒地姿态,没有恢复站立。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "反馈控制", + "source_zh": "日志", + "evidence_zh": "推断:交付控制器未能稳定本体站姿。log step 159 的生成代码用骨盆四元数分量误差与角速度组成PD反馈,只调整双侧髋、踝俯仰;log step 182 以同一控制律作会话诊断时,输出头高从约1.67米降至约0.033米、骨盆高约0.29米,说明运行满步数并不代表站立稳定。视频 0.00–0.70 s 的后仰倒地直接印证这一控制失败;log step 205、212 确认最终重放的是同一平衡轨迹。现有证据不确定是哪个增益或反馈项具体有误,不进一步标参数调试。" + } + ], + "gap_zh": "观察:inputs/ 未提供状态日志或成功参考视频,评分字段也未报告 episode_return,故该原生指标保留 null,不能把 final_reward=0.15152991948500694 直接当成累计回报。log step 207、208 的说明声称平衡控制器使机器人全程直立,但视频 0.70–10.05 s 明确显示机器人倒地,按视频标注;会话诊断仅用于支持控制根因,不转换为视频时刻。没有足够证据进一步解释 agent 为何最终只交付平衡轨迹,不据此增加认知或交付根因。" + }, + "summary_zh": "评分轨迹未完成“哑铃举高保持”,最终节点为0/1:机器人没有启动抓取或举升,哑铃始终留在地面。机器人在视频约0.70 s 后仰倒地,此后未恢复;交付脚本和会话诊断支持平衡反馈控制失败,断点为举高保持的首步前。无状态日志和累计回报字段,无法确定精确 episode_return;日志中“全程直立”的自述与视频不符,以视频为准。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "视频帧显示机器人约0.7 s后仰倒地、哑铃始终在地面,节点未达成、断点首步前、没尝试+平衡与反馈控制根因均有证据支持。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:56:53Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-privileged--pok5bqz/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-privileged--pok5bqz/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "152354adb8e3e2701740d7f3f2bb51746d6609947e12476acc3feab165ebf1c5", + "video_sha256": "6bb0f36bb8e0e21c6103362498e187ba9b544d7c09a7af052cdf1bc769f8e740", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-privileged--rdl4j8y.json b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-privileged--rdl4j8y.json new file mode 100644 index 0000000000000000000000000000000000000000..5d37daa5d58595e99bbe989a93fd3b5c338abef5 --- /dev/null +++ b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-privileged--rdl4j8y.json @@ -0,0 +1,168 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-powerlift-i00-privileged--rdl4j8y", + "taskKey": "humanoidbench/manip_powerlift", + "dag_sha256": "388cfca325afca3d4eb2a3cee3aacbb2213d1a0ad552b6c823429491cd5d2e19", + "outcome": { + "success": 0, + "reward": 0.34607471294964853, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 1000, + "final_reward": 0.34607471294964853, + "sim_steps": 1000, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 13.134, + "peak_processes": 44, + "peak_disk_gb": 0.239, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-powerlift-i00-privileged--rdl4j8y/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "af485deb946878ba660d00c5209fe9fa9ae627e9d5609805e816eb7a20b77268" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "dumbbell_raised_and_held": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.0 s 哑铃两端配重落在地面,机器人站在其后方,双手在腰部以上,与哑铃分离。" + }, + { + "t": 9.5, + "log_step": null, + "text_zh": "观察:已按约视频 0.1 s 间隔检查全段,并密查视频 9.0–10.05 s;机器人仅屈伸手臂、调整站姿,双手未下探到哑铃,哑铃始终未离地,末段仍站立,没有举高保持。" + }, + { + "t": null, + "log_step": 21, + "text_zh": "观察:log step 21 保存的交付脚本只插值调整髋、膝、踝及肘部,并以质心相对双脚的位置反馈调整髋踝;记录1000个控制步,重放前后状态哈希均为 db6bdd2e6db7。评分器字段 success=0、deterministic=1、n_actions=1000,与视频中的站立轨迹一致。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "dumbbell_raised_and_held", + "t": 0.0, + "log_step": 21, + "where": "首步前", + "evidence_zh": "观察:视频 0.0–10.05 s 中机器人一直在哑铃后方调整站立,双手没有朝地面哑铃下探,哑铃未被抓取或抬起;log step 21 保存并核验的是站立平衡轨迹,log step 29 的 agent 最终自述也确认交付轨迹从未举起哑铃。举重步骤在评分回合内未启动,以视频 0.0 s 为导航锚点。" + }, + "unmet": { + "dumbbell_raised_and_held": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 21, + "evidence_zh": "观察:视频 0.0–10.05 s 可见机器人双手始终远离地面哑铃,只在站立时屈伸手臂,未启动抓取、举升;视频 9.0–10.05 s 的密集末段画面仍是同一状态。log step 21 的交付代码也只实施站立平衡,没有接近或提起哑铃的动作。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "t": null, + "log_step": 26, + "evidence_zh": "推断:会话测试显示 agent 未能构造有效举升控制:log step 18 的上举测试连零附加载荷条件下也使骨盆降至约0.27–0.30米,log step 19 的控制追踪显示上举时质心偏离双脚并伴随关节力饱和;log step 24–26 对下蹲、伸手、闭指、起身关键姿态进行参数搜索,最高哑铃高度搜索仅报0.312米。log step 21 已保存站立基线,log step 27–29 确认继续保留该基线并预期失败。这些会话证据支持举升控制未解决,不能把测试中的跌倒或轻微抬高当作评分视频中的动作,也不足以细分具体控制子类。", + "source_zh": "日志" + } + ], + "gap_zh": "未提供状态日志与成功示范,无法独立核验会话测试的接触力和所有替代举升方案;agent 在 log step 22、29 关于物理不可行的论断未排除所有合法方案,因此不标任务或环境根因。评分器仅提供 final_reward=0.34607471294964853,未直接报告 episode_return;不把该字段或 agent 的代理奖励估算当作累计回报,原生指标值保留null。此缺口不影响视频明确未举升及评分失败的判断。" + }, + "summary_zh": "评分轨迹仅维持站立并屈伸手臂,视频 0.0–10.05 s 中双手未接近地面哑铃,哑铃未离地,哑铃举高保持节点未达到(0/1)。断点为该步骤的首步前,现象为没尝试;会话中的上举失稳及关键姿态搜索未获有效举升支持控制根因,agent 最终交付经确定性重放核验的站立基线。缺少状态日志及直接累计回报,不能核实物理不可行论断或精确累计得分。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "视频全程机器人站立、屈伸手臂,哑铃未离地,节点未达、断点首步前、没尝试与 CASES FAIL-13/FAIL-22 一致;根因标控制来自会话证据并已声明局限,可接受。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:42:51Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-privileged--rdl4j8y/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-privileged--rdl4j8y/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "c91bdb0a6e7193d9e4b3277fca316f1d9b6ccb44fd004feb0de2f08d59f12a18", + "video_sha256": "af485deb946878ba660d00c5209fe9fa9ae627e9d5609805e816eb7a20b77268", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-privileged--wptqajf.json b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-privileged--wptqajf.json new file mode 100644 index 0000000000000000000000000000000000000000..5e75113a9ca7bebbc957ef9e11a57b1139e771fc --- /dev/null +++ b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-privileged--wptqajf.json @@ -0,0 +1,188 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-powerlift-i00-privileged--wptqajf", + "taskKey": "humanoidbench/manip_powerlift", + "dag_sha256": "388cfca325afca3d4eb2a3cee3aacbb2213d1a0ad552b6c823429491cd5d2e19", + "outcome": { + "success": 0, + "reward": 0.18944544909653407, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 1000, + "final_reward": 0.18944544909653407, + "sim_steps": 1000, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 17.006, + "peak_processes": 29, + "peak_disk_gb": 0.907, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-powerlift-i00-privileged--wptqajf/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "c31c79183a16180203b77b5372eef0df9aef6cd2b1d62d3f1ff7fbfa1431fbe4" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "dumbbell_raised_and_held": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.5, + "log_step": null, + "text_zh": "观察:视频 0.0–1.5 s,机器人由站姿下蹲,两手伸向地面哑铃的横杆;视频 2.4–3.9 s,双手随杆回移,哑铃沿地面向双脚靠近,尚未举高。" + }, + { + "t": 4.2, + "log_step": null, + "text_zh": "观察:视频 4.05–4.35 s,机器人双臂突然离杆上摆,躯干后仰、骨盆向后下方落下;视频 4.8–5.1 s 已呈坐地姿态,哑铃仍低于躯干。" + }, + { + "t": 5.4, + "log_step": null, + "text_zh": "观察:视频 5.25–8.5 s,机器人坐地屈腿,横杆被弯曲的腿从下方承托而离开地面,未随手臂升至高处;视频 9.0–10.05 s,哑铃仍横在腿前,机器人没有起身或完成举高保持。" + }, + { + "t": null, + "log_step": null, + "text_zh": "观察:inputs/run.json 的评分器字段 success=0、all_success=0、deterministic=1、n_actions=1000,报告 final_reward=0.18944544909653407,未单独报告 episode_return。推断:结合视频中仅低位承托的过程及该单目标回合的失败判定,累计奖励达到800的事件未发生;不把短暂离地计为节点完成。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "未知", + "note_zh": "观察:视频 1.5–3.9 s,左右手共同接触哑铃横杆,哑铃随双手向双脚方向沿地面移动;原视频为256×256像素,无法辨清手指是否形成对向包络夹持,故不将拖杆直接等同于抓握。" + }, + { + "who": "腿/躯干", + "how": "托/舀", + "note_zh": "观察:视频 5.25–8.5 s,坐地机器人的屈膝腿部位于横杆下方,将哑铃从地面托起至低位;视频 9.0–10.05 s 横杆仍靠在腿前。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–3.9 s,双脚接地支承下蹲和拉杆;视频 4.35–10.05 s,机器人转为骨盆近地的坐姿,由下肢及坐地部位支承身体。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "dumbbell_raised_and_held", + "t": 4.1, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频 3.9 s 机器人仍在蹲姿接触哑铃横杆;视频 4.05–4.35 s 双臂离杆上摆、躯干后仰,随后坐倒。视频 5.25–10.05 s 虽出现腿部低位托杆,但没有完成举高保持,因此该举重步骤始终未修复。" + }, + "unmet": { + "dumbbell_raised_and_held": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 4.1, + "log_step": null, + "evidence_zh": "观察:视频 4.05–4.35 s,机器人在哑铃旁由双脚支承的蹲姿突然后仰下落,双臂甩起;视频 4.8–5.1 s 已坐地,此后至视频 10.05 s 未恢复站立,举重只能停留在坐姿腿部承托。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "视频 / 日志", + "evidence_zh": "推断:log step 121 的脚本在下蹲拉杆前缀后,搜索髋、膝、踝及双臂的关节位置目标序列,并按哑铃高度等代理量选取动作;log step 145 的打包脚本从此类候选中选取交付轨迹。视频 4.05–4.35 s 的坐倒和视频 5.25–10.05 s 始终无法起身举高,表明交付动作未实现举重所需的全身协调与承重控制;现有证据不足以进一步区分接触布局、轨迹生成或反馈控制子类。" + } + ], + "gap_zh": "无状态日志和成功示范;视频分辨率为256×256像素,双手与横杆接触及拖动可见,但手指对向夹持细节无法辨清,手部接触方式记未知。评分器未提供 episode_return 或逐步奖励,不能把 final_reward 擅自换算成累计分;没有逐帧接触力或跟踪误差,根因仅归到控制。log step 149 仅返回执行日期,没有最终候选名或打包脚本的完成输出,因此不把会话内各候选的精确高度、接触力及测试结果当作评分视频的测量。" + }, + "summary_zh": "机器人完成下蹲接触和沿地面拉杆,但视频约 4.1 s 向后坐倒;随后仅用弯曲的腿低位承托哑铃,直到视频 10.05 s 仍未完成唯一节点“哑铃举高保持”,终态为0/1。失败现象为全身平衡,现有视频与动作搜索日志支持控制能力不足,尚不能确定具体控制子类。缺少状态日志、累计奖励及清晰手指接触细节,手部接触方式与精确累计分数保留未知。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "节点未达成、失败点(约4.1 s 后仰坐倒)、平衡现象与控制根因均有视频支持;未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:51:34Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-privileged--wptqajf/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-privileged--wptqajf/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "ebf1bbee2d75cc9004bd3d577cfe76e44276815ff6df2ab50b80d8b5ee7c4b24", + "video_sha256": "c31c79183a16180203b77b5372eef0df9aef6cd2b1d62d3f1ff7fbfa1431fbe4", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard--3c8mpsq.json b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard--3c8mpsq.json new file mode 100644 index 0000000000000000000000000000000000000000..2ca25baea70af6d6ba1d833d599fc5ee427ab2d2 --- /dev/null +++ b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard--3c8mpsq.json @@ -0,0 +1,197 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-powerlift-i00-standard--3c8mpsq", + "taskKey": "humanoidbench/manip_powerlift", + "dag_sha256": "388cfca325afca3d4eb2a3cee3aacbb2213d1a0ad552b6c823429491cd5d2e19", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 233, + "sim_steps": 13815, + "calls": 13801, + "rows": 13774, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1692.6, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.21, + "peak_processes": 8, + "peak_disk_gb": 0.262, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-powerlift-i00-standard--3c8mpsq/replay.mp4", + "duration_s": 2.4, + "fps": 20.0, + "frames": 48, + "sha256": "28f7f5b9b83c85262b0784b8ed02cae650f41d8dfb7c155f4da93ed76d98ad33" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "dumbbell_raised_and_held": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00 s,机器人站在地面哑铃后方,双手空着前伸,哑铃两端盘均落地。" + }, + { + "t": 1.4, + "log_step": null, + "text_zh": "观察:视频 1.00–1.55 s,机器人屈膝俯身、双手下探哑铃,哑铃随触碰在地面滚动;未见双手包握承重或哑铃随手臂举高。" + }, + { + "t": 2.35, + "log_step": 133, + "text_zh": "观察:视频 1.60–2.35 s,机器人翻倒在哑铃旁并继续滚转,哑铃仍处于地面附近,没有举高保持。log step 133 的机器人服务状态为 episode=51、control_steps=233、done=True、reason=failed;评分器字段 live_success=0、graded_by_live=1。" + }, + { + "t": null, + "log_step": 132, + "text_zh": "观察:log step 131 执行最后一次实际机器人动作并报告 failed;log step 132 的后续方案在 reset 阶段失败,log step 133 明确 reset budget exhausted。因此后续修改未进入评分回合,较早测试的起身举臂也不计为本轨迹进展。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "推/扫", + "note_zh": "观察:视频 1.20–1.55 s,两只灵巧手共同下探哑铃杆及端盘附近,触碰过程中将哑铃推滚;未见手指围合夹紧哑铃并承重,不标抓握。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–1.45 s,机器人以双脚接触地面支承下蹲;视频约 1.60 s 起躯干倒向地面,随后失去站立支撑。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "dumbbell_raised_and_held", + "t": 1.5, + "log_step": 131, + "where": "操作中", + "evidence_zh": "观察:视频约 1.45–1.65 s,机器人正在俯身伸手操作哑铃,躯干迅速倾倒、头部降到地面附近;视频 1.70–2.35 s 继续倒地滚转,始终没有完成举高保持。log step 131 是最后实际执行的回合,log step 133 服务确认该回合以 failed 结束。" + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 1.5, + "log_step": 131, + "evidence_zh": "观察:视频 1.45–1.70 s,机器人由双足下蹲姿态向前侧方翻倒,头与上身落到地面附近;视频 1.75–2.35 s 身体在地面滚转,未恢复站立。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志 / 视频", + "evidence_zh": "推断:log step 129 修改下蹲目标、重心目标及手臂角度,log step 131 以 cdB=0.075、K=2.0、kroll=1.0 执行该方案;脚本原本依次安排下蹲、合指、起身、举高,但该次输出在 grip 阶段已为 failed。结合视频 1.45–1.70 s 在伸手下蹲过程中倒地,说明该执行方案未能维持本体稳定,归为控制;现有证据不足以区分接触布局、轨迹跟踪或反馈参数的具体责任。" + } + ], + "gap_zh": "没有状态日志、接触对或逐步累计奖励,评分器未报告 episode_return 数值,因此该指标保留 null;视频分辨率为256×256,不能核实每根手指的接触位置或把失衡归到某个增益。未建立服务 control_steps 与视频时钟的对应关系,所有视频时刻直接取自已检查的48帧;较早回合与 agent 最终自述的空手举臂不作为本回合节点达成证据。", + "unmet": { + "dumbbell_raised_and_held": false + } + }, + "summary_zh": "评分回合中机器人下蹲伸手触碰并推滚地面哑铃,未完成“哑铃举高保持”,节点最终为0/1。视频约 1.5–1.7 s,机器人在该步操作中失衡倒地,随后滚转直到结尾;日志确认最后回合失败,根因可归为控制,但无法细分失稳机制。没有状态日志或累计奖励数值,手指具体接触也受低分辨率限制。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "节点未达成、断点(操作中,哑铃举高保持)、平衡现象与控制根因均与视频和日志一致;时间在约一秒容差内,无时钟换算问题。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:51:33Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard--3c8mpsq/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard--3c8mpsq/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "af828d6eb417df66cdf6530ec78fcf020c90b4a99b86e3f937c656ef201aba7c", + "video_sha256": "28f7f5b9b83c85262b0784b8ed02cae650f41d8dfb7c155f4da93ed76d98ad33", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard--4qhptvn.json b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard--4qhptvn.json new file mode 100644 index 0000000000000000000000000000000000000000..e087475da2e7d7f184e7ff8e621eff6a698da16f --- /dev/null +++ b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard--4qhptvn.json @@ -0,0 +1,194 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-powerlift-i00-standard--4qhptvn", + "taskKey": "humanoidbench/manip_powerlift", + "dag_sha256": "388cfca325afca3d4eb2a3cee3aacbb2213d1a0ad552b6c823429491cd5d2e19", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 185, + "sim_steps": 589, + "calls": 50, + "rows": 589, + "resets": 3, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3519.9, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-powerlift-i00-standard--4qhptvn/replay.mp4", + "duration_s": 1.95, + "fps": 20.0, + "frames": 39, + "sha256": "8ef00f203307040b2bd03da55e45b4c2bfe30b61b476d046b2efd35ba15617d8" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "dumbbell_raised_and_held": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00 s 哑铃的两端配重落在地面,机器人站在其后,双手尚未接近杆身;逐帧检查视频 0.00–1.90 s,未见哑铃随双手举到高处并保持。" + }, + { + "t": 0.6, + "log_step": null, + "text_zh": "观察:视频 0.30–0.60 s 机器人屈髋屈膝并将双手伸向哑铃杆,视频 0.60–1.30 s 保持弯腰伸手姿态;视频 1.35–1.90 s 上身继续下坠,双腿向后滑伸,最终塌倒在哑铃后方,哑铃配重仍贴近地面。" + }, + { + "t": null, + "log_step": 90, + "text_zh": "观察:log step 3 的服务说明只有最后一回合受评;log step 90 重置至 episode 4 并执行接近、勾取和闭手命令,log step 93 继续调整双手,最后报告 control_steps=185,与评分器 n_actions=185 一致;此前回合不计入节点进度。" + }, + { + "t": 1.9, + "log_step": null, + "text_zh": "推断:视频全段没有举高保持,且评分器 live_success=0、replay_success=0,因此单一累计验收节点未达到;评分器未报告 episode_return,不能用目录 reward=0.0 代替累计奖励数值。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "压/擦", + "note_zh": "推断:视频 0.60–1.30 s 左右手同时从上方搭压哑铃杆的两处,结合视频 0.50 s、1.00 s、1.30 s 放大画面及前后帧判断为手掌/手指搭压;未确认各手形成包围式抓握,未见承重举起。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–1.30 s 机器人双脚接触地面,腿和躯干承载弯腰伸手动作;视频 1.35–1.90 s 足部滑移、支撑丧失,身体塌下。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "dumbbell_raised_and_held", + "t": 1.4, + "log_step": 93, + "where": "操作中", + "evidence_zh": "观察:视频 1.30 s 机器人仍在哑铃杆附近调整双手,视频约 1.40 s 起头部与胸部继续向前下方坠落、双脚向后滑伸,视频 1.75–1.90 s 身体塌倒;全过程未把哑铃举高。log step 93 是最后一回合的双手重新定位和闭手命令,之后没有新的机器人动作或重置,失败未修复。" + }, + "unmet": { + "dumbbell_raised_and_held": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 1.4, + "log_step": 93, + "evidence_zh": "观察:视频 1.35–1.90 s 机器人在弯腰接近哑铃时持续前倾,头胸下降,双腿向后滑伸,最后塌倒在地面附近;哑铃没有随机器人升起。这是举重操作中失去支撑,不把同一塌倒另标违反约束。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "推断:log step 93 的机器人控制脚本将全身关节目标按余弦插值组成三段批量位置命令并依次提交,每段结束才读取关节和重力观测,未在段内用姿态反馈修正目标;该步输出的踝关节实际角度在 back 段为 -0.864 rad,而目标为 -0.26 rad。结合视频 1.35–1.90 s 的支撑丧失,执行的全身控制未能稳定维持取物姿态;接触几何与伺服跟踪各自的贡献未能分离,根因保留在控制大类。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "没有状态日志、接触力或接触对记录,也没有成功示范;视频分辨率为 256×256,双手搭压可结合前后帧推断,但手指是否短暂包围杆身不能确认,不标抓握。评分器未报告 episode_return,无法还原累计积分;日志控制步与视频时间的映射未核实,所有视频时刻均直接取自帧。log step 123 出现提供方过滤拦截,但评分尝试在此前已有视频可见的物理塌倒,按已有物理失败标注,不将过滤作为机器人现象或根因。" + }, + "summary_zh": "评分尝试中机器人弯腰将双手伸向哑铃并在杆身附近调整,未举高保持,最终节点为 0/1。视频约 1.40 s 起机器人失去支撑并塌倒,断在哑铃举高保持的操作中,日志中的批量全身关节控制及跟踪偏差支持控制归因。无状态日志和累计奖励读数,低分辨率画面无法确认包围式抓握;log step 123 的提供方过滤发生在已观察到的物理失败之后。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "视频 1.95 s、39 帧与文件一致:机器人弯腰伸手,约 1.0 s 后逐渐前倾,在 1.4–1.9 s 塌倒,哑铃始终在地面,节点 0/1、断点和平衡现象与帧相符。log step 90/93 对应最后一回合(episode 4,control_steps 185)。根因“控制”为推断,已标注,未发现可证伪之处。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:46:53Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard--4qhptvn/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard--4qhptvn/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "888c6b751fe3b2256c3a6756149a4a8b20880dfcf240055c1ed2f24728503d1c", + "video_sha256": "8ef00f203307040b2bd03da55e45b4c2bfe30b61b476d046b2efd35ba15617d8", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard--jbfycor.json b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard--jbfycor.json new file mode 100644 index 0000000000000000000000000000000000000000..0e4fb574aca36acf4c7a622f99e01aff621244fa --- /dev/null +++ b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard--jbfycor.json @@ -0,0 +1,204 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-powerlift-i00-standard--jbfycor", + "taskKey": "humanoidbench/manip_powerlift", + "dag_sha256": "388cfca325afca3d4eb2a3cee3aacbb2213d1a0ad552b6c823429491cd5d2e19", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 66, + "sim_steps": 3151, + "calls": 246, + "rows": 3151, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1167.6, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-powerlift-i00-standard--jbfycor/replay.mp4", + "duration_s": 0.75, + "fps": 20.0, + "frames": 15, + "sha256": "e88408a2e2eaca0e3aef1cb648819a012cfd0f7426da4f76eae6b457243ada02" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "dumbbell_raised_and_held": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频0.00–0.70 s的全部15帧中,机器人有屈膝下蹲和手臂动作,但双手始终与哑铃分离,哑铃一直停在前方地面,没有离地举高或高处保持;视频约0.35 s后机器人逐渐后仰,视频0.65–0.70 s已仰倒、背部贴近地面。" + }, + { + "t": null, + "log_step": 272, + "text_zh": "观察:log step 194、198的机器人服务分别返回episode=51、control_steps=25和50;其间log step 196脚本提交保持蹲姿并伸手向杆的关节轨迹,服务返回done=False。log step 199提交握紧、起身及后续举高保持的分块轨迹,服务在control_steps=66返回done=True、reason=failed,log step 200确认episode=51;不能将整批后续举高保持动作视为已经执行。log step 272运行最终脚本时仍为该已结束回合且未再提交动作,评分器字段live_success=0。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频0.00–0.35 s机器人双脚与地面接触,支撑屈膝下蹲;随后机器人向后仰倒,视频0.65–0.70 s背部贴近地面。视频0.00–0.70 s双手始终未接触哑铃,未形成手部抓握。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "unit": "reward", + "threshold": 800, + "direction": "higher_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "dumbbell_raised_and_held", + "t": 0.4, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频0.00–0.35 s机器人屈膝下蹲,双手仍在哑铃上方且未触杆;视频约0.40 s起明显后仰,至视频0.65–0.70 s仰倒,哑铃仍在原地。log step 196、199记录episode 51中的伸手和起身轨迹提交,log step 200服务确认该回合control_steps=66、done=True、reason=failed。推断:机器人已开始举重步骤的下蹲伸手准备,在操作中失去平衡,后续没有恢复或完成;日志阶段不强配至具体视频帧。" + }, + "unmet": { + "dumbbell_raised_and_held": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 0.4, + "log_step": null, + "evidence_zh": "观察:视频0.35–0.60 s机器人躯干持续向后倾倒、骨盆下降;视频0.65–0.70 s背部接近地面、双臂朝上,未重新站起。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志", + "evidence_zh": "推断:评分回合episode 51中,log step 196脚本读取当前关节状态后,整批提交保持深蹲并伸手的预定插值轨迹,批次内未根据机体姿态反馈调整;返回done=False,imu_gravity_x=-0.589、imu_gyro_y=-1.677。log step 199又整批提交握紧、起身及后续举高保持轨迹,服务在control_steps=66返回done=True、reason=failed,log step 200确认同一回合失败。结合视频约0.40–0.70 s持续后倒,支持该回合的运动控制未维持平衡;不能据此认定具体参数错误或某一种控制子类,也不把未执行完的举高保持阶段视为实际动作。" + } + ], + "gap_zh": "无状态日志,也无成功参考视频;评分器未报告episode_return,原生累计奖励值保留null,不能把success=0当成累计奖励为0。log step 9给出的control_dt=0.02 s与log step 200的66个控制步相乘为仿真1.32 s,而视频仅0.75 s,缺少可靠的逐步对应关系;视频时间全部取自实看帧,不将日志阶段强配到画面,也不把其他重置回合的抓取测试计入进度。没有足够证据细分失稳的控制根因。" + }, + "summary_zh": "评分回合中机器人已开始下蹲伸手的举重准备,但双手未触及哑铃便在操作中向后仰倒;哑铃始终在地面,举高保持节点未达成,最终0/1。log step 196、199的关节轨迹提交及log step 200的episode 51失败读数,结合视频后倒,支持控制层面的失稳;后写入的完整举重脚本未再执行。无状态日志和累计奖励读数,且日志控制步时长与视频时长不一致,故节点和跌倒时刻以视频为准,控制根因未细分,置信度保持7。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 2, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "fix", + "conf": 7, + "summary_zh": "视频读数(哑铃始终在地、节点未达成、约0.4 s起后倒)与progress一致;但标签把评分回合当成“举重步骤未启动”,与log中评分回合(episode 51,control_steps=66)已执行下蹲伸手/起身分块相矛盾,没尝试和首步前不成立;另缺contact_actual。", + "n_issues": 4, + "n_major": 3 + }, + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 2, + "verdict": "agree", + "conf": 8, + "summary_zh": "第1轮问题已修正:现象仅保留全身·平衡,where为操作中,contact_actual已补;视频15帧所见(哑铃始终在地、约0.4 s起后仰、0.65–0.70 s仰倒)和log step 196/199/200(episode 51、control_steps=66、failed)与标签一致,控制根因以推断标注且来源为评分回合服务读数,未发现需修改的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:04:44Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard--jbfycor/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard--jbfycor/04_review_r1" + }, + { + "stage": "reconcile", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard--jbfycor/05_reconcile_r1" + }, + { + "stage": "traj_review", + "round": 2, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard--jbfycor/04_review_r2" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "5574061c077873174a21cb9752feb7699c952f8403983f3d6a8cb6903b59d362", + "video_sha256": "e88408a2e2eaca0e3aef1cb648819a012cfd0f7426da4f76eae6b457243ada02", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard--m2pxaam.json b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard--m2pxaam.json new file mode 100644 index 0000000000000000000000000000000000000000..2ac2fe736d23bfe09da2a445245234aea9d724cc --- /dev/null +++ b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard--m2pxaam.json @@ -0,0 +1,181 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-powerlift-i00-standard--m2pxaam", + "taskKey": "humanoidbench/manip_powerlift", + "dag_sha256": "388cfca325afca3d4eb2a3cee3aacbb2213d1a0ad552b6c823429491cd5d2e19", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 975, + "sim_steps": 8979, + "calls": 474, + "rows": 8979, + "resets": 45, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3424.0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-powerlift-i00-standard--m2pxaam/replay.mp4", + "duration_s": 9.85, + "fps": 20.0, + "frames": 197, + "sha256": "7d21990a6e6fabbe22ada8f0c63eb6f4c71f12f1bb2ed747cdcd27de2360b2ea" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "dumbbell_raised_and_held": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 104, + "text_zh": "观察:log step 104 中agent的脚本把原有抓取、起身和头顶保持航点替换为两个相同的站姿航点,双手grip采用默认0;log step 108 恢复温和髋踝反馈并reset到第46回合后运行该站姿脚本,log step 109确认此后不再运行机器人。视频0.00–0.80 s机器人双手空着停在身体前方,没有主动下探握杆或举重序列。" + }, + { + "t": 1.2, + "log_step": null, + "text_zh": "观察:视频0.95–1.30 s机器人躯干迅速前倾,双手落向地面、双腿失去原站立支承,身体伏倒在哑铃附近;哑铃随倒地发生滚动,没有被手掌合拢抓住并随手臂举高。" + }, + { + "t": 9.8, + "log_step": 108, + "text_zh": "观察:视频1.75–9.80 s机器人持续伏地,哑铃留在低处,末帧仍未形成持举;run.json评分器字段success=0、live_success=0。log step 108服务状态为episode=46、control_steps=975、done=False;agent在log step 109/112自述“完整1000步”与该计数不符,不据此判为完成。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "dumbbell_raised_and_held", + "t": 0.0, + "log_step": 108, + "where": "首步前", + "evidence_zh": "观察:log step 104的脚本修改删除举重动作序列,只留下站姿;log step 108将此配置用于最后的第46回合。视频0.00–0.80 s机器人空手站立,视频0.95–1.30 s前倒,此后直到视频9.80 s都未抓起并举高哑铃;评分回合没有启动该举重步骤。" + }, + "unmet": { + "dumbbell_raised_and_held": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 108, + "evidence_zh": "观察:log step 104将下探、握杆、起身和举高航点全部换为站姿,log step 108运行的最后第46回合沿用此修改。视频0.00–9.80 s仅见空手站姿、前倒及伏地,没有执行抓取后举高的动作;倒地过程中的身体与哑铃相遇不构成主动举重尝试。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "log_step": 100, + "source_zh": "日志 / 视频", + "evidence_zh": "推断:agent没有得到可用的举重控制器,最终退回站姿脚本。依据是log step 100/101的握杆后起身试验均由服务返回done=True、reason=failed,IMU重力x分量接近1;log step 102/103的修改试验也均返回failed,重力x分量接近-1;随后log step 104实际删除举重航点,log step 108选用站姿配置。最后评分视频0.95–1.30 s仍出现快速前倒且到视频9.80 s未恢复。此前试验仅作为控制能力的会话证据,不计为最后回合的抓取进度;现有证据不足以确定某个增益或关节符号是唯一错误。" + } + ], + "gap_zh": "无状态日志、逐步奖励或episode_return数值,无法还原累计回报和精确接触对;原生指标留null,节点未达由视频未见持举并结合评分器失败判定。agent在log step 112声称所有平衡姿态都够不着,但缺少穷尽可达性证据,不能据此归因为够不着或任务环境不可行。日志中的仿真时长不直接转换为视频时长,所有视频锚点均取自已检查帧。" + }, + "summary_zh": "最后评分的是第46回合,哑铃举高保持未达,最终0/1;agent已将举重序列改为站姿脚本,故断点为该节点的首步前,现象为没尝试。机器人在视频约0.95–1.30 s前倒,此后伏地至视频9.80 s;此前多次举重控制试验失败支持控制根因,但不计入评分回合进度。缺少状态日志和累计回报值,无法确定具体控制参数错误,也无法验证agent关于所有平衡姿态不可达的断言。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "评分回合第46回合视频中机器人空手站立约1 s后前倒并伏地至结尾,哑铃未被举起,节点未达;agent在最后回合换成站姿脚本,断点“首步前/没尝试”与根因“控制”(以会话试验为证据)符合FAIL-22的处理,未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:07:31Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard--m2pxaam/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard--m2pxaam/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "8f9132338e7af1051955c11f81833d9469a0edaa067f705dba0c362e627dbd9e", + "video_sha256": "7d21990a6e6fabbe22ada8f0c63eb6f4c71f12f1bb2ed747cdcd27de2360b2ea", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard--mpeadfr.json b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard--mpeadfr.json new file mode 100644 index 0000000000000000000000000000000000000000..21541fb5c0079e291f57894db6a48fefd9a1499b --- /dev/null +++ b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard--mpeadfr.json @@ -0,0 +1,194 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-powerlift-i00-standard--mpeadfr", + "taskKey": "humanoidbench/manip_powerlift", + "dag_sha256": "388cfca325afca3d4eb2a3cee3aacbb2213d1a0ad552b6c823429491cd5d2e19", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 51, + "sim_steps": 2975, + "calls": 336, + "rows": 2950, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1781.1, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-powerlift-i00-standard--mpeadfr/replay.mp4", + "duration_s": 0.6, + "fps": 20.0, + "frames": 12, + "sha256": "b72502af6835a01a5721b5f091d6b94fc2896be95051491ad1ea95b83303060a" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "dumbbell_raised_and_held": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.25, + "log_step": null, + "text_zh": "观察:逐帧检查视频 0.00–0.55 s 的全部12帧;机器人下蹲、双手向地面哑铃伸下,视频 0.20–0.30 s 双手仍在杆上方,未见手指包围夹住杆。随后双手空抬,哑铃两端配重始终留在地面。" + }, + { + "t": 0.5, + "log_step": null, + "text_zh": "观察:视频 0.35–0.50 s 机器人向后倾倒,视频 0.50–0.55 s 已仰倒在地,双手在身体上方,哑铃没有随手臂举起或保持在高处。" + }, + { + "t": null, + "log_step": 189, + "text_zh": "观察:log step 189 的调用先执行最后一次 reset,再调用 gen.attempt_powerlift(),该步记录1次重置和4次 submit;此前回合已被重置。评分器字段 graded_by_live=1、resets=50、n_actions=51、live_success=0,说明按最后实时回合判负;log step 190 中 agent 自述“哑铃仍在地面,最后回合51个控制步后失败”。该自述与视频中未举起的结果相符,未据其推算视频时刻。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.30 s 机器人双脚与地面接触,腿部承重下蹲;视频 0.35–0.55 s 向后失稳,最终背部落地。双手虽同时伸向哑铃并合拢,但未见夹住哑铃,故不记双臂抓握接触。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "dumbbell_raised_and_held", + "t": 0.25, + "log_step": 189, + "where": "操作中", + "evidence_zh": "观察:视频 0.20–0.30 s 机器人已下蹲伸手操作哑铃,但双手合拢未夹住杆;视频 0.35–0.55 s 双手空抬、身体后倒,哑铃仍在地面。log step 189 是最后一次重置后的举重调用,评分器 live_success=0;本步已启动但从未修复抓取失败,也未完成举高保持。" + }, + "unmet": { + "dumbbell_raised_and_held": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "没建立接触", + "t": 0.25, + "log_step": null, + "evidence_zh": "观察:视频 0.20–0.30 s 机器人双手在哑铃杆上方合拢,未包围夹住杆;视频 0.35–0.55 s 双手向上空抬,哑铃一直留在原来的地面位置,未建立举重所需的有效抓取。" + }, + { + "cat": "全身", + "item": "平衡", + "t": 0.35, + "log_step": null, + "evidence_zh": "观察:视频 0.30–0.45 s 机器人下蹲后躯干持续向后倾,骨盆向地面下降;视频 0.50–0.55 s 背部落地、身体仰倒,未恢复站立。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "视频", + "evidence_zh": "推断:视频 0.20–0.30 s 机器人执行下蹲和双手合拢,但手部未对哑铃形成有效抓取;视频 0.35–0.55 s 继续空抬双臂时又后倒。执行结果表明此次关节控制未同时实现手部接触与全身稳定;没有足够轨迹、接触力或反馈记录将其进一步归到某一控制子类。" + } + ], + "gap_zh": "没有状态日志、接触对或累计 episode_return 数值,不能还原精确手物接触力及累计奖励,也不能判定具体控制子类。视频仅0.60 s;log step 190 中 agent 自述最后回合51个控制步,按 log step 10 的 control_dt=0.02 s 对应仿真1.02 s,两者时长不一致,且 log step 189 的服务返回正文为空,缺乏可核对的逐帧映射,故所有导航时刻均直接取自视频,不由控制步数换算。最终驱动函数的完整修改内容未显示(log step 188 的 edit 仅保留片段),不能用前面测试回合的具体关节参数替代最后评分回合。" + }, + "summary_zh": "最后评分回合中机器人下蹲伸向哑铃,但双手合拢后空抬,哑铃始终在地面,举高保持节点未达成(0/1)。断点位于举重操作中,先未建立有效抓取,随后在视频 0.35–0.55 s 后倒;根因保守标为控制。缺少状态和累计奖励记录,且视频时长与日志自述控制步时长不一致,未进行两者的时间换算。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "视频12帧可见机器人下蹲、双手未碰到哑铃、随后后倒仰卧,哑铃留在地面;节点未达成、断点与两个现象有帧支持,根因保守标控制且标为推断,可接受。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:05:28Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard--mpeadfr/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard--mpeadfr/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "bcde7db56089fc2dd49906746d6c37554987883120e9c139061f88e92c5d9818", + "video_sha256": "b72502af6835a01a5721b5f091d6b94fc2896be95051491ad1ea95b83303060a", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard--oghvydp.json b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard--oghvydp.json new file mode 100644 index 0000000000000000000000000000000000000000..28734364bf792660a95d424b7163955cc0104d70 --- /dev/null +++ b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard--oghvydp.json @@ -0,0 +1,187 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-powerlift-i00-standard--oghvydp", + "taskKey": "humanoidbench/manip_powerlift", + "dag_sha256": "388cfca325afca3d4eb2a3cee3aacbb2213d1a0ad552b6c823429491cd5d2e19", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 1000, + "sim_steps": 6331, + "calls": 6131, + "rows": 6331, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1712.3, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-powerlift-i00-standard--oghvydp/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "a6b83267cd0d3e6fa998ebf355eb94bb2791c970174f7e8a932dde1dfd2bdcb2" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "dumbbell_raised_and_held": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 137, + "text_zh": "观察:视频 0.0–0.6 s 机器人站在地面哑铃后方,双手悬空;视频 0.7–1.0 s 机器人向前翻倒,未见握住哑铃或带着哑铃向上举起。log step 137 的最后回合脚本只调整踝关节与髋俯仰以维持站姿,没有执行抓取或举重序列。" + }, + { + "t": 10.05, + "log_step": 141, + "text_zh": "观察:视频 1.1–10.05 s 机器人持续俯卧,哑铃两端仍由地面支承,没有高举保持;片尾最后一秒亦无恢复。log step 141 的服务状态为 episode=51、control_steps=1000、done=True、reason=failed;评分器 live_success=0、replay_success=0,与画面一致。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–0.6 s 机器人以双脚接触地面支承身体、尝试维持站姿;视频约 1.0 s 后躯干及上肢落地承重。未观察到灵巧手包夹哑铃,不标双手抓握。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "dumbbell_raised_and_held", + "t": 0.8, + "log_step": 137, + "where": "首步前", + "evidence_zh": "观察:视频 0.7–1.0 s 机器人在尚未抓起哑铃时向前倒地,此后直到视频 10.05 s 仍俯卧。log step 137 是最后回合的站姿稳定性验证,未包含抓取、站起举物或上举动作;log step 139 的下一次 reset 被服务以重置预算耗尽拒绝,log step 141 确认最终评分回合为 episode 51。因此举重步骤尚未启动,失败位置为首步前。" + }, + "unmet": { + "dumbbell_raised_and_held": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 137, + "evidence_zh": "观察:视频 0.0–0.6 s 双手停在哑铃上方而未抓取;之后机器人翻倒,直到视频 10.05 s 均无抓举动作。log step 137 的最后回合代码仅用 IMU 俯仰与角速度修正髋、踝目标,其余保持名义姿态,没有启动举重序列。" + }, + { + "cat": "全身", + "item": "平衡", + "t": 0.8, + "log_step": 137, + "evidence_zh": "观察:视频 0.7 s 机器人开始明显前俯,视频 0.9 s 躯干接近水平,视频 1.0–1.1 s 向前倒地且双腿抬离原来的站立支撑姿态;视频 9.1–10.05 s 仍俯卧,未恢复站立。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "反馈控制", + "log_step": 137, + "evidence_zh": "推断:log step 137 的控制器以 IMU 俯仰和角速度反馈修正双踝及双髋目标、保持其余名义姿态;同一步输出显示控制步索引 0 时俯仰约 0.95°,索引 100 时已为 112.47°,后续仍超过 100°。视频 0.7–1.1 s 直接显示向前失稳并倒地,支持该站姿反馈未能稳定全身,使抓举序列未能开始;不据此推断具体增益错误、执行器力矩上限或所有合法控制均不可行。", + "source_zh": "日志" + } + ], + "gap_zh": "未提供状态日志,也未报告 episode_return 数值,故不能重建逐步累计奖励或量化与800分阈值的差距。视频中翻倒时手部与哑铃局部重叠,无法判定瞬间擦碰,但没有可见的包夹或举起。日志控制步对应的仿真总时长与视频时长不同,未建立二者映射;所有视频锚点均直接取自所检查的帧,未换算日志时间。" + }, + "summary_zh": "评分回合只执行站姿反馈控制,视频约 0.7–1.0 s 机器人在抓举启动前向前倒地,随后持续俯卧至视频 10.05 s,哑铃举高保持节点未达到,最终为0/1。最后回合的控制代码及俯仰反馈支持站立反馈控制失稳;没有执行后续抓取和举重序列。缺少状态日志及原生累计奖励数值,无法量化累计得分或确认翻倒时的瞬间手物擦碰。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "视频确认 0.7–1.1 s 机器人向前倒地并俯卧至片尾,哑铃始终在地面;最终评分回合(episode 51)即 log step 137 的站姿稳定性测试,未执行抓举,故首步前/没尝试、全身·平衡、反馈控制的标注均有证据支持。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:47:26Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard--oghvydp/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard--oghvydp/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "d91793df5556b756dd736449caf0cf9047beb22b59a2ba18bda47f1e9bc6aa9e", + "video_sha256": "a6b83267cd0d3e6fa998ebf355eb94bb2791c970174f7e8a932dde1dfd2bdcb2", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard--pj6tdsi.json b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard--pj6tdsi.json new file mode 100644 index 0000000000000000000000000000000000000000..cedaf4d8d62186810821e74f8c6c6156d95ec752 --- /dev/null +++ b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard--pj6tdsi.json @@ -0,0 +1,196 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-powerlift-i00-standard--pj6tdsi", + "taskKey": "humanoidbench/manip_powerlift", + "dag_sha256": "388cfca325afca3d4eb2a3cee3aacbb2213d1a0ad552b6c823429491cd5d2e19", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 185, + "sim_steps": 10280, + "calls": 921, + "rows": 10280, + "resets": 32, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 2602.9, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.2, + "peak_processes": 11, + "peak_disk_gb": 0.41, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-powerlift-i00-standard--pj6tdsi/replay.mp4", + "duration_s": 1.95, + "fps": 20.0, + "frames": 39, + "sha256": "5f00da031351a1f45d2bf756be5d1378a8bd533725d4485eebb534961c806cd4" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "dumbbell_raised_and_held": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00 s,哑铃横放在机器人前方地面,机器人屈膝站立,双手空着。" + }, + { + "t": 1.25, + "log_step": null, + "text_zh": "观察:逐帧检查视频 0.00–1.90 s,机器人放低双臂、屈膝并调整站姿,双手未包住哑铃杆;视频 1.25 s 和 1.60 s 可见手仍在膝部附近,哑铃没有随手升高。" + }, + { + "t": 1.9, + "log_step": null, + "text_zh": "观察:视频 1.70–1.90 s,机器人向后翻倒,哑铃仍留在地面,未出现举高保持;inputs/run.json 的评分器字段 success=0、live_success=0。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–1.60 s,机器人双脚在地面承重,腿和躯干进行屈膝与站姿调整;视频 1.70–1.90 s 失去站立支撑并倒地。未观察到双手抓握哑铃。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "dumbbell_raised_and_held", + "t": 1.7, + "log_step": null, + "where": "首步前", + "evidence_zh": "观察:视频 0.00–1.60 s,机器人仅调整空手站姿及屈膝幅度,未伸手包住哑铃或启动提举;视频约 1.70 s 开始明显向后倾倒,视频 1.85–1.90 s 已倒地,之后没有恢复或举起哑铃。" + }, + "unmet": { + "dumbbell_raised_and_held": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": null, + "evidence_zh": "观察:视频 0.00–1.90 s,机器人双手始终空着,未对地面哑铃实施抓握或提举;可见动作停留在本体屈膝、调姿,举重步骤未启动。" + }, + { + "cat": "全身", + "item": "平衡", + "t": 1.7, + "log_step": null, + "evidence_zh": "观察:视频 1.65–1.80 s,机器人躯干和骨盆连续向后倾斜下落;视频 1.85–1.90 s,背部与骨盆落到地面附近、双腿翘起,未恢复站立。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "t": 1.7, + "log_step": 160, + "evidence_zh": "推断:视频 1.65–1.90 s 的空手调姿随后失衡,说明本体姿态控制未能维持站立支撑,举重尚未启动即失败。会话中的 log step 158–160 记录 agent 反复修改起身前倾目标和执行节奏,返回结果均为 failed;这些是控制失败的补充会话证据,未将其分块数或估计高度映射为视频时间,也不据此认定某个伺服参数或硬件限制作祟。", + "source_zh": "视频 / 日志" + } + ], + "gap_zh": "输入没有状态日志或成功参考视频,也没有评分器报告的 episode_return,无法重建累计奖励或确认具体手足接触对。log step 160 的脚本输出为12个动作分块,log step 165 的 agent 最终自述称站立约3秒,而输入视频只有1.95秒;二者未有可验证的逐帧时钟映射,进度与断点均以视频为准。伺服下垂、力矩余量和具体控制参数的致因缺少独立核验,根因只标到控制。" + }, + "summary_zh": "评分轨迹中,机器人空手屈膝并调整站姿,未抓握或举起地面哑铃,唯一节点未达成(0/1)。视频约 1.70 s 机器人开始明显后倒,视频 1.85–1.90 s 已倒地,断在举重首步前;现象为没尝试与平衡,根因归为本体控制未稳定。无状态日志和累计奖励读数,日志末次动作分块与视频时长不能可靠对应,不能进一步核实具体伺服或参数原因。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "唯一节点未达成(视频全程双手空着、哑铃留在地面,约1.7 s后机器人后倒)与评分器success=0一致;断点首步前、现象没尝试+平衡、根因控制均有视频支持,与同任务先例一致,未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:50:16Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard--pj6tdsi/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard--pj6tdsi/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "58036714fe63d079a6fcaabef0e59365a07d1988ab204ad1436d45609e1f8402", + "video_sha256": "5f00da031351a1f45d2bf756be5d1378a8bd533725d4485eebb534961c806cd4", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard--skkcrxh.json b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard--skkcrxh.json new file mode 100644 index 0000000000000000000000000000000000000000..c108d9c0f188f142ec97c648c847d340d8aedaf6 --- /dev/null +++ b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard--skkcrxh.json @@ -0,0 +1,219 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-powerlift-i00-standard--skkcrxh", + "taskKey": "humanoidbench/manip_powerlift", + "dag_sha256": "388cfca325afca3d4eb2a3cee3aacbb2213d1a0ad552b6c823429491cd5d2e19", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 1000, + "sim_steps": 11718, + "calls": 11433, + "rows": 11718, + "resets": 39, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3566.7, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-powerlift-i00-standard--skkcrxh/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "dd34a734b50dc9dbf5d627445c9b4441cf2a22fbedb5d2680e4f1e6808c97354" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "dumbbell_raised_and_held": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频0.0 s机器人站在地面哑铃后方;视频0.5–1.0 s机器人下蹲,两臂伸向杠杆。" + }, + { + "t": 1.9, + "log_step": null, + "text_zh": "观察:视频1.7–2.4 s机器人在杠杆处尝试提拉,可见一只手在画面中杠杆左端附近抓取,另一只手的接触受遮挡;配重仍贴近地面,没有升至高处。推断:结合log step 151的双侧抬臂目标及log step 152的执行记录,两手可能共同参与抓握和提拉,无法从画面确认两手各自环抱。" + }, + { + "t": 2.7, + "log_step": null, + "text_zh": "观察:视频2.5–2.8 s机器人向前翻倒,头和上身落到杠杆前方,双脚翘离原站立支撑;哑铃未被举高。" + }, + { + "t": 10.05, + "log_step": 153, + "text_zh": "观察:视频3.0–10.05 s机器人持续倒伏在哑铃旁,视频9.1–10.05 s逐帧检查未见恢复或高举;log step 153服务状态为episode=40、control_steps=1000、done=true、reason=failed,评分字段live_success=0。推断:结合全片没有高举保持和该回合失败,累计验收节点未达到。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "观察:视频1.0–2.4 s可见一只手在画面中杠杆左端附近抓取;另一只手受躯干及配重遮挡,无法确认其环抱杆段或持续承重。推断:log step 151的final_run.py对双侧肩肘设置抬臂目标,log step 152记录final_lift、final_raise已执行,两手可能共同抓握并尝试提拉;此处双臂协作×抓握为推断,不能据此确认两手均已建立承重抓握。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频0.0–2.4 s双脚接地,腿和躯干支撑下蹲及提拉;视频2.5–2.8 s支撑丧失并向前翻倒,随后上身倒伏在地面及哑铃附近。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "dumbbell_raised_and_held", + "t": 2.5, + "log_step": 152, + "where": "操作中", + "evidence_zh": "观察:视频1.7–2.4 s机器人在杠杆处尝试提拉,一只手的抓取可见,另一只手接触受遮挡;视频约2.5 s躯干快速前倾,至视频2.7–2.8 s头和上身落地、双脚翘起。视频末帧10.05 s仍倒伏,举高保持始终未完成;log step 151–153记录最后回合的提拉、抬臂和随后保持命令,服务最终判失败。" + }, + "unmet": { + "dumbbell_raised_and_held": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 2.5, + "log_step": 152, + "evidence_zh": "观察:视频2.4 s机器人仍以双脚下蹲支撑;视频2.5–2.8 s向前翻倒,头和上身着地,双脚抬起。视频3.0–10.05 s没有重新站起,哑铃始终未被高举保持。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "推断:log step 151的最后回合控制脚本在合手后改变双侧髋膝目标,再改变肩肘目标抬臂,并持续调用force_move反馈控制;log step 152记录这些final_lift、final_raise命令已执行。对应视频2.4–2.8 s,机器人在提拉转抬臂过程中失去站立支撑并前翻,表明该控制序列未能稳定本体和负载;现有证据不足以细分为接触设计、轨迹或反馈控制的某一种。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "未提供状态日志或逐步累计奖励,评分字段也没有episode_return数值,因此该原生指标留空;低分辨率及躯干、配重遮挡限制了各手指接触和滑移的判断;另一只手是否环抱并承重无法确认,双手共同抓握仅为结合控制日志的推断,不能独立确认是否先滑脱再摔倒。log step 153的1000个控制步和log step 7的control_dt=0.02对应仿真20 s,视频长10.1 s,未建立控制步到视频的映射;本标注所有视频时刻均直接取自视频帧,未用agent时间换算。" + }, + "summary_zh": "最后评分回合中,机器人下蹲并在杠杆处尝试提拉,但哑铃没有被举高保持,完成节点0/1。视频约2.5 s机器人在提拉转抬臂过程中向前翻倒,此后直到视频10.05 s仍倒伏;结合最后回合控制脚本,根因归为控制。无状态日志和累计奖励数值,另一只手及手指局部接触受遮挡,双手共同抓握仅为推断,无法进一步确认失衡与滑脱的先后或具体控制子类。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 2, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "fix", + "conf": 6, + "summary_zh": "节点未达、断点(视频约2.5 s前翻)、平衡现象和控制根因基本与视频一致;唯一问题是抓握接触被写成“观察”的两手各自环抱,但帧中只清楚看到一只手臂在杆左端,另一只被躯干遮挡。", + "n_issues": 1, + "n_major": 0 + }, + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 2, + "verdict": "agree", + "conf": 8, + "summary_zh": "节点未达、断点(视频约2.5 s前翻)、平衡现象、控制根因及接触的观察/推断区分均与视频一致;上一轮问题已修正,未发现需改动的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:21:17Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard--skkcrxh/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard--skkcrxh/04_review_r1" + }, + { + "stage": "reconcile", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard--skkcrxh/05_reconcile_r1" + }, + { + "stage": "traj_review", + "round": 2, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard--skkcrxh/04_review_r2" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "9763c75982a8ec59021d2f84adc039829939ed7524a521f3d1d38fd50a9a68fc", + "video_sha256": "dd34a734b50dc9dbf5d627445c9b4441cf2a22fbedb5d2680e4f1e6808c97354", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard-realtime--4txhevf.json b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard-realtime--4txhevf.json new file mode 100644 index 0000000000000000000000000000000000000000..83ada9cf378cd0af434aa62c160837cb84c7fd0d --- /dev/null +++ b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard-realtime--4txhevf.json @@ -0,0 +1,203 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-powerlift-i00-standard-realtime--4txhevf", + "taskKey": "humanoidbench/manip_powerlift", + "dag_sha256": "388cfca325afca3d4eb2a3cee3aacbb2213d1a0ad552b6c823429491cd5d2e19", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 1000, + "sim_steps": 15962, + "calls": 1916, + "rows": 19352, + "resets": 25, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 2215.7, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 318.841, + "max_lag_s": 0.0333, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 0, + "late_steps": 0, + "requests_refused": 25, + "request_busy_s": 2.035, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-powerlift-i00-standard-realtime--4txhevf/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "1c92919729d268c323e197929079ecd642df5d78d3a62d45c9dd42c84411f577" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "dumbbell_raised_and_held": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00 s 哑铃两端配重落在地面,机器人双足站立、双手空着;视频 0.40–0.80 s 机器人屈膝并摆动双臂,未见手指合拢包握哑铃杆。" + }, + { + "t": 1.2, + "log_step": null, + "text_zh": "观察:视频 0.85–1.20 s 机器人上身向哑铃方向倾倒,双脚失去原有支撑,身体横倒在哑铃附近;哑铃未随双手举到高处。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 2.50–10.05 s 机器人一直倒卧,哑铃配重仍在地面附近;已逐帧检查视频 9.00–10.05 s,未见起身或持铃举高保持。" + }, + { + "t": null, + "log_step": 59, + "text_zh": "推断:log step 58 启动最后的 run 13,log step 59 明确自述最后评分回合为 run 13 且失败;评分器 reward.live_success=0、graded_by_live=1。结合视频全过程未出现举高保持,判定该累计验收节点未达到;reward 未报告 episode_return 数值。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.80 s,机器人双足与地面接触,腿部在屈膝和摆臂准备举重时承重;视频 1.20 s 后身体倒卧并由地面承托。未见双手包握哑铃,不将预设的双手抓握方案计作实际接触。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "dumbbell_raised_and_held", + "t": 1.0, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频 0.40–0.80 s 机器人已屈膝、摆臂准备举重;视频 0.85–1.20 s 双臂甩动,上身持续向哑铃方向倒下,双脚失去支撑。此后直到视频 10.05 s 都未恢复站立或举高哑铃,因此断在正在执行的哑铃举高保持步骤。" + }, + "unmet": { + "dumbbell_raised_and_held": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 1.0, + "log_step": null, + "evidence_zh": "观察:视频 0.85–1.20 s 机器人由屈膝站姿转为上身前倾、双脚离开原支撑位置并倒地;视频 1.20–2.40 s 四肢继续摆动,随后保持倒卧,视频 10.05 s 仍未起身。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志", + "evidence_zh": "推断:log step 51 的控制器修改将 fb_on 设为 0,log step 55 与 log step 58 继续修改关节关键帧、膝踝前馈补偿及屈膝到起身的时序,最后启动 run 13;log step 58 的回传中重力方向分量 g 很快接近 +0.9 并维持,log step 59 亦自述该回合倒地。视频 0.85–1.20 s 直接显示动作执行过程中支撑失稳,说明这套关节目标与补偿未能维持举重准备动作的全身稳定性;现有证据只支持控制层归因。" + } + ], + "gap_zh": "没有状态日志、接触对、接触力或逐步奖励累计值;细微手物碰触及具体失稳机制无法核实,episode_return 留空。log step 58 的回合控制时间延伸至仿真约 15.60 s,评分器 n_actions=1000、任务控制周期为 0.02 s,而视频仅 10.10 s;未建立可靠的控制步到视频时钟映射,所有导航时刻均直接取自视频。agent 在 log step 59 关于伺服下沉、延迟和踝关节饱和的解释未由本回合状态日志独立验证,不细分控制根因。" + }, + "summary_zh": "评分回合中机器人开始屈膝和摆臂准备举重,但视频约 1.00–1.20 s 即失衡倒地,随后一直倒卧至视频 10.05 s;哑铃举高保持未达到,最终为 0/1。最后控制程序的关节目标和前馈补偿未能维持全身稳定,根因标为控制;没有状态日志或累计奖励值,无法进一步核实失稳机制及精确分数。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "视频 0–10.05 s 与标注一致:机器人在约 0.8–1.2 s 失衡倒地,哑铃始终在地面,节点未达成,断点与平衡/控制归因有证据支持。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:13:32Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard-realtime--4txhevf/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard-realtime--4txhevf/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "115ba29be6ed569ccd783173a6263ee8b31fda1fa1213d888e245e426d445feb", + "video_sha256": "1c92919729d268c323e197929079ecd642df5d78d3a62d45c9dd42c84411f577", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard-realtime--a2w6wz6.json b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard-realtime--a2w6wz6.json new file mode 100644 index 0000000000000000000000000000000000000000..61f7626204faaee50842abb8a45da0eb68933974 --- /dev/null +++ b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard-realtime--a2w6wz6.json @@ -0,0 +1,217 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-powerlift-i00-standard-realtime--a2w6wz6", + "taskKey": "humanoidbench/manip_powerlift", + "dag_sha256": "388cfca325afca3d4eb2a3cee3aacbb2213d1a0ad552b6c823429491cd5d2e19", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 359, + "sim_steps": 29204, + "calls": 34361, + "rows": 34302, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3365.7, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 583.429, + "max_lag_s": 0.0863, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 2816, + "late_steps": 0, + "requests_refused": 50, + "request_busy_s": 26.81, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.158, + "peak_processes": 11, + "peak_disk_gb": 0.957, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-powerlift-i00-standard-realtime--a2w6wz6/replay.mp4", + "duration_s": 3.65, + "fps": 20.0, + "frames": 73, + "sha256": "b1ebf3dc792cdbd89b710967add76793cd1a95a6837f3f8958f118d901347fd3" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "dumbbell_raised_and_held": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.8, + "log_step": null, + "text_zh": "观察:视频 0–0.8 s,机器人由站姿下蹲,双手伸向地面哑铃的杆部;视频 0.8–2.2 s,双手在杆部两处附近合拢、随哑铃在地面的小幅滚移调整位置,未见哑铃随手升离地面。" + }, + { + "t": 2.3, + "log_step": null, + "text_zh": "观察:视频 2.2–2.7 s,机器人起身并收抬双臂,双手与哑铃分离,哑铃仍由两端配重支承在地面;视频 2.7–3.6 s,机器人空手抬臂,随后向后倒下。全片未出现举高并保持哑铃。" + }, + { + "t": 3.6, + "log_step": 191, + "text_zh": "观察:log step 191 中机器人服务报告最后一回合 episode=51、control_steps=359、done=true、reason=failed、resets_left=0;评分器字段 live_success=0、graded_by_live=1。推断:该单节点绑定的累计奖励验收未达成;未将此前重置回合的动作计入节点。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "推/扫", + "note_zh": "推断:视频 0.8–2.2 s,左右灵巧手同时在哑铃杆部两处接触尝试,手部动作伴随哑铃在地面滚移,按双手推拨记录;视频 2.2–2.4 s 双手离开而哑铃留在地面。画面不能核实手指是否曾形成对向夹持,故不记抓握。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0–3.2 s,机器人双脚接地,腿和躯干支承下蹲、起身及抬臂动作;视频 3.35–3.6 s 向后失衡,骨盆降向地面。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "dumbbell_raised_and_held", + "t": 2.25, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频 2.2–2.4 s,机器人由蹲姿开始起身,双手离开哑铃而未将它带起;到视频 3.6 s 仍未重新抓取或举高,且机器人已向后倒下。推断:失败发生在正在执行的哑铃举高保持步骤,空手起身后的举重失败未被修复。" + }, + "unmet": { + "dumbbell_raised_and_held": false + }, + "phenomena": [], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "推断:视频 0.8–2.4 s,机器人双手已到哑铃杆部附近,却未能使哑铃随起身动作升起;视频 3.35–3.6 s 又未能维持身体支撑。log step 186 记录 agent 的末轮控制脚本继续调整手高、手指屈曲与拇指目标并执行举重序列,log step 191 的服务结果确认最后一回合失败;这些证据支持抓取及全身动作控制未奏效,但不足以确定具体接触几何、逆解或反馈参数是哪一项错误。", + "source_zh": "视频 / 日志" + } + ], + "gap_zh": "观察:inputs/ 只有运行记录和外部视角视频等输入,没有状态日志或成功示范;评分器未报告 episode_return 数值。视频手部像素较少,杆部局部遮挡;视频 2.2–2.4 s 可见双手空载离开,但不能可靠区分此前从未形成对向抓握,还是形成短暂抓握后立即滑脱,故按 docs/03_trajectory.md Step 3.2 及 TAXONOMY_v2.md §3 将现象层留空,也不细分控制根因。视频 3.45–3.6 s 的后倒晚于视频 2.2–2.4 s 已出现且未修复的空手起身失败,不以这一末段事件替代断点上的接触失败现象;倒地经过保留在断点证据中(DECISIONS.md S14/S15)。log step 191 的359个控制步按任务50 Hz为仿真7.18 s,与视频3.65 s不等长;所有视觉锚点直接按视频帧取值,未将控制步或agent时间换算为视频时间。" + }, + "summary_zh": "观察:评分的最后一回合中,机器人下蹲并以双手接近、推动哑铃,随后空手起身和抬臂;哑铃一直留在地面,哑铃举高保持节点未完成,最终进度为0/1。观察:视频约 3.45 s 机器人向后失衡倒下;推断:抓取和全身控制未奏效,但视频不足以区分未形成抓握与瞬间滑脱,故现象层留空,末段倒地只作为经过说明;无状态日志和累计奖励数值可核对。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 2, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "fix", + "conf": 6, + "summary_zh": "节点(未达成,0/1)、断点(哑铃举高保持,操作中)和时钟处理都与视频相符。问题在现象层:唯一的现象“平衡”是回合末的倒地,而倒地是骨盆过低终止的结果(S14/S15),不应单独作为现象;真正失败的动作(双手没能把哑铃带离地面)反而没有现象。", + "n_issues": 1, + "n_major": 1 + }, + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 2, + "verdict": "agree", + "conf": 7, + "summary_zh": "节点未达成(0/1)、断点(哑铃举高保持,操作中)与视频一致;哑铃全程留在地面,末段空手抬臂后向后倒下。现象层按证据不足留空并写明 gap_zh,根因停在控制,均符合文档规则。第1轮的问题已处理。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:22:00Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard-realtime--a2w6wz6/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard-realtime--a2w6wz6/04_review_r1" + }, + { + "stage": "reconcile", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard-realtime--a2w6wz6/05_reconcile_r1" + }, + { + "stage": "traj_review", + "round": 2, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard-realtime--a2w6wz6/04_review_r2" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "e40a2cfc6a105701ef9c010ec95bd46d4c73fa0e2e35a9fb448de69a41e1379f", + "video_sha256": "b1ebf3dc792cdbd89b710967add76793cd1a95a6837f3f8958f118d901347fd3", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard-realtime--eufepzd.json b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard-realtime--eufepzd.json new file mode 100644 index 0000000000000000000000000000000000000000..9b27fca85f6acf72a0ec51c1558624d68abfbf32 --- /dev/null +++ b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard-realtime--eufepzd.json @@ -0,0 +1,229 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-powerlift-i00-standard-realtime--eufepzd", + "taskKey": "humanoidbench/manip_powerlift", + "dag_sha256": "388cfca325afca3d4eb2a3cee3aacbb2213d1a0ad552b6c823429491cd5d2e19", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 112, + "sim_steps": 17489, + "calls": 11130, + "rows": 11068, + "resets": 43, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3514.0, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 349.123, + "max_lag_s": 0.0926, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 6743, + "late_steps": 0, + "requests_refused": 48, + "request_busy_s": 6.689, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-powerlift-i00-standard-realtime--eufepzd/replay.mp4", + "duration_s": 1.2, + "fps": 20.0, + "frames": 24, + "sha256": "0c697794be05a6fc82ebf3d27c84633193da6e4f0abb716a46520a6766f62a18" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "dumbbell_raised_and_held": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.35, + "log_step": null, + "text_zh": "观察:视频 0.00–0.35 s,机器人从站姿屈膝下蹲,左右手同时向地面哑铃的杆身伸下,已开始举重步骤。" + }, + { + "t": 0.85, + "log_step": null, + "text_zh": "观察:逐帧检查视频 0.35–0.95 s,双手到达杆身附近并随机器人起身移动,哑铃在脚前低处发生位移,未出现举到躯干或头顶高度并保持的过程。" + }, + { + "t": 1.15, + "log_step": 666, + "text_zh": "观察:视频 1.00–1.15 s,机器人躯干前倾,向前(朝哑铃杆方向)扑倒,末帧俯卧、双腿翘起,哑铃仍在地面附近;log step 666 的机器人服务状态为 episode=44、control_steps=112、done=True、reason=failed,评分字段 success=live_success=replay_success=0。" + }, + { + "t": null, + "log_step": null, + "text_zh": "推断:可见举重过程未实现高处保持,且本任务唯一目标绑定累计奖励验收,评分器判失败,因此该事件节点未达到;不把会话中的其他重置回合计入节点进度。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "未知", + "note_zh": "观察:视频 0.35–0.90 s,左右手同时下探哑铃杆身并在杆旁合拢,哑铃随起身动作在地面附近移动;256×256画面中手指与杆身相互遮挡,无法可靠辨别是否形成对向夹握,故不确定抓握、推拨或压触的具体方式。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.90 s,机器人双脚接触地面,腿和躯干支撑下蹲及起身动作;视频约 0.95 s 后失去直立支撑并向前扑倒。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "dumbbell_raised_and_held", + "t": 0.95, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频 0.35–0.90 s,机器人已下蹲伸手到哑铃杆身并尝试起身;视频约 0.90–0.95 s 起身体前倾失衡,随后向前扑倒、俯卧落地,至视频 1.15 s 未恢复,也未完成举高保持。" + }, + "unmet": { + "dumbbell_raised_and_held": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 0.95, + "log_step": null, + "evidence_zh": "观察:视频 0.90–1.15 s,机器人在哑铃旁起身时向前扑倒,双脚离开原支撑位置,俯卧着地、双腿翘起;该跌倒发生在举重操作中。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "t": 0.95, + "log_step": 661, + "source_zh": "视频 / 日志", + "evidence_zh": "推断:视频 0.70–1.15 s 可见机器人尝试伸腿起身却未保持支撑,最终向前扑倒,表明该举重动作的全身控制未能稳定机器人。会话证据中,log step 658 的控制脚本使用关节姿态插值和髋部姿态反馈,log step 661 执行时报告 gx=0.696 而 tgt_gx=0.004,随后服务以 failed 结束;这些记录支持控制执行存在明显姿态偏差,但不足以确定接触设计、反馈律或某个参数是哪一具体根因。" + } + ], + "gap_zh": "输入无状态日志和成功参考视频,评分字段未报告 episode_return 的数值,不能重建精确累计奖励或骨盆终止高度;手指与哑铃杆在低分辨率画面中重叠,无法可靠区分夹握与压推,故接触方式保留未知,也不标没建立接触或没保持住。log step 661 为最后一次已执行控制,log step 666 确认其结束;log step 664 之后只有脚本修改、语法检查和状态查询,未见再次执行。服务的 112 个控制步按日志 control_dt=0.02 s 为 2.24 s,与视频 1.20 s 不符,未建立控制步到视频时间的映射;节点和断点时间均取自视频帧,日志仅作会话及结果证据,控制子类仍不能确定。" + }, + "summary_zh": "机器人下蹲、双手伸到哑铃杆身并尝试起身,未完成“哑铃举高保持”,最终节点为0/1。视频约 0.95 s 在举重操作中失衡向前扑倒,结合控制日志中的姿态偏差,根因标为控制。缺少状态与累计奖励记录,手部接触方式及控制步与视频时间的对应关系无法精确确定。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 2, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "fix", + "conf": 8, + "summary_zh": "节点0/1、断点 dumbbell_raised_and_held、平衡现象与控制根因的类别均成立,但文件把倒地方向写反:视频 0.85–1.15 s 机器人是躯干前倾、脸朝下向前扑倒(末帧俯卧、双腿翘起在身后),不是“向后倒、背部着地”;与日志中 gx 持续约 0.6–0.85(躯干前倾)一致。需改正多处观察性描述。", + "n_issues": 4, + "n_major": 3 + }, + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 2, + "verdict": "agree", + "conf": 8, + "summary_zh": "第1轮指出的倒地方向已改正;节点0/1、断点 dumbbell_raised_and_held(操作中)、平衡现象与控制根因(inference 标注、gap 说明)与视频一致,未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:21:03Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard-realtime--eufepzd/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard-realtime--eufepzd/04_review_r1" + }, + { + "stage": "reconcile", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard-realtime--eufepzd/05_reconcile_r1" + }, + { + "stage": "traj_review", + "round": 2, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard-realtime--eufepzd/04_review_r2" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "a033fb53927aeac121d245ee2323c8cd9a0f9539ddacf5f8e7569a610bbf3932", + "video_sha256": "0c697794be05a6fc82ebf3d27c84633193da6e4f0abb716a46520a6766f62a18", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard-realtime--qn35emf.json b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard-realtime--qn35emf.json new file mode 100644 index 0000000000000000000000000000000000000000..2ae4c01b694925ae17ac2e09f1b4d4204b8f6591 --- /dev/null +++ b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard-realtime--qn35emf.json @@ -0,0 +1,212 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-powerlift-i00-standard-realtime--qn35emf", + "taskKey": "humanoidbench/manip_powerlift", + "dag_sha256": "388cfca325afca3d4eb2a3cee3aacbb2213d1a0ad552b6c823429491cd5d2e19", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 1000, + "sim_steps": 20267, + "calls": 27201, + "rows": 17087, + "resets": 24, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 2345.4, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 404.954, + "max_lag_s": 0.0297, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 6609, + "late_steps": 0, + "requests_refused": 0, + "request_busy_s": 11.228, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.212, + "peak_processes": 10, + "peak_disk_gb": 0.523, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-powerlift-i00-standard-realtime--qn35emf/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "c42ac1e6f5c5e8b31a609d214e4852895249064025ee59abfeeb8f684972c11d" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "dumbbell_raised_and_held": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.0 s 机器人站在哑铃后方,哑铃两端配重落地;视频 1.0–3.5 s 机器人屈膝俯身,双手伸向杆部。" + }, + { + "t": 4.2, + "log_step": null, + "text_zh": "观察:视频 3.5–4.7 s 双手在杆部附近上移,哑铃沿地面滚动,配重未随双手抬离地面;视频 5.0–8.5 s 机器人起身并把空手举过头顶,哑铃仍在脚前地面。" + }, + { + "t": 10.05, + "log_step": 90, + "text_zh": "观察:视频 8.9–9.4 s 机器人向前倾倒,视频 9.5–10.05 s 俯卧在地,哑铃仍落地;log step 90 中 agent 也自述未举起哑铃,但其“全程未摔倒”的说法与视频末段不符。" + }, + { + "t": null, + "log_step": null, + "text_zh": "推断:视频全程未见哑铃被举高或持续持举,且评分器 reward.success、live_success、replay_success 均为0,因此该累计奖励验收节点未达到;评分器未报告 episode_return 的具体数值。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "推/扫", + "note_zh": "观察:视频 3.5–4.2 s 左右手共同在哑铃杆附近弯指、向上移动,哑铃被拨动后沿地面滚开;未见两手各自形成夹住杆部的可靠对向接触,不据闭指动作标为抓握。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–8.8 s 双脚支撑机器人完成下蹲、起身和举臂;视频 9.4–10.05 s 摔倒后躯干落地。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "dumbbell_raised_and_held", + "t": 4.2, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频 3.5–4.7 s 机器人尝试抬手带起哑铃,但哑铃仍由地面支承并滚开;视频 5.0–8.5 s 机器人空手起身、举臂,此次持举失败始终未修复,视频 8.9–9.4 s 又向前摔倒。" + }, + "unmet": { + "dumbbell_raised_and_held": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "没建立接触", + "t": 4.2, + "log_step": null, + "evidence_zh": "观察:视频 3.5–4.7 s 机器人双手从杆附近抬起,哑铃却沿地面滚开;视频 5.0–6.5 s 双手与哑铃已明显分离,双手为空。所需承载抓持未建立,虽然手与杆发生过拨碰,也没有先稳定持起再掉落的可见阶段。" + }, + { + "cat": "全身", + "item": "平衡", + "t": 9.2, + "log_step": null, + "evidence_zh": "观察:视频 8.9 s 起机器人举臂前倾,视频 9.1–9.4 s 身体越过双脚支承范围并向前倒下,视频 9.5–10.05 s 躯干俯卧地面,未再站起。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志、视频", + "evidence_zh": "推断:接触执行与回合末段的持续稳定控制未实现。视频 3.5–4.7 s 手部抬升未带动哑铃离地;log step 87 的控制程序安排弯指托举再闭指,其自行打印的关节读数显示抬升阶段手指偏离弯曲姿态,但不足以区分承载几何和执行器能力。另据 log step 16 的循环退出条件及 log step 87 的 T=17.0,控制程序会提前退出;log step 87 输出 END 852 False None,表示程序退出时回合仍未结束,随后 log steps 88–90 未再发送控制,视频 8.9–9.4 s 出现失衡倒地。" + } + ], + "gap_zh": "无状态日志或成功示范;视频为单一256×256外部视角,细小手指接触存在遮挡,无法确认逐指接触对、握力或准确载荷,故不把 agent 在 log step 90 所称“手指电机承载不足”当作已证实机制,控制根因不细分。评分器未提供 episode_return,不能还原精确累计奖励;未将日志控制步或 agent 工作时间换算为视频时间。log step 90 自述全程未摔倒与视频末段矛盾,以视频为准。" + }, + "summary_zh": "评分回合中机器人下蹲接近哑铃,但未建立承载抓持,随后空手起身并举臂,哑铃始终留在地面,“哑铃举高保持”节点完成0/1。视频 8.9–9.4 s 机器人又向前摔倒;日志支持接触执行与持续稳定控制存在问题,但未提供接触力、状态日志或累计奖励,无法确定更细的抓持失效机制。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "节点未达成(哑铃始终在地面)、断点(操作中,约 4.2 s)、没建立接触与摔倒(平衡)均与视频一致;根因停在“控制”并已标推断,日志 step 87 的 END 852 与视频 8.5–9 s 起倒的时间吻合,未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:08:55Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard-realtime--qn35emf/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard-realtime--qn35emf/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "49f599503321a6d3cf9313fa77c75ff8cea6bdeab6e9aa170c256fd46390fe7e", + "video_sha256": "c42ac1e6f5c5e8b31a609d214e4852895249064025ee59abfeeb8f684972c11d", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard-realtime--qzsxmuw.json b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard-realtime--qzsxmuw.json new file mode 100644 index 0000000000000000000000000000000000000000..c91305c01ae891e45ee19fb0eff3d75434e75f43 --- /dev/null +++ b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard-realtime--qzsxmuw.json @@ -0,0 +1,190 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-powerlift-i00-standard-realtime--qzsxmuw", + "taskKey": "humanoidbench/manip_powerlift", + "dag_sha256": "388cfca325afca3d4eb2a3cee3aacbb2213d1a0ad552b6c823429491cd5d2e19", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 68, + "sim_steps": 12669, + "calls": 12123, + "rows": 11984, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 2888.9, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 252.589, + "max_lag_s": 0.0136, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 757, + "late_steps": 0, + "requests_refused": 44, + "request_busy_s": 5.436, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.252, + "peak_processes": 10, + "peak_disk_gb": 0.84, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-powerlift-i00-standard-realtime--qzsxmuw/replay.mp4", + "duration_s": 0.75, + "fps": 20.0, + "frames": 15, + "sha256": "e2fc6cee90487d6d40b40baec99abc4f7beb79b7467ac6767ac0d611d9cbd835" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "dumbbell_raised_and_held": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:逐帧检查视频 0.00–0.70 s 的全部15帧,哑铃始终在机器人前方地面,双手始终高于哑铃杆且未接触它;机器人只在屈膝站立准备后向后倒下,没有举起或高处保持。" + }, + { + "t": 0.7, + "log_step": 241, + "text_zh": "观察:视频 0.70 s 机器人已仰倒在地,哑铃仍在原地。log step 240 开始最后的第51回合;log step 241 的服务状态为 episode=51、control_steps=68、done=true、reason=failed、resets_left=0,评分器 live_success 与 replay_success 均为0。此前重置掉的测试不计入本节点。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "dumbbell_raised_and_held", + "t": 0.35, + "log_step": 240, + "where": "首步前", + "evidence_zh": "观察:视频 0.00–0.35 s 机器人双手停在腰前,未伸到哑铃杆;视频 0.35 s 已明显后倾,视频 0.70 s 仰倒,哑铃全程未离地。log step 240 的控制计划先执行 balance_bias 和 shallow_squat,再执行 forward_reach、grip 和举升;log step 241 确认最后回合在 control_steps=68 时已失败。推断:机器人在站姿/屈膝准备阶段失衡,尚未启动哑铃的抓取举升,按 CASES.md FAIL-13 标为首步前。" + }, + "unmet": { + "dumbbell_raised_and_held": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 240, + "evidence_zh": "观察:视频 0.00–0.70 s 全部帧中,机器人双手均未接触哑铃,哑铃一直落地;机器人仅改变腿部站姿后倒下。推断:最后受评回合未进入抓取举升,log step 240 中排在屈膝之后的伸手、抓握及上举阶段没有执行到;依据 CASES.md FAIL-13,这种举重前仅维持平衡即倒下的轨迹属于没尝试举重。" + }, + { + "cat": "全身", + "item": "平衡", + "t": 0.35, + "log_step": 241, + "evidence_zh": "观察:视频 0.35 s 机器人躯干向后倾斜,视频 0.50–0.65 s 后倾持续增大,视频 0.70 s 背部落地、双臂朝上,未恢复站立;哑铃仍在前方地面。log step 241 的服务返回最后回合已结束且 reason=failed。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "反馈控制", + "evidence_zh": "推断:机器人屈膝准备时的姿态反馈未能抑制后倾,导致举升前倒地。依据为 log step 240:agent 的控制脚本启用踝部前馈与俯仰反馈(kp_pitch=4.0、kd_pitch=0.35、max_pitch_correction=0.15),同一步输出在 shallow_squat 阶段记录俯仰设定值为−0.10 rad、估计俯仰为−0.208 rad、gyro_y 为−0.835;log step 241 的脚本读数进一步为 pitch=−1.571,服务同时报告回合已失败。视频 0.35–0.70 s 直接显示相应的持续后倾直至仰倒。这支持平衡反馈控制未稳住本体,不能据此确定是哪一个增益或前馈参数错误。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "未提供状态日志;评分器未报告累计 episode_return,因此原生指标值留空。最后回合日志记录68控制步,任务文本控制率为50 Hz,但视频只有0.75 s,未建立逐控制步到视频的可靠映射;所有视频时间均直接取自已检查的帧,不把日志控制步换算为视频时间。部分日志输出自身已截断,无法核验每个控制周期或确定具体错误参数。" + }, + "summary_zh": "受评的最后回合未举起哑铃,目标节点最终为0/1;视频全程显示双手未接触哑铃,机器人在屈膝和平衡准备阶段后倾倒地,断点为哑铃举高保持的首步前。现象为没尝试举重和全身失衡,控制记录与画面支持平衡反馈未能稳住本体;先前重置的回合不计入进度。无状态日志和累计奖励,且控制步与短视频未建立可靠时间映射。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "标签与视频一致:0.75 s 的15帧显示机器人屈膝后向后倒下,哑铃始终未被接触;断点为首步前,现象为没尝试加平衡,符合 CASES.md FAIL-13,根因反馈控制有日志支持。未发现需要修改的问题。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:12:27Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard-realtime--qzsxmuw/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard-realtime--qzsxmuw/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "5225adf6b28ebccdcf01e78a73b4780f34b26e6ea6e42f6c22039d585ae59e8d", + "video_sha256": "e2fc6cee90487d6d40b40baec99abc4f7beb79b7467ac6767ac0d611d9cbd835", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard-realtime--sm5tysz.json b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard-realtime--sm5tysz.json new file mode 100644 index 0000000000000000000000000000000000000000..cd9a6a2a6988a8e8668c0291890a39f95efea347 --- /dev/null +++ b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard-realtime--sm5tysz.json @@ -0,0 +1,198 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-powerlift-i00-standard-realtime--sm5tysz", + "taskKey": "humanoidbench/manip_powerlift", + "dag_sha256": "388cfca325afca3d4eb2a3cee3aacbb2213d1a0ad552b6c823429491cd5d2e19", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 1000, + "sim_steps": 11190, + "calls": 10541, + "rows": 10512, + "resets": 28, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 2388.0, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 223.358, + "max_lag_s": 0.0624, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 703, + "late_steps": 0, + "requests_refused": 20, + "request_busy_s": 10.112, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.204, + "peak_processes": 9, + "peak_disk_gb": 0.246, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-powerlift-i00-standard-realtime--sm5tysz/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "475add4bd7b66ec05f0a76830186ff295f7ff7fffabbbf239c76b46b1ea2c000" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "dumbbell_raised_and_held": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.2, + "log_step": null, + "text_zh": "观察:视频 0.8–1.6 s,机器人屈膝向地面哑铃伸出双手,手指在杆上方闭合,未包住杆;哑铃仍在原地。" + }, + { + "t": 4.0, + "log_step": null, + "text_zh": "观察:视频 2.0–4.0 s 机器人起身并抬臂,双手为空,哑铃没有随手臂离地;视频 4.0–10.05 s 哑铃继续留在地面,末帧机器人仍站立、双手举在身前高处。" + }, + { + "t": null, + "log_step": 161, + "text_zh": "观察:log step 160 启动最后一次回合,log step 161 的机器人服务状态为 episode 29、control_steps 1000、done True、reason failed;run.json 的评分器字段 success、live_success、replay_success 均为 0,未报告 episode_return。视频未见哑铃举起,与该失败结果一致。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–10.05 s,机器人双脚接触地面,在下蹲、起身和空手抬臂时支撑本体;视频 0.8–1.6 s 双手虽执行闭合动作,但未包住哑铃杆,不记为实际抓握。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "dumbbell_raised_and_held", + "t": 1.2, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频约 1.2 s,机器人正在下蹲取哑铃,双手在杆上方闭合,未夹住杆;视频 2.0–4.0 s 随后起身抬臂时没有带起哑铃,直到视频 10.05 s 仍未修复这次抓取失败。" + }, + "unmet": { + "dumbbell_raised_and_held": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "没建立接触", + "t": 1.2, + "log_step": null, + "evidence_zh": "观察:视频 0.8–1.6 s 的双手与哑铃杆之间有可见间隔,手指闭合成空拳,未形成夹持;视频 3.5–4.0 s 两臂空手上抬,哑铃始终在地面,没有先被抓住再滑落的过程。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "抓取和接触设计", + "alt_feasible": "未知", + "t": 1.2, + "log_step": 160, + "source_zh": "视频 / 日志", + "evidence_zh": "推断:所选下蹲和伸臂姿态使双手闭合位置高于哑铃杆,未形成承重抓握。视频 0.8–1.6 s 可见双手在杆上方空合;log step 153 写入浅蹲、闭指、起身、抬臂的固定关键帧,log step 160 将最后回合恢复到该配置,视频中的动作顺序与之相符。agent 在 log step 162 也自述浅蹲时手指在杆上方或未到杆处闭合;其声称的工作空间硬限制未独立验证,不能据此排除可行替代姿态。" + } + ], + "gap_zh": "输入没有状态日志和成功示范,评分器也未报告 episode_return,无法恢复累计得分或精确手物距离、接触力。log step 161 报告最后回合 1000 个控制步,任务文本给出的控制周期为仿真 0.02 s,而视频长 10.1 s;未建立逐控制步到视频的映射,所有标注时间均直接取自视频帧,未将日志仿真时间换算为视频时间。" + }, + "summary_zh": "机器人完成下蹲、双手闭合、起身和空手抬臂,但哑铃一直留在地面,哑铃举高保持节点未达到,最终进度为 0/1。断点在视频约 1.2 s 的取物操作中:双手在杆上方空合,所选抓取姿态未形成抓握,随后也未修复;评分器判定失败。无状态日志且未报告累计得分,无法量化精确接触几何或奖励差距。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "哑铃全程留在地面,节点未达成、进度0/1、断点与『没建立接触』『抓取和接触设计』均与帧相符;根因已标为推断且声明了未验证的限制。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:09:43Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard-realtime--sm5tysz/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard-realtime--sm5tysz/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "5b8efae4054554a3ef99b1de0e4b711bdc23cf0ef1270fc9d92225f1e8a4a3ef", + "video_sha256": "475add4bd7b66ec05f0a76830186ff295f7ff7fffabbbf239c76b46b1ea2c000", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard-realtime--woveggc.json b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard-realtime--woveggc.json new file mode 100644 index 0000000000000000000000000000000000000000..764fbac23b28e0b3b7c736811e88b5eca725e977 --- /dev/null +++ b/phase2/humanoidbench/manip_powerlift/trajectories/humanoidbench-manip-powerlift-i00-standard-realtime--woveggc.json @@ -0,0 +1,192 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-powerlift-i00-standard-realtime--woveggc", + "taskKey": "humanoidbench/manip_powerlift", + "dag_sha256": "388cfca325afca3d4eb2a3cee3aacbb2213d1a0ad552b6c823429491cd5d2e19", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 109, + "sim_steps": 17911, + "calls": 14762, + "rows": 26748, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1393.6, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 357.564, + "max_lag_s": 0.0265, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 2444, + "late_steps": 0, + "requests_refused": 43, + "request_busy_s": 7.53, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.332, + "peak_processes": 10, + "peak_disk_gb": 1.045, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-powerlift-i00-standard-realtime--woveggc/replay.mp4", + "duration_s": 1.15, + "fps": 20.0, + "frames": 23, + "sha256": "03ce4018e8249408ddf80b04526ff739218f83232634e1c788609f3dd5389324" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "dumbbell_raised_and_held": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00–1.10 s 的全部23帧中,哑铃两端配重始终靠地,未随机器人手臂升高;机器人双手未包握哑铃杆。机器人仅调整站姿,随后向后倾倒,未启动抓举动作,故不记录该节点的抓握接触。" + }, + { + "t": 1.1, + "log_step": 158, + "text_zh": "观察:视频 1.10 s 时机器人已经倒地,哑铃仍留在原处;log step 158 的机器人服务状态为 episode=51、control_steps=109、done=true、reason=failed、resets_left=0,评分器 live_success=0、replay_success=0。节点未达到;判断并非只依据零分。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 800, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "dumbbell_raised_and_held", + "t": 0.2, + "log_step": null, + "where": "首步前", + "evidence_zh": "观察:视频 0.10–0.35 s 机器人躯干逐渐向后倾斜,骨盆降低;视频 0.40–1.10 s 持续失去站立姿态并倒地。双手尚未伸到并握住地面哑铃,整个视频没有抓举,失败发生在举重首步启动前。" + }, + "unmet": { + "dumbbell_raised_and_held": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": null, + "evidence_zh": "观察:视频 0.00–1.10 s 机器人只做站姿调整及跌倒后的肢体运动,没有朝哑铃杆建立抓取或举起动作,双手一直空着,哑铃始终在地面。这里的未尝试指受评回合的抓举动作,不包括此前会话测试。" + }, + { + "cat": "全身", + "item": "平衡", + "t": 0.2, + "log_step": null, + "evidence_zh": "观察:视频 0.15–0.35 s 机器人向后倾斜且骨盆快速下沉;视频 0.40–0.85 s 已无法恢复直立,视频 0.90–1.10 s 躯干倒向地面、双腿翻起,直到结束都未站回。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "t": 0.2, + "log_step": 152, + "source_zh": "日志 / 视频", + "evidence_zh": "推断:agent 在 log step 81 编写由骨盆重力方向和角速度驱动髋、踝目标的平衡控制,log step 152 最后一个候选站姿测试只记录到 control step 108,倾角约82.1度;log step 158 服务确认最后回合结束于 control step 109。结合视频 0.15–1.10 s 站姿调整后持续倾倒,支持站立控制未能维持支撑,导致尚未抓举就失败;不把其他回合的稳定站立或抓举测试计入本回合,也不据此认定某个具体增益有错。" + } + ], + "gap_zh": "无状态日志、无成功示范,评分器未报告 episode_return,不能给出累计奖励或精确手脚接触对。log step 154、156 读取的 graded 文件仍是 control step 61、resets_left=7 的旧回合,与 log step 158 服务给出的最终 episode=51、control_steps=109 不同,因此无法核实 agent 最后自述的控制配置;原始逐步控制文件未随输入提供。按日志控制周期0.02 s,109步为仿真 2.18 s,而视频时长仅1.15 s,缺少可靠映射,所有导航时刻均直接读取视频帧,未用控制步或agent时间换算。" + }, + "summary_zh": "受评视频中机器人未启动抓举,哑铃始终在地面,“哑铃举高保持”最终为0/1。机器人在视频约0.20 s 开始明显向后失衡,随后倒地,断点为该节点的首步前,根因仅归到站立控制失败。没有状态日志或累计奖励值,且最后启动命令读取了旧回合日志,具体受评控制配置与仿真到视频的时钟映射无法核实。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "视频全程23帧机器人向后倾倒、哑铃始终在地面,节点未达、断点首步前、没尝试+平衡、根因控制均有证据支持;时钟未换算,未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:08:20Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard-realtime--woveggc/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-powerlift-i00-standard-realtime--woveggc/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "bf6ce04ee450c553fa2faa124582aca9c403bb55b5cf02cbe34826048c1e1f3a", + "video_sha256": "03ce4018e8249408ddf80b04526ff739218f83232634e1c788609f3dd5389324", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-privileged--8krmvxv.json b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-privileged--8krmvxv.json new file mode 100644 index 0000000000000000000000000000000000000000..0f95a559c29ccfa0b2487697652144d9627f5465 --- /dev/null +++ b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-privileged--8krmvxv.json @@ -0,0 +1,201 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-room-i00-privileged--8krmvxv", + "taskKey": "humanoidbench/manip_room", + "dag_sha256": "081a0a8f4158f72c3be894442d203666d7e71a9e8771af0ea997f42197bdbd1c", + "outcome": { + "success": 0, + "reward": 0.19459464323402992, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 1000, + "final_reward": 0.19459464323402992, + "sim_steps": 1000, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-room-i00-privileged--8krmvxv/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "03abbaeb7d014dd4e59792f897f5cde725d3875145629ab0eca49850d6163a22" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "objects_gathered": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.0 s 的机器人双脚着地、膝部弯曲、双臂停在身前;全段粗采样及视频 9.0–10.05 s 的逐帧检查显示它一直在原地微调站姿,没有走向散物、抓取、推拢或搬移动作,也没有跌倒。相机主要覆盖机器人,不能直接核实六件物品的全景分布。" + }, + { + "t": 10.05, + "log_step": 166, + "text_zh": "观察:log step 166 中,agent 用最终参数重新生成并保存 1000×61 动作的交付轨迹,重放返回 deterministic=true、sim_steps=1000;视频 10.05 s 仍是原地站姿,与交付脚本的站立行为一致。" + }, + { + "t": null, + "log_step": null, + "text_zh": "推断:评分器字段 success=0、all_success=0、deterministic=1、grader_error=0,结合 DAG 将唯一目标直接绑定累计奖励达到 400 的事件,判定本节点全程未达标;这不是依据分数为零推断未操作,未启动归拢的判断来自视频动作与最终交付脚本(log step 160、166)。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "unit": "奖励", + "threshold": 400, + "direction": "higher_better", + "value": null + }, + { + "name": "max_xy_variance", + "unit": "平方米", + "threshold": null, + "direction": "lower_better", + "value": null + }, + { + "name": "room_object_organized", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "stand_reward", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "small_control", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "per_timestep_reward", + "unit": "奖励", + "threshold": null, + "direction": "higher_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "objects_gathered", + "t": 0.0, + "log_step": 166, + "where": "首步前", + "evidence_zh": "观察:视频 0.0–10.05 s,机器人仅在原地维持站姿,归拢所需的接近、抓取或推移动作没有开始。log step 160 展示的交付脚本仅调节踝、髋、膝和躯干以保持初姿;log step 166 用最终参数生成并验证该轨迹,log step 169 的最终报告也自述物品未动、未能获得行走控制器。" + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 166, + "evidence_zh": "观察:视频 0.0 s、全段粗采样以及视频 9.0–10.05 s 的末秒逐帧检查均显示机器人原地站立,双手没有向物品伸取或推拢,双脚没有启动接近物品的行走;log step 166 交付的是只保持初始站姿的最终脚本所录轨迹。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志 / agent 自述", + "evidence_zh": "推断:会话控制试验未形成可持续行走的控制器,导致最终只交付站立轨迹。log step 83 的步态试验在控制步 100 后骨盆水平位置从约 1.56 m 仅增至控制步 375 的约 1.82 m,同时骨盆高度降至约 0.72 m;这些是会话试验数据,不是视频时刻或评分回合的跌倒。agent 在 log step 169 明确自述未能建立行走策略,因而改为最大化站立项;log step 160 的最终脚本和 log step 166 的保存记录支持这条因果链。现有证据不足以进一步区分轨迹生成、反馈或参数问题。" + } + ], + "gap_zh": "观察:inputs/ 只有运行记录、元数据和回放视频,没有状态日志或成功示范;视频相机主要覆盖机器人,不能逐件核查房间全景与接触对。评分器只给出 final_reward=0.19459464323402992 及通过标志,没有报告 DAG 所列的累计奖励、最大水平坐标方差及各奖励分项,故六项原生指标均留 null;不把会话脚本计算的两轴方差均值 4.102 当作评分器的 max_xy_variance,也不将 final_reward 擅自转换成累计奖励。归拢从未启动,因此省略 contact_actual;画面中的双脚静态支撑不算该步的实际操作接触。", + "unmet": { + "objects_gathered": false + } + }, + "summary_zh": "评分回放中机器人全程原地维持站姿,没有启动六件物品归拢,唯一累计验收节点未达成,最终进度为 0/1。断点为 objects_gathered 的首步前,现象为没尝试;会话步态试验及最终交付记录支持控制能力不足导致只交付站立控制,不能把会话中的失稳标成评分回合跌倒。没有状态日志或房间全景视图,无法精确核算累计奖励和各分项指标。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "回放中机器人全程原地站立,未接近或归拢任何物品,objects_gathered 未达成、首步前/没尝试、根因控制(推断,来源 log step 83、169)均有证据支持;时钟均为视频或 log 步,未发现需改动之处。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:52:45Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-privileged--8krmvxv/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-privileged--8krmvxv/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "0ed19c31fc490218df505e33f615cc8c83d029259990b34e7c183b1219805c87", + "video_sha256": "03abbaeb7d014dd4e59792f897f5cde725d3875145629ab0eca49850d6163a22", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-privileged--hwdasxt.json b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-privileged--hwdasxt.json new file mode 100644 index 0000000000000000000000000000000000000000..13796fb13ef345ad7902dc5becd07a393d56ef18 --- /dev/null +++ b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-privileged--hwdasxt.json @@ -0,0 +1,212 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-room-i00-privileged--hwdasxt", + "taskKey": "humanoidbench/manip_room", + "dag_sha256": "081a0a8f4158f72c3be894442d203666d7e71a9e8771af0ea997f42197bdbd1c", + "outcome": { + "success": 0, + "reward": 0.13413108285674308, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 1000, + "final_reward": 0.13413108285674308, + "sim_steps": 1000, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 2.091, + "peak_processes": 10, + "peak_disk_gb": 0.833, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-room-i00-privileged--hwdasxt/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "8201cbc9ec9995d657f0c940b8f71ebd7ed351f6f7ed85caf5be96b8e928f4c6" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "objects_gathered": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.0–2.0 s 密查帧中,H1 在原地屈膝、前后调整躯干和脚部支撑,双手空置,未见朝散物伸手、抓取、搬运或推拢动作。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频全段以约 0.5 s 间隔检查,视频 9.0–10.05 s 逐帧检查;机器人保持近原处的屈膝站姿,双手仍空置,未出现物体搬移或归拢动作,也未见倒地。镜头集中在机器人,未提供六件散物的全景。" + }, + { + "t": null, + "log_step": 112, + "text_zh": "观察:log step 112 打印的交付生成器使用骨盆俯仰、俯仰速度和前后位置反馈修改双腿目标,其余身体关节保持初始目标;log step 164 的交付说明仍将其描述为 1000 步站姿控制器。" + }, + { + "t": null, + "log_step": 170, + "text_zh": "观察:log step 170 中 agent 明确保留对称反馈站姿轨迹并重新生成、调用重放;log step 173 自述物品仍散开,未声称归拢成功。该交付内容与视频中的原地站姿相符。" + }, + { + "t": null, + "log_step": null, + "text_zh": "推断:评分器字段 success=0、all_success=0、deterministic=1、sim_steps=1000、grader_error=0,结合本节点绑定整段累计奖励达到 400 的验收,判定该累计事件未达到。reward.final_reward=0.13413108285674308 是末步奖励(log step 127 的 status 定义将其取自 per_timestep_reward),不能当作累计回报。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "unit": "奖励", + "threshold": 400, + "direction": "higher_better", + "value": null + }, + { + "name": "max_xy_variance", + "unit": "平方米", + "threshold": null, + "direction": "lower_better", + "value": null + }, + { + "name": "room_object_organized", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "stand_reward", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "small_control", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "per_timestep_reward", + "unit": "奖励", + "threshold": null, + "direction": "higher_better", + "value": 0.13413108285674308 + } + ], + "failure": { + "breakpoint": { + "ref": "objects_gathered", + "t": 0.0, + "log_step": 170, + "where": "首步前", + "evidence_zh": "观察:视频 0.0–2.0 s 只见机器人原地调整支撑,随后至视频 10.05 s 保持屈膝站姿,没有启动抓取、搬运或推拢物品;log step 170 明确保留并重新生成站姿轨迹,log step 173 自述散物仍未归拢。评分回合未包含实际归拢动作。" + }, + "unmet": { + "objects_gathered": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 170, + "evidence_zh": "观察:视频 0.0–10.05 s 机器人只在原处调整并维持支撑,未见针对散物的接近、抓取或推拢;log step 112 的生成器为站姿反馈,log step 164 的交付说明和 log step 170 的最终选择仍为该站姿控制器。归拢步骤在评分轨迹中没有启动。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "log_step": 31, + "source_zh": "日志", + "evidence_zh": "推断:会话试验中,agent 在 log step 31 将交替髋、膝、踝运动叠加到站姿反馈,七个非零迈步候选均在 52–76 个仿真控制步结束,骨盆高度降至约 0.27–0.30 米,物品水平位移近零;log step 38 的单腿迈步候选也在 155–199 个仿真控制步结束。agent 在 log step 51 自述迈步使机器人在到达物品前失稳;log step 107、170 据此保留站姿轨迹。支持移动与平衡控制未解决,使交付停留在站姿基线;这些会话试验仅用于根因,未当作评分视频中的跌倒现象。" + } + ], + "gap_zh": "视频主要显示机器人,六件散物没有同时进入可辨识全景;没有状态日志和成功参考视频,无法独立复核六物坐标、接触对及方差变化。评分器未报告 episode_return、max_xy_variance、room_object_organized、stand_reward 或 small_control,故这些原生指标留 null;累计门槛未达由评分器失败结果与节点验收绑定推断。部分交付重放调用输出为空,哈希相同和物品散布代理值来自 agent 自述,未作为状态日志或原生指标。控制根因由会话试验及最终交付选择推断,尚不能确定更细控制子类。" + }, + "summary_zh": "评分轨迹只完成原地站姿调整并维持支撑,未启动六件物品归拢,节点最终为 0/1,评分器判定未通过累计奖励门槛。断点为归拢步骤的首步前,现象为没尝试;会话中的迈步试验反复失稳,agent 最终保留站姿基线,支持控制根因。视频未覆盖六物全景且没有状态日志或累计回报数值,物品具体接触与指标无法独立复核。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "视频全程只见 H1 原地屈膝站姿,无抓取/推拢;断点首步前+没尝试,控制根因由会话试验推断且已标明,与 log step 31/38/51/107/164/170/173 及评分器 success=0 相符。未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:49:58Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-privileged--hwdasxt/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-privileged--hwdasxt/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "5ae34799e84555f6b65445ab5d55bcbe0b5ffd64aacdf8275018dd4a271905bd", + "video_sha256": "8201cbc9ec9995d657f0c940b8f71ebd7ed351f6f7ed85caf5be96b8e928f4c6", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-privileged--mez9qpe.json b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-privileged--mez9qpe.json new file mode 100644 index 0000000000000000000000000000000000000000..d37cbb522a4d589de1a3ce395186faa79f62c71d --- /dev/null +++ b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-privileged--mez9qpe.json @@ -0,0 +1,201 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-room-i00-privileged--mez9qpe", + "taskKey": "humanoidbench/manip_room", + "dag_sha256": "081a0a8f4158f72c3be894442d203666d7e71a9e8771af0ea997f42197bdbd1c", + "outcome": { + "success": 0, + "reward": 0.14575902334907173, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 1000, + "final_reward": 0.14575902334907173, + "sim_steps": 1000, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 6.702, + "peak_processes": 25, + "peak_disk_gb": 0.239, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-room-i00-privileged--mez9qpe/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "6c856a07543b84a53befd949c4702ce0f25337c15cdfb85f654c4bdff30ed4e6" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "objects_gathered": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.0–10.05 s 按约 0.25 s 间隔检查,机器人一直在同一地砖附近屈膝支撑,双臂调整后保持在身旁或身前,没有走向物品、俯身取物或持续推拢的动作;视频 9.0–10.05 s 逐帧检查仍如此,也没有跌倒。画面以机器人为中心,未提供六件物品的完整全景。" + }, + { + "t": null, + "log_step": 51, + "text_zh": "观察:log step 51 重新生成最终交付轨迹,记录 1000 个动作,重复重放的末态哈希均为 1f2bf4997b79;log step 54 再次验证同一哈希。log step 38、44 的生成脚本只执行本体姿态反馈与原地周期性腿部调整,没有物品目标或搬移阶段,和视频中的原地站立一致。" + }, + { + "t": null, + "log_step": null, + "text_zh": "推断:run.json 的评分器字段 success=0、all_success=0、deterministic=1、grader_error=0;本任务唯一目标节点以累计奖励达到 400 验收,因此该累计事件未达到,区间为空、终值为 false。此结论按评分验收判定;未尝试的判断另有视频和生成脚本依据。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "unit": "奖励", + "threshold": 400, + "direction": "higher_better", + "value": null + }, + { + "name": "max_xy_variance", + "unit": "平方米", + "threshold": null, + "direction": "lower_better", + "value": null + }, + { + "name": "room_object_organized", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "stand_reward", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "small_control", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "per_timestep_reward", + "unit": "奖励", + "threshold": null, + "direction": "higher_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "objects_gathered", + "t": 0.0, + "log_step": 51, + "where": "首步前", + "evidence_zh": "观察:视频 0.0–10.05 s 机器人只在原处屈膝站立及调整双臂,始终未开始面向物品的接近、抓取或推拢。log step 51、54 确认交付的是这条 1000 动作的确定性站立轨迹;agent 在 log step 53 的说明中明确承认物品没有归拢。首步未启动,视频 0.0 s 作为导航锚点。" + }, + "unmet": { + "objects_gathered": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 51, + "evidence_zh": "观察:视频 0.0、1.0、2.0、4.0、6.0、8.0 s 及视频 9.0–10.05 s 的密集帧均显示机器人停留在原地,只有站姿与手臂调整,没有启动任何取物、搬移或推拢动作;log step 38、44 的最终生成脚本只含站姿控制。会话中的早期跌倒测试不属于这段交付视频。" + } + ], + "root_causes": [ + { + "cat": "认知", + "item": "推理/规划", + "source_zh": "日志", + "evidence_zh": "推断:log step 37、43 的搜索代码把适应度设为每步存活奖励减去俯仰、侧倾、速度和相对出生点位移的惩罚,没有六件物品坐标或归拢收益;log step 46 沿用此目标继续优化,log step 49、51 根据该适应度提高而选定并交付轨迹。该代理目标鼓励停在出生点,无法引导机器人完成归拢;视频 0.0–10.05 s 的原地站立符合这一结果。归因依据实际优化目标与策略选择,而非仅依据 agent 停止或预算不足;log step 53、55 也承认最终方案只保留站立收益。" + } + ], + "gap_zh": "视频只有跟随机器人的局部视角,无法逐件核实六件散物的完整位置;inputs/ 中无状态日志及成功参考视频。评分器仅报告 final_reward=0.14575902334907173 等汇总字段,没有明确报告 episode_return、max_xy_variance、room_object_organized、stand_reward、small_control 或 per_timestep_reward;final_reward 的累计/末步口径未在本输入中明确,故不将其换算或直接映射到这些原生指标。log step 52 的物品方差是 agent 自己的会话检查,未作为评分器数值填写;视频时间也未由 agent 工作时间或控制步数换算。" + }, + "summary_zh": "观察:视频 0.0–10.05 s 中机器人维持原地屈膝站立,未启动六件物品归拢,唯一目标节点最终为 false(0/1),评分器 success=0。推断:断点在归拢首步前;log step 37、43、46 的搜索以存活和原地稳定为优化目标,缺少物品归拢项,最终选出了站立轨迹。观察:视频缺少六物体全景且没有状态日志,累计奖励和奖励分项未被明确报告,不能给出其数值。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "视频全程机器人原地屈膝站立,未去归拢物品,节点 false、断点首步前、没尝试,与日志(log step 37/43/46/51/53 的站立适应度与交付说明)一致,未发现需修改的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:46:35Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-privileged--mez9qpe/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-privileged--mez9qpe/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "5f5e0fd4f27d1789e5ea256c7bea6d23b8457930c693aeb882c3cc55f44b46f1", + "video_sha256": "6c856a07543b84a53befd949c4702ce0f25337c15cdfb85f654c4bdff30ed4e6", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-privileged--nlpmrt5.json b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-privileged--nlpmrt5.json new file mode 100644 index 0000000000000000000000000000000000000000..361b8680804b594143f4b1aebfd0a0621cb49f5f --- /dev/null +++ b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-privileged--nlpmrt5.json @@ -0,0 +1,202 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-room-i00-privileged--nlpmrt5", + "taskKey": "humanoidbench/manip_room", + "dag_sha256": "081a0a8f4158f72c3be894442d203666d7e71a9e8771af0ea997f42197bdbd1c", + "outcome": { + "success": 0, + "reward": 0.1955988559034519, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 1000, + "final_reward": 0.1955988559034519, + "sim_steps": 1000, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 6.932, + "peak_processes": 26, + "peak_disk_gb": 0.24, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-room-i00-privileged--nlpmrt5/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "c7ecf9e4cdbb3ef17138b74cfeacb1aa4959d4e14ac86f653e5024049a6cf118" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "objects_gathered": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0, + "log_step": null, + "text_zh": "观察:视频 0.00–0.65 s 机器人双脚在原处支撑、双臂前伸;视频每隔约 0.5 s 的抽帧至视频 9.50 s,以及视频 9.00–10.05 s 逐帧检查,机器人始终停在原处,没有走向物品、抓取或推拢的动作。画面主要覆盖机器人,不能直接查看全部六件物品的位置。" + }, + { + "t": 10.05, + "log_step": 12, + "text_zh": "观察:log step 12 的最终生成脚本只设置双腿轻蹲和踝关节平衡反馈,其余执行器保持初始目标,保存并重放了 1000 个动作;log step 20 再次生成后逐项比较动作完全相同。视频 10.05 s 机器人仍站在原处,双手未搬运物品,与该站立基线一致。" + }, + { + "t": 10.05, + "log_step": 21, + "text_zh": "推断:完整视频未显示归拢作业;agent 在 log step 21 也自述交付轨迹没有移动六件物品且预计达不到累计奖励 400。评分器 reward.success=0、all_success=0、deterministic=1、n_actions=1000;结合唯一节点按累计奖励达标验收的规则,判定该事件未达成。评分器的 final_reward=0.1955988559034519 未说明为累计回报,不能当作 episode_return。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "unit": "奖励", + "threshold": 400, + "direction": "higher_better", + "value": null + }, + { + "name": "max_xy_variance", + "unit": "平方米", + "threshold": null, + "direction": "lower_better", + "value": null + }, + { + "name": "room_object_organized", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "stand_reward", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "small_control", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "per_timestep_reward", + "unit": "奖励", + "threshold": null, + "direction": "higher_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "objects_gathered", + "t": 0, + "log_step": 12, + "where": "首步前", + "evidence_zh": "观察:视频 0.00–10.05 s 机器人一直原地站立,没有启动归拢所需的接近、抓取、搬移或推拢动作。log step 12 保存的交付脚本只控制蹲姿和平衡,log step 20 确认重生成动作一致;会话另行测试的行走没有进入最终文件。" + }, + "unmet": { + "objects_gathered": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0, + "log_step": 12, + "evidence_zh": "观察:视频 0.00–0.65 s、贯穿视频的约 0.5 s 抽帧及视频 9.00–10.05 s 的全部末段帧中,机器人都在原位置站立,未迈步接近物品,双臂未执行物品操作;log step 12 的最终脚本只录制站立平衡动作,归拢步骤从未启动。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "log_step": 17, + "source_zh": "日志", + "evidence_zh": "推断:归拢作业未启动背后有会话中的移动控制失败证据。log step 15 的参数化步态搜索调整摆腿、抬腿和踝部反馈;log step 17 用搜索所得步态测试时,机器人骨盆坐标从约 [0,0,0.97] 变到控制步 400 的 [1.50,-0.41,0.75],控制步 422 已降至 [1.78,-0.59,0.30] 并终止,头高 0.49、躯干直立投影 0.27,未获得持续稳定的行走。log step 21 的 agent 最终说明因该步态不能支持绕到物品外侧而保留站立基线;log step 12、20 的保存及复现记录和视频原地站立相符。因此归为控制能力不足;测试中的跌倒仅用于解释根因,不是本次交付视频的失败现象。" + } + ], + "gap_zh": "输入没有状态日志和成功 oracle;视频为跟随机器人的局部视角,不能逐件核实六物体坐标或接触对。评分器仅给 success、final_reward 等汇总字段,未报告 DAG 所列累计奖励、水平最大方差及奖励分项,六个 native_metrics 均保留 null;log step 13 的自测方差采用两轴方差之和,不能替代评分器要求的 max_xy_variance。根因由会话步态测试与最终交付记录推断,无法确定更细的控制子类或证明可行替代方案。" + }, + "summary_zh": "评分交付轨迹中,机器人在视频 0.00–10.05 s 始终原地站立,没有启动六件物品归拢,唯一节点最终为假(0/1),断点为归拢步骤的首步前,现象为没尝试。会话步态测试在 log step 17 失稳终止,最终生成及复现记录(log step 12、20)表明仅交付站立基线,支持控制层面的根因。视频未覆盖全部物品,且没有状态日志、成功示范或评分器累计回报及分项数值,因此不能给出精确物体方差或奖励轨迹。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "视频 0–10.05 s 机器人始终原地站立,唯一节点 false、断点首步前/没尝试、根因控制(依据 log step 17 的会话步态失稳)均与 FAIL-22 的处理一致,未见需改之处。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:51:32Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-privileged--nlpmrt5/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-privileged--nlpmrt5/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "9412c26c141eac7d4b18bfb9b48328c02bcc43f2152c30ac2dcc512e79405f05", + "video_sha256": "c7ecf9e4cdbb3ef17138b74cfeacb1aa4959d4e14ac86f653e5024049a6cf118", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-privileged--nvpsyw7.json b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-privileged--nvpsyw7.json new file mode 100644 index 0000000000000000000000000000000000000000..e7aadb8846a5b5917622c54fe1058a767384ce7f --- /dev/null +++ b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-privileged--nvpsyw7.json @@ -0,0 +1,209 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-room-i00-privileged--nvpsyw7", + "taskKey": "humanoidbench/manip_room", + "dag_sha256": "081a0a8f4158f72c3be894442d203666d7e71a9e8771af0ea997f42197bdbd1c", + "outcome": { + "success": 0, + "reward": 0.19673033750136038, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 1000, + "final_reward": 0.19673033750136038, + "sim_steps": 1000, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 14.3, + "peak_processes": 26, + "peak_disk_gb": 0.87, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-room-i00-privileged--nvpsyw7/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "a977eda70298593534d418d95aedadb7588b248450c25d808177986486bd8895" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "objects_gathered": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.0–1.0 s,机器人双脚留在原地,身体略微抬高、手臂随站姿调整;未出现朝散物行走、伸手抓取或推拢的动作。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 1.0–10.05 s(含视频 9.0–10.05 s 末段),机器人继续原地站立,手臂保持前伸,没有搬移或推拢物品的动作,也没有跌倒。" + }, + { + "t": null, + "log_step": 99, + "text_zh": "观察:log step 93 的交付生成脚本只循环调用站姿控制器并调整质心目标;log step 96、99 继续微调站姿后覆盖保存,log step 100、102 确认最终文件含 1000 个动作且重放哈希一致。agent 在 log step 103 自述“维持站立,但未完成归拢”。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:结合视频 0.0–10.05 s 未启动归拢动作及评分器 success=0、all_success=0、deterministic=1,唯一累计验收节点 objects_gathered 未达到;不是仅凭零成功分数判断没有尝试。该节点从未实际开展,不附加抓握、推扫或搬运接触。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "unit": "奖励", + "threshold": 400, + "direction": "higher_better", + "value": null + }, + { + "name": "max_xy_variance", + "unit": "平方米", + "threshold": null, + "direction": "lower_better", + "value": null + }, + { + "name": "room_object_organized", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "stand_reward", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "small_control", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "per_timestep_reward", + "unit": "奖励", + "threshold": null, + "direction": "higher_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "objects_gathered", + "t": 0.0, + "log_step": 93, + "where": "首步前", + "evidence_zh": "观察:视频 0.0–10.05 s,机器人仅在原地调整并保持站姿,始终未开始抓、推或搬运散物;log step 93 写出的交付脚本只执行站姿控制,log step 96、99 只修改站姿参数,log step 102 确认最终文件确定性重放。归拢步骤在交付回合中未启动,视频 0.0 s 用作首步前导航锚点。" + }, + "unmet": { + "objects_gathered": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": null, + "evidence_zh": "观察:视频 0.0–10.05 s,机器人双脚一直留在同一区域,只调整身体高度和手臂姿态,没有针对散物执行接近、抓取、推拢或搬移;视频 9.0–10.05 s 仍如此。交付回合没有启动六件物品归拢。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "t": null, + "log_step": 92, + "source_zh": "日志", + "evidence_zh": "推断:agent 未能得到可用的移动与伸手控制,最终只交付站姿轨迹。会话证据为 log step 91 的伸手拉动物品试验脚本与 log step 92 的输出:所示试验在 887–891 个控制步结束,机器人根部高度约 0.26–0.29 米,物品水平坐标方差代理仍约 4.102–4.103;log step 96 中 agent 自述这些伸手试验跌倒且未减少物品分散程度。log step 93、96、99 因而生成并保存仅维持站姿的最终文件;这些试验只用于解释控制困难,不作为交付视频中的跌倒或接触现象。", + "note_zh": "未独立证实具体接触几何、逆解、反馈或参数错误,根因停留在控制。" + } + ], + "gap_zh": "视频为机器人近景,六件散物未全部入镜,不能逐件核实全局位置;本目录没有状态日志和成功参考视频。评分器未报告 episode_return、物体坐标方差或奖励分项,final_reward 字段与原生累计回报的换算关系未给出,故六项 native_metrics 均为 null。控制根因来自会话试验及交付脚本,无法进一步确定控制子类。" + }, + "summary_zh": "交付视频中机器人原地调整后保持站立,没有启动六件物品归拢,唯一节点最终为 0/1。断点在归拢首步前,现象为没尝试;会话中的行走、伸手控制未能形成可用归拢动作,最终保存的是确定性站姿轨迹。视频未覆盖全部散物,且缺少状态日志及原生累计回报,控制归因只能依据会话试验,不能细分子类。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "视频 0–10 s 机器人原地站立、未做归拢,节点 0/1、断点首步前/没尝试、根因停在控制,均与视频和 log step 93–102 的交付脚本一致;未发现需改动的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:55:08Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-privileged--nvpsyw7/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-privileged--nvpsyw7/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "99a954aa59eb7ff8806a01f1b6c239f53efb99b78912ac09d3f9acbf7485ed0f", + "video_sha256": "a977eda70298593534d418d95aedadb7588b248450c25d808177986486bd8895", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-privileged--pb9jup2.json b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-privileged--pb9jup2.json new file mode 100644 index 0000000000000000000000000000000000000000..e96f73ca22566f2fa5d42309ac9d8298c37a75bd --- /dev/null +++ b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-privileged--pb9jup2.json @@ -0,0 +1,201 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-room-i00-privileged--pb9jup2", + "taskKey": "humanoidbench/manip_room", + "dag_sha256": "081a0a8f4158f72c3be894442d203666d7e71a9e8771af0ea997f42197bdbd1c", + "outcome": { + "success": 0, + "reward": 0.1931999057461626, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 1000, + "final_reward": 0.1931999057461626, + "sim_steps": 1000, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 3.025, + "peak_processes": 24, + "peak_disk_gb": 0.239, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-room-i00-privileged--pb9jup2/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "47fb19436f5db418d559ba771afe6d933567f1dfc0f873f40f48c8357c227759" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "objects_gathered": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00–10.00 s 按约 0.5 s 间隔检查,并逐帧检查视频 0.00–1.00 s 与视频 9.05–10.05 s;H1 始终双脚着地、双臂前伸,在原处轻微调整姿态,没有迈步接近、抓取、搬移或推拢物品的动作。镜头只覆盖机器人及近处地面,没有六件物品的全景。" + }, + { + "t": null, + "log_step": 68, + "text_zh": "观察:log step 68 写入最终生成脚本,固定初始关节目标,仅依重心偏差调整双踝及髋侧倾,录制 1000 个控制步后保存;log step 83 再生成及重放均为 1000 步、同一末态哈希,log step 99 最后重放仍为 deterministic=true。该脚本与评分视频中原地站立的行为一致;会话中另行测试的行走轨迹没有提交。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:inputs/run.json 的评分器字段 success=0、all_success=0、deterministic=1、grader_error=0、n_actions=1000;本节点按整段累计奖励达到 400 验收,累计达标后不撤销,因此评分失败结合已验证的站立交付表明该事件未达成,区间为空、终值为假。final_reward=0.1931999057461626 未提供与累计奖励的对应说明,不将其换算为累计回报。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "unit": "奖励", + "threshold": 400, + "direction": "higher_better", + "value": null + }, + { + "name": "max_xy_variance", + "unit": "平方米", + "threshold": null, + "direction": "lower_better", + "value": null + }, + { + "name": "room_object_organized", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "stand_reward", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "small_control", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "per_timestep_reward", + "unit": "奖励", + "threshold": null, + "direction": "higher_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "objects_gathered", + "t": 0.0, + "log_step": 68, + "where": "首步前", + "evidence_zh": "观察:视频 0.00–10.05 s 中 H1 只在原地维持站立,未启动接近或归拢物品的动作;log step 68 的最终脚本仅控制站立,log step 83 与 log step 99 的重放核验对应此站立交付。故断点为六件物品归拢,视频 0.00 s 仅作未启动步骤的导航锚点。" + }, + "unmet": { + "objects_gathered": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 68, + "evidence_zh": "观察:视频 0.00–10.00 s 的全段抽帧及视频 9.05–10.05 s 的逐帧检查均显示机器人保持同一站位,没有启动移向物品或操作物品的动作;log step 68 保存的生成脚本也只有原地平衡控制。该现象指评分交付没有开始归拢,不指 agent 未做过会话测试。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志", + "evidence_zh": "推断:未能获得稳定的行走控制,使 agent 最终只能交付站立基线。log step 61 的测试脚本将行走终止步记为 fell,其参数搜索输出均提前终止;log step 67 的最好输出仅前进 1.27 m、测试控制步 162 终止。log step 94 的独立诊断中,机器人重心横向坐标由测试控制步 32 的 0.107 m 增至测试控制步 128 的 1.011 m,随后测试控制步 131 以骨盆高度 0.281 m 终止;log step 97 修改控制后仍只前进约 1.06–1.25 m。agent 在 log step 103 明确自述因行走测试反复失稳而保存站立轨迹,且最终脚本及重放记录(log step 68、83、99)支持此交付选择。上述跌倒仅属会话中的根因证据,不是评分视频的现象;具体控制子类缺少充分验证,停在控制。" + } + ], + "gap_zh": "视频视角只显示机器人和近处地面,不能逐件直接核实六件散物的位置、接触或水平坐标方差;inputs/ 无状态日志及成功示范。评分器报告了 final_reward=0.1931999057461626,但未明确报告 DAG 所列的累计 episode_return、max_xy_variance、room_object_organized、stand_reward、small_control 或 per_timestep_reward,故这些原生指标均留 null,不自行换算。控制根因来自会话测试与最终交付记录,缺少测试视频及独立控制诊断,无法进一步确认具体子类;不存在对评分视频跌倒的断言。" + }, + "summary_zh": "评分轨迹中 H1 全程原地站立,没有启动六件物品归拢,唯一节点未达成,最终进度为 0/1,断点在首步前。会话行走测试反复失稳,agent 最终交付了经重放验证的站立基线,根因据日志归为控制;评分视频中没有跌倒。视频缺少房间全景且没有状态日志,无法逐件核实物品位置或计算原生奖励分量,累计节点按评分器未通过的结果判假。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "视频中 H1 全程原地站立,未启动归拢,唯一节点为假、断点首步前/没尝试、根因控制(依会话证据,符合 FAIL-22 的处理)均有证据支持;未发现需改动之处。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:53:57Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-privileged--pb9jup2/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-privileged--pb9jup2/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "fc36f33805d0445459d82a8e09e75e85e2aae3072db999aca2e8512a88295fcf", + "video_sha256": "47fb19436f5db418d559ba771afe6d933567f1dfc0f873f40f48c8357c227759", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-privileged--pcg6bhn.json b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-privileged--pcg6bhn.json new file mode 100644 index 0000000000000000000000000000000000000000..2231fe4e93aecb82ccc5144059dc1e4b35fef73d --- /dev/null +++ b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-privileged--pcg6bhn.json @@ -0,0 +1,201 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-room-i00-privileged--pcg6bhn", + "taskKey": "humanoidbench/manip_room", + "dag_sha256": "081a0a8f4158f72c3be894442d203666d7e71a9e8771af0ea997f42197bdbd1c", + "outcome": { + "success": 0, + "reward": 0.19445730974314476, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 1000, + "final_reward": 0.19445730974314476, + "sim_steps": 1000, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.877, + "peak_processes": 10, + "peak_disk_gb": 0.253, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-room-i00-privileged--pcg6bhn/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "afb16a400a98be4e256ad4c360c46cbaf300d8e0146867c288cb12104bc297be" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "objects_gathered": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00–1.00 s 逐帧、视频中段约每 0.5 s 抽帧及视频 9.10–10.05 s 逐帧均显示 H1 在原地作轻微站姿调整,双脚未迈步,双手保持前伸,未见接近目标、抓搬或推拢物品。画面没有覆盖六件散物的完整布局,因此不据此估算方差。" + }, + { + "t": null, + "log_step": 174, + "text_zh": "观察:log step 174 写入的最终 gen.py 在全部 1000 次控制循环中只根据质心相对双脚的偏差调整两个踝关节和两个髋侧摆关节,其余关节保持初始目标;随后保存录制动作。log step 183 再次读出的脚本相同,未包含行走、伸手取物或推拢阶段,与视频中的原地站立一致。" + }, + { + "t": 10.05, + "log_step": 15, + "text_zh": "推断:评分器字段 success=0、all_success=0、deterministic=1、grader_error=0,结合视频未启动归拢动作,判定累计验收事件未达,节点全程不成立。评分器 final_reward=0.19445730974314476;log step 15 展示的 Session.status 代码将 final_reward 取自 per_timestep_reward,所以该值填入末步奖励指标,不充当 episode_return。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "unit": "奖励", + "threshold": 400, + "direction": "higher_better", + "value": null + }, + { + "name": "max_xy_variance", + "unit": "平方米", + "threshold": null, + "direction": "lower_better", + "value": null + }, + { + "name": "room_object_organized", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "stand_reward", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "small_control", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "per_timestep_reward", + "unit": "奖励", + "threshold": null, + "direction": "higher_better", + "value": 0.19445730974314476 + } + ], + "failure": { + "breakpoint": { + "ref": "objects_gathered", + "t": 0.0, + "log_step": 174, + "where": "首步前", + "evidence_zh": "观察:视频 0.00 s 起机器人仅维持站姿,直到视频 10.05 s 仍无行走、抓搬或推拢动作;log step 174 保存的最终脚本也只有原地平衡反馈。六件物品归拢的操作从未启动,故用视频起点作导航锚点。" + }, + "unmet": { + "objects_gathered": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 174, + "evidence_zh": "观察:视频 0.00–1.00 s 与视频 9.10–10.05 s 逐帧以及中间抽帧只见 H1 原地站立、双手空悬;log step 174 的交付脚本全程维持初始上肢目标,未进入物品操作。视频中机器人没有跌倒,日志早期调试中的跌倒不属于本次交付轨迹。" + } + ], + "root_causes": [ + { + "cat": "认知", + "item": "推理/规划", + "source_zh": "日志", + "evidence_zh": "推断:agent 在 log step 52 已计算将散物搬到一组可降低方差,log step 142 仍把物品操作列为后续工作;但 log step 160 的调优目标 stand_r 只包含站立、直立、静止与控制惩罚,不含物品分布,log step 174 最终保存的策略只调整踝关节与髋侧摆关节,全部 1000 次循环都没有转入归拢阶段。结合视频 0.00–10.05 s 的原地站立,最终策略停留在站立子目标、遗漏归拢阶段,支持跨阶段规划不完整;此归因依据具体调优目标和交付程序,而非仅由停止或零进度反推。" + } + ], + "gap_zh": "观察:inputs/ 没有状态日志或成功示范;视频视角只展示机器人附近,无法逐件核对六件散物的位置、接触对或精确方差。评分器未报告 episode_return、max_xy_variance、room_object_organized、stand_reward、small_control,故这些原生指标保留 null;final_reward 仅为末步奖励,不能还原整段累计值。节点未通过由评分器 verdict 与已检查的无归拢动作共同判断,不将会话测试坐标或跌倒当作评分轨迹事实。" + }, + "summary_zh": "评分视频中 H1 全程在原地保持站立,未启动搬移或推拢,六件物品归拢节点未达,最终进度为 0/1。断点位于首步前,现象为没尝试;日志显示 agent 用站立指标调优并交付只有平衡反馈的脚本,最终策略遗漏了归拢阶段。视频未覆盖六件物品全景,且没有状态日志和累计回报数值,无法量化其方差或距 400 门槛的差额。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "视频 0–10 s 抽帧显示 H1 全程原地站立、无行走或推拢,objects_gathered 未达、断点首步前/没尝试与日志中的只含平衡反馈的 gen.py 一致;根因推理/规划有 log step 52/160/174 的具体调优目标支撑。未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:46:04Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-privileged--pcg6bhn/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-privileged--pcg6bhn/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "f4357defbc82e5966b56791fe7c888ecf41184607dd5b0bba7e3f14d643c3dec", + "video_sha256": "afb16a400a98be4e256ad4c360c46cbaf300d8e0146867c288cb12104bc297be", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-privileged--u2vmtdw.json b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-privileged--u2vmtdw.json new file mode 100644 index 0000000000000000000000000000000000000000..62af1c616cf7043df453be117265c63ab69e8af2 --- /dev/null +++ b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-privileged--u2vmtdw.json @@ -0,0 +1,224 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-room-i00-privileged--u2vmtdw", + "taskKey": "humanoidbench/manip_room", + "dag_sha256": "081a0a8f4158f72c3be894442d203666d7e71a9e8771af0ea997f42197bdbd1c", + "outcome": { + "success": 0, + "reward": 0.17510965192111586, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 1000, + "final_reward": 0.17510965192111586, + "sim_steps": 1000, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 1.023, + "peak_processes": 22, + "peak_disk_gb": 0.253, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-room-i00-privileged--u2vmtdw/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "452c4c9d2cdd1d23c98aee49d11cec9e43589411a0c6f8e5a9f05e02920293a3" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "objects_gathered": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.0–3.2 s,机器人转身并走向椅子一侧;椅子与地上的红色耳机相隔,未形成整体聚团。" + }, + { + "t": 3.5, + "log_step": null, + "text_zh": "观察:视频 3.25–3.75 s,机器人靠椅背的一只手掌贴过椅背上沿,腿脚与椅子下部接近并蹭推,椅子有小幅位移;机器人随后从椅子旁离开,没有把椅子搬到其余物品处。" + }, + { + "t": 4.8, + "log_step": null, + "text_zh": "观察:视频 4.4–5.3 s,机器人双脚交替跨过、拨动红色耳机,使其移动并翻转;视频 6.5 s 耳机仍独立留在机器人身后的地面,椅子在墙边。" + }, + { + "t": 10.0, + "log_step": null, + "text_zh": "观察:视频 9.0–10.05 s 机器人仍站立行走;椅子、耳机、奖杯和包裹分处画面不同位置,没有显示六件物品归为一组。评分器记录 success=0、all_success=0、deterministic=1、sim_steps=1000;按本节点绑定的累计验收,归拢事件未达成。final_reward=0.17510965192111586 不作为 episode_return 使用。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "推/扫", + "note_zh": "观察:视频 3.25–3.75 s,靠椅背的一只 Shadow 手掌贴过椅背上沿并随机器人前行,无包围夹持。" + }, + { + "who": "腿/躯干", + "how": "推/扫", + "note_zh": "观察:视频 3.4–3.75 s 腿脚蹭推椅子下部;视频 4.4–5.3 s 双脚交替拨推地上的红色耳机,耳机发生平移和翻转。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–10.05 s,H1 双腿双脚在地面交替承重,支撑绕行和拨推物品;已检查末段,未见倒地。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "unit": "奖励", + "threshold": 400, + "direction": "higher_better", + "value": null + }, + { + "name": "max_xy_variance", + "unit": "平方米", + "threshold": null, + "direction": "lower_better", + "value": null + }, + { + "name": "room_object_organized", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "stand_reward", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "small_control", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "per_timestep_reward", + "unit": "奖励", + "threshold": null, + "direction": "higher_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "objects_gathered", + "t": 3.8, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频 3.5–4.0 s,机器人仅蹭动椅子便离开,椅子留在墙边;视频 4.4–5.3 s 又拨动耳机,但视频 6.5 s 耳机仍单独留在身后,直到视频 10.05 s 结束也未修复整体分散。断点取开始离开未归拢椅子的约视频 3.8 s,属于同一个整体归拢步骤,非跌倒终止。" + }, + "phenomena": [ + { + "cat": "运动", + "item": "位姿不准", + "t": 4.8, + "log_step": null, + "evidence_zh": "观察:视频 3.5–4.0 s 椅子被蹭推后仍在墙边;视频 4.4–5.3 s 机器人脚步使耳机移动、翻转,却未将其送到其他物品旁,视频 6.5 s 耳机留在机器人身后。视频 9.0–10.05 s 可见椅子、耳机、奖杯及包裹仍相互分离;实际拨推后的物品位置没有形成要求的归拢关系。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "log_step": 127, + "source_zh": "日志 / 视频", + "evidence_zh": "推断:log step 127 的最终生成脚本用机器人位置、朝向、质心速度和躯干姿态控制普通交替摆腿,前一阶段仅以物体水平位置构造机体行走目标,后两阶段改用固定机体路点;没有独立控制脚与物体的接触位置或物体实际搬移量。视频 3.25–5.3 s 对应表现为蹭过椅子、反复跨拨耳机,随后留下两件物品继续走,说明该控制实现未把行走转化为足够的定向归拢。log step 71–74 的会话试验也反复记录物体几乎不动或只移动少量,作为控制能力的辅助证据,不计入评分轨迹进度。现有证据不足以细分接触设计与反馈控制。" + } + ], + "gap_zh": "没有状态日志或成功参考视频。视频 5.4–6.3 s 机器人和耳机被家具大幅遮挡,无法确定该段每次接触及左右脚细节;结合前后画面只能确认耳机被拨动后仍留在身后。评分器未报告 episode_return、max_xy_variance、room_object_organized、stand_reward、small_control 或 per_timestep_reward,故这些原生指标均填 null,不能从 final_reward 擅自换算累计回报;失败节点由可见分散结果与评分判定共同确定。控制根因是结合最终脚本和画面的推断,未确定更细子类。", + "unmet": { + "objects_gathered": false + } + }, + "summary_zh": "机器人保持站立,走到椅子旁蹭推,随后用脚拨动耳机,但最终六件物品仍未归拢,唯一节点 objects_gathered 未达成(0/1)。断点在约视频 3.8 s 离开未归拢的椅子,后续耳机拨推也未修复;最终脚本与画面支持行走及接触控制未形成有效定向搬移的推断。缺少状态日志和累计奖励细项,视频 5.4–6.3 s 的家具遮挡限制了接触细节判断。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "单节点 objects_gathered 未达成(评分器 success=0、视频可见椅子、耳机、奖杯、包裹分散)与画面一致;contact_actual(手掌贴椅背、腿脚蹭推椅子和拨耳机、腿支撑)、断点、位姿不准现象和控制根因均有画面支持,未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:54:28Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-privileged--u2vmtdw/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-privileged--u2vmtdw/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "8b68e92df8bef825238485daad540aeb6c11fd7a4c9a843bb2fc5f216e5dcd3b", + "video_sha256": "452c4c9d2cdd1d23c98aee49d11cec9e43589411a0c6f8e5a9f05e02920293a3", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard--4kbm8i6.json b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard--4kbm8i6.json new file mode 100644 index 0000000000000000000000000000000000000000..3c9dc0aff184af68675db94b6e7244749b480077 --- /dev/null +++ b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard--4kbm8i6.json @@ -0,0 +1,224 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-room-i00-standard--4kbm8i6", + "taskKey": "humanoidbench/manip_room", + "dag_sha256": "081a0a8f4158f72c3be894442d203666d7e71a9e8771af0ea997f42197bdbd1c", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 111, + "sim_steps": 5548, + "calls": 1436, + "rows": 5435, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1033.3, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-room-i00-standard--4kbm8i6/replay.mp4", + "duration_s": 1.2, + "fps": 20.0, + "frames": 24, + "sha256": "f7dc51319db6ca998d5e0ab17ede477a61928eff668c2baadacac512a19e72b8" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "objects_gathered": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 163, + "text_zh": "观察:视频 0.00–1.15 s 全部24帧中,机器人仅维持屈膝站姿并随后向后倒下,没有向物品行走、伸手抓取、搬移或推拢的动作;可见椅子未被搬移。log step 163 的最终脚本只调节双踝和髋部横滚,其余关节保持初始目标,没有物品操作指令,归拢步骤未启动,故不填该节点的 contact_actual。" + }, + { + "t": 1.15, + "log_step": 164, + "text_zh": "观察:视频 0.80–1.15 s 机器人后倾加剧、骨盆下降,最后身体接近水平。log step 164 在最后一次重置后报告 episode=51、control_steps=111、done=True、reason=failed;log step 166 确认该回合已结束且 resets_left=0,未产生新回合。" + }, + { + "t": 1.15, + "log_step": 164, + "text_zh": "推断:该任务只有一个绑定累计奖励达到400的目标节点;评分器 success=0、live_success=0、graded_by_live=1,结合最后现场回合失败及画面中未启动归拢,可判该累计事件未达到,final=false。评分器没有报告累计奖励数值,不能由失败分数反推具体奖励或各物体方差。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "unit": "奖励", + "threshold": 400, + "direction": "higher_better", + "value": null + }, + { + "name": "max_xy_variance", + "unit": "平方米", + "threshold": null, + "direction": "lower_better", + "value": null + }, + { + "name": "room_object_organized", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "stand_reward", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "small_control", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "per_timestep_reward", + "unit": "奖励", + "threshold": null, + "direction": "higher_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "objects_gathered", + "t": 0.8, + "log_step": 164, + "where": "首步前", + "evidence_zh": "观察:视频 0.00–1.15 s 机器人没有启动任何物品归拢动作,约视频 0.80 s 开始明显后仰,视频 1.00–1.15 s 向后倒下,未恢复站立或操作物品。log step 163 的交付脚本只有站姿调节,log step 164 的最后现场回合以 reason=failed 结束,log step 166 未再启动回合。" + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 163, + "evidence_zh": "观察:视频 0.00–1.15 s 双手未接近并操作物品,机器人未执行搬移或推拢;log step 163 的最终脚本将动作限制在踝关节俯仰和髋关节横滚的站姿调节,没有归拢动作。" + }, + { + "cat": "全身", + "item": "平衡", + "t": 0.8, + "log_step": 164, + "evidence_zh": "观察:视频约 0.80 s 起机器人躯干持续后倾,视频 1.00–1.15 s 骨盆向地面下降、身体接近水平,未恢复支撑姿态;log step 164 的最后回合报告 done=True、reason=failed。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "反馈控制", + "log_step": 164, + "source_zh": "日志", + "evidence_zh": "推断:log step 163 的实际执行脚本以双踝平均角度构造比例修正,并用 imu_gravity_y 调节髋部横滚;log step 164 中服务返回的 imu_gravity_x 从 control step 100 的 -0.7660 发展到 control step 111 的 -0.9641,随后回合失败。结合视频 0.80–1.15 s 同方向加剧的后倾,闭环站姿调节未能抑制俯仰失稳,使机器人在物品操作前倒下;这些控制步计数仅作日志证据,不换算为视频时间。", + "note_zh": "没有证据进一步判定某个增益、关节目标或动力学模型单独导致失稳,故不标参数调试。" + } + ], + "gap_zh": "没有状态日志和成功参考视频;评分器未报告 episode_return、物体方差及其他原生奖励分项,六件物品也未同时完整出现在评分视频视野中,无法精确重建物体布局或累计奖励。log step 164 记录最后回合仿真时长2.22 s,而视频总长1.20 s,缺少可信的仿真到视频映射;所有视频时刻均直接取自所检帧,不将控制步计数或agent工作时间换算为视频时间。控制失败的具体参数与动力学原因仍不能确定。", + "unmet": { + "objects_gathered": false + } + }, + "summary_zh": "评分回合未启动六件物品归拢,唯一目标节点未达到,最终进度0/1。机器人只运行站姿反馈调节,约视频 0.80 s 开始明显后仰,至视频 1.15 s 向后倒下;最终现场回合在 log step 164 返回失败。缺少状态日志、累计奖励数值和可信的仿真到视频时间映射,无法进一步确定具体控制参数问题。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "单节点 objects_gathered=false、最终0/1,断点为首步前、现象 没尝试+平衡(与 powerlift 先例一致),根因为反馈控制(推断已标注、控制步计数未换算为视频时间)。视频仅1.2 s,机器人站立后约0.8 s起后仰、1.15 s倒下,与标注相符。未发现需修改的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:57:31Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-standard--4kbm8i6/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-standard--4kbm8i6/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "07ac6f831d41d91ba1e160a1f337fcd4bd03929ad1764fecc2c1c4611c5e224f", + "video_sha256": "f7dc51319db6ca998d5e0ab17ede477a61928eff668c2baadacac512a19e72b8", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard--fsrugge.json b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard--fsrugge.json new file mode 100644 index 0000000000000000000000000000000000000000..6e91c47a11e6a0ace3416df0f962016a218eaae2 --- /dev/null +++ b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard--fsrugge.json @@ -0,0 +1,218 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-room-i00-standard--fsrugge", + "taskKey": "humanoidbench/manip_room", + "dag_sha256": "081a0a8f4158f72c3be894442d203666d7e71a9e8771af0ea997f42197bdbd1c", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 185, + "sim_steps": 8326, + "calls": 1066, + "rows": 8326, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 613.3, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.31, + "peak_processes": 11, + "peak_disk_gb": 0.829, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-room-i00-standard--fsrugge/replay.mp4", + "duration_s": 1.95, + "fps": 20.0, + "frames": 39, + "sha256": "f86acf6b00912a9176676ec8bd557f5cd6d272b0c96acf793a84da3b70037730" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "objects_gathered": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 69, + "text_zh": "观察:视频 0.00–1.90 s 的全部 39 帧中,机器人先在原地维持屈膝站姿,双手空置在身前,随后后仰跌倒;没有伸手抓取、推拢或搬运物品的动作。log step 69 的最后回合脚本仅重复发送同一个站姿关节目标,没有物品操作指令。" + }, + { + "t": 1.85, + "log_step": 70, + "text_zh": "观察:视频 1.85–1.90 s 机器人已向后倒至接近地面;log step 70 的机器人服务状态为 episode=51、control_steps=185、done=true、reason=failed,评分器字段 live_success=0。推断:按任务的累计达标即成功结束规则,最后回合未触发累计奖励达到 400 的归拢验收;画面不覆盖全部六物体,未据此估算方差或累计奖励。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "unit": "奖励", + "threshold": 400, + "direction": "higher_better", + "value": null + }, + { + "name": "max_xy_variance", + "unit": "平方米", + "threshold": null, + "direction": "lower_better", + "value": null + }, + { + "name": "room_object_organized", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "stand_reward", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "small_control", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "per_timestep_reward", + "unit": "奖励", + "threshold": null, + "direction": "higher_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "objects_gathered", + "t": 1.5, + "log_step": 69, + "where": "首步前", + "evidence_zh": "观察:视频 0.00–1.30 s 机器人原地站立,未开始抓取或推拢物品;视频约 1.50 s 起躯干持续后仰,至视频 1.85 s 身体近乎水平、骨盆下降,未恢复。log step 69 仅执行固定站姿目标,log step 70 确认这是最后现场回合 episode 51,服务已判 failed。推断:因此断在尚未启动的六件物品归拢步骤。" + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 69, + "evidence_zh": "观察:视频 0.00–1.90 s 机器人没有走向散物、伸手抓取或推拢物品,双手一直空置;log step 69 的最后回合只有站姿维持命令,归拢动作没有启动。" + }, + { + "cat": "全身", + "item": "平衡", + "t": 1.5, + "log_step": 69, + "evidence_zh": "观察:视频约 1.50–1.85 s 机器人躯干逐渐向后倾倒、骨盆下沉,双腿未能恢复支撑,末帧身体接近水平;log step 70 的机器人服务确认该回合已结束且 reason=failed。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "log_step": 69, + "source_zh": "日志", + "evidence_zh": "推断:最后回合的站姿控制未能稳定机器人,导致物品操作尚未开始即跌倒。log step 69 的脚本只在重置后构造一次关节目标(髋偏置 -0.05 rad、踝偏置 0.0245 rad),随后以每批 50 个控制步重复发送同一目标;虽读取本体感知,但未据反馈修改命令。该步输出的俯仰估计由 control step 100 的 -0.029 rad 增至 control step 150 的 -0.223 rad、control step 185 的 -1.28 rad,角速度幅值也增加;视频 1.50–1.85 s 可见相符的后仰跌倒。" + } + ], + "gap_zh": "无状态日志、成功示范或原生累计奖励及分项数值,视频视野未同时覆盖全部六件物品,不能逐物体核实位置或计算归拢指标。log step 69–70 的最后回合为 185 个控制步,按日志给出的控制步长 0.02 s 对应仿真 3.70 s,而视频总长仅 1.95 s;未提供逐帧对应关系,故不将控制步直接换算为视频时间,所有视频锚点均取自已检查的帧。", + "unmet": { + "objects_gathered": false + } + }, + "summary_zh": "最后现场回合只维持原地站姿,未启动六件物品归拢,累计验收未通过,节点完成 0/1。机器人在视频约 1.50 s 起向后失衡,视频 1.85 s 已倒至接近地面;log step 69 的固定站姿命令未随本体反馈调整,支持控制失败的归因。缺少原生奖励及状态日志,且最后回合的控制步时长与视频时长不同,无法精确对应两者或计算物品方差。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "标签与视频和日志一致:最后现场回合(episode 51)机器人只站立,视频约 1.5 s 起后仰跌倒,未开始归拢;断点、没尝试+平衡、控制根因均有 log step 69 的俯仰读数与视频支撑,未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:08:01Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-standard--fsrugge/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-standard--fsrugge/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "18058270c0c1cd0a9c2cf5d37d4580c330266c4529db55f6ea78aaddd1cf22c4", + "video_sha256": "f86acf6b00912a9176676ec8bd557f5cd6d272b0c96acf793a84da3b70037730", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard--gziafse.json b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard--gziafse.json new file mode 100644 index 0000000000000000000000000000000000000000..6475f14436be0385b76713a071aca3698e51f7fe --- /dev/null +++ b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard--gziafse.json @@ -0,0 +1,216 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-room-i00-standard--gziafse", + "taskKey": "humanoidbench/manip_room", + "dag_sha256": "081a0a8f4158f72c3be894442d203666d7e71a9e8771af0ea997f42197bdbd1c", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 1000, + "sim_steps": 11676, + "calls": 11592, + "rows": 11676, + "resets": 35, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 520.2, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-room-i00-standard--gziafse/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "4d87d920bd2b85b39fce23dee24a460415ff1e3c40b5fd2a3f898955e5883eb4" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "objects_gathered": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.0–1.0 s 的逐帧检查显示机器人由略屈膝姿态调整为较直立的站姿,双脚留在原处,双臂前伸;未见机器人向散物行走或开始抓取、推拢。" + }, + { + "t": 10.05, + "log_step": 27, + "text_zh": "观察:视频 1.0–9.0 s 的每半秒抽帧及视频 9.05–10.05 s 的逐帧检查均显示机器人继续原地站立,未出现搬移或推拢动作。log step 27 是最后一次 reset 后执行 gen.py 的现场回合,服务返回 episode=36、control_steps=1000、done=True、reason=failed。" + }, + { + "t": 10.05, + "log_step": 27, + "text_zh": "推断:该节点绑定累计奖励达到 400 的事件;任务规则规定达到门槛即结束并返回 success,而最后评分回合的服务结果为 failed,且评分器字段 live_success=0、graded_by_live=1,因此该累计事件未达成。具体累计奖励数值未报告,不把评分 reward=0 当作原生 episode_return。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "奖励", + "threshold": 400, + "direction": "higher_better" + }, + { + "name": "max_xy_variance", + "value": null, + "unit": "平方米", + "threshold": null, + "direction": "lower_better" + }, + { + "name": "room_object_organized", + "value": null, + "unit": "无量纲", + "threshold": null, + "direction": "higher_better" + }, + { + "name": "stand_reward", + "value": null, + "unit": "无量纲", + "threshold": null, + "direction": "higher_better" + }, + { + "name": "small_control", + "value": null, + "unit": "无量纲", + "threshold": null, + "direction": "higher_better" + }, + { + "name": "per_timestep_reward", + "value": null, + "unit": "奖励", + "threshold": null, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "objects_gathered", + "t": 0.0, + "log_step": 27, + "where": "首步前", + "evidence_zh": "观察:视频 0.0–10.05 s 中机器人只调整并维持原地站姿,没有启动接近散物、抓取或推拢。log step 23 保存的 gen.py 默认执行站立控制,log step 27 将其用于最终第 36 回合并跑满 1000 控制步,服务返回 failed;归拢步骤始终未启动。" + }, + "unmet": { + "objects_gathered": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 27, + "evidence_zh": "观察:视频 0.0–10.05 s 仅见机器人在同一位置站立,手臂没有朝物品伸取或推扫;log step 27 执行的最终脚本仅运行站立控制。log step 28 中 agent 也自述最终回合没有归拢物品。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "log_step": 25, + "source_zh": "日志", + "evidence_zh": "推断:会话中的移动控制未能建立,是最后只保留站立基线的能力障碍。log step 25 的脚本先逐渐施加髋滚转、再屈髋屈膝抬右脚,该试验在 169 控制步返回 failed,记录的 imu_gravity_y 从约 -0.013 发展至 -0.827;log step 26 改用相反髋滚转后两次试验分别在 155、151 控制步返回 failed。随后 log step 27 重置并执行原地站立脚本;log step 28 中 agent 明确将停止行走尝试归于这些重心转移及抬脚失败。上述试验只是根因的会话证据,不作为最终评分回合的跌倒现象;控制子类和可行替代步态尚不能核实。" + } + ], + "gap_zh": "观察:inputs 中没有状态日志或成功参考视频;所提供视频是机器人侧面近景,未覆盖六件物品的完整分布,不能从画面量化其方差或逐件位置。run.json 的评分字段及最终服务状态未报告六项原生奖励指标,故指标值均为 null。最终回合服务记录为 1000 控制步、任务控制周期为仿真 0.02 s,而视频总长 10.1 s;缺少可核对的运动事件,未将控制步换算为视频时间。推断:控制根因来自已重置的会话试验与最终控制脚本,无法仅凭评分视频确定具体失稳机制,也不能认定机器人不存在合法可行步态。" + }, + "summary_zh": "评分回合仅完成站姿调整与原地维持,未启动六件物品归拢,最终节点为 0/1;视频 0.0–10.05 s 未见行走、抓取或推拢,log step 27 的最终第 36 回合跑满 1000 控制步后返回 failed。断点位于归拢首步前,现象为没尝试;会话中重心转移与抬脚控制反复失败后,agent 在最终回合只运行站立控制,根因据此推断为控制。视频没有房间物品全景,且没有状态日志或原生奖励数值,无法进一步核实物品方差及具体控制机制。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "视频 0–10.05 s 只见机器人原地站立,节点未达、断点首步前、没尝试,均与视频和 log step 27 的服务结果(episode 36、1000 步、failed)一致。根因控制以会话试验(log 25–26)为证并标明为推断,符合 FAIL-22;未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:51:26Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-standard--gziafse/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-standard--gziafse/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "9c9b1ef03974729b0c9d3afb5cb9cc98c5ab83bd521b686c5ef63cce50cecf09", + "video_sha256": "4d87d920bd2b85b39fce23dee24a460415ff1e3c40b5fd2a3f898955e5883eb4", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard--rmjsy2v.json b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard--rmjsy2v.json new file mode 100644 index 0000000000000000000000000000000000000000..340d79ade2c254228440646ecbbe826b8083899e --- /dev/null +++ b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard--rmjsy2v.json @@ -0,0 +1,217 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-room-i00-standard--rmjsy2v", + "taskKey": "humanoidbench/manip_room", + "dag_sha256": "081a0a8f4158f72c3be894442d203666d7e71a9e8771af0ea997f42197bdbd1c", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 87, + "sim_steps": 5583, + "calls": 2943, + "rows": 5583, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 486.1, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-room-i00-standard--rmjsy2v/replay.mp4", + "duration_s": 0.95, + "fps": 20.0, + "frames": 19, + "sha256": "f5cf761ac6ff3fe90b49dec8c8318007a4e543b818639e0d04045d71c245fec2" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "objects_gathered": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 52, + "text_zh": "观察:已逐帧检查视频 0.00–0.90 s 的全部19帧,机器人仅改变站姿并抬起双臂,双手空着,未见抓取、搬运或推拢物品;可见的椅子和红色物品未被搬到一起。log step 48 的控制脚本及 log step 52 最后一次调用只执行髋、膝、踝站姿与平衡反馈,没有归拢动作,因此不填写该未启动步骤的 contact_actual。" + }, + { + "t": 0.9, + "log_step": 54, + "text_zh": "观察:视频 0.75–0.90 s,机器人双腿向前伸出、骨盆降至地面附近;log step 54 服务状态为 episode=51、control_steps=87、done=true、reason=failed、resets_left=0,评分器 live_success=0 且 graded_by_live=1。推断:按任务 DAG 绑定的累计奖励达400即成功的事件验收,最后现场回合从未达到该节点,final=false;此结论结合服务终止判定,并非仅由零分推断没有动作。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "unit": "奖励", + "threshold": 400, + "direction": "higher_better", + "value": null + }, + { + "name": "max_xy_variance", + "unit": "平方米", + "threshold": null, + "direction": "lower_better", + "value": null + }, + { + "name": "room_object_organized", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "stand_reward", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "small_control", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "per_timestep_reward", + "unit": "奖励", + "threshold": null, + "direction": "higher_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "objects_gathered", + "t": 0.7, + "log_step": 52, + "where": "首步前", + "evidence_zh": "观察:视频 0.00–0.60 s,机器人只做站姿调整和抬臂,没有接近、抓取或推拢散物;视频约 0.65–0.90 s 双腿逐渐伸出、骨盆下落,未恢复站立。log step 52 的最后一次站立控制测试结束,log step 54 确认最后现场回合 episode 51 已失败,log step 55 确认无法继续提交或重置。推断:归拢工作尚未启动即因失衡结束,断点属于六件物品归拢的首步前;视频约 0.70 s 仅为失衡导航锚点。" + }, + "unmet": { + "objects_gathered": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 52, + "evidence_zh": "观察:视频 0.00–0.90 s,机器人未对任何散物执行抓、搬或推,双手一直空着;log step 48 的脚本仅调节站姿与髋踝反馈,log step 52 最后运行同一站立测试。推断:评分回合未启动归拢步骤,符合首步前的没尝试,而不是已开始抓取却未接触目标。" + }, + { + "cat": "全身", + "item": "平衡", + "t": 0.7, + "log_step": 54, + "evidence_zh": "观察:视频约 0.65 s 起机器人身体支撑姿态明显恶化,视频 0.75–0.90 s 双腿伸出、骨盆落至地面附近;log step 54 服务报告最后回合 done=true、reason=failed,未恢复。该现象取自最后回合视频,不取此前调试回合的跌倒。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "反馈控制", + "source_zh": "日志 / 视频", + "evidence_zh": "推断:站立平衡反馈未能稳定机器人,使其在归拢前失去支撑。log step 21 的运行器逐控制步读取本体感知并更新动作,log step 48 的控制器以骨盆重力方向与陀螺仪构造髋、踝位置反馈;log step 52 最后执行该控制器(KH=4.0、KDH=0.8、KPA=2.0、KDA=0.4),输出 survived 86,log step 54 确认最后回合87个控制步即失败,视频约 0.65–0.90 s 直接显示支撑丧失。证据支持闭环站立控制失稳,尚不足以确定具体增益、反馈符号或执行器能力中哪一项是更深层原因。" + } + ], + "gap_zh": "没有状态日志或成功示范;视频为机器人近景,未覆盖全部六件物品,不能从画面计算水平坐标方差、站立奖励或累计回报,run.json 也未报告六项原生指标,故其值均为null。log step 54 的最后回合87个控制步按日志控制周期0.02 s对应仿真1.74 s,而视频总长只有0.95 s,缺少经可见事件校准的时钟映射;所有导航时间均直接取自视频帧,不将控制步换算为视频时间。具体失稳机理缺少最后回合的逐步本体与接触记录;log step 57 的网络探测错误发生在 log step 54 已确认回合结束之后,不作为外因中断。" + }, + "summary_zh": "最后评分回合(episode 51)只执行站姿与平衡控制,尚未开始六件物品归拢,最终节点完成数为0/1。视频约 0.65–0.90 s 可见支撑丧失及骨盆下落,服务确认回合失败;控制代码与调用日志支持平衡反馈未能稳定机器人。没有状态日志、全房间视角或原生奖励数值,且控制步时长与短视频时长未能校准,不能进一步确定具体失稳参数或累计奖励。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "最后评分回合 episode 51 实为站立平衡测试,视频19帧(0–0.9 s)显示机器人抬臂后失衡、双腿前伸、骨盆落地,没有任何归拢动作;节点 final=false、首步前/没尝试 + 平衡、反馈控制根因均有视频与服务读数支持,时钟未做换算,未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:52:11Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-standard--rmjsy2v/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-standard--rmjsy2v/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "1362d530b0b450969f221d2d00418cbaeb3af6e54cea54145b9c8bcaa10923af", + "video_sha256": "f5cf761ac6ff3fe90b49dec8c8318007a4e543b818639e0d04045d71c245fec2", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard--ygqxrnq.json b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard--ygqxrnq.json new file mode 100644 index 0000000000000000000000000000000000000000..525071d51b20cce3a9b5ed5de40aaa906d851cc4 --- /dev/null +++ b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard--ygqxrnq.json @@ -0,0 +1,220 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-room-i00-standard--ygqxrnq", + "taskKey": "humanoidbench/manip_room", + "dag_sha256": "081a0a8f4158f72c3be894442d203666d7e71a9e8771af0ea997f42197bdbd1c", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 243, + "sim_steps": 15394, + "calls": 14817, + "rows": 15394, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3483.2, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-room-i00-standard--ygqxrnq/replay.mp4", + "duration_s": 2.5, + "fps": 20.0, + "frames": 50, + "sha256": "b72d01b81445f2d78d9471b3157700eac01d1ad40ecdc0bbc6e96925b8104fd8" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "objects_gathered": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:逐帧检查视频 0.00–2.45 s,机器人由屈膝站立逐渐伸腿、抬起双臂并转动躯干;画面中的椅子及腿旁红色物体没有被双手拿起,未见六件物品归成一组。视频视野未覆盖全部物品,不能由此计算水平坐标方差。" + }, + { + "t": 2.3, + "log_step": 117, + "text_zh": "观察:视频约 2.0 s 起机器人倾斜加重,视频 2.30–2.45 s 身体迅速扭转、下坠,未恢复站立;log step 117 的机器人服务回报最后回合 episode=51、control_steps=243、done=true、reason=failed,log step 118 核对最终回合命令为 243×61。评分字段 live_success=0、replay_success=0、deterministic=1。" + }, + { + "t": 2.45, + "log_step": 118, + "text_zh": "推断:本任务只有累计奖励达到 400 才算归拢事件,最后回合服务以 failed 结束且现场与重放均未通过,结合视频未见成功搬移、末段失衡,判定该事件未达到。评分中的 reward=0.0 不是原生累计奖励,不据此赋值 episode_return;更早回合的行走和桌边伸手不计入本次进度。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–2.20 s,H1 双脚接触地面,双腿在抬臂、转躯干和倾身尝试中承重,随后失衡下坠。双手动作可见,但未确认手与目标物的接触,不标抓握。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "unit": "奖励", + "threshold": 400, + "direction": "higher_better", + "value": null + }, + { + "name": "max_xy_variance", + "unit": "平方米", + "threshold": null, + "direction": "lower_better", + "value": null + }, + { + "name": "room_object_organized", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "stand_reward", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "small_control", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "per_timestep_reward", + "unit": "奖励", + "threshold": null, + "direction": "higher_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "objects_gathered", + "t": 2.3, + "log_step": 117, + "where": "操作中", + "evidence_zh": "观察:视频 0.55–1.20 s 机器人抬起双臂、转动躯干,随后倾身;视频 2.30–2.45 s 身体快速下坠并扭转,归拢尚未完成。log step 116 的最终脚本在重置后执行双臂伸出、躯干转动、倾身和按倾角触发的合指动作;log step 117 的服务确认此最后回合在 control step 243 以 failed 终止,log step 118 确认没有后续重置或修复。" + }, + "unmet": { + "objects_gathered": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 2.3, + "log_step": 117, + "evidence_zh": "观察:视频 1.75–2.20 s 机器人躯干逐渐倾斜,视频 2.30–2.45 s 支撑姿态崩溃、身体急剧下落和转动,未恢复;log step 117 中机器人服务的最后回合 IMU 重力 z 分量由 control step 200 的约 −0.970 变为 control step 243 的约 −0.407,随后服务回报 done=true、reason=failed。此处只以视频画面定位,不把控制步数直接换算成视频时间。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "log_step": 116, + "source_zh": "日志", + "evidence_zh": "推断:log step 116 的最终控制脚本同时改变腿部站姿、双臂和躯干,并加入倾身;其平衡修正将估计的前后误差与速度用于双踝目标。log step 117 的服务 IMU 显示最后回合姿态持续偏离直立,视频 2.0–2.45 s 对应可见的倾斜加剧与下坠,说明该组合动作的全身控制未能维持支撑,归拢因失衡中断。没有证据证明某个增益、逆解或接触方案本身不可行,故不指定控制子类。" + } + ], + "gap_zh": "没有状态日志或成功参考视频;当前视频只显示机器人及近旁部分物体,不能精确确认全部目标物接触或计算归拢方差、各奖励分项。评分字段没有报告 DAG 中六项原生连续指标,均保留 null。log step 117 的最后回合为仿真 4.86 s(243 个控制步,每步 0.02 s),视频总长仅 2.50 s,二者不能直接等同,未建立精确映射,断点仅用已检查的视频帧定位;原生跌倒阈值和具体控制子类不能由现有材料确定。" + }, + "summary_zh": "最后一次回合中,机器人抬臂、转动躯干并倾身,未见可确认的物品搬移或归拢,累计事件未达,最终节点为 0/1。视频约 2.30 s 的伸手尝试因全身失衡中断,最终脚本的姿态控制未维持站立,机器人服务也确认最后回合 failed。缺少状态日志及原生奖励数值,视频未覆盖全部物品,且视频时长与日志仿真时长不一致,接触细节和更细根因无法确定。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "视频全程可见机器人站立、抬臂、倾身后在约 2.3 s 失衡下坠,未见物品被搬移,节点未达、断点与平衡现象、控制根因均有帧和 log 支持;时钟未把控制步换算成视频时间。未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:53:32Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-standard--ygqxrnq/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-standard--ygqxrnq/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "9871931becdbbcfde57901604a74fa9f14498670f92813e5aed634b3f0540769", + "video_sha256": "b72d01b81445f2d78d9471b3157700eac01d1ad40ecdc0bbc6e96925b8104fd8", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard-realtime--3ga5l2p.json b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard-realtime--3ga5l2p.json new file mode 100644 index 0000000000000000000000000000000000000000..b70e66c4e7e9dedd1237240475b8f095f98921ff --- /dev/null +++ b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard-realtime--3ga5l2p.json @@ -0,0 +1,230 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-room-i00-standard-realtime--3ga5l2p", + "taskKey": "humanoidbench/manip_room", + "dag_sha256": "081a0a8f4158f72c3be894442d203666d7e71a9e8771af0ea997f42197bdbd1c", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 349, + "sim_steps": 7105, + "calls": 7202, + "rows": 7097, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1270.4, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 141.438, + "max_lag_s": 0.0712, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 602, + "late_steps": 0, + "requests_refused": 52, + "request_busy_s": 6.651, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.312, + "peak_processes": 11, + "peak_disk_gb": 0.831, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-room-i00-standard-realtime--3ga5l2p/replay.mp4", + "duration_s": 3.55, + "fps": 20.0, + "frames": 71, + "sha256": "d9641668f65ae25bcb8465a3a23b7aa23b630f06c4578274dbb420aaf2571886" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "objects_gathered": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:逐帧检查视频 0.00–3.50 s,机器人始终空手,双臂保持向前的站姿并伴随下肢调整,未见伸手抓取、推拢或搬运散物;视频约 3.25–3.50 s 机器人下坠倒地,未进入归拢操作。因此本节点未开始,不填写操作物品的 contact_actual。" + }, + { + "t": 3.5, + "log_step": 224, + "text_zh": "观察:log step 221 的服务重置结果为 episode=51、resets_left=0;log step 222 启动带 --late-balance 和 --ankle-bias 0.0250165 的控制器;log step 224 服务报告最后回合 control_steps=349、done=True、reason=failed。inputs/run.json 的评分字段 live_success=0、replay_success=0、deterministic=1,与视频末段倒地相符。" + }, + { + "t": 3.5, + "log_step": null, + "text_zh": "推断:该节点按整段累计奖励达到 400 验收;评分器判最后回合未通过,且视频没有归拢动作,故累计事件没有达到,final=false、intervals=[]。此判断不是把目录 reward=0.0 当作原生 episode_return 数值。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "unit": "奖励", + "threshold": 400, + "direction": "higher_better", + "value": null + }, + { + "name": "max_xy_variance", + "unit": "平方米", + "threshold": null, + "direction": "lower_better", + "value": null + }, + { + "name": "room_object_organized", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "stand_reward", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "small_control", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "per_timestep_reward", + "unit": "奖励", + "threshold": null, + "direction": "higher_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "objects_gathered", + "t": 3.25, + "log_step": 224, + "where": "首步前", + "evidence_zh": "观察:视频 0.00–3.20 s 机器人仅调整站姿,没有对散物开始抓取或推拢;视频约 3.25 s 双腿支撑开始明显塌落,视频 3.50 s 机器人躯干已接近水平、双手着地。log step 224 的服务报告最后回合结束为 failed;失败落在尚未开始的“六件物品归拢”,不是把跌倒约束另作步骤。" + }, + "unmet": { + "objects_gathered": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 222, + "evidence_zh": "观察:视频 0.00–3.50 s 中机器人空手维持和调整站姿,未见朝散物发起抓取、搬运或推拢;log step 222 最后一回合运行的是站姿偏置与迟发平衡修正控制器。归拢步骤在评分轨迹里没有启动。" + }, + { + "cat": "全身", + "item": "平衡", + "t": 3.25, + "log_step": null, + "evidence_zh": "观察:视频约 2.90 s 机器人躯干后倾并调整双腿,视频约 3.25–3.50 s 骨盆下降、双腿向后伸开,躯干向前倒下,末帧双手落到地面,机器人失去双足站立支撑。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志、视频", + "evidence_zh": "推断:站立控制未能稳定本体,阻止了归拢操作的启动。log step 220 中 agent 自述要在最后一次重置测试迟发平衡修正;log step 221–224 记录了该控制器在最后回合实际运行并结束为 failed,视频约 2.90–3.50 s 对应可见的姿态摆动、支撑塌落和倒地。依据控制器运行记录与实际失稳归因到控制;没有证据确定某一增益、关节偏置或反馈算法是具体错误。" + } + ], + "gap_zh": "没有状态日志或成功示范;视频为跟随机器人近景,六件散物没有同时完整入镜,不能逐物体核实其坐标与接触。评分字段未报告 episode_return、max_xy_variance、room_object_organized、stand_reward、small_control 或 per_timestep_reward,原生指标均填 null。log step 224 的最后回合为 349 控制步,按任务 50 Hz 对应仿真 6.98 s,而所给视频长 3.55 s,缺少可靠的仿真到视频映射;本文件全部视频时间直接取所检查的帧,不将控制步换算为视频时间。日志中的控制器修改内容多处被截断,不能进一步确定控制失稳的具体子类或参数原因。" + }, + "summary_zh": "评分轨迹未启动抓取或推拢散物,“六件物品归拢”累计验收未达成,最终进度为 0/1。机器人只调整站姿,视频约 3.25–3.50 s 失去双足支撑并倒地,断点在归拢步骤的首步前,根因可支持到控制层。缺少状态日志、原生累计奖励及完整房间视野,且控制步与视频时长不直接对应,时间锚点均来自视频帧。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "视频 0–3.5 s 机器人空手站姿,约 3.25 s 起倒地,未进行归拢;节点 false、断点首步前、没尝试+平衡、根因控制均有帧和服务读数支持,未发现需改动的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:14:01Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-standard-realtime--3ga5l2p/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-standard-realtime--3ga5l2p/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "973bf605e2b924c68304828899af425f0c1fb3411e29e2b6433d13d35d082326", + "video_sha256": "d9641668f65ae25bcb8465a3a23b7aa23b630f06c4578274dbb420aaf2571886", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard-realtime--dhek6oq.json b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard-realtime--dhek6oq.json new file mode 100644 index 0000000000000000000000000000000000000000..7f277b978bb0577f2c775d71eddadb76e4e56382 --- /dev/null +++ b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard-realtime--dhek6oq.json @@ -0,0 +1,241 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-room-i00-standard-realtime--dhek6oq", + "taskKey": "humanoidbench/manip_room", + "dag_sha256": "081a0a8f4158f72c3be894442d203666d7e71a9e8771af0ea997f42197bdbd1c", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 969, + "sim_steps": 22652, + "calls": 30937, + "rows": 22555, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3418.8, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 452.274, + "max_lag_s": 0.1006, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 2857, + "late_steps": 0, + "requests_refused": 43, + "request_busy_s": 17.076, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.302, + "peak_processes": 12, + "peak_disk_gb": 0.863, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-room-i00-standard-realtime--dhek6oq/replay.mp4", + "duration_s": 9.75, + "fps": 20.0, + "frames": 195, + "sha256": "bc3cddd86fe61f5cbca201f0eadccdfc002d31e21dee6b4205e9c5ad8b9e883c" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "objects_gathered": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 140, + "text_zh": "观察:视频 0.0–8.5 s,机器人从初始位置侧向行走,双手未持物;可见椅子与红色散物仍分开。log step 140 启动最后一次重置后的控制程序,先前回合均已重置,不计入本段视频的进展。" + }, + { + "t": 9.15, + "log_step": null, + "text_zh": "观察:逐帧查看视频 8.5–9.7 s,机器人在奖杯旁抬脚扫动;视频 9.05–9.2 s 足部与奖杯底座区域相接,奖杯短距离滑移,仍直立,未见被抓起或搬走。视频约 9.3 s 起机器人倾倒,视频 9.7 s 已翻倒;可见物品没有形成一组。" + }, + { + "t": 9.7, + "log_step": 142, + "text_zh": "推断:本节点按累计奖励达到 400 验收;run.json 的 graded_by_live=1、live_success=0、finalized=1,且 log step 142 的服务最终状态为 episode=51、control_steps=969、done=true、reason=failed,与视频末段翻倒一致,因此累计事件未达成,final=false;不是用零分推断机器人未尝试。评分字段未报告累计奖励的数值。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.0–9.3 s,H1 双脚交替接触地面并支撑机器人侧向行走,随后支撑失稳并翻倒。" + }, + { + "who": "腿/躯干", + "how": "推/扫", + "note_zh": "观察:视频 8.7 s 左右机器人在奖杯旁抬脚扫动;视频 9.05–9.2 s,靠近奖杯的脚在底座区域接触并将奖杯推移一小段。该视角不能可靠判别每次短促接触的左右脚及接触力,未见手部抓握。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "unit": "奖励", + "threshold": 400, + "direction": "higher_better", + "value": null + }, + { + "name": "max_xy_variance", + "unit": "平方米", + "threshold": null, + "direction": "lower_better", + "value": null + }, + { + "name": "room_object_organized", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "stand_reward", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "small_control", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "per_timestep_reward", + "unit": "奖励", + "threshold": null, + "direction": "higher_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "objects_gathered", + "t": 9.3, + "log_step": 142, + "where": "操作中", + "evidence_zh": "观察:视频 8.7–9.2 s,机器人在奖杯旁扫脚并短距离推移奖杯;视频约 9.3 s 开始明显倾倒,视频 9.5–9.7 s 躯干下落、双腿翻起,未恢复站立或继续归拢。log step 142 的服务记录最后回合 episode=51 以 reason=failed 结束,resets_left=0;失败发生在归拢物品的脚部操作及恢复支撑过程中。" + }, + "unmet": { + "objects_gathered": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 9.3, + "log_step": 142, + "evidence_zh": "观察:视频 9.0–9.3 s 机器人双脚间距扩大,随后躯干持续倾斜;视频 9.5–9.7 s 机器人失去足底稳定支撑、身体翻倒,未修复。该现象来自最后评分回合的画面,不取自此前调试回合的跌倒。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "t": 9.3, + "log_step": 136, + "source_zh": "日志", + "evidence_zh": "推断:log step 133 的脚本将预设扫腿轨迹镜像到另一条腿,并为支撑腿设置髋踝反馈;log step 136 将扫腿改为分段的髋、膝目标并在扫动结束后恢复步态;log step 140 仅调整导航目标及侧向速度后启动最终回合。视频 8.7–9.7 s 显示该控制程序执行扫脚后脚距扩大、身体倾倒,未能恢复稳定支撑,支持归因为全身动作控制未稳住机器人。", + "note_zh": "未提供目标与实际关节轨迹的完整记录及足底力,不能进一步证实是具体反馈增益、轨迹跟踪还是支承切换问题,保留在“控制”。" + } + ], + "gap_zh": "没有状态日志或成功参考视频;本视频为跟随机器人的单视角,不能同时核实六件物品的全部位置,足部与奖杯的短促接触也有遮挡,接触力和各次接触的左右脚无法确定。评分字段未报告 episode_return、max_xy_variance、room_object_organized、stand_reward、small_control 或 per_timestep_reward,六项数值均置 null。视频时长为 9.75 s,而 log step 142 的末回合 sim_time_s 为仿真 19.38 s;本文所有视频时刻直接取自帧时间,未将 agent 时间或仿真时间换算为视频时刻。" + }, + "summary_zh": "最后评分回合中,机器人侧向走到奖杯旁,抬脚扫动并使奖杯短距离滑移,未达成“六件物品归拢”的累计验收,节点最终为 0/1。视频约 9.3 s 起机器人失去平衡并翻倒,断点为归拢操作中,日志中的扫腿及支撑控制程序支持根因标为控制。缺少状态日志、完整房间视图及累计奖励数值,不能细分控制失稳的具体机制。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "单节点未达成、断点在归拢操作中、平衡现象和控制根因均有视频与日志支撑(视频 9.3 s 起翻倒,log step 142 的 pitch/roll 与脚距骤变一致)。未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:13:59Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-standard-realtime--dhek6oq/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-standard-realtime--dhek6oq/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "5920ab341cb971bce1427d7e3ad21da17911b1c6fb4a5e86ef1435caa5083cee", + "video_sha256": "bc3cddd86fe61f5cbca201f0eadccdfc002d31e21dee6b4205e9c5ad8b9e883c", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard-realtime--ekbtle5.json b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard-realtime--ekbtle5.json new file mode 100644 index 0000000000000000000000000000000000000000..a24736aa0cdf74effa04dadc0fb363fd62c506eb --- /dev/null +++ b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard-realtime--ekbtle5.json @@ -0,0 +1,242 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-room-i00-standard-realtime--ekbtle5", + "taskKey": "humanoidbench/manip_room", + "dag_sha256": "081a0a8f4158f72c3be894442d203666d7e71a9e8771af0ea997f42197bdbd1c", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 102, + "sim_steps": 5341, + "calls": 40113, + "rows": 41290, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 750.0, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 106.029, + "max_lag_s": 0.003, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 181, + "late_steps": 0, + "requests_refused": 46, + "request_busy_s": 12.505, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-room-i00-standard-realtime--ekbtle5/replay.mp4", + "duration_s": 1.1, + "fps": 20.0, + "frames": 22, + "sha256": "4ef8f47ee844e5e7b5d673852c7621bc94e828ac8e309fb75ec15416f8d41a07" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "objects_gathered": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:已逐帧检查视频 0.00–1.05 s 的全部 22 帧;H1 双脚着地、屈膝原地站立,双手伸在身体前方,未见抓取、携带或推拢散物的动作;视频 0.80–1.05 s 向后倾倒,骨盆降至地面附近。镜头只覆盖机器人附近,不能据此逐件测量六物体间距。" + }, + { + "t": 1.05, + "log_step": 22, + "text_zh": "观察:log step 20、22 的机器人服务状态均为 episode=51、control_steps=102、done=True、reason=failed、resets_left=0;log step 22 的 reset 被拒绝,评分器字段 graded_by_live=1、live_success=0、finalized=1。" + }, + { + "t": 1.05, + "log_step": 22, + "text_zh": "推断:按任务 DAG 以累计奖励达到 400 为唯一目标验收,最后现场回合已终止且未触发成功,因此 objects_gathered 未达成;结合视频只有站立和跌倒的过程,不把此前会话中的测试或结束后编写的控制器算作归拢进展。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.75 s,H1 双脚与地面接触承重,仅进行了归拢动作前的站立准备;视频 0.80–1.05 s 支撑失稳并向后跌落。未见手部抓握或推扫六件散物,不按 DAG 的典型方案补记物体接触。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "unit": "奖励", + "threshold": 400, + "direction": "higher_better", + "value": null + }, + { + "name": "max_xy_variance", + "unit": "平方米", + "threshold": null, + "direction": "lower_better", + "value": null + }, + { + "name": "room_object_organized", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "stand_reward", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "small_control", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "per_timestep_reward", + "unit": "奖励", + "threshold": null, + "direction": "higher_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "objects_gathered", + "t": 0.8, + "log_step": 20, + "where": "首步前", + "evidence_zh": "观察:视频 0.00–0.75 s 机器人只在原地屈膝站立,未启动取物或推拢;视频 0.80 s 起躯干明显后仰,至视频 1.05 s 骨盆已接近地面,未恢复。log step 20 的服务状态确认最后的 episode 51 已 failed;log step 22 再次确认该回合终止。" + }, + "unmet": { + "objects_gathered": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 18, + "evidence_zh": "观察:视频 0.00–1.05 s 机器人始终留在起始位置附近,双手未执行取物或推拢动作;log step 18 运行的是站姿反馈测试脚本,未含物品归拢阶段。推断:评分回合在实际归拢动作开始前已结束,符合首步前的没尝试。" + }, + { + "cat": "全身", + "item": "平衡", + "t": 0.8, + "log_step": 20, + "evidence_zh": "观察:视频 0.75 s 机器人尚由双脚支撑;视频 0.80–0.95 s 躯干持续后仰、骨盆下降;视频 1.00–1.05 s 身体向后落下,骨盆接近地面,未见恢复。log step 20 的服务状态显示最后回合 done=True、reason=failed。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "log_step": 18, + "source_zh": "日志", + "note_zh": "控制脚本异常:异常处理将 submit 失败转为反复 reset,未形成有效的持续控制;现有证据不足以细分物理失稳的反馈控制参数。", + "evidence_zh": "观察:log step 18 中 agent 将控制循环改成 submit 抛异常后调用 fresh()(执行 reset)并重启循环,输出连续出现 EpisodeOver 和递减的 resets_left;log step 20、22 的服务状态随后固定为 episode 51、102 控制步、failed,且 reset 被拒绝。log step 21 才编写最终 gen.py,log step 30 执行它时服务仍返回 episode finished。推断:agent 自己的控制脚本错误处理造成反复重启,最终未在被评分回合中实现有效的站立控制或归拢动作;该因果依据具体代码与服务反馈,不将预算耗尽本身当作根因,也不把 agent 自述的服务竞态当作已证实的环境故障。" + } + ], + "gap_zh": "没有状态日志及成功示范;视频为机器人近景,六物体不在完整视野中,无法独立恢复其坐标、接触力或各项原生奖励,评分器也未报告六项 native_metrics 的数值,均留 null。log step 22 报告 102 个控制步,按任务文本 0.02 s/步为仿真约 2.04 s,而视频长 1.10 s;缺少可靠逐帧映射,因此全部导航时间直接使用视频帧,未把服务步数或 agent 时间换算为视频时间。重试造成持续控制中断有代码和反馈支持,但最终回合每步实际目标及具体失稳机制未完整记录。" + }, + "summary_zh": "最后评分回合未启动六件物品归拢,唯一节点 objects_gathered 未达成,最终进度为 0/1;机器人在视频约 0.80 s 开始明显后仰,至视频 1.05 s 向后跌落。日志显示 agent 的提交异常重试不断触发重置,最终失败回合被保留,之后编写的控制器未能进入新回合,根因归为控制。缺少状态日志与房间全景,无法恢复各物体接触和原生奖励数值,且不将服务控制步数直接映射到这段短视频。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "节点未达、首步前/没尝试 + 全身·平衡(视频 0.8 s 起后倾、1.05 s 跌落)、根因停在“控制”均与视频和 log step 18–22(episode 51、102 控制步、failed、resets_left 0)一致;时钟未换算,未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:13:29Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-standard-realtime--ekbtle5/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-standard-realtime--ekbtle5/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "ab528fe90e44dfba43b4f2a0d03162d38b960fe4b059f6a7e1a1f92ca2f4d950", + "video_sha256": "4ef8f47ee844e5e7b5d673852c7621bc94e828ac8e309fb75ec15416f8d41a07", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard-realtime--h2juv3h.json b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard-realtime--h2juv3h.json new file mode 100644 index 0000000000000000000000000000000000000000..8138fc03fdd1571cf44a645383010f78de99656d --- /dev/null +++ b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard-realtime--h2juv3h.json @@ -0,0 +1,237 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-room-i00-standard-realtime--h2juv3h", + "taskKey": "humanoidbench/manip_room", + "dag_sha256": "081a0a8f4158f72c3be894442d203666d7e71a9e8771af0ea997f42197bdbd1c", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 105, + "sim_steps": 5243, + "calls": 14512, + "rows": 262957, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 550.0, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 104.114, + "max_lag_s": 0.0101, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 210, + "late_steps": 3780, + "requests_refused": 45, + "request_busy_s": 6.551, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.17, + "peak_processes": 9, + "peak_disk_gb": 0.239, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-room-i00-standard-realtime--h2juv3h/replay.mp4", + "duration_s": 1.15, + "fps": 20.0, + "frames": 23, + "sha256": "6e138666a8cd590dc6dbb765865e457a10560b47b02983e1c4982ddc56968a28" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "objects_gathered": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:已检查视频 0.00–1.10 s 的全部 23 帧;机器人从屈膝站立逐渐倾倒,双手未见抓取、提起或推送房间散物的动作,视频 1.05–1.10 s 已倒地,未出现物品归成一组的结果。" + }, + { + "t": 1.1, + "log_step": 60, + "text_zh": "观察:log step 60 的机器人服务状态显示最后回合 episode=51、control_steps=105、done=true、reason=failed、resets_left=0;inputs/run.json 的评分字段 success=0、live_success=0、replay_success=0。" + }, + { + "t": 1.1, + "log_step": null, + "text_zh": "推断:本节点按累计奖励达到 400 验收,评分器未通过,且视频没有归拢动作或结果,因此累计达标事件未发生;失败分数未用于反推某次抓取是否发生。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.75 s,H1 双脚在地面承重,机器人只进行归拢前的站姿维持;随后失衡倒地。未见灵巧手与散物建立抓握或推拢接触。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "unit": "奖励", + "threshold": 400, + "direction": "higher_better", + "value": null + }, + { + "name": "max_xy_variance", + "unit": "平方米", + "threshold": null, + "direction": "lower_better", + "value": null + }, + { + "name": "room_object_organized", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "stand_reward", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "small_control", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "per_timestep_reward", + "unit": "奖励", + "threshold": null, + "direction": "higher_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "objects_gathered", + "t": 0.8, + "log_step": 60, + "where": "首步前", + "evidence_zh": "观察:视频 0.00–0.75 s 机器人只屈膝站立,没有开始抓取或推拢物品;视频约 0.80 s 躯干倾斜明显增大,视频 1.05–1.10 s 身体落地,直到片尾未恢复。log step 60 确认最后回合已失败结束,故断点为尚未启动的六件物品归拢。" + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 0.8, + "log_step": null, + "evidence_zh": "观察:视频 0.80–1.10 s,机器人躯干由倾斜变为接近水平,膝腿折下、身体落地,双脚未能维持站立支撑;片尾没有起身。" + }, + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 59, + "evidence_zh": "观察:视频 0.00–1.10 s 没有朝散物进行抓取、搬移或推拢;log step 59 最后运行的是站姿平衡增益扫描,随后 log step 60 报最后回合已失败。归拢动作在该评分回合中没有启动。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "反馈控制", + "source_zh": "视频 / 日志", + "evidence_zh": "推断:最后回合的站姿反馈未能稳定机器人,使其在归拢开始前倒地。视频 0.60–1.10 s 显示倾斜持续扩大直至落地;log step 58 将平衡脚本提交改为不传 at 后,各测试虽 late=0 仍以 failed 结束;log step 59 延续该脚本测试踝部 IMU 反馈(kx=6、kwy=1.2 等组合),log step 60 服务确认最后回合失败。只能据此归因于站姿反馈未稳定,不把早期回合的迟到指令或错误足底公式直接当作最后回合的已证实原因。" + } + ], + "gap_zh": "无状态日志和成功参考视频;评分视频只有跟随机器人视角,未完整展示六件物品,不能量化物品坐标方差、接触力或累计回报,六项原生指标均未报告。log step 60 的最后回合有 105 个控制步,任务文本给出每步 0.02 s,名义仿真时长为 2.10 s,而视频仅 1.15 s,缺少可靠映射;所有导航时刻均直接取自视频,未用控制步换算。log step 59 的输出被重置预算耗尽异常占据,没有最后回合的完整反馈与执行器轨迹,因此无法确定具体失稳增益或关节饱和原因。", + "unmet": { + "objects_gathered": false + } + }, + "summary_zh": "评分回合未完成“六件物品归拢”,最终进度为 0/1;机器人只维持屈膝站姿,视频约 0.80 s 明显失衡,视频 1.05–1.10 s 已倒地,归拢动作尚未启动。结合 log step 58–60 的平衡脚本和最后回合服务状态,根因判断为站姿反馈未能稳定机器人。没有状态日志和原生奖励数值,且最后回合控制步数与视频时长缺少可靠映射,不能进一步确定具体控制参数问题。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "标签与视频、日志一致:最后回合机器人屈膝站立后约0.7–0.8 s(视频时钟)开始倾倒,1.05 s 已倒地,从未接触散物;节点未达、断点首步前、平衡+没尝试、反馈控制根因均有证据支持(与 FAIL-13 powerlift 同型)。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:13:08Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-standard-realtime--h2juv3h/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-standard-realtime--h2juv3h/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "460c559f95c1d347d2dbb388e5bc0230b880cdd3eb1e2764caa9a669e7b9238d", + "video_sha256": "6e138666a8cd590dc6dbb765865e457a10560b47b02983e1c4982ddc56968a28", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard-realtime--inlztte.json b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard-realtime--inlztte.json new file mode 100644 index 0000000000000000000000000000000000000000..723df6b5793029aad93f6dd11635237c2a87c428 --- /dev/null +++ b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard-realtime--inlztte.json @@ -0,0 +1,216 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-room-i00-standard-realtime--inlztte", + "taskKey": "humanoidbench/manip_room", + "dag_sha256": "081a0a8f4158f72c3be894442d203666d7e71a9e8771af0ea997f42197bdbd1c", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 1000, + "sim_steps": 1566, + "calls": 1852, + "rows": 1844, + "resets": 5, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1368.9, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 31.24, + "max_lag_s": 0.0211, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 41, + "late_steps": 0, + "requests_refused": 7, + "request_busy_s": 0.969, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-room-i00-standard-realtime--inlztte/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "bc5b142629bc44ea816ba9c58184af5b16e4f1714de01305606c5f2ef26b6915" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "objects_gathered": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.0 s 起,H1 双脚着地、双臂伸在身前;按视频每 0.5 s 查看至视频 10.0 s,并逐帧查看视频 9.0–10.05 s,机器人始终原地站立,未见走向物品、抓取、推拢或搬运。当前侧视画面不能同时展示六件物品的全局布局。" + }, + { + "t": null, + "log_step": 70, + "text_zh": "观察:log step 70 的最后一版现场控制脚本只计算双腿站立与重心反馈,其他关节目标保持零;log step 71 启动此脚本,log step 72 的 agent 自述也是等待站立回合完成。其后 log step 77–129 是本地 MuJoCo 行走或转向测试,未见再次向现场服务部署归拢控制器;这些测试不作为视频中的动作。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:run.json 评分器字段 finalized=1、live_success=0、replay_success=0、deterministic=1,结合视频仅站立的过程,判定唯一绑定累计验收的“六件物品归拢”未达到门槛,final=false、无达成区间;不把目录 reward=0.0 当成原生累计奖励数值,也不据此推定具体接触。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "unit": "奖励", + "threshold": 400, + "direction": "higher_better", + "value": null + }, + { + "name": "max_xy_variance", + "unit": "平方米", + "threshold": null, + "direction": "lower_better", + "value": null + }, + { + "name": "room_object_organized", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "stand_reward", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "small_control", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "per_timestep_reward", + "unit": "奖励", + "threshold": null, + "direction": "higher_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "objects_gathered", + "t": 0.0, + "log_step": 70, + "where": "首步前", + "evidence_zh": "观察:视频 0.0–10.05 s 机器人维持原地站立,未启动抓、推、搬移物品的归拢动作;log step 70–72 记录最后现场脚本只执行站立控制。推断:评分回合停留在归拢动作开始之前,视频 0.0 s 仅作为导航锚点。" + }, + "unmet": { + "objects_gathered": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 70, + "evidence_zh": "观察:视频 0.0 s、视频 0.5–9.0 s 的抽帧及视频 9.0–10.05 s 的逐帧画面均只见 H1 站立,没有面向目标的移动或操作;log step 70 的现场脚本只有站立控制,log step 71 启动后没有部署物品归拢动作的记录。" + } + ], + "root_causes": [], + "gap_zh": "观察:输入没有状态日志或成功参考视频,侧视视频不能展示全部物品布局;评分器没有报告 episode_return、方差及奖励分项,故六项原生指标均留 null。log step 77–129 记录大量本地行走、转向控制测试,包含失败,也有 log step 122 的直行测试完成;这些是 agent 自建仿真中的会话证据,不能确认现场为何始终没有启动归拢,因此根因层留空。log step 72 的 agent 自述现场回合为 20 秒,而提供的视频长 10.1 秒,未建立控制步到视频时间的可靠映射,所有视频锚点均取自已查看帧。" + }, + "summary_zh": "观察:评分视频中 H1 全程维持原地站立,未启动物品归拢操作,最终节点为 0/1,评分器记录失败。推断:断点为“六件物品归拢”的首步前,现象为没尝试;此前现场控制测试和此后本地行走测试不计入该视频的进展。缺少全局物品视图、状态日志和原生累计奖励,且日志不足以确定未启动归拢的具体根因。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "标注与视频和日志一致:视频 0–10.05 s H1 始终原地站立,未走向或触碰物品;归拢节点 final=false、断点首步前/没尝试,符合 CASES.md FAIL-22 的处理;根因层留空并有 gap_zh,可接受。log step 引用(70–72、77–129)经核对存在且内容与证据文字相符。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:12:29Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-standard-realtime--inlztte/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-standard-realtime--inlztte/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "7f1c61955fca413578182b659a00cc96702b0c64c636768a09baab1b9b7bef10", + "video_sha256": "bc5b142629bc44ea816ba9c58184af5b16e4f1714de01305606c5f2ef26b6915", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard-realtime--tbqqjzi.json b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard-realtime--tbqqjzi.json new file mode 100644 index 0000000000000000000000000000000000000000..7cedecb2465621db4603e2f9ee1966ad15c1428d --- /dev/null +++ b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard-realtime--tbqqjzi.json @@ -0,0 +1,224 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-room-i00-standard-realtime--tbqqjzi", + "taskKey": "humanoidbench/manip_room", + "dag_sha256": "081a0a8f4158f72c3be894442d203666d7e71a9e8771af0ea997f42197bdbd1c", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 1000, + "sim_steps": 16357, + "calls": 47165, + "rows": 47020, + "resets": 48, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1118.0, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 326.501, + "max_lag_s": 0.1671, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 1096, + "late_steps": 0, + "requests_refused": 50, + "request_busy_s": 19.154, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.315, + "peak_processes": 10, + "peak_disk_gb": 0.997, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-room-i00-standard-realtime--tbqqjzi/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "6413f5333938d363ae9c8c66850de5693b2a3c75a8dafbd76019fb06c9174439" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "objects_gathered": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00–1.00 s 的逐帧画面中,机器人从屈膝姿态伸腿站起;视频 1.0–9.5 s 每隔约 0.5 s 的画面及视频 9.00–10.05 s 的逐帧画面中,双脚保持在原处,双手空置、前伸,没有向散物行走、抓取或推拢的动作。归拢步骤未启动,因此不列该步骤的 contact_actual;原地站立的脚底支撑不等于尝试归拢。" + }, + { + "t": null, + "log_step": 156, + "text_zh": "观察:log step 155 启动最终站立控制脚本,log step 156 的服务状态为 episode=49、control_steps=1000、done=True、reason=failed、resets_left=2;log step 159–169 的后续状态检查仍为第 49 回合且已失败。log step 103、105 的脚本只执行站立并在跑满 1000 步后保留该回合;不是此前步态测试回合。" + }, + { + "t": null, + "log_step": null, + "text_zh": "推断:评分字段 live_success=0、replay_success=0,结合唯一目标的累计奖励达到 400 即通过并结束回合这一判据,最终未达到 objects_gathered 的累计验收条件;没有用零分反推未接触,未启动的判断来自视频和最终控制脚本。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "unit": "奖励", + "threshold": 400, + "direction": "higher_better", + "value": null + }, + { + "name": "max_xy_variance", + "unit": "平方米", + "threshold": null, + "direction": "lower_better", + "value": null + }, + { + "name": "room_object_organized", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "stand_reward", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "small_control", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "per_timestep_reward", + "unit": "奖励", + "threshold": null, + "direction": "higher_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "objects_gathered", + "t": 0.0, + "log_step": 155, + "where": "首步前", + "evidence_zh": "观察:视频 0.00–10.05 s 中机器人只起身并原地调节站姿,未开始向物品移动或归拢;log step 155–156 保留的最终第 49 回合由站立脚本产生,服务报告跑满 1000 控制步后 failed。此前步态测试被重置,不属于本次评分回合。" + }, + "unmet": { + "objects_gathered": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 156, + "evidence_zh": "观察:视频 0.00–1.00 s 仅伸腿起立,视频 1.0–9.5 s 的连续抽帧及视频 9.00–10.05 s 的逐帧检查均显示机器人原地站立、双手空置,未实施物品归拢。log step 103、105 的最终回合脚本仅用本体反馈维持站立,log step 156 确认保留跑满 1000 步的失败回合。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "log_step": 113, + "source_zh": "日志", + "evidence_zh": "推断:控制器未形成可供搬运物品使用的步态,是最终只保留站立回合的能力缺口。独立的会话测试记录支持这一判断:log step 113 的行走测试到 control step 74 即结束,脚本统计 maxp=1.29 rad;log step 133 的移重心/抬脚测试到 control step 84 结束,maxr=1.101 rad;log step 153 的缓慢侧向移重心测试到 control step 245 结束,maxp=1.252 rad、maxr=1.02 rad。agent 在 log step 158 的报告中自述未找到可用步态;最终脚本(log step 103、139)仅保留站立控制。上述测试只用于根因推断,不把测试中的失衡标为最终视频的现象,也不足以证明机器人结构上不可行。" + } + ], + "gap_zh": "视频只有机器人附近的侧视画面,没有覆盖六件物品的房间全景;inputs/ 无状态日志,评分字段也未报告 episode_return、物品方差及奖励分项,故全部原生指标填 null。log step 156 的 1000 控制步按服务 control_dt=0.02 对应仿真 20 s,而视频仅 10.1 s,缺少经可见事件核实的映射,不将控制步或 agent 会话时间换成视频时间。控制根因来自会话测试与最终脚本,具体是步态设计、轨迹还是反馈参数的问题尚无法区分。" + }, + "summary_zh": "评分回合中机器人仅起身并维持原地站立,未启动六件物品归拢,最终节点为 0/1;log step 156 确认第 49 回合跑满 1000 控制步后失败。断点为 objects_gathered 的首步前,现象是没尝试;会话内多次步态测试失稳、最终仅交付站立控制,支持保守归因于控制能力不足。视频缺少房间全景,且没有状态日志或原生累计奖励数值,控制步与视频时钟映射亦未核实。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "评分回合(第49回合)视频中机器人只起身原地站立、双手空置,节点未达、断点首步前/没尝试均有帧和 log 156 支持;根因控制依据会话步态测试,符合 FAIL-22 的允许写法,且已写明不确定性。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:10:22Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-standard-realtime--tbqqjzi/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-standard-realtime--tbqqjzi/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "9ba147d12a8ef5bd195775bfe9c0f2ca47934501881efcab0b3ee8e7c35312d3", + "video_sha256": "6413f5333938d363ae9c8c66850de5693b2a3c75a8dafbd76019fb06c9174439", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard-realtime--uw3f35y.json b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard-realtime--uw3f35y.json new file mode 100644 index 0000000000000000000000000000000000000000..b83c9352e9d2aad47d576d9ddbfc36d1634e76da --- /dev/null +++ b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard-realtime--uw3f35y.json @@ -0,0 +1,224 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-room-i00-standard-realtime--uw3f35y", + "taskKey": "humanoidbench/manip_room", + "dag_sha256": "081a0a8f4158f72c3be894442d203666d7e71a9e8771af0ea997f42197bdbd1c", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 1000, + "sim_steps": 8957, + "calls": 14724, + "rows": 14664, + "resets": 29, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 593.6, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 178.678, + "max_lag_s": 0.0037, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 0, + "late_steps": 0, + "requests_refused": 30, + "request_busy_s": 6.083, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.156, + "peak_processes": 11, + "peak_disk_gb": 0.239, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-room-i00-standard-realtime--uw3f35y/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "acf0d96ad151c5f6b5da578c956bb47e9574fd7898de59a42cca9ef5caa8e2fa" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "objects_gathered": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.3, + "log_step": 33, + "text_zh": "观察:视频 0.0–0.3 s 中机器人由屈膝、前臂向前变为较直立且手臂下垂;此后逐半秒检查至视频 9.5 s,并密查视频 9.0–10.05 s,机器人始终留在原处,没有走向物品或伸手抓推的动作。log step 33 显示 agent 恢复并启动站立脚本,与画面相符;归拢操作未启动,因此不列该节点的任务接触。" + }, + { + "t": 10.05, + "log_step": 34, + "text_zh": "观察:视频 10.05 s 最后一帧机器人仍站立。log step 34 的机器人服务状态为 episode=30、control_steps=1000、done=True、reason=failed,calls=14724、rows=14664、control_steps_total=8957 与评分器 reward 字段一致,确认最后这次站立回合受评。" + }, + { + "t": 10.05, + "log_step": 34, + "text_zh": "推断:该节点按累计奖励达到 400 验收;评分器 live_success=0、replay_success=0,且最终服务回合为 failed,故此累计事件没有达标,final=false。视角未覆盖六件物品的全貌,不能据画面估计其方差;不将目录 reward=0.0 当作原生累计奖励。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "unit": "奖励", + "threshold": 400, + "direction": "higher_better", + "value": null + }, + { + "name": "max_xy_variance", + "unit": "平方米", + "threshold": null, + "direction": "lower_better", + "value": null + }, + { + "name": "room_object_organized", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "stand_reward", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "small_control", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "per_timestep_reward", + "unit": "奖励", + "threshold": null, + "direction": "higher_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "objects_gathered", + "t": 0.0, + "log_step": 33, + "where": "首步前", + "evidence_zh": "观察:视频 0.0–10.05 s 仅见机器人起身、双脚支撑并维持原地站立,没有向散物移动或抓推物品的动作。log step 33 启动的是恢复后的站立控制器,log step 34 确认最后回合以 failed 结束;六件物品归拢这一步始终没有开始,视频 0.0 s 仅作为起点导航锚。" + }, + "unmet": { + "objects_gathered": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 33, + "evidence_zh": "观察:视频 0.0–0.3 s 机器人伸直双腿并放下手臂,此后到视频 10.05 s 保持在原处,未出现朝物品行进或进行抓取、搬移、推拢的动作;log step 33 的最后控制器只执行站立与踝部平衡反馈。早先步态测试不属于此评分回合。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "log_step": 31, + "source_zh": "日志", + "evidence_zh": "推断:会话测试显示 agent 未形成可用于接近物品的稳定步态控制器。log step 29 的交替抬腿脚本记录重力侧向分量升至 0.89;log step 30–31 的多种横向反馈和抬腿设置均在 control step 242–284 附近结束,脚本估计的横向重心偏移绝对值达 0.85–0.92 m,而 log step 31 关闭抬腿的对照可运行至 control step 999、偏移约 0.05 m。log step 33 随后恢复纯站立控制器;这些独立测试支持行走控制能力不足,解释为何评分回合未进入物品操作,但不足以区分具体控制子类。测试计数不映射为评分视频时刻,也不将测试中的失稳标作评分视频的现象。" + } + ], + "gap_zh": "视频仅提供跟随机器人的局部外视角,未覆盖六件散物全貌;inputs/ 无状态日志、成功参考视频或原生累计奖励及分项读数,无法计算物品方差与奖励缺口。log step 34 记录最终回合 1000 个控制步,任务规定每步仿真 0.02 s,即仿真 20 s,而视频长 10.1 s;缺少可核对的仿真到视频映射,本文只用实际帧时间,不换算控制步定位。控制根因来自会话测试及最终脚本选择的联合推断,尚不能定位到某个反馈参数或接触设计。" + }, + "summary_zh": "评分视频中机器人起身后全程原地站立,未启动六件物品归拢,节点最终为 0/1;log step 34 的最终回合 failed 与评分结果一致。会话步态测试反复失稳,agent 在 log step 33 恢复站立脚本,支持控制能力不足这一根因,评分片段本身没有跌倒现象。局部视角、缺少状态与原生奖励读数,以及视频和仿真时长映射未明,限制了数值与具体控制子类判断。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "视频全程机器人原地站立,未向物品移动,节点 objects_gathered 未达成、断点首步前/没尝试与帧和 log step 33–35 一致;根因标为推断并说明会话测试不映射评分视频,时钟处理合规。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:07:30Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-standard-realtime--uw3f35y/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-standard-realtime--uw3f35y/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "7622027ef18a23b43d2865544b676cd979b0de61aeeaea069f5c96d194b34e7f", + "video_sha256": "acf0d96ad151c5f6b5da578c956bb47e9574fd7898de59a42cca9ef5caa8e2fa", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard-realtime--vjxstyq.json b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard-realtime--vjxstyq.json new file mode 100644 index 0000000000000000000000000000000000000000..05e2eb53ac80479511ffedbd94d51ccbbb5333a1 --- /dev/null +++ b/phase2/humanoidbench/manip_room/trajectories/humanoidbench-manip-room-i00-standard-realtime--vjxstyq.json @@ -0,0 +1,232 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-room-i00-standard-realtime--vjxstyq", + "taskKey": "humanoidbench/manip_room", + "dag_sha256": "081a0a8f4158f72c3be894442d203666d7e71a9e8771af0ea997f42197bdbd1c", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 101, + "sim_steps": 4067, + "calls": 3817, + "rows": 11332, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 590.6, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 80.639, + "max_lag_s": 0.0659, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 762, + "late_steps": 0, + "requests_refused": 37, + "request_busy_s": 3.091, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.187, + "peak_processes": 6, + "peak_disk_gb": 0.251, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-room-i00-standard-realtime--vjxstyq/replay.mp4", + "duration_s": 1.1, + "fps": 20.0, + "frames": 22, + "sha256": "3ef70d6b1896f57c9e0f74121017759cd4f0862adfc52885d4af2718404c56e1" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "objects_gathered": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:逐帧检查视频 0.00–1.05 s,机器人起初屈膝站立、两臂前伸,随后身体倾倒;未见机器人走向物品、抓取、搬运或推拢物品,视频 1.05 s 时骨盆已落近地面。" + }, + { + "t": 1.05, + "log_step": 92, + "text_zh": "推断:inputs/run.json 的评分器字段 success、live_success、replay_success 均为 0;log step 92 的机器人仿真服务状态为最后回合 episode=51、done=True、reason=failed。该 DAG 唯一目标绑定累计奖励达到 400,故最终未达标;这不是仅从零分推断机器人没有动作。" + }, + { + "t": null, + "log_step": 9, + "text_zh": "观察:log step 9 的服务规则说明只评分最后一次回合;log step 90、log step 92 确认最终为 episode=51、resets_left=0,较早重置回合的测试不计入节点进度。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "unit": "奖励", + "threshold": 400, + "direction": "higher_better", + "value": null + }, + { + "name": "max_xy_variance", + "unit": "平方米", + "threshold": null, + "direction": "lower_better", + "value": null + }, + { + "name": "room_object_organized", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "stand_reward", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "small_control", + "unit": "无量纲", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "per_timestep_reward", + "unit": "奖励", + "threshold": null, + "direction": "higher_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "objects_gathered", + "t": 0.85, + "log_step": null, + "where": "首步前", + "evidence_zh": "观察:视频 0.00–0.75 s 中机器人仅原地站立和调整肢体,未启动抓取、搬移或推拢;视频约 0.85 s 起失去站立支撑,视频 0.95–1.05 s 双脚前伸、骨盆下落,直到片尾未恢复,归拢步骤始终未启动。" + }, + "unmet": { + "objects_gathered": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": null, + "evidence_zh": "观察:视频 0.00–1.05 s 的全部 22 帧中,机器人未朝任何散物实施抓取、推拢或搬移;两臂随站姿调整和跌倒运动,没有开始归拢动作。" + }, + { + "cat": "全身", + "item": "平衡", + "t": 0.85, + "log_step": null, + "evidence_zh": "观察:视频 0.80–0.90 s 机器人躯干倾斜、双腿失去原先支撑姿态;视频 0.95–1.05 s 双脚向前抬起、骨盆下降到地面附近,身体倾倒且未恢复。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "反馈控制", + "t": 0.85, + "log_step": 84, + "source_zh": "日志", + "evidence_zh": "推断:log step 84 保存的控制脚本以重力方向、角速度反馈调整髋、膝、踝和肩关节以维持站立;log step 90 的最后回合记录中,控制步 50 时重力方向为 (0.027,0.021,-0.999),到控制步 100 变为 (0.781,-0.009,-0.624),滤波俯仰角速度达 4.958,髋与踝修正分别达到脚本限幅 0.220、0.150,随后服务回报 failed。结合视频 0.85–1.05 s 的实际跌倒,判断站立反馈未能抑制失稳,阻止了归拢动作启动;这些控制步读数仅作根因证据,不换算为视频时间。" + } + ], + "gap_zh": "无状态日志或成功示范;视频只有跟随机器人的局部视角,不能同时核实六件散物的完整位置,也不能量化物体方差、执行器力或累计奖励,六项原生指标在评分器字段中均未报告。log step 92 报告最终回合有 101 个控制步,log step 7 给出控制周期 0.02 s,二者对应仿真 2.02 s,而视频仅 1.10 s;缺少可验证的两者时间映射,所有视频锚点均直接取自所检查帧,日志数值不投射到视频时钟。" + }, + "summary_zh": "评分回合未启动六件物品的抓取、搬移或推拢,唯一目标“六件物品归拢”最终未达标(0/1)。机器人在首步前原地维持站姿,于视频约 0.85 s 失衡并在视频 1.05 s 降至地面附近;控制脚本和末回合反馈支持站立反馈控制失稳。没有状态日志与累计奖励数值,视频亦未覆盖全房间,且其时长与日志控制步时长无法直接对齐。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "评分回合(episode 51,log step 90 末段)机器人只站立后倒下,未启动归拢;节点未达、断点(首步前)、没尝试+平衡、反馈控制根因均与视频帧和日志读数相符,时钟处理(日志控制步不换算视频时间)合规。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:12:09Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-standard-realtime--vjxstyq/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-room-i00-standard-realtime--vjxstyq/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "05918f0d28fdbeef510e00ec3221f5b6b58f7cd7c79da5a591819506db7f6410", + "video_sha256": "3ef70d6b1896f57c9e0f74121017759cd4f0862adfc52885d4af2718404c56e1", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-privileged--9mkr2ud.json b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-privileged--9mkr2ud.json new file mode 100644 index 0000000000000000000000000000000000000000..6a25a7ad55f825b86aa99716515ecbbee116d3a8 --- /dev/null +++ b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-privileged--9mkr2ud.json @@ -0,0 +1,240 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-window-i00-privileged--9mkr2ud", + "taskKey": "humanoidbench/manip_window", + "dag_sha256": "4dc821ae18297046c83611bc223535d1570899c28d34142481b029cb27e902ac", + "outcome": { + "success": 0, + "reward": 0.23043885800015218, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 1000, + "final_reward": 0.23043885800015218, + "sim_steps": 990, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-window-i00-privileged--9mkr2ud/replay.mp4", + "duration_s": 10.0, + "fps": 20.0, + "frames": 200, + "sha256": "b8035fb5d0ae30541ad70e12c5952f2f9f4f6d832a7047774d5e019f9b69c5aa" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "wiping_tool_in_hand": { + "method": "visual", + "intervals": [ + [ + 0.35, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.35, + "log_step": null, + "text_zh": "观察:视频 0.00–0.30 s,悬在双手前的银杆先下落、擦头下摆;视频 0.35–0.60 s,机器人弯曲手指并抬起双手,银杆随手上升,工具没有继续自由下落。以视频约 0.35 s 作为拿稳的首次可见时刻,精度约 0.1 s。" + }, + { + "t": 4.0, + "log_step": null, + "text_zh": "观察:视频 1.00–5.00 s,双手在杆身下方及两侧承接工具,工具保持在胸前;视频 7.00–9.95 s,身体虽前倾下沉,杆仍随双手移动,未脱手落地。推断:按 DAG 不强制完整包围抓法的拿稳判据,此事件已完成并锁存;画面不能证明有牢固的对向夹持。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "托/舀", + "note_zh": "观察:视频 0.30–0.60 s,两只 Shadow 手的掌部和弯曲手指共同从下方承接、抬起银色工具杆;后续继续托持。可见承托,未能确认手指与拇指完整夹住杆身,故不记抓握。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.60 s,机器人双脚接触地面,支撑身体接住并抬起工具。" + } + ] + }, + "window_wiped": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.5, + "log_step": null, + "text_zh": "观察:视频 0.50–1.50 s,机器人将工具送向窗面,黑色擦头靠上青色玻璃;视频 1.50、2.00、2.50、3.00、3.50、4.00、4.50、5.00 s,擦头基本保持在同一小块区域,未见持续的上下擦动。视频 5.50–9.95 s,工具转动并随身体下沉,仍没有形成稳定的竖直擦拭行程。" + }, + { + "t": 9.95, + "log_step": 130, + "text_zh": "观察:评分器字段 success=0、all_success=0、deterministic=1、sim_steps=990;log step 125 保存最终动作文件,log step 130 对该文件复验,均得到相同末态哈希 11c5306348f1 和 990 个控制步。推断:结合视频中的擦动不足与受评失败,擦窗的累计回报验收未通过,不能把贴窗本身当作整步完成。" + } + ], + "contact_actual": [ + { + "who": "手持工具", + "how": "压/擦", + "note_zh": "观察:视频约 1.50–5.00 s,机器人双手承托银杆,使黑色擦头贴在青色窗面附近并压靠,小幅摆动;主要表现为停留压靠,未完成稳定的上下擦拭。同一工具到玻璃的接触链只记一次。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.50–6.50 s,双脚在地面支撑窗前操作;视频约 7.00 s 起,一侧腿向后滑伸,机器人身体前倾并下沉,另一脚仍着地承重。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 650, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "window_wiped", + "t": 1.5, + "log_step": 127, + "where": "操作中", + "evidence_zh": "观察:视频约 1.50 s,机器人已把擦头送到窗面,却在视频 1.50–5.00 s 基本保持同一姿态,之后仍未完成稳定上下擦拭;log step 127 对最终交付的说明明确写成接住、抬臂、保持固定姿态,并写明交付轨迹不擦动。推断:首次未修复的问题在擦窗步骤,视频 1.50 s 是该停滞的导航锚点;后段身体下沉仍发生在同一步操作中。" + }, + "phenomena": [ + { + "cat": "其他", + "item": "擦拭运动不足", + "t": 1.5, + "log_step": null, + "evidence_zh": "观察:视频 1.50–5.00 s,黑色擦头长时间停留在同一块窗面,银杆只小幅晃动;视频 5.50–9.95 s 主要是杆身翻转和随躯干下沉,未见稳定的上下擦动。", + "note_zh": "候选名称:擦拭运动不足。机器人已经接触并操纵工具,故不是没尝试;现有位姿不准、动态时机等条目不能准确表达贴窗后长期静止、擦拭行程不足的现象。" + }, + { + "cat": "全身", + "item": "平衡", + "t": 7.0, + "log_step": null, + "evidence_zh": "观察:视频 6.60–7.30 s,机器人一侧腿向后滑伸,失去原有双脚站姿,躯干前倾、骨盆下沉;视频 9.00–9.95 s 身体继续塌低,没有恢复原先站姿,双手和工具也随之下移。此处标失衡,不推定精确终止高度或触发约束的时刻。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "抓取和接触设计", + "alt_feasible": "已验证有", + "log_step": 127, + "source_zh": "日志 / agent 自述 / 视频", + "evidence_zh": "观察:log step 107 的动作生成代码将 OSC 设为空字典,log step 117 仅修正手腕目标,log step 125 保存该静态方案并复验,log step 130 再验得到相同末态哈希 11c5306348f1。log step 123 的会话试验输出中,肩部振荡幅度 0.05、0.10、0.18 rad 对应自算擦头接触占比约 0.16、0.15、0.37,低于静态方案的约 0.43;随后肘部试验因 KeyError 中断。agent 在 log step 127 自述杆只靠掌面摩擦、需要玻璃承载,并明确写明因振荡破坏接触而选择固定姿态、不擦动。视频 1.50–5.00 s 可见机器人托持杆身、擦头停留在同一小块窗面,视频 7.00–9.95 s 杆随身体前倾下移。推断:所选掌面承托、依赖贴窗承载的接触布局没有支持稳定竖直擦动,agent 因而交付不含擦拭行程的方案;这支持接触设计归因,不能仅表述为有擦拭目标而控制器未能跟踪。inputs/oracle/oracle.json 的 success=1 与参考视频 0.50–9.50 s 的持杆贴窗操作,验证本任务存在受评成功的替代方案。", + "note_zh": "替代可行性指本任务存在成功路线,不声称参考运行采用已核实的包围抓握,也不把非包围承托本身判为失败。玻璃承担载荷的说法来自 agent 自述;本次画面与会话振荡试验共同支持接触布局不适于擦动的推断,不能精确分解载荷。会话试验只用于归因,不替代本次节点值;不另将同一接触困难导致的静态选择重复标为规划根因。" + } + ], + "gap_zh": "本次没有状态日志,视频只有 256×256 的单一外部视角,无法核实具体手指接触对、法向力及擦头贴窗的精确起止时刻;拿稳按可见承托和随手移动判定,不声称形成对向抓握。run.json 没有报告 episode_return;final_reward=0.23043885800015218 未被明确标为累计回报,故原生累计回报值留 null,不换算为 650 门槛下的数值。没有将会话诊断的控制步数换算为视频时间,也未据 agent 在 log step 131 的“骨盆未下降”自述否定视频后段可见的下沉。 身体后段下沉的具体成因无法由现有日志确定;未将其直接归因于玻璃承载或某个反馈控制缺陷。log step 123 只输出了肩部振荡试验结果,agent 所称肘、膝各幅度均已试验并破坏接触(log step 127/131)未获该输出完整验证。", + "unmet": { + "window_wiped": false + } + }, + "summary_zh": "机器人在视频约 0.35 s 用双手承接并抬起擦窗工具,完成拿稳工具,最终节点进度为 1/2。擦窗在视频约 1.50 s 后停滞:擦头长期停在同一区域,后段机器人失衡下沉;最终代码与会话试验支持接触设计归因,所选承托、贴窗布局难以兼顾擦动,因而交付固定姿态且不含擦拭行程的方案,评分器判定失败。成功参考运行验证本任务存在可行替代;本次没有状态日志及累计回报数值,精确接触载荷、终止高度与身体下沉原因不可核实。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 2, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "fix", + "conf": 6, + "summary_zh": "WHERE 标签(拿稳 0.35 s、擦窗未达、断点 window_wiped、平衡现象)与视频一致;但根因只写了笼统的“控制”(开环控制器未能实现),而 log 107/123/127 显示是交付方案本身的选择:依赖玻璃承重、不擦动,属于方案/接触设计层面,应改判别次序下的根因。", + "n_issues": 1, + "n_major": 1 + }, + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 2, + "verdict": "agree", + "conf": 7, + "summary_zh": "节点(拿稳约 0.35 s、擦窗未达)、断点 window_wiped、现象(擦拭运动不足、平衡)与第二轮修订后的根因(抓取和接触设计,alt_feasible 已验证有)均与视频和日志一致,未发现需修改之处。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:06:30Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-privileged--9mkr2ud/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-privileged--9mkr2ud/04_review_r1" + }, + { + "stage": "reconcile", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-privileged--9mkr2ud/05_reconcile_r1" + }, + { + "stage": "traj_review", + "round": 2, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-privileged--9mkr2ud/04_review_r2" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "f8dba82696bb12dd3ffbdb70eae2a6a54672766c2e0034af5ad2be35c5e3e21b", + "video_sha256": "b8035fb5d0ae30541ad70e12c5952f2f9f4f6d832a7047774d5e019f9b69c5aa", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-privileged--eovprfp.json b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-privileged--eovprfp.json new file mode 100644 index 0000000000000000000000000000000000000000..1d2b3cd91baa42ba99cb6e5a1c00a1fdad86417a --- /dev/null +++ b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-privileged--eovprfp.json @@ -0,0 +1,214 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-window-i00-privileged--eovprfp", + "taskKey": "humanoidbench/manip_window", + "dag_sha256": "4dc821ae18297046c83611bc223535d1570899c28d34142481b029cb27e902ac", + "outcome": { + "success": 0, + "reward": 0.23510533567488356, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 869, + "final_reward": 0.23510533567488356, + "sim_steps": 869, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 2.238, + "peak_processes": 9, + "peak_disk_gb": 0.833, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-window-i00-privileged--eovprfp/replay.mp4", + "duration_s": 8.75, + "fps": 20.0, + "frames": 175, + "sha256": "22147a037122363d17cee828e89097c5fe8a8c33b251ffe9ece0daaced606f0f" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "wiping_tool_in_hand": { + "method": "visual", + "intervals": [ + [ + 0.15, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.15, + "log_step": null, + "text_zh": "观察:视频 0.00–0.30 s,机器人双手在银色杆两处合拢,长杆随双手上抬而非继续自由下落;视频 0.15 s 已可确认工具受到手部控制,起始时刻约有 0.1 s 不确定性。" + }, + { + "t": 1.5, + "log_step": null, + "text_zh": "观察:视频 1.50–5.00 s,银色杆横在双手处并由手掌、弯曲手指拿住;视频 6.50–8.70 s 工具虽翻转摆动,仍在双手附近受控,没有整件脱手坠地。此节点为 event,拿稳一次后锁存。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "机器人两只灵巧手分别合拢在银色工具杆的两处,共同接住并持杆;视频 0.15–0.30 s 可见杆随手上抬,视频 1.50 s 可见手掌与弯曲手指围住杆。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "机器人双脚在地面承重,支撑双臂接取工具;随后身体前倾、腿部逐渐下沉。" + } + ] + }, + "window_wiped": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.5, + "log_step": null, + "text_zh": "观察:视频 1.30–5.20 s,机器人把黑色擦头保持在青色窗面下缘附近,工具大体横置且上下移动很小;视频 5.30–6.30 s 可见小幅摆动,确实启动了持工具擦窗的尝试。" + }, + { + "t": 6.5, + "log_step": null, + "text_zh": "观察:视频 6.40–7.50 s,黑色擦头明显翻转,长杆斜向下摆,擦头离开原先贴窗位置;视频 7.75–8.70 s,工具继续大幅转动,末帧擦头已摆到机器人头侧,未保持沿窗面稳定上下擦动。" + }, + { + "t": 8.7, + "log_step": 119, + "text_zh": "推断:inputs/run.json 的评分器字段 success=0、all_success=0,且 deterministic=1、n_actions=sim_steps=869;log step 119 核对交付文件为 869 行。按 DAG 的累计回报达到 650 才完成的验收规则,本回合擦窗节点未达标。final_reward=0.23510533567488356 未标为累计回报,不能当作 episode_return 使用。" + } + ], + "contact_actual": [ + { + "who": "手持工具", + "how": "压/擦", + "note_zh": "机器人经双手控制银色长杆,使黑色擦头在视频 1.30–5.20 s 接触青色玻璃下部,并于视频 5.30–6.30 s 小幅摆动;后续擦头翻转离开原位置,接触未持续保持。工具到玻璃的接触链只记一次。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "机器人双脚承重支撑持杆擦窗,身体持续前倾并下沉;没有轮式底盘运动。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 650, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "window_wiped", + "t": 6.5, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频 6.40–7.50 s,正在操纵工具擦窗的机器人使长杆翻转并向下斜摆,黑色擦头离开此前贴住窗面下部的位置;视频 7.75–8.70 s 仍大幅转动,直到结束也未恢复稳定擦窗。评分器 success=0。" + }, + "unmet": { + "window_wiped": false + }, + "phenomena": [ + { + "cat": "运动", + "item": "位姿不准", + "t": 6.5, + "log_step": null, + "evidence_zh": "观察:视频 1.50–5.00 s 工具近乎横向停在窗面下缘;视频 6.50 s 擦头翻面,视频 7.25–7.50 s 杆斜向下且擦头摆到玻璃下方,视频 8.70 s 擦头又转到头侧。工具仍被拿着,但擦头没有维持贴合窗面、沿竖直方向擦动的位姿。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "t": 6.5, + "log_step": 117, + "source_zh": "日志 / 视频", + "evidence_zh": "推断:log step 84 的生成器以固定手指闭合和按控制步变化的肩、肘关节目标持杆;log step 89、100、113、117 对其逐次修改,最终加入定时肩部调整和肘部余弦往复,没有依据擦头相对玻璃的位姿或接触反馈修正动作。视频 6.40–8.70 s 的工具翻转、摆离窗面说明该关节控制序列未能稳定控制擦头贴合和擦动;具体是握持约束还是运动跟踪占主因无法区分,停留在“控制”。" + } + ], + "gap_zh": "输入没有状态日志,也没有本回合逐步或累计回报字段,因此 episode_return 记为 null,无法确定接触和运动各回报项的贡献、精确丢失接触时刻,以及终止究竟由骨盆还是工具高度触发。视频分辨率为 256×256,部分手指和远侧手被遮挡;能够确认工具被双手拿住和后续摆离窗面,但不能据此确定指间滑动或接触力。log step 115 的接触计数属于会话测试,只用作控制调试背景,不代替视频判定节点。" + }, + "summary_zh": "机器人约在视频 0.15 s 拿稳擦窗工具,完成 1/2 个节点;随后使擦头接触窗面下部,但较长时间几乎不擦动,视频约 6.50 s 起工具明显翻转、摆离窗面,擦窗累计回报验收未通过。定时肩肘关节控制未能维持擦头位姿,根因归为控制,无法细分握持与跟踪问题。没有状态日志或累计回报记录,无法确认精确终止触发项,原生累计回报指标留空。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "节点、时间、接触与断点与视频一致:视频0.0-0.3 s双手已持杆,1.5-5 s擦头贴窗下缘,6.5 s起工具翻转摆离;评分器success=0。位姿不准与控制根因有证据支持,未发现需要改动的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:49:28Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-privileged--eovprfp/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-privileged--eovprfp/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "c1b287fba3fdff00a9c128809ed6a8574ff88f8bfe8188fe77c635e6e2430a95", + "video_sha256": "22147a037122363d17cee828e89097c5fe8a8c33b251ffe9ece0daaced606f0f", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-privileged--fhskrta.json b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-privileged--fhskrta.json new file mode 100644 index 0000000000000000000000000000000000000000..9de3b4302680e385b2fe26ecd8af207e7dac0921 --- /dev/null +++ b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-privileged--fhskrta.json @@ -0,0 +1,214 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-window-i00-privileged--fhskrta", + "taskKey": "humanoidbench/manip_window", + "dag_sha256": "4dc821ae18297046c83611bc223535d1570899c28d34142481b029cb27e902ac", + "outcome": { + "success": 0, + "reward": 0.22149682327560533, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 1000, + "final_reward": 0.22149682327560533, + "sim_steps": 1000, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 1.351, + "peak_processes": 10, + "peak_disk_gb": 0.262, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-window-i00-privileged--fhskrta/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "49bd6398e810bdfd0beb74bdd16c6cc4bf36be4e3a55cd8c4a09ca165e5d826a" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "wiping_tool_in_hand": { + "method": "visual", + "intervals": [ + [ + 0.35, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.35, + "log_step": null, + "text_zh": "观察:视频 0.00–0.25 s 银色杆和黑色擦头先下降、翻动;视频 0.30–0.60 s 工具随机器人弯臂抬起,停止自由下落。视频 1.00 s 手指位于杆的上下两侧,之后工具一直随双臂运动,至视频 10.05 s 仍在手边,未脱手落地;拿稳时刻约为视频 0.35 s。" + }, + { + "t": 1.0, + "log_step": 403, + "text_zh": "推断:视频 0.35–1.00 s 可见近侧手围住杆,另一手在躯干前配合控杆;结合 log step 403 交付脚本对左右 Shadow 手均设置拇指对掌及多指屈曲,按双臂协作抓握标注。远侧手指有局部遮挡,不据此断言每根手指的接触。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "双侧灵巧手在胸腹前合指控住银色工具杆,工具停止下落并随两臂抬升;远侧手的细部被躯干局部遮挡。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "机器人双脚在地面支撑拿杆、抬臂时的身体。" + } + ] + }, + "window_wiped": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.5, + "log_step": null, + "text_zh": "观察:视频 1.00–4.50 s 机器人将黑色擦头送到青色窗面,随后擦头大体停留在同一局部,主要是小幅转动;视频 4.90–5.25 s 随机器人前倾,擦头下移至窗面低处,视频 5.50–10.05 s 又基本停在该处,没有形成持续的上下擦动。全片保持持杆和脚下支撑,未见工具掉落或机器人摔倒。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:inputs/run.json 的评分器字段 success=0、all_success=0、n_actions=1000、deterministic=1、grader_error=0,final_reward=0.22149682327560533;未报告名为 episode_return 的累计回报。" + }, + { + "t": 10.05, + "log_step": 421, + "text_zh": "推断:视频显示机器人已尝试贴窗,但擦动很少;评分器在无报错且确定性回放成立时判失败,依 DAG 的累计回报验收,擦拭窗户节点未达成。log step 421 是 agent 对最终交付文件的自编诊断,报告接触期间平均竖直速率 0.0118 m/s,支持运动不足;其中 surrogate return=577.1 是自定义替代分数,不是评分器累计回报。" + } + ], + "contact_actual": [ + { + "who": "手持工具", + "how": "压/擦", + "note_zh": "机器人经双手控制银色杆,将黑色擦头贴向青色玻璃,发生局部压靠及少量下擦;长时间停留,未完成持续上下擦拭。握杆到擦头的同一接触链仅在此记一次。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚持续接触地面,支撑机器人伸臂、前倾操纵擦头。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 650, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "window_wiped", + "t": 1.5, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频 1.00–2.50 s 机器人拿着工具伸臂贴向玻璃后,擦头停在局部而未继续形成上下行程;视频 4.90–5.25 s 下移后,至视频 10.05 s 仍停在窗面低处。评分器 success=0,擦窗步骤之后没有完成;以视频 1.50 s 的停滞作为导航锚点。" + }, + "unmet": { + "window_wiped": false + }, + "phenomena": [ + { + "cat": "运动", + "item": "位姿不准", + "t": 1.5, + "log_step": null, + "evidence_zh": "观察:视频 1.50–4.50 s 擦头几乎停在一个高度,只小幅转动;视频 4.90–5.25 s 擦头下移后,在视频 5.50–10.05 s 基本停于窗面低处。机器人保持着工具,但工具没有沿窗面形成持续上下擦动,失败表现为擦拭轨迹未实现。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "t": null, + "log_step": 403, + "evidence_zh": "推断:log step 403 的最终生成脚本在合指后只将肩俯仰、肘目标单向渐变到固定值,并在骨盆降低时调膝踝,没有随后驱动工具上下擦动的目标序列或工具速度反馈。log step 421 对该交付文件的自编诊断报告贴窗期间平均竖直速率仅 0.0118 m/s,和视频 1.50–4.50 s、5.50–10.05 s 的局部停滞一致;因此支持控制层未实现所需擦动。会话其他试验不能替代本次执行,未进一步归为某一力控或逆解子类。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "没有状态日志及评分器逐步回报,episode_return 未单独报告,不能从 final_reward 或 agent 的 surrogate return 换算累计回报,也不能精确定位每次擦头接触。远侧手指局部遮挡,抓握由工具随手运动、可见近侧合指及双手控制脚本交叉判断;控制不足有证据,但缺少最终交付的逐关节跟踪与接触力记录,无法确认更细的控制子类。" + }, + "summary_zh": "机器人约在视频 0.35 s 拿稳擦窗工具,完成 1/2 个节点;之后伸臂将擦头贴向窗面,但大部分时间停在局部,约视频 5 s 下移后再次停滞,擦窗累计验收未通过。最终脚本只有伸臂渐变和姿态支撑调节,交付文件诊断也显示竖直运动很小,支持控制层未实现所需擦动。无状态日志或评分器累计回报明细,具体回报数值及更细的接触力原因不能确定。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "视频中工具自约0.3 s起被持住并始终在手,擦头长时间停在窗面局部,节点值、断点和现象与帧一致;未发现需改动的内容。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:43:57Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-privileged--fhskrta/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-privileged--fhskrta/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "8b4966c6a1d731d82574592b688f93f6ac4843704a4a37214974238f40d0ee22", + "video_sha256": "49bd6398e810bdfd0beb74bdd16c6cc4bf36be4e3a55cd8c4a09ca165e5d826a", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-privileged--hdkztci.json b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-privileged--hdkztci.json new file mode 100644 index 0000000000000000000000000000000000000000..28cf1c61fd87c2b1afd1679b98df79c7a6b459ef --- /dev/null +++ b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-privileged--hdkztci.json @@ -0,0 +1,202 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-window-i00-privileged--hdkztci", + "taskKey": "humanoidbench/manip_window", + "dag_sha256": "4dc821ae18297046c83611bc223535d1570899c28d34142481b029cb27e902ac", + "outcome": { + "success": 0, + "reward": 0.6931176124727428, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 1000, + "final_reward": 0.6931176124727428, + "sim_steps": 1000, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 3.865, + "peak_processes": 34, + "peak_disk_gb": 0.841, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-window-i00-privileged--hdkztci/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "019ec86d36b25ced5e3700d89e6ad5f6999536a09e6ddc01c9e38cc490c4391c" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "wiping_tool_in_hand": { + "method": "visual", + "intervals": [ + [ + 0.3, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.3, + "log_step": null, + "text_zh": "观察:视频 0.00–0.35 s,银色工具杆先下沉,随后保持在机器人双手所在高度;视频 0.70–1.00 s,工具随屈肘动作上移,没有继续自由下落。拿稳时刻约为视频 0.3 s,误差约 0.2 s。" + }, + { + "t": 9.5, + "log_step": null, + "text_zh": "观察:视频 1.05–10.05 s,近侧手的手指弯在杆周围,杆随手做小幅运动,末段视频 8.90–10.05 s 逐帧检查未见工具脱手或身体倒地。" + }, + { + "t": 0.5, + "log_step": 133, + "text_zh": "推断:视频 0.50 s 可见两只手位于杆的不同位置,之后远侧握点被躯干部分遮挡;结合工具与双臂同步运动,以及 log step 133 所列最终交付说明“两手保留手柄”,判为双手协同持杆。该日志仅辅助解释握法,节点成立依据是视频中的持杆和随手移动。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "机器人两只 Shadow 手分别围握银色长杆的不同位置,协同接住并转动工具;后续远侧握点部分被身体遮挡。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "机器人双脚接触地面,屈膝站立支撑接杆及抬杆动作。" + } + ] + }, + "window_wiped": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.0, + "log_step": null, + "text_zh": "观察:视频 0.70–1.00 s,机器人抬杆并转动黑色擦头;视频 1.05–8.85 s 及末段视频 8.90–10.05 s,擦头在青色玻璃处小幅上下往复并摆动,工具始终留在手中。可见机器人实际尝试擦窗,不能标为未尝试。" + }, + { + "t": 9.5, + "log_step": 133, + "text_zh": "观察:log step 133 的最终文件核验输出列出 1000 个动作、913 个玻璃接触步、平均工具竖直速度 0.38071848405755976 m/s,且两次回放哈希一致;log step 138 明确说明实际回报不可用。这些是 agent 脚本的测量,不是评分器的累计回报。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "推断:inputs/run.json 的评分器字段 success=0、all_success=0、deterministic=1、sim_steps=1000、grader_error=0,结合视频完整擦拭且未掉杆或摔倒,按本任务累计回报达到 650 的验收条件,此步仍未达标。final_reward=0.6931176124727428 未标为累计回报,不能当作 episode_return 或乘以步数反推。" + } + ], + "contact_actual": [ + { + "who": "手持工具", + "how": "压/擦", + "note_zh": "机器人双手控杆,使黑色擦头在青色玻璃上接触并做短行程往复;按工具—玻璃这一接触链记一次。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "机器人双脚在地面支撑身体,在原处屈膝站立完成擦动,没有轮式底盘。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 650, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "window_wiped", + "t": null, + "log_step": 133, + "where": "操作中", + "evidence_zh": "观察:视频 1.05–10.05 s,机器人已持杆在玻璃处持续做小幅擦动,直到末帧仍持杆站立;log step 133 确认最终交付已完成 1000 步并通过确定性回放检查,评分器却给出 success=0。推断:未完成的是擦拭步骤的累计回报验收;视频中没有可定位的单次不可逆失误,故断点时间留空。" + }, + "unmet": { + "window_wiped": false + }, + "phenomena": [], + "root_causes": [], + "gap_zh": "观察:inputs/ 提供完整受评视频、运行日志及成功参考视频,但没有状态日志、逐步回报或 episode_return。视频 0.00–10.05 s 可见拿稳工具并重复短行程擦动,末段逐帧未见掉杆、摔倒或明确阻挡;成功参考视频也有小幅擦动,不能仅凭动作幅度判定本次错误。log step 133 的接触计数与速度是 agent 自行测量,不能说明评分器中哪项奖励不足。推断:可以确定擦拭步骤未达标,但无法可靠区分接触姿态、运动速度、站姿等软奖励的贡献,也无法确定对应的可见失败现象和根因,故两层留空;同样不能将 final_reward 当作累计回报填写原生指标。" + }, + "summary_zh": "机器人约在视频 0.3 s 拿稳擦窗工具,随后持续持杆在玻璃处做小幅往复,直至视频 10.05 s 仍站立且未掉杆,最终完成 1/2 个节点。评分器判定失败,断点为擦拭窗户的操作中,累计回报验收未达到;现有画面和 agent 的最终测量无法确定具体奖励缺项,失败现象与根因留空。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "拿稳工具约视频 0.3 s 成立,全程未掉杆/摔倒;擦拭节点 final=false 与评分器 success=0 一致;断点 window_wiped、t 留空、现象与根因留空并附 gap_zh,证据(无状态日志、final_reward 非累计回报、agent 自测不可当评分)支持,未发现需要修改的内容。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:44:09Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-privileged--hdkztci/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-privileged--hdkztci/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "cad261d75e46d9abb2abc50c12463c4dc4bfb2680f1512817d1d85f2361d987a", + "video_sha256": "019ec86d36b25ced5e3700d89e6ad5f6999536a09e6ddc01c9e38cc490c4391c", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-privileged--jwh7cag.json b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-privileged--jwh7cag.json new file mode 100644 index 0000000000000000000000000000000000000000..afdefbf4cb1ddb2410de2fc7d5b455c93a7e23c5 --- /dev/null +++ b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-privileged--jwh7cag.json @@ -0,0 +1,207 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-window-i00-privileged--jwh7cag", + "taskKey": "humanoidbench/manip_window", + "dag_sha256": "4dc821ae18297046c83611bc223535d1570899c28d34142481b029cb27e902ac", + "outcome": { + "success": 0, + "reward": 0.23251715661381323, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 912, + "final_reward": 0.23251715661381323, + "sim_steps": 912, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 2.042, + "peak_processes": 11, + "peak_disk_gb": 0.505, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-window-i00-privileged--jwh7cag/replay.mp4", + "duration_s": 9.2, + "fps": 20.0, + "frames": 184, + "sha256": "af6883d4656f36705e9d82458269dad1a13a2497bb30d2d0a2a1487d0c3d9fd4" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "wiping_tool_in_hand": { + "method": "visual", + "intervals": [ + [ + 0.3, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.3, + "log_step": null, + "text_zh": "观察:视频 0.00–0.80 s 的逐帧画面中,机器人双手在银杆两段处合拢承接工具,银杆短暂下移后停在手边;视频 0.30 s 起工具随手向窗面移动,未继续自由下落,满足拿稳判据(时刻约有 0.2 s 不确定性)。" + }, + { + "t": 9.15, + "log_step": null, + "text_zh": "观察:视频 8.00–9.15 s 机器人身体下沉、近侧手臂垂下,工具倾斜并与原先手掌位置分离;拿稳工具是 event 节点,早先已经达成,后续脱离不撤销该节点。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "双手在银色工具杆的两段处合手承接并握持,近侧手可见手指弯合在杆周围,远侧手部分被躯干遮挡;随后共同控杆前移。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "机器人双脚接触地面,支撑接住工具和前伸双臂时的身体。" + } + ] + }, + "window_wiped": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 2.1, + "log_step": null, + "text_zh": "观察:视频 1.60–3.10 s 机器人持杆将黑色擦头送到青色玻璃处,擦头间歇贴靠并反复俯仰摆动;视频 3.10–7.35 s 仍有小幅摆动,身体逐渐前倾,不能从这些画面确认累计回报已达 650。成功参考视频 0.50–9.50 s 显示机器人在窗前维持较稳定站姿和工具运动,仅用来对照几何。" + }, + { + "t": 9.15, + "log_step": 179, + "text_zh": "观察:视频 7.60–9.15 s 机器人持续下沉前倾,末帧双腿撑开、近侧手臂垂下,未恢复站姿继续擦拭。log step 179 重新生成交付文件并输出“steps 912 proxy TOTAL 855.5”;评分器报告 n_actions=912、deterministic=1、success=0。推断:交付与该基线相符;代理分数不是评分器累计回报,按受评失败结果与画面判定本节点未达成。" + } + ], + "contact_actual": [ + { + "who": "手持工具", + "how": "压/擦", + "note_zh": "机器人经银色工具杆使黑色擦头在青色玻璃处间歇贴靠、小幅摆动擦动;没有把双手握杆重复记入同一工具接触链。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚在地面支撑窗前操纵工具,后半段腿部逐渐撑开、身体前倾下沉,未维持稳定支撑姿态。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 650, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "window_wiped", + "t": 7.6, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频 2.10 s 起机器人在窗前持工具操作;视频 6.10–7.60 s 躯干继续前倾、骨盆下沉,视频 7.60–9.15 s 下沉加重、双腿撑开且手臂离开原先握杆位置,直至录像结束未恢复。推断:拿稳工具已完成,未修复的失败发生在擦窗操作中;视频 7.60 s 为明显失稳段的导航锚点。" + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 7.6, + "log_step": null, + "evidence_zh": "观察:视频 6.10–9.15 s 机器人上身越来越向窗面倒伏,头与骨盆降低,双腿向后方撑开,近侧手臂最终垂下;末帧仍在塌落而非稳定站立擦窗。推断:机器人失去稳定支撑,属于平衡失败;不声称已看到躯干触地。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "轨迹或逆解", + "evidence_zh": "观察:log step 75 的 pelvis_ik 对骨盆调用 mj_jacBody,再取腿关节列来求逆解;log step 95 把当时的控制器内嵌到 gen.py,log step 179 运行此文件重新保存 912 步轨迹。log step 165 的会话诊断打印骨盆雅可比的腿关节 x、z 行全零、位移修正 dq 全零、cond(J)=inf;log step 166 随后才修改控制器,未更新内嵌旧代码的 gen.py。推断:交付基线的骨盆目标误差不能通过该逆解转为腿关节纠偏,支持其在视频 6.10–9.15 s 持工具时无法制止身体前倾下沉的控制根因。", + "source_zh": "日志" + } + ], + "gap_zh": "没有状态日志,单一低分辨率视角不能精确还原手指接触对、擦头接触力及骨盆触发终止阈值的时刻;未把会话测试中的状态数值映射为本视频时刻。评分器只给 final_reward=0.23251715661381323 等结果字段,没有独立报告 episode_return,故原生累计回报值留空;log step 179 的 proxy TOTAL 855.5 是 agent 脚本的代理指标,不能作为累计达标证据。", + "unmet": { + "window_wiped": false + } + }, + "summary_zh": "受评录像中机器人约在视频 0.30 s 接住并拿稳擦窗工具,随后将擦头送到玻璃处小幅摆动,但擦窗未通过累计回报验收,完成 1/2 个节点。视频 6.10–9.15 s 身体持续前倾下沉并失稳;会话代码与诊断支持交付基线的骨盆逆解无有效腿关节纠偏这一控制根因。没有状态日志或正式累计回报读数,接触力和精确终止时刻无法确认。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "节点判定(拿稳工具已达成、擦窗未达成)、断点 window_wiped、平衡现象和骨盆逆解根因均与视频及日志一致,未发现需要修改的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:55:44Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-privileged--jwh7cag/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-privileged--jwh7cag/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "139c4be3d04d1ac3c5d7d04bdedeb7377568971a37e0922dcc9d02db72237163", + "video_sha256": "af6883d4656f36705e9d82458269dad1a13a2497bb30d2d0a2a1487d0c3d9fd4", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-privileged--wbhe9qa.json b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-privileged--wbhe9qa.json new file mode 100644 index 0000000000000000000000000000000000000000..342203bc28300ad3ece526307a69cf61862a2450 --- /dev/null +++ b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-privileged--wbhe9qa.json @@ -0,0 +1,218 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-window-i00-privileged--wbhe9qa", + "taskKey": "humanoidbench/manip_window", + "dag_sha256": "4dc821ae18297046c83611bc223535d1570899c28d34142481b029cb27e902ac", + "outcome": { + "success": 0, + "reward": 0.6269085904328326, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 0, + "deterministic": 1, + "n_actions": 1000, + "final_reward": 0.6269085904328326, + "sim_steps": 1000, + "grader_error": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 1.282, + "peak_processes": 24, + "peak_disk_gb": 0.239, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-window-i00-privileged--wbhe9qa/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "3e2c1c63c0fd297ebe999becbf1ce2ab3f7797507812da6b79a9639498496ed3" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "wiping_tool_in_hand": { + "method": "visual", + "intervals": [ + [ + 0.2, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.2, + "log_step": null, + "text_zh": "观察:视频 0.00 s 银色工具杆悬在手掌上方;视频 0.10–0.20 s 机器人近镜头的左手手指弯曲包住杆,工具随左手抬起并转动,不再持续自由下落。拿稳时刻约为视频 0.2 s,精度约 ±0.1 s。" + }, + { + "t": 0.5, + "log_step": null, + "text_zh": "观察:视频 0.30–0.60 s 机器人把杆放低后再次带起,左手仍环握杆;视频 1.6–9.5 s 及逐帧检查的视频 9.00–10.05 s,工具始终横在手前,没有脱手落地。此节点为 event,达到后锁存。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "抓握", + "note_zh": "观察:视频 0.15–0.50 s 近镜头的左手手指环握银色工具杆并随手带动工具;另一手局部被躯干和杆遮挡,不据此认定双手同时抓握。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.70 s 机器人双脚着地支撑接杆与移杆,随后分腿站立支撑身体。" + } + ] + }, + "window_wiped": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.6, + "log_step": null, + "text_zh": "观察:视频 0.7–1.5 s 机器人前倾并伸出持杆手,黑色擦头转向窗面;视频约 1.6 s 起擦头贴在青色玻璃下沿附近,之后基本维持该处的压持姿态。" + }, + { + "t": 3.0, + "log_step": null, + "text_zh": "观察:视频 3.00–3.95 s 以每帧 0.05 s 检查,黑色擦头相对窗下沿仅有轻微起伏;视频 2.0–9.5 s 的全段抽帧也主要是同处压持,机器人身体持续明显前倾。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 9.00–10.05 s 逐帧检查,机器人仍持杆在玻璃下沿附近,没有明显增加上下擦动,也没有跌倒或掉杆。评分器 reward.success=0、all_success=0、deterministic=1、grader_error=0,完整执行 n_actions=1000。" + }, + { + "t": null, + "log_step": 55, + "text_zh": "推断:DAG 将整步绑定于本回合累计回报达到 650;评分器对完整、确定性的交付回放判为失败,因此该累计验收未达到,不能仅凭已贴窗判完成。log step 55 的 proxy R=4798.9 是 agent 自建代理回报;log step 56 的正式回放接口明确返回 scored=false,均不构成实际累计达标证据。" + } + ], + "contact_actual": [ + { + "who": "手持工具", + "how": "压/擦", + "note_zh": "观察:视频约 1.6–10.05 s 机器人经银色长杆把黑色擦头压在青色玻璃下沿附近,主要持续压持并伴轻微擦动;不把握杆接触重复记在同一工具—玻璃链路中。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.7–10.05 s 双脚分开着地,腿和躯干支撑前倾身体操纵工具,末尾仍站立。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 650, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "window_wiped", + "t": 2.0, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频约 1.6–2.0 s 机器人完成贴窗后停留在窗下沿附近的压持姿态;视频 3.00–3.95 s 的密集帧和视频 9.00–10.05 s 的末段帧均显示竖直擦动很小,之后没有修复为足够的擦拭表现。评分器对完整的 1000 步回放判 success=0;拿稳工具已实现,未完成的是擦窗的累计回报验收。" + }, + "unmet": { + "window_wiped": false + }, + "phenomena": [ + { + "cat": "运动", + "item": "其他", + "t": 2.0, + "log_step": null, + "note_zh": "候选名称:擦拭运动不足。工具已经贴窗且保持持握,主要问题是近乎原地压持、竖直运动很少;不是未接触、脱手、静态目标位姿偏差或对运动目标的时机错误,现有条目没有直接描述这种持续运动不足。", + "evidence_zh": "观察:视频 1.6–10.05 s 机器人把擦头主要保持在玻璃下沿附近;视频 3.00–3.95 s 和视频 9.00–10.05 s 的逐帧画面只有微小起伏,未形成明显持续的上下擦动。机器人一直保持工具和足部支撑。" + } + ], + "root_causes": [ + { + "cat": "认知", + "item": "推理/规划", + "source_zh": "日志 / agent 自述 / 视频", + "evidence_zh": "推断:agent 将偏重贴窗的代理目标当作实际累计验收的充分替代,选择了牺牲运动质量的策略。log step 18 的自建 proxy_reward 默认给接触权重 4、竖直运动权重 1;log step 55 最终保存的轨迹诊断为 contact_steps=931、mv_avg=0.05。agent 在 log step 61 明确写入说明,优化器以竖直运动和头部位置项换取近乎永久接触;在 log step 63 又声称这种接触比例足以在任何合理权重下超过 650。视频 2.0–10.05 s 的持续低位压持与这一策略一致,但正式评分器 success=0,表明该替代策略不足以完成验收。" + } + ], + "gap_zh": "没有状态日志,也没有逐步真实回报或 episode_return 总和;评分器只报告 final_reward=0.6269085904328326,不能把它当作累计回报或擅自乘步数,故原生累计指标留 null。画面分辨率为 256×256,背侧手部局部遮挡,无法确定是否两手始终同时包握;持稳节点由清楚可见的左手和工具随动判断。竖直运动不足及明显前倾可见,但缺少真实回报分项,不能量化各项对未达 650 的贡献,也不把代理诊断 mv_avg 当作实测速度。" + }, + "summary_zh": "机器人在视频约 0.2 s 拿稳工具,随后前倾并于视频约 1.6 s 把擦头压到窗面,最终完成 1/2 个节点。断点为擦拭窗户:后续主要近乎原地压持、竖直擦动很小;日志显示 agent 用接触占优的代理回报选择了牺牲运动质量的策略,实际完整回放未通过累计验收。没有状态日志或真实累计回报分项,无法量化回报缺口,背侧手部接触也有遮挡。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "节点值、时间、断点与根因均有视频和日志支持。拿稳工具约 0.2–0.8 s,擦窗节点未达成(评分器 success=0,视频只见近乎原地压持)。未发现需要修改的内容。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:49:48Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-privileged--wbhe9qa/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-privileged--wbhe9qa/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "1c5b24082f85fa4a2e775b225f6eda1698d905edbc0ca8767ea3d63356c18ec2", + "video_sha256": "3e2c1c63c0fd297ebe999becbf1ce2ab3f7797507812da6b79a9639498496ed3", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard--5xxbfmm.json b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard--5xxbfmm.json new file mode 100644 index 0000000000000000000000000000000000000000..b5f2835389511dab65e9d347f672623341b9d5ed --- /dev/null +++ b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard--5xxbfmm.json @@ -0,0 +1,233 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-window-i00-standard--5xxbfmm", + "taskKey": "humanoidbench/manip_window", + "dag_sha256": "4dc821ae18297046c83611bc223535d1570899c28d34142481b029cb27e902ac", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 928, + "sim_steps": 15635, + "calls": 3051, + "rows": 15635, + "resets": 38, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1307.7, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.226, + "peak_processes": 9, + "peak_disk_gb": 0.284, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-window-i00-standard--5xxbfmm/replay.mp4", + "duration_s": 9.35, + "fps": 20.0, + "frames": 187, + "sha256": "7a64f89d383cb5b39bb716375d92c8bcba220f47205a3eb2979e8484916da20e" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "wiping_tool_in_hand": { + "method": "visual", + "intervals": [ + [ + 0.25, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.25, + "log_step": null, + "text_zh": "观察:视频 0.00–0.20 s 银色工具杆先下降;视频约 0.25 s 开始,杆停在机器人双手处并随手抬动,视频 0.25–0.75 s 未继续自由下落,拿稳时刻约有 0.2 s 误差。" + }, + { + "t": 5.2, + "log_step": null, + "text_zh": "观察:视频 4.80–5.10 s 杆仍在双手附近;视频 5.20–5.45 s 双手抬到杆上方,杆滑至腕部/前臂附近,视频 8.00–9.30 s 双手与杆仍明显分离。此节点为 event,初段拿稳后锁存,后续失握不撤销已完成事件。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "观察:视频约 0.25–0.75 s,机器人两只灵巧手在银色长杆的不同位置合拢并共同拿住,杆随双手移动。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.25–0.75 s,机器人双脚接触地面,双腿支撑拿杆时的身体。" + } + ] + }, + "window_wiped": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.5, + "log_step": null, + "text_zh": "观察:视频 0.50–4.75 s,机器人持杆使黑色擦头贴着青色窗面小幅移动、转动;视频 5.10–5.45 s 擦头明显上移。机器人确实尝试擦窗,并非未接触或未尝试。" + }, + { + "t": 5.2, + "log_step": null, + "text_zh": "观察:视频约 5.20 s,机器人双手升到银杆上方,杆不再被手掌包住;视频 6.50–9.30 s,杆留在腕/前臂附近,机器人身体继续向窗面倾倒、骨盆下沉,未恢复直立。" + }, + { + "t": 9.3, + "log_step": 126, + "text_zh": "推断:log step 126 的服务最终状态为 episode=39、control_steps=928、done=True、reason=failed,与评分器 n_actions=928、graded_by_live=1、live_success=0 一致;任务规定累计回报达标立即成功,因此末回合未达到擦窗节点所需累计 650 回报。画面中的接触擦动不足以单独完成此节点;评分器未报告 episode_return 数值。" + } + ], + "contact_actual": [ + { + "who": "手持工具", + "how": "压/擦", + "note_zh": "观察:视频 0.50–5.10 s,机器人经银杆控制黑色擦头贴住青色玻璃并小幅擦动;视频约 5.20 s 后手掌与杆分离,工具移到腕/前臂附近,擦头仍在窗面附近摆动。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.50–9.30 s,机器人双脚与地面接触承重;擦拭过程中身体逐渐前倾下沉,未见成功换步恢复站姿。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 650, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "window_wiped", + "t": 5.2, + "log_step": 126, + "where": "操作中", + "evidence_zh": "观察:视频约 5.20 s,机器人抬臂时银杆滑到手掌下方,随后双手与杆分离;视频 6.50–9.30 s 身体继续前倾、骨盆下沉,末段未恢复。机器人此前已拿住工具,正在擦窗;log step 126 的最后现场回合以 failed 结束,擦窗节点始终未达标。" + }, + "unmet": { + "window_wiped": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "没保持住", + "t": 5.2, + "log_step": null, + "evidence_zh": "观察:视频 0.25–0.75 s 银杆被双手拿住并随手移动;视频 5.10–5.45 s 抬臂过程中,双手上移而银杆留在其下方的腕/前臂处。视频 8.00–9.30 s 双手仍举在杆上方,没有重新握住;工具未掉到地上不等于保持了抓握。" + }, + { + "cat": "全身", + "item": "平衡", + "t": 6.5, + "log_step": null, + "evidence_zh": "观察:视频 5.00、6.50、8.00、9.30 s,机器人躯干向窗面倾斜逐渐加重,双腿斜伸在身后,骨盆持续下降,身体经工具顶在窗前,直至结束未恢复站姿;不是一次受控擦动后的稳定站立。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志", + "evidence_zh": "推断:log step 116 的控制代码根据 pitch 修正髋、膝与踝目标;log step 126 的最终脚本仍执行双手擦动与姿态反馈,step_end=-1 禁用此前的换步分支,其末段自算 pitch 从 0.968 增至 0.971、踝积分 I[ank]=1.00,手部位置也未跟上目标。结合视频 5.20–9.30 s 的失握和不断前倾,持杆及全身伺服控制未能维持所需接触和身体稳定;证据不能可靠区分接触设计、轨迹跟踪或反馈参数,故停在控制大类。" + } + ], + "gap_zh": "没有状态日志,256×256 外部视角不能分辨逐指接触力或准确滑脱机制;失握约在视频 5.20 s,时刻仅为视觉锚点。服务只报告 failed,未报告具体终止触发项或累计 episode_return,不能确定是骨盆高度还是工具高度触发。log step 126 的 928 控制步按 log step 3 的 control_dt=0.02 对应仿真 18.56 s,视频则长 9.35 s;未把日志打印时间直接当作视频时间。agent 在 log step 129 所称初始位置过远、踝关节机械饱和不足以证明环境不可解,未采用为任务或环境根因。" + }, + "summary_zh": "机器人约在视频 0.25 s 拿稳工具,完成 1/2 个节点,并使擦头在窗面移动,但末回合没有达到累计 650 回报的擦窗验收。视频约 5.20 s 工具滑离手掌,随后身体持续前倾下沉;持杆与全身控制未能维持,断点为擦拭窗户。没有状态日志或累计回报读数,具体终止触发项及控制子类不能确定。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "抽查视频帧与标注一致:约0.3 s拿稳工具,擦动后约5.2 s双手上抬、杆滑离手掌,身体持续前倾下沉;断点、现象与根因(控制)有帧支持,未发现需改动的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:54:55Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard--5xxbfmm/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard--5xxbfmm/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "89d24c902ccde0837a1d18acf5846b2550bf72e8a623c53514cec7c2c78f3f19", + "video_sha256": "7a64f89d383cb5b39bb716375d92c8bcba220f47205a3eb2979e8484916da20e", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard--fei3twd.json b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard--fei3twd.json new file mode 100644 index 0000000000000000000000000000000000000000..8ab16e60929b374b9087511213931db87c6a4f87 --- /dev/null +++ b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard--fei3twd.json @@ -0,0 +1,225 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-window-i00-standard--fei3twd", + "taskKey": "humanoidbench/manip_window", + "dag_sha256": "4dc821ae18297046c83611bc223535d1570899c28d34142481b029cb27e902ac", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 291, + "sim_steps": 13124, + "calls": 12589, + "rows": 13124, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1942.3, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.328, + "peak_processes": 9, + "peak_disk_gb": 1.317, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-window-i00-standard--fei3twd/replay.mp4", + "duration_s": 3.0, + "fps": 20.0, + "frames": 60, + "sha256": "4fe0948d15be7385c08be580f29be56d5820f8539ef3278f633502b7b2522cd5" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "wiping_tool_in_hand": { + "method": "visual", + "intervals": [ + [ + 0.15, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.15, + "log_step": null, + "text_zh": "观察:视频 0.00 s 银杆悬在张开的左手上方;视频 0.10–0.30 s 左手手指闭合在杆附近,工具短暂下降后留在手中并转动。视频 0.40–0.65 s 杆随左手移动,未继续自由落下;首次拿稳约在视频 0.15 s,时间精度约 0.1 s。" + }, + { + "t": 2.95, + "log_step": null, + "text_zh": "观察:视频 0.70–2.95 s 左手继续握着银杆,黑色擦头随手反复摆动;末帧工具仍与左手一起移动,未见工具离手坠地。右手未参与持杆。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "抓握", + "note_zh": "左侧 Shadow 手闭合手指与拇指握住银色工具杆,之后单手控杆;右手未共同拿工具。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "机器人双脚接触地面,支撑接住并持稳工具时的身体。" + } + ] + }, + "window_wiped": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.7, + "log_step": null, + "text_zh": "观察:视频 0.40–0.65 s 黑色擦头在玻璃下缘附近转动;视频约 0.70 s 机器人上身倾斜、左臂下垂,擦头降到窗台下方。视频 1.00–2.95 s 擦头主要在灰色下墙板前往复,未见恢复到玻璃面上的持续上下擦动。成功参考视频 0.70–3.00 s 的擦头则保持在青色玻璃区域;参考仅用于辨认应有的空间关系。" + }, + { + "t": 2.95, + "log_step": 187, + "text_zh": "观察:log step 185 最后实际运行的 tL 变体执行 291 控制步并返回 failed,随后重置因额度耗尽被拒;log step 187 服务状态为 episode=51、control_steps=291、done=true、reason=failed,与评分器 success=0 一致。视频末帧仍是低位持杆摆动。" + }, + { + "t": null, + "log_step": 187, + "text_zh": "推断:本任务唯一受评目标是累计回报达到 650,服务已以 failed 结束最后回合,且评分器报告失败,因此擦拭节点未达标;不能把未报告的累计回报当作零。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "抓握", + "note_zh": "左手实际握着工具杆并带动擦头摆动;画面能确认手—杆抓握,不能可靠确认黑色擦头对玻璃的受压接触,故不追加手持工具×压/擦。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "机器人双脚着地支撑持杆摆动,上身姿态明显改变;视频末帧未见身体倒地。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 650, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "window_wiped", + "t": 0.7, + "log_step": 185, + "where": "操作中", + "evidence_zh": "观察:视频约 0.70 s 在已持住工具后,机器人左臂与擦头降到玻璃下缘以下,随后至视频 2.95 s 仍在低位持杆往复,没有恢复玻璃面上的擦拭位置;log step 185、187 确认最后受评回合结束为 failed。" + }, + "unmet": { + "window_wiped": false + }, + "phenomena": [ + { + "cat": "运动", + "item": "位姿不准", + "t": 0.7, + "log_step": null, + "evidence_zh": "观察:视频 0.70–2.95 s 机器人握着工具摆动,但黑色擦头大部分位于青色玻璃下方的灰色墙板前;视频 2.50–2.95 s 逐帧仍如此,未回到玻璃面上的工作位置。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "t": 0.7, + "log_step": 185, + "source_zh": "日志 / 视频", + "evidence_zh": "推断:log step 169 的控制脚本在初始关节目标上给左肩俯仰叠加三角波;log step 179 去掉速度阻尼,log step 181 延后开始,log step 185 最后实际运行的 tL 使用偏置 -0.08、幅度 0.18、周期 1.0 的变体,腿部增益为 0,没有按擦头相对玻璃的位置修正轨迹。视频 0.70–2.95 s 上身姿态改变后,手臂仍带工具在玻璃下方摆动,支持控制未维持工作位姿;证据不足以进一步确定是抓握几何、逆解还是某个参数单独造成。" + } + ], + "gap_zh": "没有状态日志,服务仅给出 failed,未报告累计回报、玻璃接触对或具体终止条件。视频早段工具与玻璃下缘的深度关系及压力不清,无法确认是否有短暂擦头—玻璃接触;末段未见倒地,不能把 agent 在 log step 195、196 所称“骨盆过低/平衡是主要阻碍”当作本回合已证实的终止原因。log step 187 的 291 控制步与 log step 6 所列每步仿真 0.02 s 相乘为仿真 5.82 s,与视频总长 3.00 s 不一致,故所有节点与断点时间只按已检查的视频帧给出,不据此换算。" + }, + "summary_zh": "受评回合在视频约 0.15 s 单手拿稳擦窗工具,最终完成 1/2 个节点,未达到累计回报 650 的擦拭验收。视频约 0.70 s 起擦头降至玻璃下方,之后低位往复至结束;断点为擦拭窗户,现象为位姿不准,控制脚本未维持擦头相对玻璃的工作位姿。没有状态日志或精确回报,无法确认短暂玻璃接触及服务判负的具体终止条件。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "节点(拿稳工具约0.15 s、擦拭未达成)、断点(window_wiped,操作中)、位姿不准与控制根因均与视频帧一致;未发现需要改动的内容。后期帧中右手是否也碰到杆分辨率不足,无法据以推翻“单手抓握”。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:48:35Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard--fei3twd/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard--fei3twd/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "091e7164578f1be1fa086a1ecc8efbdef12c2a9d6fd3c0a6f248081e9c5db131", + "video_sha256": "4fe0948d15be7385c08be580f29be56d5820f8539ef3278f633502b7b2522cd5", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard--gdqycqd.json b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard--gdqycqd.json new file mode 100644 index 0000000000000000000000000000000000000000..0515b4e62b9b922a6cdde4a02e9922c7ae5903db --- /dev/null +++ b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard--gdqycqd.json @@ -0,0 +1,228 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-window-i00-standard--gdqycqd", + "taskKey": "humanoidbench/manip_window", + "dag_sha256": "4dc821ae18297046c83611bc223535d1570899c28d34142481b029cb27e902ac", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 454, + "sim_steps": 18007, + "calls": 949, + "rows": 17974, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 974.5, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-window-i00-standard--gdqycqd/replay.mp4", + "duration_s": 4.6, + "fps": 20.0, + "frames": 92, + "sha256": "f202f3b40dd3f1104780a00685ae581843297cc48d846a0430f3bfe5972c4f63" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "wiping_tool_in_hand": { + "method": "visual", + "intervals": [ + [ + 0.15, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.15, + "log_step": null, + "text_zh": "观察:视频 0.00 s 银色工具杆位于张开的手掌上方;视频 0.10–0.50 s,近镜头左手的手指从杆的上下两侧合拢,银杆及黑色擦头随该手抬高、降低,没有继续自由下落,故约在视频 0.15 s 达到拿稳工具。" + }, + { + "t": 3.8, + "log_step": null, + "text_zh": "观察:视频 0.75–3.80 s 工具仍保持在手部附近并反复随手运动;视频 4.15–4.55 s 杆相对左手发生滑转,擦头下垂。此节点为 event,首次拿稳后锁存,不因后续滑移撤销。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "抓握", + "note_zh": "视频 0.10–0.50 s 可见近镜头左手的手指包围银色工具杆并带动工具;远侧右手开局被躯干遮挡,不把控制脚本的双手闭合命令直接当作双手抓握事实。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "拿杆时机器人双脚接触地面,支撑屈膝站立的身体。" + } + ] + }, + "window_wiped": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.45, + "log_step": null, + "text_zh": "观察:视频 0.10–1.20 s,机器人持杆反复升降黑色擦头,擦头沿青色玻璃附近上下移动并不断转动朝向;视频 1.40–3.80 s,擦头逐渐靠近玻璃下边缘。推断:擦头贴窗运动支持已尝试擦拭,但画面不能计算累计回报,不能据此判定整步验收完成。" + }, + { + "t": 4.25, + "log_step": null, + "text_zh": "观察:视频 4.00–4.55 s 机器人仍在操纵工具,银杆逐渐斜转,黑色擦头下降到青色玻璃下缘以下的灰色板前,结束前没有恢复到玻璃上。" + }, + { + "t": null, + "log_step": 118, + "text_zh": "观察:评分器字段 success=0、live_success=0、graded_by_live=1;log step 118 的机器人服务状态为 episode=51、control_steps=454、done=True、reason=failed,log step 120 再次确认相同终局。推断:受评的最后一个现场回合没有达到本节点要求的累计回报 650;没有把早期重置回合的测试结果计入进度。" + } + ], + "contact_actual": [ + { + "who": "手持工具", + "how": "压/擦", + "note_zh": "视频 0.10–1.20 s 机器人经银色杆带动黑色擦头沿玻璃上下运动,贴合窗面的几何支持压擦接触;后段擦头斜转并降到玻璃下方。无接触日志,具体接触力和持续时间不能量化。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "操纵擦头期间双脚与地面接触并支撑身体;末段身体倾斜、双腿张开,未据此额外断言已倒地。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 650, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "window_wiped", + "t": 4.15, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频 0.15 s 已能持杆,随后机器人反复抬落擦头;视频约 4.15 s 起,银杆相对近镜头左手明显斜转,视频 4.25–4.55 s 黑色擦头降到玻璃下方,结束前没有恢复。推断:未修复的失败发生在持工具擦窗的操作中,不能把已拿稳过工具的前置事件标为未达成。" + }, + "unmet": { + "window_wiped": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "没保持住", + "t": 4.15, + "log_step": null, + "evidence_zh": "观察:视频 0.10–0.50 s 左手包住银杆,工具随手运动;视频 4.15–4.55 s,银杆在左手处滑转,近侧持握关系未保持,黑色擦头下垂到玻璃下方,远端杆露到右肩旁。推断:这是已有持握后的滑移,非从未建立接触;不据此断言工具已完全脱离两手。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "t": 4.15, + "log_step": 113, + "source_zh": "日志 / 视频", + "evidence_zh": "观察:log step 113 的最终脚本给手指固定闭合目标,以肩肘周期位置命令驱动工具,反馈只使用机体 IMU 来调髋踝;log step 115 执行该脚本,log step 118 显示最后回合失败。视频 4.15–4.55 s 可见杆相对手部滑转、擦头下垂,控制未将其恢复。推断:执行控制没有持续稳定工具与手的相对姿态及擦头位置;证据不足以进一步区分握持几何、力不足或身体运动耦合,故仅标控制。" + } + ], + "gap_zh": "没有状态或接触日志,远侧手的大部分接触过程被躯干遮挡,不能确认两手完整持握时序、末段是否完全脱手,也不能确定骨盆或工具哪一项高度阈值触发终止。episode_return 未报告,保留 null;未达擦窗验收由受评失败结果与视频过程共同判断。log step 120 给出仿真 9.08 s,而视频文件长 4.60 s(末帧视频 4.55 s),没有直接套用控制步与仿真时间来定位视频,本文所有视频锚点均取自所检帧。" + }, + "summary_zh": "机器人约在视频 0.15 s 拿稳工具,随后反复操纵擦头,两个节点中仅拿稳工具达成,擦窗累计回报验收未通过。视频约 4.15 s 起工具在手部滑转,擦头降到玻璃下方且未恢复,断点为擦拭窗户,现象为没保持住,根因仅支持控制层。缺少接触与累计回报记录,远侧手受遮挡,且日志仿真时长与视频时长不一致,无法细分滑移机制或确认精确终止触发项。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "节点值(拿稳工具达成、擦窗未达成)、断点(window_wiped,操作中,视频约4.15 s)、现象(没保持住)与根因(控制,证据不足时不细分)与视频和日志一致,时钟均为视频时间,未发现需要修改的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:51:07Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard--gdqycqd/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard--gdqycqd/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "71d211d29dbf4d805bd73d577b1e557a869e0febacc25940d59dc3b0bf810fe8", + "video_sha256": "f202f3b40dd3f1104780a00685ae581843297cc48d846a0430f3bfe5972c4f63", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard--hnrbe2e.json b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard--hnrbe2e.json new file mode 100644 index 0000000000000000000000000000000000000000..a0c4e2f21ee7d0725c68796f9bf69f33931247d2 --- /dev/null +++ b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard--hnrbe2e.json @@ -0,0 +1,226 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-window-i00-standard--hnrbe2e", + "taskKey": "humanoidbench/manip_window", + "dag_sha256": "4dc821ae18297046c83611bc223535d1570899c28d34142481b029cb27e902ac", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 970, + "sim_steps": 26411, + "calls": 473, + "rows": 26411, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3373.9, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.339, + "peak_processes": 9, + "peak_disk_gb": 0.932, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-window-i00-standard--hnrbe2e/replay.mp4", + "duration_s": 9.8, + "fps": 20.0, + "frames": 196, + "sha256": "3f76ff84e838e929c38272d30ce8e1d60d419d8a26d7f0090266bc3f2c915bf7" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "wiping_tool_in_hand": { + "method": "visual", + "intervals": [ + [ + 0.2, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.2, + "log_step": null, + "text_zh": "观察:视频 0.0 s 银色杆悬在手掌上方;视频 0.1–0.5 s 近侧左手的手指围在杆周围,杆随屈肘先抬高再下降,没有继续自由下坠。视频 0.6–1.3 s 工具继续随左手移动和转动,满足拿稳工具的可见判据;首次成立约为视频 0.2 s。" + }, + { + "t": 9.75, + "log_step": null, + "text_zh": "观察:视频 1.5–8.7 s 工具保持在左手前方,视频 8.8–9.75 s 身体前倾加深时黑色擦头仍在窗前、银杆仍连到近侧手部,未见工具掉到地面。此节点为 event,首次拿稳后锁存。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "抓握", + "note_zh": "视频 0.2–0.6 s 可见近侧左手的手指包围银色工具杆并使其随手抬动;远侧右手的杆上接触被身体遮挡,不据 agent 的双手抓握自述追加双臂协作。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "视频 0.0–1.3 s 机器人双脚接触地面,支撑抓杆和抬动工具时的身体。" + } + ] + }, + "window_wiped": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.3, + "log_step": null, + "text_zh": "观察:视频 0.2–1.3 s 左手带动工具升降,但黑色擦头同时明显翻转;视频 1.5–8.7 s 擦头停留在竖窗框附近的胸部高度,仅有很小的位移和转角变化,未见持续稳定的贴窗上下擦动;视频 8.8–9.75 s 擦头继续转斜,身体前倾加深。" + }, + { + "t": 9.75, + "log_step": 393, + "text_zh": "观察:log step 391–392 表明 agent 用最后一次重置启动平滑擦动脚本;log step 393 的服务结果记录 episode 51 在 control_steps=970、仿真 19.4 s 结束,done=true、reason=failed、resets_left=0,inputs/run.json 的 live_success=0、graded_by_live=1,未报告 episode_return。" + }, + { + "t": 9.75, + "log_step": 393, + "text_zh": "推断:本任务累计回报达到 650 会由现场服务立即判成功,最后受评回合的服务失败结果与视频中的不稳定擦动一致,故擦窗累计验收未达成;不是把总奖励 0 当作没有尝试。成功参考录像的视频 0.5–10.0 s 展示了较稳定的持杆与贴窗几何,仅用于对照,不作为本回合证据。" + } + ], + "contact_actual": [ + { + "who": "手持工具", + "how": "压/擦", + "note_zh": "视频 0.2–0.5 s 近侧左手持杆使黑色擦头在窗面附近升降;视频 1.3–9.75 s 擦头压靠在玻璃与竖框交界附近并小幅移动。单一斜后视角不能精确区分后段擦头压在玻璃还是竖框上;这里记录工具对窗体的压擦,不表示玻璃接触奖励已成立。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "视频 0.2–9.75 s 机器人脚部接触地面,支撑持杆操作;身体逐渐前倾,末段一腿向后伸出。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 650, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "window_wiped", + "t": 1.3, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:机器人已在视频 0.2 s 左右拿住工具;视频 0.6–1.3 s 擦头翻转,随后视频 1.5–8.7 s 长时间滞留在竖窗框附近,直到视频 9.75 s 也未恢复稳定上下擦动。log step 393 的最后现场回合 episode 51 结束于 failed;断点为正在操作而未完成的擦窗步骤。" + }, + "phenomena": [ + { + "cat": "运动", + "item": "位姿不准", + "t": 1.3, + "log_step": null, + "evidence_zh": "观察:视频 0.6–1.3 s 左手抬动银杆时黑色擦头显著翻转,至视频 1.5–8.7 s 擦头主要停留在竖框边的很小范围内,未形成稳定的贴窗上下运动;视频 9.0–9.75 s 擦头又随前倾身体转斜。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "推断:log step 391–393 记录 agent 在最后回合以平滑肩部擦动脚本控制机器人;视频 0.2–1.3 s 工具确实被拿住并受手驱动,但随后擦头转斜、工具运动滞留,直到视频 9.75 s 仍未转为稳定擦动。证据支持工具运动与接触控制未奏效;没有足够的最终控制器细节、接触力或跟踪误差来区分抓取设计、轨迹或反馈控制子类。", + "source_zh": "视频 / 日志" + } + ], + "gap_zh": "没有状态日志、擦头接触对、接触力或逐步累计回报;episode_return 保留 null。单一 256×256 斜后视角使远侧右手接触及擦头在玻璃与竖框间的确切接触面不清;无法量化擦动速度或确定 failed 具体由骨盆高度还是工具高度触发。视频末段前倾加深,但不足以另判已经跌倒;不从早期调试回合的跌倒或掉杆替代本回合现象。log step 393 的仿真 19.4 s 与视频总长 9.8 s 不同,所有节点与断点时间均取已查看的视频帧,没有将 agent 工作时间或仿真时间直接换算为视频时间。", + "unmet": { + "window_wiped": false + } + }, + "summary_zh": "最后受评的是 episode 51,前面的 50 次重置不计入本轨迹进度:机器人约在视频 0.2 s 拿稳工具,最终完成 1/2 个节点。擦窗时擦头翻转并长期滞留在窗框附近,未能形成稳定擦动,现场服务在 log step 393 报告该回合失败,断点为擦拭窗户,根因仅归到控制。没有状态日志和累计回报数值,无法判定精确接触面或具体高度终止条件。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "节点 wiping_tool_in_hand 已达成(视频开头即持杆)、window_wiped 未达成,断点、位姿不准与控制根因均有帧和服务读数支撑,时钟未把仿真时间换算成视频时间。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:08:05Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard--hnrbe2e/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard--hnrbe2e/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "7ede63ff9d2c7f2369df83ecc71b3526fd2b6c58a6fa0e70768ab9063409c15c", + "video_sha256": "3f76ff84e838e929c38272d30ce8e1d60d419d8a26d7f0090266bc3f2c915bf7", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard--mhvtgre.json b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard--mhvtgre.json new file mode 100644 index 0000000000000000000000000000000000000000..87df9be2240f9267e18db52a779e1c75069e72c1 --- /dev/null +++ b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard--mhvtgre.json @@ -0,0 +1,225 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-window-i00-standard--mhvtgre", + "taskKey": "humanoidbench/manip_window", + "dag_sha256": "4dc821ae18297046c83611bc223535d1570899c28d34142481b029cb27e902ac", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 355, + "sim_steps": 14257, + "calls": 14086, + "rows": 14257, + "resets": 38, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3537.6, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-window-i00-standard--mhvtgre/replay.mp4", + "duration_s": 3.65, + "fps": 20.0, + "frames": 73, + "sha256": "50418a7fdfb5f17c23e0ad524047be618c1a465b7f17730ebae663c2ed3aa161" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "wiping_tool_in_hand": { + "method": "visual", + "intervals": [ + [ + 0.3, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.3, + "log_step": null, + "text_zh": "观察:视频 0.00–0.20 s,银色长杆和黑色擦头先下降;视频 0.25–0.40 s,画面近侧灵巧手的手指围住杆身,工具停止自由下落并随手重新抬高。以视频约 0.30 s 记首次拿稳。" + }, + { + "t": 1.4, + "log_step": null, + "text_zh": "观察:视频 0.75–2.80 s,机器人抬动持杆手,银杆和黑色擦头随之升降,未继续自由坠落;视频 3.30–3.60 s 才脱手下坠。本节点为 event,首次拿稳后锁存,后续掉落不撤销已达到记录。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "抓握", + "note_zh": "观察:视频 0.25–0.40 s 及视频 0.75–2.80 s,画面近侧灵巧手的手指包住银色工具杆并带动它;远侧手被躯干遮挡,不据脚本的双手目标推定双手实际同时抓住。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–3.60 s,机器人双脚落地,腿部承重,支撑持杆和抬臂。" + } + ] + }, + "window_wiped": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 2.2, + "log_step": null, + "text_zh": "观察:视频 0.75–2.95 s,机器人在窗前升降持杆手,黑色擦头在青色玻璃的投影区域内移动并反复翻转;单一外部视角不能可靠区分擦头贴玻璃与悬在玻璃前,因此只确认开始操纵工具尝试擦窗。" + }, + { + "t": 3.5, + "log_step": 161, + "text_zh": "观察:视频 3.10–3.60 s,工具先在手中大幅转斜,随后杆身滑过手指并向下坠落,机器人仍站立,未重新抓住。log step 161 中机器人服务报告最后回合 episode=39、control_steps=355、done=True、reason=failed;评分器 live_success=0、replay_success=0。" + }, + { + "t": 3.6, + "log_step": 161, + "text_zh": "推断:结合视频末尾工具脱手、机器人服务的最后回合失败结果和评分器成功字段,累计回报验收的擦窗节点未达到;不能仅凭画面中的工具升降记整步完成。inputs/run.json 的 reward 未报告 episode_return,故不填写其数值。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察:视频 0.75–2.95 s 可见机器人持工具在玻璃前升降,但擦头与玻璃之间的深度间隙和接触压力不可辨;实际擦头—玻璃接触方式未知。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.75–3.60 s,双脚持续在地面支撑,机器人抬臂尝试擦窗时没有摔倒。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 650, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "window_wiped", + "t": 3.3, + "log_step": 161, + "where": "操作中", + "evidence_zh": "观察:视频 0.30 s 左右机器人已经接住工具,随后持续抬臂操纵它;视频 3.10–3.60 s,工具在持杆手中转斜并滑脱,擦窗动作中断,末尾未恢复。log step 159 重置并启动最后回合,log step 161 的 episode=39 在 control_steps=355 以 reason=failed 结束;更早的长回合已被重置,不计本次进度。" + }, + "unmet": { + "window_wiped": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "没保持住", + "t": 3.3, + "log_step": null, + "evidence_zh": "观察:视频 0.30–2.80 s,工具能随持杆手运动;视频 3.10–3.45 s,杆身相对手指明显转斜、滑移,视频 3.50–3.60 s 工具向下落出手掌,手停在上方。机器人失去的是已经建立的持杆接触。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "t": 3.3, + "log_step": 159, + "evidence_zh": "推断:log step 106 的控制器采用分阶段手指闭合和手臂逆解,log step 146 加入按预定时序闭合右手,log step 159 最后回合使用 elbowwave=0.12、wristgain=0.05 驱动擦动;视频 3.10–3.60 s 显示工具相对持杆手滑转并最终滑脱,身体仍有支撑。这支持抓持与手臂运动的控制未能在擦动时维持工具约束;现有证据不足以进一步区分接触布局、握力或关节跟踪问题。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "没有状态日志或接触力记录;远侧手被躯干遮挡,不能确认两手是否同时包握,也不能精确确认擦头—玻璃的实际接触。评分器未报告 episode_return,无法恢复累计回报曲线;log step 161 的仿真终点为 7.10 s,而视频全长为 3.65 s,未核实两者映射,所有节点与断点时间均直接取已检查的视频帧。控制根因保留大类,不推断未观测的握力、接触几何或反馈失稳子因。" + }, + "summary_zh": "受评回合在视频约 0.30 s 接住工具并开始抬动,拿稳工具事件已达到,最终节点进度为 1/2。视频约 3.30 s,机器人尝试擦窗时未保持住工具,长杆滑脱下坠,擦窗累计回报验收未达到;身体仍站立,根因仅能支持到控制大类。缺少状态、接触力和累计回报记录,远侧手及擦头是否真正压住玻璃无法可靠确认。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "节点值、时间、断点和现象与视频一致:视频约0.3 s起工具随手移动,3.1–3.6 s滑脱坠落,终局log step 161 done=True reason=failed 与工具掉落终止相符;控制根因保持大类并有gap说明,未发现需要改动的内容。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:55:50Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard--mhvtgre/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard--mhvtgre/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "6c3b4d09eb59719ed9a3ead740bff4f3dbdc6788ffe449ea557bd8730e0ab1e1", + "video_sha256": "50418a7fdfb5f17c23e0ad524047be618c1a465b7f17730ebae663c2ed3aa161", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard--oaqy84d.json b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard--oaqy84d.json new file mode 100644 index 0000000000000000000000000000000000000000..60d1cae62f4ba0eeeb04df1e1ebc4e52b1d1bd7f --- /dev/null +++ b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard--oaqy84d.json @@ -0,0 +1,228 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-window-i00-standard--oaqy84d", + "taskKey": "humanoidbench/manip_window", + "dag_sha256": "4dc821ae18297046c83611bc223535d1570899c28d34142481b029cb27e902ac", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 200, + "sim_steps": 1611, + "calls": 57, + "rows": 1611, + "resets": 6, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3576.2, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-window-i00-standard--oaqy84d/replay.mp4", + "duration_s": 2.1, + "fps": 20.0, + "frames": 42, + "sha256": "a9d86f442bc02f9db023fd4a468af7c57a8b0848e13673c40bd511a144f659ca" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "wiping_tool_in_hand": { + "method": "visual", + "intervals": [ + [ + 0.15, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.15, + "log_step": null, + "text_zh": "观察:视频 0.00–0.10 s,银杆位于双手前方并略有下沉;视频 0.15–0.30 s,双手在杆的两处分握,工具随抬手上升,未继续自由下落。首次拿稳时间约为视频 0.15 s。" + }, + { + "t": 2.0, + "log_step": null, + "text_zh": "观察:视频 0.60–2.05 s,工具虽反复转动,仍随机器人近侧手移动,未掉到地面;本节点为 event,首次拿稳后锁存。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "观察:视频 0.15–0.60 s,机器人两只灵巧手分别在银杆两处合指握持并抬杆;视频约 0.70 s 后远侧手逐渐下移,近侧手继续持杆。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–2.05 s,机器人双脚接地支撑持杆及身体前倾,未见跌倒。" + } + ] + }, + "window_wiped": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.65, + "log_step": null, + "text_zh": "观察:视频 0.50–0.65 s,机器人持杆前伸至窗面附近;视频 0.65–1.10 s,黑色擦头反复从近乎平放翻到竖起,银杆随之倾斜,没有形成稳定的擦面朝向。" + }, + { + "t": 2.0, + "log_step": null, + "text_zh": "观察:逐帧检查视频 1.10–2.05 s,擦头继续大幅翻转,远侧手处于杆下方,近侧手持杆在窗前摆动;末帧仍未显示稳定的贴窗上下擦动。" + }, + { + "t": null, + "log_step": 121, + "text_zh": "观察:log step 121 是最后一次 reset 后执行的回合,控制脚本返回 Done: False、Reason: None、Steps: 200;后续 log step 122–125 仅分析运动学或查看图像,没有继续提交动作。inputs/run.json 的评分器字段 live_success=0、replay_success=0、n_actions=200,未报告 episode_return。" + }, + { + "t": null, + "log_step": null, + "text_zh": "推断:DAG 将整步完成绑定本回合累计回报达到 650;评分器未验收成功,且画面只见不稳定的窗前运动,因此擦拭窗户整步未达成。此结论不把零成功值解释为完全没有接触玻璃。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察:视频 0.65–2.05 s,机器人持工具在玻璃附近摆动,擦头与窗面在单一外部视角中重叠;256×256 画面无法可靠区分真实压擦与贴近、擦过窗框,实际工具—玻璃接触方式未知。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.50–2.05 s,机器人双脚接地,以身体前倾支撑窗前持工具运动。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 650, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "window_wiped", + "t": 0.65, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频约 0.65 s 起,机器人将已拿住的工具送到窗前后,擦头开始反复明显翻转;视频 1.10–2.05 s 仍以倾斜摆动为主,没有恢复稳定贴窗擦动。拿稳工具已先达成,未完成且未修复的步骤是擦拭窗户。" + }, + "unmet": { + "window_wiped": false + }, + "phenomena": [ + { + "cat": "运动", + "item": "位姿不准", + "t": 0.65, + "log_step": null, + "evidence_zh": "观察:视频 0.65–1.10 s,黑色擦头在近乎平放、竖起及斜置之间反复变化;视频 1.30–2.05 s,银杆斜向上且擦头持续翻摆,机器人未把擦面稳定对准玻璃并保持上下擦动。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "视频", + "evidence_zh": "推断:视频 0.15–0.60 s 已显示机器人能够接住并搬动工具,但视频 0.65–2.05 s 的执行未能约束工具朝向,远侧手下移后近侧手持杆仍反复翻摆,支持工具操纵控制未稳定这一层归因;画面不足以细分抓握几何、关节跟踪或反馈控制问题。" + } + ], + "gap_zh": "观察:inputs/ 未提供状态日志;仅有单一低分辨率外部视频,无法确定擦头—玻璃的实际接触、压力及具体控制失效机制,故擦窗接触标未知,根因停在控制大类。评分器未报告 episode_return,原生指标保留 null。log step 121 的 200 个控制步按提交 dt=0.02 对应仿真 4.0 s,而视频总长 2.1 s,缺少可验证的控制步—视频映射;所有节点及断点时间直接取自视频,不换算日志时间。" + }, + "summary_zh": "受评视频中机器人约在视频 0.15 s 接住并抬动擦窗工具,完成拿稳工具,最终节点进度为 1/2。视频约 0.65 s 起擦头反复翻转,至视频 2.05 s 仍未稳定对准窗面擦动,断在擦拭窗户,支持位姿不准及控制大类归因。无状态日志和累计回报数值,具体玻璃接触及控制失效子类无法确认。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "逐帧抽查视频 0–2.1 s:双手握杆抬起、随后擦头反复翻转且未稳定贴窗,与标注一致;节点值、断点、现象与根因保守且有证据,时间均取自视频时钟。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:46:52Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard--oaqy84d/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard--oaqy84d/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "3629d602855b28ae9afd7f87f77e78b3e630c6550f57f5c554a730997e7eadb9", + "video_sha256": "a9d86f442bc02f9db023fd4a468af7c57a8b0848e13673c40bd511a144f659ca", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard--vxqffsp.json b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard--vxqffsp.json new file mode 100644 index 0000000000000000000000000000000000000000..f81af4e1be5b9df4e5c1ff05477869eb82a77603 --- /dev/null +++ b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard--vxqffsp.json @@ -0,0 +1,226 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-window-i00-standard--vxqffsp", + "taskKey": "humanoidbench/manip_window", + "dag_sha256": "4dc821ae18297046c83611bc223535d1570899c28d34142481b029cb27e902ac", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 1000, + "sim_steps": 13342, + "calls": 11702, + "rows": 13276, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 778.0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.167, + "peak_processes": 8, + "peak_disk_gb": 0.241, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-window-i00-standard--vxqffsp/replay.mp4", + "duration_s": 10.1, + "fps": 20.0, + "frames": 202, + "sha256": "a6a512b5cf5f80f88ea813c33cfd7ce79af309af52eff84b0d39bf7dfc33e95b" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "wiping_tool_in_hand": { + "method": "visual", + "intervals": [ + [ + 0.2, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.2, + "log_step": null, + "text_zh": "观察:视频 0.0–0.1 s 银杆略下落,视频 0.2–0.4 s 机器人画面左侧的灵巧手以手掌和弯曲手指夹住银杆并将其抬起;银杆随手移动,不再自由下坠。首次拿稳时间约为视频 0.2 s。" + }, + { + "t": 10.05, + "log_step": null, + "text_zh": "观察:视频 1.0–8.0 s 工具持续随持杆手摆动,视频 9.0–10.05 s 仍连在手上,未见整杆脱手落地;另一只手没有形成持续的共同握杆。此节点为 event,首次拿稳后锁存。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "抓握", + "note_zh": "视频 0.2–0.4 s,画面左侧的一只灵巧手以手掌与弯曲手指夹住银色工具杆并抬动,随后由这一只手持续控杆。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "机器人双脚接触地面,支撑接住并抬动工具时的身体。" + } + ] + }, + "window_wiped": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.0, + "log_step": null, + "text_zh": "观察:视频 0.5–5.0 s,机器人持杆将黑色擦头贴近并顶靠青色窗面,反复小幅升降和转动;可见擦窗尝试,但这些画面本身不能证明累计回报达到 650。" + }, + { + "t": 8.0, + "log_step": null, + "text_zh": "观察:视频 5.5–8.0 s 双脚逐渐向后滑、身体前倾加重,持杆手和擦头下降;视频 8.0–10.05 s 黑色擦头主要在玻璃下沿及其下方摆转,未恢复前段在玻璃中的工作位置,银杆仍在手中。" + }, + { + "t": 10.05, + "log_step": 50, + "text_zh": "推断:log step 50 是最后一次 reset 后执行的现场回合,机器人服务返回 DONE 999 failed 0;评分字段 graded_by_live=1、n_actions=1000、live_success=0,且未报基础设施错误。结合视频中的擦拭及后期下降,按此节点的累计回报验收判为未达 650;具体 episode_return 未提供,不能把 catalog reward=0 当作累计回报。" + } + ], + "contact_actual": [ + { + "who": "手持工具", + "how": "压/擦", + "note_zh": "视频前段,一只灵巧手操纵银杆,使黑色擦头顶靠青色玻璃并小幅擦动;后段擦头逐渐降至玻璃下沿以下。手握工具到擦头作用窗面的链路只记一次。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚接触地面支撑擦窗操作;后半段脚部向后滑移,身体前倾加重。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 650, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "window_wiped", + "t": 7.5, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频约 6.0–7.5 s 机器人在继续持杆摆动时,脚部后移、躯干前倾,擦头工作位置逐渐降到窗户下沿;视频 8.0–10.05 s 擦头落在玻璃下沿以下的画面持续出现,未恢复有效的窗面擦拭位置。log step 50 的最后现场回合执行满 1000 个控制步后由服务判为 failed;失败发生在擦窗步骤,拿稳工具已完成。" + }, + "unmet": { + "window_wiped": false + }, + "phenomena": [ + { + "cat": "运动", + "item": "位姿不准", + "t": 7.5, + "log_step": null, + "evidence_zh": "观察:视频 6.0–8.0 s 机器人身体和持杆手逐渐降低,黑色擦头从青色玻璃内降向下沿;视频 9.0–10.05 s 擦头在玻璃下方的灰色墙板前继续摆转,未回到窗面内上下擦动,工具仍随手运动。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "反馈控制", + "source_zh": "日志 / 视频", + "evidence_zh": "推断:log step 28 写出的腿部控制器只对固定关节目标做位置误差放大和速度阻尼;log step 39 的执行循环调用该控制器,log step 44 的姿态自适应需显式启用,而 log step 50 最终采用 K=2、Kd=0.05、未启用 adapt。手臂仍按固定肩目标和肘关节三角波执行(log step 26、49、50),没有随身体漂移修正擦头位置。视频 5.5–10.05 s 显示脚部后滑、前倾持续加重、擦头降出窗面,说明局部关节反馈未能稳定身体并维持工具相对窗面的工作位姿;具体增益是否最优和各回报分项损失无法由现有记录确定。" + } + ], + "gap_zh": "无状态日志、接触力或累计回报数值;256×256 单一外部视角不能精确核实擦头每次接触及各手指接触对,前段压擦按擦头顶靠窗面并随手运动的几何推断。控制记录中 1000 步×0.02 s 对应仿真 20 s,而视频仅 10.1 s,未建立两者映射,所有视频时刻均直接取自检查的帧,不将 log step 50 的控制步数换算为视频时间。" + }, + "summary_zh": "机器人在视频约 0.2 s 拿稳工具,完成 1/2 个节点,并在前半段持杆尝试擦窗;最后现场回合执行满 1000 个控制步后未通过累计回报 650 的验收。视频后半段脚部后滑、身体前倾加重,擦头逐渐降至玻璃下沿以下,擦窗步骤未修复;日志中的局部关节反馈未能阻止这一工作位姿漂移。没有状态日志或累计回报分项,不能精确量化接触持续时间及分项损失,视频与仿真时钟亦未建立映射。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "拿稳工具与擦窗未达标的节点判断、断点(window_wiped,脚后滑、身体前倾、擦头降出窗面)和反馈控制根因与视频一致,未发现需改动的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T03:44:28Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard--vxqffsp/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard--vxqffsp/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "fdae9c733906cb4ac4d23ed67857f68aebee1ca1a218ccecc2db95652e386e18", + "video_sha256": "a6a512b5cf5f80f88ea813c33cfd7ce79af309af52eff84b0d39bf7dfc33e95b", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard-realtime--4msmchf.json b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard-realtime--4msmchf.json new file mode 100644 index 0000000000000000000000000000000000000000..077395452f50ba6a68345a00ffc12a581162ccd3 --- /dev/null +++ b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard-realtime--4msmchf.json @@ -0,0 +1,234 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-window-i00-standard-realtime--4msmchf", + "taskKey": "humanoidbench/manip_window", + "dag_sha256": "4dc821ae18297046c83611bc223535d1570899c28d34142481b029cb27e902ac", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 751, + "sim_steps": 20530, + "calls": 20648, + "rows": 20501, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3570.5, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 409.889, + "max_lag_s": 0.0162, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 2237, + "late_steps": 0, + "requests_refused": 52, + "request_busy_s": 20.977, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.31, + "peak_processes": 10, + "peak_disk_gb": 0.839, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-window-i00-standard-realtime--4msmchf/replay.mp4", + "duration_s": 7.6, + "fps": 20.0, + "frames": 152, + "sha256": "bb328e5f073f6eb61574509191db8c9f3afd46a430e979ffa380ba597546815c" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "wiping_tool_in_hand": { + "method": "visual", + "intervals": [ + [ + 0.2, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.2, + "log_step": null, + "text_zh": "观察:视频 0.10–0.30 s,可见的一只灵巧手手指合拢在银色长杆周围,工具由初始悬空位置随手抬起;视频 0.40–6.90 s,长杆随该手下降并维持在手中,未继续自由下落。拿稳时刻约取视频 0.20 s。" + }, + { + "t": 7.55, + "log_step": null, + "text_zh": "观察:视频 7.00–7.55 s,长杆及擦头虽向下倾斜,可见手仍扣住杆的近端,工具随手摆动。此节点为 event,早先拿稳已锁存。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "抓握", + "note_zh": "观察:视频 0.10–0.30 s,靠近镜头的一只灵巧手手指合拢握住银色工具杆,之后持续控杆;未把另一只位于身体旁的手算作共同握杆。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.95 s,机器人双脚接触地面,支撑接住并拿稳工具时的身体。" + } + ] + }, + "window_wiped": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.7, + "log_step": null, + "text_zh": "观察:视频 0.00–0.45 s,黑色擦头在玻璃前短暂抬起、翻转;视频约 0.60–0.90 s,擦头降到青色玻璃下缘,之后视频 1.00–6.90 s 基本停留在窗下沿附近,未见持续在玻璃区域上下擦动。" + }, + { + "t": 7.3, + "log_step": 332, + "text_zh": "观察:视频 7.10–7.55 s,工具向下倾斜,黑色擦头明显位于玻璃下方;log step 332 的服务返回 EpisodeOver、reason failed、resets_left 0,评分器字段 success=0、live_success=0,未报告 episode_return。" + }, + { + "t": null, + "log_step": null, + "text_zh": "推断:上述画面未显示满足判据的持续贴窗擦动,且本回合未被评分器验收成功,因此擦拭整步未达到累计回报 650 的完成条件;评分器 reward=0.0 不是本回合累计回报数值。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察:视频 0.00–0.45 s,机器人以手中工具朝窗面动作,但单一外视角无法分清擦头是否短暂压到玻璃、窗框或只在前方经过;视频 0.70–6.90 s,擦头处于下沿附近,也无法确认实际接触面。故不将握杆本身重复算作工具—玻璃擦拭接触。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.40–7.55 s,机器人脚部在地面支撑身体,身体逐渐前倾,仍未见倒地。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 650, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "window_wiped", + "t": 0.7, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频约 0.60–0.90 s,机器人已经持杆尝试擦窗,但手臂和擦头下降,擦头到达窗户下缘;视频 1.00–6.90 s 一直没有回到较高的玻璃区域,视频 7.10–7.55 s 又明显垂到玻璃下方。擦拭步骤未在本回合修复。" + }, + "unmet": { + "window_wiped": false + }, + "phenomena": [ + { + "cat": "运动", + "item": "位姿不准", + "t": 0.7, + "log_step": null, + "evidence_zh": "观察:视频 0.70–6.90 s,机器人握住杆,但黑色擦头长期处于玻璃最下缘附近,视频 7.10–7.55 s 进一步降至玻璃下方;未形成窗面内持续上下擦动。视频结尾仍能看到工具跟随手移动,未用这一事件另标脱手或倒地。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志 / 视频", + "evidence_zh": "推断:控杆与姿态控制未能把擦头保持在所需的窗面轨迹。依据是视频 0.70–6.90 s 擦头长期过低且身体逐渐前倾;本次最终回合的控制器输出在 log step 330 记录 tool_head_y=161.5、162.5、168.5、169.5,而 target_y=130.0,在 log step 332 又记录 tool_head_y=192.0、183.0 而目标仍为 130.0(这些是 agent 脚本的图像纵坐标估计,数值越大越靠下)。脚本持续发出控制并以 controller_exit=0 正常结束,却未消除该偏差,支持一般控制失败;不足以进一步确定具体增益、逆解或抓握几何问题。" + } + ], + "gap_zh": "观察:输入没有状态日志,只有单一外部视角,无法确定擦头是否曾短暂接触玻璃或窗框,也无法核实接触力、终止时骨盆或工具是否越过高度阈值。评分器未报告 episode_return,不能给出累计回报数值。log step 329 确认最终重置为 episode 51,log step 332 确认该回合 failed;其控制步数 751 按日志 control_dt=0.02 对应仿真约 15.02 s,而提供视频仅 7.60 s,未获可靠逐帧映射,本文时间均直接取所看视频帧,不将控制步计数换算为视频时间。部分控制脚本修改与输出在输入日志中已经截断,因此不细分控制根因。" + }, + "summary_zh": "最终受评的是第 51 回合:机器人约在视频 0.20 s 用一只灵巧手拿稳擦窗工具,完成 1/2 个节点。擦头约从视频 0.70 s 起降至窗户下缘附近,结尾进一步低于玻璃,擦拭步骤未完成;画面与最终控制器的持续高度偏差支持控制失败。没有状态日志和本回合累计回报数值,短暂玻璃接触与精确终止条件无法核实,日志控制步与视频时长也未建立映射。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "视频帧支持:工具由手拿住,擦头停在窗下缘、未在玻璃上擦动;终止 failed。断点、现象与根因(一般控制)与证据一致,未发现需修改的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:13:18Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard-realtime--4msmchf/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard-realtime--4msmchf/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "63309560b1541b55790d20494d9ede6983a45abb1f839054718ee039a37bb7b1", + "video_sha256": "bb328e5f073f6eb61574509191db8c9f3afd46a430e979ffa380ba597546815c", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard-realtime--5nwof6z.json b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard-realtime--5nwof6z.json new file mode 100644 index 0000000000000000000000000000000000000000..a1e6cb9d04103fb54cb62cd1b8790f10ba9c5143 --- /dev/null +++ b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard-realtime--5nwof6z.json @@ -0,0 +1,239 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-window-i00-standard-realtime--5nwof6z", + "taskKey": "humanoidbench/manip_window", + "dag_sha256": "4dc821ae18297046c83611bc223535d1570899c28d34142481b029cb27e902ac", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 549, + "sim_steps": 7303, + "calls": 44696, + "rows": 45164, + "resets": 26, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 2598.4, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 145.665, + "max_lag_s": 0.0045, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 102, + "late_steps": 0, + "requests_refused": 24, + "request_busy_s": 14.348, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-window-i00-standard-realtime--5nwof6z/replay.mp4", + "duration_s": 5.55, + "fps": 20.0, + "frames": 111, + "sha256": "b7188ae215d8837126ee7acc2bf44bbfd5e295ead18058b89661f7807b06bf0e" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "wiping_tool_in_hand": { + "method": "visual", + "intervals": [ + [ + 0.25, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.25, + "log_step": null, + "text_zh": "观察:视频 0.00–0.80 s 逐帧可见机器人靠镜头一侧的左手合拢在银色工具杆周围,工具随手臂下降、回移后仍留在手中,没有继续自由下落;视频 1.00–5.50 s 杆与该手一起运动,另一只手未共同握杆。拿稳时刻约取视频 0.25 s;擦头朝向仍有转动,不等同于擦头姿态已控制好。" + }, + { + "t": 5.5, + "log_step": null, + "text_zh": "观察:视频 5.00–5.50 s 的逐帧画面中,机器人左手仍夹住杆的近端,黑色擦头未与杆一起脱手掉到地面;本节点是 event,拿稳成果按首次达成锁存。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "抓握", + "note_zh": "观察:机器人左手的手掌和弯曲手指包住银色工具杆并带动工具;右手没有同时握杆。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:机器人双脚接触地面,支撑起初合手拿杆和移动工具的身体。" + } + ] + }, + "window_wiped": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.2, + "log_step": null, + "text_zh": "观察:视频 0.25–0.80 s 擦头降到玻璃下方,视频 0.85–1.00 s 短暂升到玻璃下缘附近,视频 1.10–3.50 s 则主要沿窗下框高度小幅运动并转动;视频 4.00–5.50 s 擦头又明显低于青色玻璃。未见参考成功录像中擦头留在玻璃面内持续运动的几何关系(参考视频 0.70–10.00 s)。" + }, + { + "t": 5.5, + "log_step": 64, + "text_zh": "观察:log step 63 启动最后一回合 episode 27,log step 64 的服务状态为 control_steps=549、done=true、reason=failed,评分器字段 live_success=0、graded_by_live=1;输入没有报告 episode_return 数值。结合视频 0.00–5.50 s 的低位操杆和上述本回合失败结果,擦窗累计回报验收未达成,不能将摆动工具本身记为完成。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察:机器人左手持杆,使黑色擦头在玻璃下缘和下框前运动;视频 0.85–1.00 s 擦头与玻璃下缘在该斜后视角中重叠。缺少接触对或接触力记录,不能确定擦头是否实际压住玻璃或下框,故工具—窗面这一接触关系记未知,不把持杆摆动当作已发生持续压/擦。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:机器人双脚接触地面支撑持杆操作;后段双脚逐渐外滑、两腿张开、骨盆下沉,支撑姿态失稳。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 650, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "window_wiped", + "t": 1.2, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频约 1.20 s,机器人已拿住工具并向窗面操杆,但擦头主要停在玻璃下框而非玻璃面内;这一偏低位置在视频 1.20–5.50 s 未被修复,后段又出现双脚外滑和身体下沉。log step 64 确认最后一回合以 failed 结束,因此断点是正在进行的擦窗步骤。" + }, + "unmet": { + "window_wiped": false + }, + "phenomena": [ + { + "cat": "运动", + "item": "位姿不准", + "t": 1.2, + "log_step": null, + "evidence_zh": "观察:视频 1.10–3.50 s,机器人持杆手在腰部附近,黑色擦头在青色玻璃的下框处摆动并翻转;视频 4.00–5.50 s 擦头多位于玻璃下方,未形成玻璃面内的稳定上下擦动。" + }, + { + "cat": "全身", + "item": "平衡", + "t": 4.0, + "log_step": null, + "evidence_zh": "观察:对比视频 1.00 s 与视频 3.50–5.50 s,机器人两脚向两侧滑开,两腿张角增大,骨盆持续下沉,躯干随之偏斜;视频 4.50–5.50 s 逐帧仍在恶化,未恢复原站姿。画面未显示完整倒地,不据此断言精确终止高度已触发。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "视频 / 日志", + "evidence_zh": "推断:机器人已经握住工具,但视频 1.10–5.50 s 的操杆未把擦头维持在玻璃面内,且支撑腿不断外滑下沉。log step 53 的控制脚本使用肩部正弦指令和基于躯干倾斜的高度补偿,log step 60–63 又调整擦动幅度及髋部姿态反馈并将其用于最后一回合;执行结果仍有上述偏低和失稳,支持控制层失效。现有记录不足以区分工具接触几何、轨迹高度补偿与平衡反馈各自的贡献,因此不指定子类或某一参数错误。" + } + ], + "gap_zh": "无状态日志、接触对、接触力或本回合 episode_return 数值,单一低分辨率斜后视角不能确定擦头在窗下缘短暂重叠时是否有真实压擦;但擦头长期偏低可见,擦窗节点并非因遮挡而未知。log step 64 报告的最后一回合为 549 控制步,按任务文本的仿真每步 0.02 s 对应仿真 10.98 s,而输入视频只有 5.55 s;没有已核对的控制步—视频映射,不将日志中的仿真时刻换成视频时刻,也不把 agent 在 log step 65 自述的后续倒塌当作视频已显示完整倒地。精确终止项及控制子类不能确定。" + }, + "summary_zh": "机器人约在视频 0.25 s 单手拿稳擦窗工具,完成 1/2 个节点;随后虽持杆运动,擦头主要处于玻璃下框或其下方,擦窗节点未通过本回合验收。断点为擦拭窗户的操作中,现象是擦头位姿不准和后段支撑失稳,根因仅能归到控制层。没有接触与累计回报记录,且最后一回合控制时长与视频长度不一致,具体接触方式、终止项和控制子类仍有缺口。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "逐帧核对后,0.25 s 左手拿稳工具、擦头主要停在玻璃下框附近、后段双脚外滑下沉的描述与视频一致;断点、现象与控制层根因证据充分,视频与日志时钟未被混用。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:16:15Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard-realtime--5nwof6z/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard-realtime--5nwof6z/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "24cc4af197ccdb31f13df055047bd7964367af20fcb1df3e7e91b4f07ba7694a", + "video_sha256": "b7188ae215d8837126ee7acc2bf44bbfd5e295ead18058b89661f7807b06bf0e", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard-realtime--jfdypok.json b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard-realtime--jfdypok.json new file mode 100644 index 0000000000000000000000000000000000000000..cae28142f1d23f7f712c8a8dde0013fe7e9783d7 --- /dev/null +++ b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard-realtime--jfdypok.json @@ -0,0 +1,217 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-window-i00-standard-realtime--jfdypok", + "taskKey": "humanoidbench/manip_window", + "dag_sha256": "4dc821ae18297046c83611bc223535d1570899c28d34142481b029cb27e902ac", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 141, + "sim_steps": 21841, + "calls": 88487, + "rows": 90126, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 2261.4, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 436.244, + "max_lag_s": 0.0131, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 1875, + "late_steps": 0, + "requests_refused": 38, + "request_busy_s": 37.134, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.393, + "peak_processes": 10, + "peak_disk_gb": 1.093, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-window-i00-standard-realtime--jfdypok/replay.mp4", + "duration_s": 1.5, + "fps": 20.0, + "frames": 30, + "sha256": "9d6ddc160f66c8c28c72ec4ae416e906a0b6e4c0da40ed8d24bb120de00ec47d" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "wiping_tool_in_hand": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.2, + "log_step": null, + "text_zh": "观察:逐帧检查视频 0.00–0.85 s,机器人抬起并合拢画面左侧的手,银杆先被带起,随后横在手腕/前臂下方转动;视频 0.20–0.85 s 可见合拢的手位于杆上方,未见手指把杆包住。" + }, + { + "t": 1.3, + "log_step": null, + "text_zh": "观察:视频 0.90–1.45 s,银杆逐渐向下转成近竖直,黑色擦头降至窗台以下,末帧工具从手旁滑向画面下缘;机器人仍双脚站立。未出现符合判据的拿稳工具状态。" + }, + { + "t": null, + "log_step": 275, + "text_zh": "观察:log step 273 的控制程序接续 episode 51(control_steps=0),log step 275 的服务状态报告该局 control_steps=141、done=true、reason=failed;log step 256 的 episode 50 已被后续 reset 替换,不是受评局。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "托/舀", + "note_zh": "观察:视频 0.05–0.85 s,画面左侧的手腕/前臂在抬臂合手时短暂托碰银杆,杆绕接触处转动并滑移;未见手指形成包围夹持,不记抓握。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–1.45 s,机器人的双脚接触地面,支撑其伸手接工具,未见翻倒。" + } + ] + }, + "window_wiped": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.8, + "log_step": null, + "text_zh": "观察:视频 0.00–1.45 s,工具在机器人手臂旁翻转、下滑,未见拿稳后使擦头持续贴住青色玻璃上下擦动;视频 1.10–1.45 s 擦头降至窗台下方。成功参考录像中的工具在手前维持于玻璃区域,本局未出现对应的持续工作状态。" + }, + { + "t": null, + "log_step": 275, + "text_zh": "观察:log step 275 的服务报告最后受评 episode 51 以 failed 结束,inputs/run.json 的 reward.live_success=0、reward.success=0;未报告该局 episode_return,不能从失败值反推精确累计回报。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 650, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "wiping_tool_in_hand", + "t": 0.25, + "log_step": 273, + "where": "操作中", + "evidence_zh": "观察:视频约 0.20–0.30 s,机器人已合手抬臂,但银杆位于合拢的手下方,绕手腕/前臂转动而未被手指拿住;视频 1.10–1.45 s 工具继续向下滑落,之后没有重新拿稳。log step 273 启动的是最后受评 episode 51,log step 275 报告该局 failed。" + }, + "unmet": { + "window_wiped": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "没建立接触", + "t": 0.25, + "log_step": null, + "evidence_zh": "观察:视频 0.20–0.85 s,闭合手在银杆上方,杆只与手腕/前臂附近短暂托碰并持续转动;视频 1.10–1.45 s 杆滑向下方,未见手指包围夹住工具。所缺的是拿稳所需的抓握接触,存在托碰不等于已抓住后丢失。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "视频 / 日志", + "evidence_zh": "推断:log step 273 的受评控制配置采用 close_steps=6 并执行手臂高度控制;视频 0.20–0.85 s 实际合手位置与银杆错开,工具沿手腕/前臂翻转滑移,视频 1.10–1.45 s 继续下落。抓取控制没有形成并稳定工具的夹持;现有证据不足以进一步区分手指接触布局、握力或手臂跟踪的贡献。" + } + ], + "gap_zh": "没有状态日志或手指—工具接触对,原始画面仅 256×256,不能量化各指接触力或精确识别每个碰触的链接。log step 275 报告 141 个控制步,任务控制周期为仿真 0.02 s(共仿真 2.82 s),而所给录像只有视频 1.50 s,二者未能建立经核对的时钟映射,本文时间全部直接取视频帧;不将 agent 在 log step 284 自述的“前压推出工具”作为已验证的具体致因。评分器未报告本局 episode_return。" + }, + "summary_zh": "受评的是最后的 episode 51,两个节点均未完成(0/2):机器人合手抬臂时没有把银杆包握住,工具在手腕/前臂旁翻转后下滑。断点为拿稳擦窗工具,现象为没建立所需抓握接触,证据支持抓取控制失败;此前 episode 50 的较长测试不计入本局。没有状态日志或精确回报记录,且日志控制步数与录像时长缺少可靠映射,因此不采纳 agent 对前压致因的具体解释。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "逐帧核对:工具始终未被手指包握,只靠在手腕/前臂上翻转后下滑,两节点未达成、断点、接触与时钟处理均与画面一致;未发现需改内容的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:06:53Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard-realtime--jfdypok/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard-realtime--jfdypok/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "a8d498f98230ec8e5a7e5bca0f8dc5d10aff2cf795a2349da451ea2afc80b890", + "video_sha256": "9d6ddc160f66c8c28c72ec4ae416e906a0b6e4c0da40ed8d24bb120de00ec47d", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard-realtime--kznktxz.json b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard-realtime--kznktxz.json new file mode 100644 index 0000000000000000000000000000000000000000..a973b0daae5f14f28801420b721d125132f0e05c --- /dev/null +++ b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard-realtime--kznktxz.json @@ -0,0 +1,199 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-window-i00-standard-realtime--kznktxz", + "taskKey": "humanoidbench/manip_window", + "dag_sha256": "4dc821ae18297046c83611bc223535d1570899c28d34142481b029cb27e902ac", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 0, + "sim_steps": 0, + "calls": 1, + "rows": 0, + "resets": 0, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 869.3, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 0.0, + "max_lag_s": 0.0, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 0, + "late_steps": 0, + "requests_refused": 1, + "request_busy_s": 0.007, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.172, + "peak_processes": 7, + "peak_disk_gb": 0.249, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-window-i00-standard-realtime--kznktxz/replay.mp4", + "duration_s": 0.1, + "fps": 20.0, + "frames": 2, + "sha256": "3334a4f4004d0dd43978b15977bc8288bf62c166d560ffb299b3e95ef35e91b9" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "wiping_tool_in_hand": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察:视频 0.00 s 与 0.05 s 两帧均为相同初始姿态,银色工具杆横悬于机器人张开的手掌上方,未见手指合拢握杆或工具随手移动。" + }, + { + "t": null, + "log_step": 5, + "text_zh": "观察:log step 5 中机器人服务在 get_obs(seconds=0.0) 后报告 control_steps=0、rows=0、running=False;评分器记录 n_actions=0、sim_steps=0,整段记录没有执行拿取动作。" + } + ], + "contact_actual": [ + { + "who": "不适用", + "how": "不适用", + "note_zh": "观察:评分器记录 n_actions=0、rows=0;log step 38 的过滤器中断前未交付控制轨迹,接触标注不适用。" + } + ] + }, + "window_wiped": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.05, + "log_step": null, + "text_zh": "观察:视频 0.00 s 与末帧 0.05 s 中机器人和工具位置相同,未发生黑色擦头沿玻璃上下擦动的动作。" + }, + { + "t": null, + "log_step": 5, + "text_zh": "推断:log step 5 的机器人服务报告 control_steps=0,评分器最终 sim_steps=0、n_actions=0、live_success=0;结合两帧静止画面,受评回合没有执行擦拭,未达到本节点。" + } + ], + "contact_actual": [ + { + "who": "不适用", + "how": "不适用", + "note_zh": "观察:评分器记录 sim_steps=0、n_actions=0,未交付擦拭控制轨迹;不将初始几何邻近记为工具擦窗接触。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 650, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "unknown", + "t": null, + "log_step": 38, + "where": "首步前", + "evidence_zh": "观察:log step 38 为“This request was blocked by Gemini's filters”过滤器中断提示,之前无 submit 控制调用;评分器 n_actions=0、sim_steps=0、rows=0,视频 0.00 s 与 0.05 s 仅显示同一初态。推断:外部因素在控制轨迹交付及物理步骤启动前中断本次运行,无法定位物理失败步骤。" + }, + "unmet": { + "window_wiped": false + }, + "phenomena": [], + "root_causes": [], + "gap_zh": "观察:log step 38 明确记录 Gemini 过滤器中断,评分器 n_actions=0、sim_steps=0;已有视频仅含视频 0.00 s、0.05 s 两帧初态,没有执行中的物理轨迹,也未提供状态日志或 episode_return 字段。推断:依外部因素中断规则,本次应重测并排除于机器人失败归因,保留目录 success=0;没有可标注的物理失败现象或机器人根因,故两层为空。" + }, + "summary_zh": "本次受评记录只含视频 0.00 s 和 0.05 s 的静止初态,拿稳工具与擦窗均未达成,最终节点为 0/2。log step 38 的 Gemini 过滤器提示结合评分器 n_actions=0、sim_steps=0,表明运行在首步启动前被外部因素中断,应重测并排除于机器人失败归因,目录 success=0 保持原值。没有已执行控制轨迹、状态日志或累计回报记录,无法评价物理失败现象及机器人根因。", + "conf": 9, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "标注符合外部因素规则:log step 38 为 Gemini 过滤器中断,评分器 n_actions=0、sim_steps=0,视频仅两帧初态;断点 ref=unknown、首步前、两层为空且 gap_zh 说明中断,节点均为 false,未见错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:08:29Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard-realtime--kznktxz/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard-realtime--kznktxz/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "f91324f0d19cb7340e108a50cb29b0cbe35534a5337341559bf40dc510b6d790", + "video_sha256": "3334a4f4004d0dd43978b15977bc8288bf62c166d560ffb299b3e95ef35e91b9", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard-realtime--m9jgbw5.json b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard-realtime--m9jgbw5.json new file mode 100644 index 0000000000000000000000000000000000000000..cb56f606c293ce9aabe8a82c7468bc1bf0e222c0 --- /dev/null +++ b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard-realtime--m9jgbw5.json @@ -0,0 +1,235 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-window-i00-standard-realtime--m9jgbw5", + "taskKey": "humanoidbench/manip_window", + "dag_sha256": "4dc821ae18297046c83611bc223535d1570899c28d34142481b029cb27e902ac", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 898, + "sim_steps": 31017, + "calls": 80950, + "rows": 80581, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 2371.2, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 619.573, + "max_lag_s": 0.0151, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 3512, + "late_steps": 0, + "requests_refused": 40, + "request_busy_s": 28.242, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.257, + "peak_processes": 10, + "peak_disk_gb": 0.38, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-window-i00-standard-realtime--m9jgbw5/replay.mp4", + "duration_s": 9.05, + "fps": 20.0, + "frames": 181, + "sha256": "f243ba8349378f4f52d51370e1a72eb840106e6e2baf0cfdacbe7827f703ba58" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "wiping_tool_in_hand": { + "method": "visual", + "intervals": [ + [ + 0.1, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.1, + "log_step": null, + "text_zh": "观察:视频 0.00 s 银杆在双手前;视频 0.10–0.40 s 双手接住杆,工具虽有俯仰转动,仍停留在手上,随后随双臂移动;视频 0.80–2.00 s 杆由闭合手指和另一手承托,不再自由下落。首次拿稳约为视频 0.10 s。" + }, + { + "t": 8.9, + "log_step": null, + "text_zh": "观察:视频 8.00–9.00 s 机器人仍在窗前持杆,视频 8.90–9.00 s 一侧手臂下垂,工具仍在另一手附近。该节点是 event,先前已完成的拿稳事件保持锁存,不将末段单手变化撤销为未完成。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "视频 0.10–2.00 s 双手共同承载银色工具杆,可见手指闭合夹住杆;不把双手靠近本身当作抓握,也不声称两手均完整包握。" + }, + { + "who": "双臂协作", + "how": "托/舀", + "note_zh": "视频 1.00–4.00 s 另一只手在银杆下侧承托,与夹杆的手共同保持工具;抓握与下托为两手的同时分工。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "机器人双脚接触地面,支撑初期接住、持稳工具时的身体。" + } + ] + }, + "window_wiped": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.0, + "log_step": null, + "text_zh": "观察:视频 0.50–4.00 s 机器人持杆小幅摆动,但黑色擦头大多平放或倾斜在窗面下沿附近;视频 6.00–8.75 s 擦头进一步低于青色玻璃下沿,机器人明显前倾。全段未见成功参考视频中那样把擦头保持在玻璃区域持续擦动的形态。" + }, + { + "t": 9.0, + "log_step": 144, + "text_zh": "观察:视频 8.00–9.00 s 擦头仍在窗面下边缘附近,末段一侧手臂落下;log step 144 的机器人实时服务返回 episode=51、control_steps=898、done=True、reason=failed、resets_left=0,与 reward 中 n_actions=898 和 success=0 一致。推断:这就是最后受评回合,擦窗累计回报验收未通过;不把 success=0 当作从未尝试或从未接触的证明。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "机器人双手持杆尝试擦窗,但此单一低分辨率外部视角不能可靠区分黑色擦头与玻璃、下窗框的真实接触;视频 1.00、4.00、6.00、8.00、9.00 s 的局部放大仍不能确认压力接触,故不照抄任务计划中的手持工具×压/擦。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "视频 0.50–8.75 s 机器人双脚仍与地面接触,支撑持杆摆动和前倾操作;未把姿态前倾直接标为已摔倒。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 650, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "window_wiped", + "t": 1.0, + "log_step": 144, + "where": "操作中", + "evidence_zh": "观察:视频约 1.00 s 机器人已持住工具并摆动银杆,黑色擦头却以偏平姿态停留在窗面下沿附近;视频 4.00–8.75 s 擦头逐渐降到玻璃以下,视频 9.00 s 前仍未修复到有效擦拭姿态。log step 144 的服务状态确认最后回合以 failed 结束;失败步骤为持工具擦窗,而非已完成的拿稳工具。" + }, + "unmet": { + "window_wiped": false + }, + "phenomena": [ + { + "cat": "运动", + "item": "位姿不准", + "t": 1.0, + "log_step": null, + "evidence_zh": "观察:视频 1.00、2.00、4.00 s 黑色擦头偏平且贴近玻璃下边缘,视频 6.00、8.00 s 已落到该边缘以下;手臂虽有摆动,未把擦头持续保持在玻璃区域进行有效上下擦动。成功参考视频 1.00–9.50 s 的擦头保持在青色玻璃区域;参考只用于比较形态,不作为本回合接触事实。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "log_step": 144, + "evidence_zh": "推断:log step 144 启动最终控制器,采用固定肩部前推 push=0.6、肘部往复 amp=0.1 及双肩偏航 lyaw=ryaw=0.3;视频 1.00–8.75 s 的实际结果是持杆摆动同时擦头逐渐降到玻璃下沿以下,未能把工具末端维持于擦拭目标姿态。证据支持工具姿态控制未奏效,但不足以区分接触设计、轨迹求解与反馈控制哪个子项主导;agent 在 log step 147 关于弱踝关节和窗框阻挡的解释仍属其自述,不据此确诊。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "输入无状态日志,也无本回合 episode_return 数值,不能恢复精确累计回报或判定具体是哪一项高度阈值终止了回合。低分辨率外部视角能判工具所在区域及拿稳事件,不能可靠确认擦头是否偶尔接触玻璃、是否被窗框阻挡,或末段手指的精确接触变化;因此擦窗接触记未知,不标被挡/碰撞或没建立接触。日志有 50 次重置,早先回合仅作会话背景;本文件只标 log step 144 对应的末回合,所有时间取自视频帧,未将 control_steps 或 agent 时间换算为视频时间。" + }, + "summary_zh": "受评末回合在视频约 0.10 s 拿稳擦窗工具,节点最终完成 1/2;擦窗始终未通过累计回报验收。机器人持杆摆动时擦头长期偏平并停在窗面下沿附近,随后降至玻璃以下,断点为擦拭窗户的操作中,现象为位姿不准,根因保留为控制。无状态日志和累计回报数值,不能确认擦头的真实表面接触及具体终止阈值。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "拿稳工具约0.1 s、擦窗未达成、断点window_wiped操作中、位姿不准、根因停在控制,均与视频帧和log step 144的服务读数一致,未发现需改动的错误。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:06:40Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard-realtime--m9jgbw5/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard-realtime--m9jgbw5/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "c1c887aa78ea11ee2068a3233d81ea8add5b562104651cfb7bdb3ca628bedf0b", + "video_sha256": "f243ba8349378f4f52d51370e1a72eb840106e6e2baf0cfdacbe7827f703ba58", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard-realtime--vvxdpag.json b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard-realtime--vvxdpag.json new file mode 100644 index 0000000000000000000000000000000000000000..c1726cc27912e752f4a5688b19f33b0b430b8c64 --- /dev/null +++ b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard-realtime--vvxdpag.json @@ -0,0 +1,287 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-window-i00-standard-realtime--vvxdpag", + "taskKey": "humanoidbench/manip_window", + "dag_sha256": "4dc821ae18297046c83611bc223535d1570899c28d34142481b029cb27e902ac", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 664, + "sim_steps": 12104, + "calls": 75748, + "rows": 439535, + "resets": 39, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3558.4, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 241.504, + "max_lag_s": 0.0038, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 74, + "late_steps": 0, + "requests_refused": 37, + "request_busy_s": 24.606, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.192, + "peak_processes": 8, + "peak_disk_gb": 0.242, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-window-i00-standard-realtime--vvxdpag/replay.mp4", + "duration_s": 6.7, + "fps": 20.0, + "frames": 134, + "sha256": "49ff721c05b8a9d7b81a9fff4e81477fda470cb96ffdeca7bc6ebdb8fb44ca7c" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "wiping_tool_in_hand": { + "method": "visual", + "intervals": [ + [ + 0.3, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.3, + "log_step": null, + "text_zh": "观察:视频 0.00–0.30 s,银色杆和黑色擦头先向下移动,手指随后合在杆身周围;视频 0.40–0.65 s,机器人屈肘将工具重新向上带起,工具没有继续自由下落,故拿稳工具约在视频 0.3 s 达到。" + }, + { + "t": 3.0, + "log_step": null, + "text_zh": "观察:视频 0.80–4.20 s,工具杆持续留在机器人手部并随手小幅运动;该节点为 event,首次拿稳后锁存,后段工具下转不撤销已经达到的节点。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "抓握", + "note_zh": "观察:视频 0.30–1.50 s,近侧灵巧手握住银色工具杆,工具随手抬起并继续留在手部;远侧手的接触部位受躯干遮挡,是否同时握杆未知。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.80 s,机器人双脚接触地面,支撑闭手接住工具及随后的屈肘抬杆。" + } + ] + }, + "window_wiped": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.55, + "log_step": null, + "text_zh": "观察:视频 0.40–0.80 s,机器人持杆使黑色擦头在青色玻璃前上抬再下移;视频 1.00–4.20 s,擦头大多停留在玻璃下缘附近,仅有小幅摆动,未看到成功参照中持续保持在玻璃区域的工具姿态。" + }, + { + "t": 4.4, + "log_step": null, + "text_zh": "观察:视频 4.25–4.85 s,银色杆从近水平姿态向下转动,黑色擦头降到玻璃下方的灰色墙板区域;视频 5.70–6.65 s,擦头仍在低处并接近地面,未恢复到玻璃上。" + }, + { + "t": null, + "log_step": 113, + "text_zh": "观察:log step 112 的服务输出确认最终重置为 episode 40;log step 113 的服务状态为 episode 40、control_steps=664、done=True、reason=failed,评分字段 live_success=0、replay_success=0。推断:本任务累计回报达到 650 即成功结束,因此受评回合未达到擦窗节点的累计验收;未用总体 reward=0 推断此前没有擦拭。" + } + ], + "contact_actual": [ + { + "who": "手持工具", + "how": "压/擦", + "note_zh": "观察:视频 0.40–0.80 s,机器人手控银色杆,使黑色擦头在青色窗面处上下短幅运动;随后擦头贴近玻璃下缘小幅擦动,视频约 4.3 s 后下转离开玻璃。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.80–6.65 s,机器人双脚持续着地承重,躯干逐渐倾斜,配合手臂操纵工具;没有画面显示机器人倒地。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 650, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "window_wiped", + "t": 4.3, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:机器人已经接住工具并进行短幅擦动;视频 4.25–4.85 s,杆身明显向下转,黑色擦头离开玻璃区域,之后直到视频 6.65 s 仍未恢复,擦窗步骤未完成。log step 113 的服务状态确认最终 episode 40 以 failed 结束。" + }, + "unmet": { + "window_wiped": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "没保持住", + "t": 4.3, + "log_step": null, + "evidence_zh": "观察:视频 0.30–4.20 s,机器人已用手拿住工具并使擦头在玻璃下缘附近短幅运动;视频 4.25–4.85 s,机器人手部仍在胸腹前,银色杆却相对手部由近水平转为陡斜向下,杆后端上翘越过肩部,黑色擦头落到灰色墙板前;视频 5.70–6.65 s,擦头降至近地面,未回到玻璃。推断:先前建立的稳定握持未能保持;远侧手遮挡使是否完全脱手无法确定,但杆身相对手部的滑转已支持没保持住。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "观察:log step 105–106 的 calls[].code 记录 agent 将肘关节擦动改为三角波并叠加上抬斜坡,log step 108 的代码恢复腿部偏置,log step 112 的调用启动该控制脚本用于最终 episode 40;视频 4.25–6.65 s,机器人未能维持工具握持,杆身相对手部下转,擦头离开玻璃且未恢复。推断:持杆和全身协同控制未能维持擦拭所需的工具接触与姿态。已考虑抓取和接触设计:握法或闭手方式不适配可能导致滑转,但远侧手及逐指接触受遮挡,现有证据不能确认接触几何或承载布局有何具体缺陷,也不能证明接触方案可行而仅反馈跟踪失稳;因此按 docs/03_trajectory.md Step 3.3 保留上位根因控制,不强分抓取和接触设计或反馈控制。可行替代握法未验证;log step 114 的踝关节力矩饱和仅为 agent 自述,不作为根因依据。", + "source_zh": "日志 / 视频", + "alt_feasible": "未知" + } + ], + "gap_zh": "未提供状态日志,视频原始分辨率为 256×256,远侧手及逐指接触受躯干遮挡。视频中杆身相对手部滑转、擦头持续下落支持稳定握持未保持,但不能确定是否完全脱手,也不能区分抓取和接触设计不适配、关节跟踪或接触受力问题;可行替代握法未验证。无法验证 agent 在 log step 114 自述的踝关节力矩饱和。评分字段和最终服务输出未报告 episode_return 数值,不能给出精确累计回报或确定是哪项高度阈值终止回合;log step 113 的 664 控制步与服务 control_dt=0.02 对应的控制时长约 13.28 s,而视频只有 6.7 s,未建立映射,所有视频时刻均直接取自帧时间。" + }, + "summary_zh": "机器人约在视频 0.3 s 接住擦窗工具,完成拿稳工具节点,并进行了短幅持工具擦动,最终为 1/2 个节点。视频约 4.3 s 后,工具相对手部滑转、擦头持续下落至近地面且未恢复,现象为没保持住;最终回合未通过累计回报 650 的验收,断点仍为擦拭窗户。根因保守记为控制,已考虑抓取和接触设计及反馈控制,但遮挡和缺少状态日志使具体机制及完全脱手与否无法确定;未将日志控制时长换算成视频时刻。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 3, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "fix", + "conf": 6, + "summary_zh": "节点值(拿稳=true、擦窗=false)与断点 window_wiped 基本成立;但视频 4.25–5.0 s 帧显示杆身后端越过肩部上翘、擦头迅速落向地面并停在低处,是工具从手中滑脱,不是“位姿不准”,应改为 没保持住。", + "n_issues": 2, + "n_major": 1 + }, + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 2, + "verdict": "fix", + "conf": 6, + "summary_zh": "节点值(拿稳=true、擦窗=false)、断点 window_wiped、现象“接触·没保持住”和根因“控制”与视频一致;仅拿稳节点的 contact_actual 把“双臂协作·抓握”写成事实,画面只能看到近侧单手握杆,属于由 agent 日志推断。", + "n_issues": 1, + "n_major": 0 + }, + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 3, + "kind": "confirm", + "verdict": "agree", + "summary_zh": "I1 的接受已在 prev/trajectory.json 中完整应用:who 改为手指/夹爪、how 保持抓握,note 只记近侧手可见握杆、远侧手被遮挡未知,双臂协作推断已删除。", + "n_issues": 0, + "n_checks": 1 + } + ], + "open_zh": [], + "confirm": { + "verdict": "confirmed", + "summary_zh": "I1 的接受已在 prev/trajectory.json 中完整应用:who 改为手指/夹爪、how 保持抓握,note 只记近侧手可见握杆、远侧手被遮挡未知,双臂协作推断已删除。", + "checks_ok": "1/1", + "n_factual_errors": 0 + } + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:23:46Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard-realtime--vvxdpag/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard-realtime--vvxdpag/04_review_r1" + }, + { + "stage": "reconcile", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard-realtime--vvxdpag/05_reconcile_r1" + }, + { + "stage": "traj_review", + "round": 2, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard-realtime--vvxdpag/04_review_r2" + }, + { + "stage": "reconcile", + "round": 2, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard-realtime--vvxdpag/05_reconcile_r2" + }, + { + "stage": "confirm", + "round": 3, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard-realtime--vvxdpag/06_confirm_r3" + } + ], + "confirm_prompt": "prompts/06_confirm.md", + "confirm_prompt_sha256": "b9f56c095f7ea5e609e58bb21e076f565766c370333d71c7a3512cffb4d78267", + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "3564548c7d7e8dabb9f2930ac49ae81cb699ef79764efdabeaf6208a4abd7360", + "video_sha256": "49ff721c05b8a9d7b81a9fff4e81477fda470cb96ffdeca7bc6ebdb8fb44ca7c", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard-realtime--w4hg7ei.json b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard-realtime--w4hg7ei.json new file mode 100644 index 0000000000000000000000000000000000000000..be161d4c28db6d0d52baa11846e8f4c67d8d3f3f --- /dev/null +++ b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard-realtime--w4hg7ei.json @@ -0,0 +1,251 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-window-i00-standard-realtime--w4hg7ei", + "taskKey": "humanoidbench/manip_window", + "dag_sha256": "4dc821ae18297046c83611bc223535d1570899c28d34142481b029cb27e902ac", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "graded_by_live": 1, + "replays": 0, + "live_rule": 1, + "live_verdict_error": 0, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "instance_mismatch": 0, + "n_actions": 607, + "sim_steps": 23558, + "calls": 16807, + "rows": 16759, + "resets": 46, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3490.1, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 470.585, + "max_lag_s": 0.148, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 6940, + "late_steps": 0, + "requests_refused": 47, + "request_busy_s": 33.59, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.235, + "peak_processes": 10, + "peak_disk_gb": 1.116, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-window-i00-standard-realtime--w4hg7ei/replay.mp4", + "duration_s": 6.15, + "fps": 20.0, + "frames": 123, + "sha256": "a6c94905beaaf41c2ef56da679f76e4f414de6b32e92adc4b50d50798277b0fe" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "wiping_tool_in_hand": { + "method": "visual", + "intervals": [ + [ + 0.05, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.05, + "log_step": null, + "text_zh": "观察:逐帧检查视频 0.00–0.55 s,视频 0.00 s 近侧手掌摊开,银色长杆在掌面上方,尚不能辨认该手包围杆;视频 0.05 s 手指已围住杆,视频 0.10–0.55 s 杆保持在手中并改变倾角。视频 0.75–1.50 s 工具继续随双手移动,直到视频 5.50 s 仍被拿在手中。推断:以首个可辨认握杆的帧,即视频 0.05 s 作为拿稳工具的到达时刻;20 fps 画面不能进一步定位首帧间的接触建立时刻,不把工具早期向下转动断言为自由下落。此节点为 event,末段失衡不撤销已达到事件。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "观察:视频 0.05–0.75 s 可见机器人手指围住银色长杆,双手共同控杆;视频 1.50 s 仍见近擦头侧的手与另一侧手共同控杆。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.75 s 机器人双脚接触地面,屈膝支撑拿稳工具时的身体。" + } + ] + }, + "window_wiped": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.5, + "log_step": null, + "text_zh": "观察:视频 1.25–5.50 s 机器人持杆向窗面伸出,黑色擦头靠近青色玻璃下部,随手有小幅升降及转动,未见成功参考视频中持续稳定的持杆姿态;视频 5.75–6.10 s 身体迅速倾斜下坠,工具随手翻转。" + }, + { + "t": 6.1, + "log_step": 174, + "text_zh": "观察:log step 172 启动最后的 trial47;log step 174 中基准服务返回 episode=47、control_steps=607、done=True、reason=failed,其 calls=16807、rows=16759、control_steps_total=23558 与受评分器字段吻合。受评分器 success=0、live_success=0,未报告 episode_return;结合视频中的未完成操作,擦窗累计回报达到 650 的验收没有通过,不把持杆靠窗本身记为整步完成。" + } + ], + "contact_actual": [ + { + "who": "手持工具", + "how": "压/擦", + "note_zh": "推断:视频 1.25–5.50 s 机器人双手控制银色长杆,黑色擦头贴靠青色玻璃下部并小幅移动,按工具对玻璃的压擦接触标注;斜视低分辨率画面不能量化压力或逐帧核实接触点。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 1.25–5.50 s 机器人双脚在地面承重,支撑双手持杆操作;视频 5.75–6.10 s 支撑姿态失稳,身体向侧前方塌落。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 650, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "window_wiped", + "t": 5.75, + "log_step": 174, + "where": "操作中", + "evidence_zh": "观察:视频 5.50 s 机器人仍持工具在窗前操作;视频 5.75–6.10 s 躯干与骨盆快速侧倾下沉,一腿向侧后方伸展,工具随双手倾转,直到末帧未恢复站立。log step 174 的最后回合服务状态为 reason=failed;失败发生在已拿住工具后的擦窗操作中。" + }, + "unmet": { + "window_wiped": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 5.75, + "log_step": null, + "evidence_zh": "观察:视频 5.50–5.75 s 机器人开始明显偏斜;视频 5.95–6.10 s 躯干继续倒向窗侧、骨盆明显降低,双腿不能保持原有支撑姿态,持杆操作被失衡打断,视频结束前没有恢复。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "t": 5.75, + "log_step": 172, + "source_zh": "日志 / 视频", + "evidence_zh": "推断:log step 172 中 agent 的最后回合控制脚本在持杆动作中向双髋俯仰目标加入正弦运动(hipwave=0.22、freq=0.55),并沿用上一回合的站姿和重心控制。视频 1.25–5.50 s 机器人已持工具在窗前,视频 5.75–6.10 s 却未能稳住承载工具时的身体而持续倾倒,支持全身操纵与平衡控制未协调好的归因;没有足够证据把失败单独归于某个增益、逆解或抓握几何。" + } + ], + "gap_zh": "没有状态日志,也未报告最后回合的 episode_return、骨盆高度、工具高度及接触力,不能确定触发终止的精确高度条件。256×256 单一斜视画面中的擦头贴窗接触属于上下文推断,无法核实持续接触时长或压力;拿稳工具与末段失衡可直接辨认。log step 174 的 607 个控制步按任务说明每步仿真 0.02 s 为仿真 12.14 s,而视频仅 6.15 s,未建立可靠时间映射,所有节点时刻与断点只取已检查的视频帧,不把控制步或 agent 时间直接换成视频时间。" + }, + "summary_zh": "最后受评回合在视频 0.05 s 首次可辨认握住擦窗工具,随后持杆靠近玻璃下部并作小幅运动,节点最终达到 1/2,擦窗累计回报验收未通过。视频约 5.75 s 起机器人在擦窗操作中失衡倾倒,结合最后回合的双髋运动控制,归因停留在控制层。缺少状态日志与累计回报明细,擦头接触方式由贴窗画面推断,不能判断精确终止高度或具体控制参数的责任。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 2, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "fix", + "conf": 7, + "summary_zh": "节点、断点(window_wiped,失衡)与根因总体与视频一致;唯一问题是 wiping_tool_in_hand 的证据描述与开头帧不符:视频 0 s 起双手已握杆,并非‘悬在手掌上方后落下’。", + "n_issues": 1, + "n_major": 0 + }, + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 2, + "verdict": "agree", + "conf": 6, + "summary_zh": "节点(拿稳工具达到、擦窗未达到)、断点 window_wiped 操作中、平衡现象与控制根因与视频一致;工具自开头帧即在手中,0.05 s 与 0 s 差异不改变结论。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:21:45Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard-realtime--w4hg7ei/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard-realtime--w4hg7ei/04_review_r1" + }, + { + "stage": "reconcile", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard-realtime--w4hg7ei/05_reconcile_r1" + }, + { + "stage": "traj_review", + "round": 2, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard-realtime--w4hg7ei/04_review_r2" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "96a07e5da480861e7db6122ec2b7e13d743770e89e9456c4b83195608cf9cd40", + "video_sha256": "a6c94905beaaf41c2ef56da679f76e4f414de6b32e92adc4b50d50798277b0fe", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard-realtime--zqoq3ch.json b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard-realtime--zqoq3ch.json new file mode 100644 index 0000000000000000000000000000000000000000..0d2ba0a2624e4726f37379966a55f9ab8e4dbf7f --- /dev/null +++ b/phase2/humanoidbench/manip_window/trajectories/humanoidbench-manip-window-i00-standard-realtime--zqoq3ch.json @@ -0,0 +1,225 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "humanoidbench-manip-window-i00-standard-realtime--zqoq3ch", + "taskKey": "humanoidbench/manip_window", + "dag_sha256": "4dc821ae18297046c83611bc223535d1570899c28d34142481b029cb27e902ac", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "live_success": 0, + "replay_success": 0, + "deterministic": 1, + "finalized": 1, + "missing_log": 0, + "infra_error": 0, + "grader_error": 0, + "sim_error": 0, + "sim_crashed_on_call": 0, + "n_actions": 811, + "sim_steps": 25852, + "calls": 21312, + "rows": 21192, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3450.2, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 516.276, + "max_lag_s": 0.0136, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 6424, + "late_steps": 0, + "requests_refused": 47, + "request_busy_s": 10.3, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 1, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.0, + "peak_processes": 0.0, + "peak_disk_gb": 0.0, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/humanoidbench-manip-window-i00-standard-realtime--zqoq3ch/replay.mp4", + "duration_s": 8.2, + "fps": 20.0, + "frames": 164, + "sha256": "22707dafe9e36b8bbbbbd246b89311a6b82b24b3a4a20bbf753edfd4e98d63c3" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "wiping_tool_in_hand": { + "method": "visual", + "intervals": [ + [ + 0.3, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.3, + "log_step": null, + "text_zh": "观察:视频 0.00–0.20 s 银杆先下沉,视频 0.30–0.75 s 机器人手指合在杆上,银杆与黑色擦头随手向上、向下和转动,不再自由下落;视频 1.20–8.15 s 工具仍留在手边。拿稳时刻约为视频 0.30 s;该节点为 event,后续姿态变低不撤销已达到的拿稳事件。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "观察:视频 0.30–0.75 s 机器人近侧手的手掌和弯曲手指包住银杆,远侧手位于杆的另一段、部分被躯干遮住;推断:双臂共同控杆,工具随两臂动作而移动,后续一直未脱离手边。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.00–0.75 s 机器人双脚着地,屈膝支撑拿杆动作。" + } + ] + }, + "window_wiped": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.7, + "log_step": null, + "text_zh": "观察:视频 0.30–0.80 s 机器人持杆使黑色擦头摆动、翻转;视频 1.20–1.80 s 擦头逐渐降到青色玻璃下沿,视频 2.00–8.15 s 大多处于下沿或下方的灰色墙板前,未见持续沿玻璃上下擦动。成功参考录像的参考视频 0.70–9.50 s 则把擦头保持在青色窗面范围内;参考仅用于辨认几何。" + }, + { + "t": 8.15, + "log_step": 577, + "text_zh": "观察:log step 574 启动最终 Episode 51;log step 577 中机器人服务返回 episode=51、control_steps=811、done=True、reason=failed,评分器 success=0、live_success=0。推断:按 DAG 的累计回报达到 650 才完成的验收规则,本回合擦窗节点未达到;不能把评分器 success=0 当作完全没有擦动或没有触碰的证据。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察:视频 0.30–0.80 s 机器人持工具在窗前摆动,视频 1.20–8.15 s 黑色擦头靠近窗下沿;单一低分辨率外部视角不能确证擦头与玻璃之间是否曾有短暂压擦,故不把典型方案中的手持工具×压/擦直接记为已发生。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察:视频 0.30–8.15 s 机器人双脚支承持杆动作,身体逐步前倾,远侧腿向后伸展。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "episode_return", + "value": null, + "unit": "reward", + "threshold": 650, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "window_wiped", + "t": 1.6, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察:视频 0.30–0.75 s 机器人已拿住并移动工具;视频 1.20–1.80 s 擦头降到窗下沿,视频 2.00–8.15 s 持续停留在玻璃下沿附近或下方,机器人身体继续前倾,未把擦头恢复到持续擦玻璃的运动中。log step 577 的最终 Episode 51 服务状态为 failed;失败发生在持工具擦窗这一步,非拿稳工具之前。" + }, + "phenomena": [ + { + "cat": "运动", + "item": "位姿不准", + "t": 1.6, + "log_step": null, + "evidence_zh": "观察:视频 1.20–1.80 s 机器人把仍在手边的银杆和黑色擦头降到青色玻璃下沿,视频 2.00–8.15 s 擦头大多位于玻璃下方,杆的高度变化很小,没有形成持续的窗面上下擦动;视频 7.20–8.15 s 逐帧检查仍为这一姿态。工具没有明显脱手下坠,机器人也未清楚倒地,所见错误是擦拭位置与运动不到位。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "推断:log step 542 保存并在 log step 574 启动的最终控制器用固定腿部目标和按步数生成的肩肘正弦目标执行擦窗;它读取本体感觉,但重力分量只用于打印,未用于纠正身体倾斜或擦头位置,且 submit 使用 cameras=[]。结合视频 1.20–8.15 s 身体前倾、擦头降到窗下沿后没有恢复,支持控制未维持所需擦拭位姿;没有足够证据进一步区分接触几何、关节跟踪或具体参数问题。", + "source_zh": "日志", + "log_step": 542 + } + ], + "gap_zh": "没有状态日志,也没有本回合 episode_return 数值或擦头接触力、接触对记录;单一 256×256 外部视频不能确定早期是否曾短暂压住玻璃,亦不能判定最终究竟是哪项高度条件触发终止。log step 8 报告 control_dt=0.02,log step 577 的 811 控制步按此为仿真 16.22 s,而视频总长为 8.20 s,缺少已核对的逐步映射,因此全部节点与断点时间直接取视频,不采用 agent 在 log step 578 自述的时长来换算。", + "unmet": { + "window_wiped": false + } + }, + "summary_zh": "受评的是最后启动的 Episode 51:机器人约在视频 0.30 s 拿稳擦窗工具,完成 1/2 个节点,但未达到累计回报验收的擦窗节点。视频约 1.60 s 擦头降至玻璃下沿,此后身体前倾、擦头长期处于窗下沿或下方,未恢复持续擦窗;固定目标控制未纠正这一偏差,服务最终返回 failed。缺少状态日志、精确接触与累计回报数据,控制步时长与视频时长也未建立可靠映射。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "视频全程工具在手、擦头停在玻璃下沿,节点1达成、节点2未达成、断点与位姿不准均与帧一致;时钟全部取视频时间,未换算agent时间。", + "n_issues": 0, + "n_major": 0 + } + ], + "open_zh": [] + }, + "provenance": { + "prompt": "prompts/03_trajectory.md", + "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-30T04:09:05Z", + "review_prompt": "prompts/04_trajectory_review.md", + "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "traj", + "round": 0, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard-realtime--zqoq3ch/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/humanoidbench-manip-window-i00-standard-realtime--zqoq3ch/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "2e6db639b81233b11c4d48c4e6f9d942418a640fb651cd4be72af51cfbe301c5", + "video_sha256": "22707dafe9e36b8bbbbbd246b89311a6b82b24b3a4a20bbf753edfd4e98d63c3", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +}