{ "schema": "efe-task-dag/1", "taxonomy": "v2", "taskKey": "bigym2/move_two_plates", "suite": "bigym2", "base": "bigym2-move-two-plates-i00", "title_zh": "两盘移入右架", "summary_zh": "两只盘子从左架移到右架槽位并按指定朝向竖立、触架不触桌、均离手,两盘共同保持50个控制步,且机器人全程未摔倒、盘子未触地、骨盆未越界。", "embodiment_zh": "三种模式均为Unitree G1双足人形机器人,左右各一条七关节机械臂、各装二指夹爪,双腿和腰部由GR00T全身控制器驱动;机器人可行走、转身、蹲身和前倾,无轮式底盘。", "task_class": "compositional", "analysis_level": "step", "spec": { "version": "1.2", "task": "bigym2-move-two-plates-i00", "bench": "bigym2", "progress": "live", "entities": {}, "milestones": [ { "id": "plates_in_right_rack_1", "zh": "第1盘入右架", "mode": "state", "when": null, "after": [], "hold_s": 0, "freeze_held": false, "goal": true, "criterion_zh": "双外部视角可见,原先空的木色齿槽架(机器人初始右侧)中有至少一只白盘以盘沿向下竖立:每盘中心距该架六个槽位点之一在0.05米以内,盘轴与机器人初始右向夹角在20度以内且不能反面,接触目标架而不接触桌面;悬在架上方、仍在原左架或平躺桌上均不算。右架中不再有任何一只盘子满足上述几何条件即失去本步。最终联合验收还要求两盘同时满足这些条件、两夹爪的手指均不接触任一盘,连续50个0.02秒控制步(任务称1秒);松爪但手指仍贴盘不算离手,共同离手及保持窗口归入整体成功而不另建节点。", "step": { "cat": "关系", "item": "放入" }, "contact": [ { "who": "手指/夹爪", "how": "抓握", "note_zh": "左手或右手的二指夹爪夹住一只盘子的边沿,将盘子移入右架槽位;不限定哪只手或哪只盘子先做。", "plan": 1 }, { "who": "腿/躯干", "how": "支撑", "note_zh": "双腿和足部支撑人形机器人,搬盘时可迈步或转身调整相对盘架的位置。", "plan": 1 }, { "who": "双臂协作", "how": "抓握", "note_zh": "两手的二指夹爪分别夹住同一盘子的不同边沿,共同移入右架;这是单手搬盘的替代方案。", "plan": 2 }, { "who": "腿/躯干", "how": "支撑", "note_zh": "双腿和足部在双手共同搬盘时支撑人形机器人,并可迈步调整位置。", "plan": 2 } ], "optional": false, "verify_zh": null, "sources": [ "inputs/task/privileged/instruction.md:6-9 任务句与两盘开局位于左架", "inputs/task/privileged/instruction.md:15 逐盘槽位距离、轴向、触架与不触桌;两盘同时离手并连续50控制步", "inputs/grader/efe_bigym2-files/privileged/environment/rcb_bigym2.py:147-157 调用上游is_success及episode_fell", "inputs/grader/efe_bigym2-files/sim/server.py:1-20 三种模式共用Episode物理规则", "inputs/oracle/replay.mp4:0.00,16.00,36.48 初态、一个盘入目标架、两个盘入目标架" ], "conf": 6, "note_zh": "两盘同类且目标不固定配对,按当前满足入架几何关系的盘子数计数;计数边不指定搬盘实例顺序,允许同时完成。取出左架、抬高、搬运、调整盘面朝向均并入入架步骤。共同离手和50步保持为整体成功验收,故本步hold_s为0且freeze_held为false,几何入架可在持盘时成立,移出时也应失去。" }, { "id": "plates_in_right_rack_2", "zh": "第2盘入右架", "mode": "state", "when": null, "after": [ "plates_in_right_rack_1" ], "hold_s": 0, "freeze_held": false, "goal": true, "criterion_zh": "双外部视角可见,原先空的木色齿槽架(机器人初始右侧)中有两只白盘以盘沿向下竖立:每盘中心距该架六个槽位点之一在0.05米以内,盘轴与机器人初始右向夹角在20度以内且不能反面,接触目标架而不接触桌面;悬在架上方、仍在原左架或平躺桌上均不算。任一盘移出目标槽位、倒伏、反向或触桌即失去本步;不要求固定的实例或槽位配对。最终联合验收还要求两盘同时满足这些条件、两夹爪的手指均不接触任一盘,连续50个0.02秒控制步(任务称1秒);松爪但手指仍贴盘不算离手,共同离手及保持窗口归入整体成功而不另建节点。", "step": { "cat": "关系", "item": "放入" }, "contact": [ { "who": "手指/夹爪", "how": "抓握", "note_zh": "左手或右手的二指夹爪夹住一只盘子的边沿,将盘子移入右架槽位;不限定哪只手或哪只盘子先做。", "plan": 1 }, { "who": "腿/躯干", "how": "支撑", "note_zh": "双腿和足部支撑人形机器人,搬盘时可迈步或转身调整相对盘架的位置。", "plan": 1 }, { "who": "双臂协作", "how": "抓握", "note_zh": "两手的二指夹爪分别夹住同一盘子的不同边沿,共同移入右架;这是单手搬盘的替代方案。", "plan": 2 }, { "who": "腿/躯干", "how": "支撑", "note_zh": "双腿和足部在双手共同搬盘时支撑人形机器人,并可迈步调整位置。", "plan": 2 } ], "optional": false, "verify_zh": null, "sources": [ "inputs/task/privileged/instruction.md:6-9 任务句与两盘开局位于左架", "inputs/task/privileged/instruction.md:15 逐盘槽位距离、轴向、触架与不触桌;两盘同时离手并连续50控制步", "inputs/grader/efe_bigym2-files/privileged/environment/rcb_bigym2.py:147-157 调用上游is_success及episode_fell", "inputs/grader/efe_bigym2-files/sim/server.py:1-20 三种模式共用Episode物理规则", "inputs/oracle/replay.mp4:0.00,16.00,36.48 初态、一个盘入目标架、两个盘入目标架" ], "conf": 6, "note_zh": "两盘同类且目标不固定配对,按当前满足入架几何关系的盘子数计数;计数边不指定搬盘实例顺序,允许同时完成。取出左架、抬高、搬运、调整盘面朝向均并入入架步骤。共同离手和50步保持为整体成功验收,故本步hold_s为0且freeze_held为false,几何入架可在持盘时成立,移出时也应失去。" } ], "constraints": [ { "id": "plate_touched_floor", "zh": "盘子触地", "never": null, "criterion_zh": "任一白盘与地板接触即失败并结束,不把正常的机器人足部触地算作违规。", "consequence": "立即终止", "sources": [ "inputs/task/privileged/instruction.md:15", "inputs/grader/efe_bigym2-files/privileged/environment/rcb_bigym2.py:14-16" ] }, { "id": "robot_fell", "zh": "机器人摔倒", "never": null, "criterion_zh": "机器人骨盆相对竖直倾斜超过约53度或高度低于0.30米即记为摔倒;之后即使恢复站立并放好盘子也不能成功,摔倒本身不立即结束回合。", "consequence": "历史否决", "sources": [ "inputs/task/privileged/instruction.md:15", "inputs/grader/efe_bigym2-files/privileged/environment/rcb_bigym2.py:14-16", "inputs/grader/efe_bigym2-files/privileged/environment/rcb_bigym2.py:156-157" ] }, { "id": "pelvis_out_of_bounds", "zh": "骨盆越界", "never": null, "criterion_zh": "骨盆距场景原点超过10米触发任务失败并终止;输入未给出该距离计算的上游实现,视频只能在明显离场时判断。", "consequence": "立即终止", "sources": [ "inputs/task/privileged/instruction.md:17-19", "inputs/grader/efe_bigym2-files/privileged/environment/rcb_bigym2.py:14-16" ] } ], "success": null, "atoms": [ { "atom": "plates_in_right_rack_count", "level": "new_code", "source": "需记录task.plates逐盘中心到task.rack_target六槽位点的最近距离、盘轴相对初始右向的夹角、触目标架及触桌标志;据instruction.md:15计算合格盘数。" }, { "atom": "plates_joint_released_hold", "level": "new_code", "source": "需记录上游逐盘is_gripper_holding_object两侧结果以及两盘几何合格且离手的连续控制步计数(50步);可另记录每帧原生成功状态。" }, { "atom": "plate_touched_floor", "level": "new_code", "source": "需记录逐盘与地板的上游碰撞判据。" }, { "atom": "robot_fell", "level": "new_code", "source": "需记录Episode.fell()/episode_fell的历史标志及骨盆高度和倾斜;rcb_bigym2.py:156-157。" }, { "atom": "pelvis_out_of_bounds", "level": "new_code", "source": "需记录上游骨盆离原点的任务失败判据,核对其距离计算方式。" } ] }, "native_metrics": [ { "name": "plate_nearest_target_slot_distance", "unit": "m", "threshold": 0.05, "direction": "lower_better" }, { "name": "plate_axis_angle_to_initial_robot_right", "unit": "deg", "threshold": 20, "direction": "lower_better" }, { "name": "both_plates_qualified_and_released_control_steps", "unit": "control_steps", "threshold": 50, "direction": "higher_better" } ], "open_questions_zh": [ "inputs/states/index.json的logs为空;按DECISIONS.md P8,所有节点when、约束never及整体success均为null,无法用状态回放核对阈值、历史失败和最终成功,需由视频及原生评分结果判断。", "输入评分器通过rcb_bigym2.py:147-157调用外部bigym.loco.eval.is_success与episode_fell;bench_src未含上游move_plates任务检查函数,oracle/run.json中步骤id=9的源码输出也截断了检查主体。因此0.05米、20度、六槽位、接触判据和50步窗口依据三模式instruction.md的明确成功说明,距离与角度边界的严格或非严格比较、摔倒精确角度及越界距离计算方式尚不能从完整源码复核;没有自定阈值。", "按CASES.md OPEN-5及DECISIONS.md D5当前处理,“从左架取出”并入各盘入右架步骤:给出的成功说明只评右架终态,未显示独立取出目标或原生取出阶段;taxonomy中以本任务举例的取出与调整位姿类型不据此增加节点,盘面朝向归入入架验收。", "整体共同离手与50步保持窗口放在success的人工判据中,两个计数节点只表达逐盘入架几何关系;遵循Step 3.4,避免搬下一盘时靠近已放盘导致逐盘进度误降。无状态字段可绑定整体窗口时,必须同时参照两节点criterion_zh的联合验收说明。", "privileged评分器grade.py:103-114要求两次独立重放均成功且终态哈希相同;standard评分器v2_1/grade.py:93-100采用实时回合结果,standard-realtime复用同一适配器。两次重放为同一轨迹的验证,不是两个独立任务场景,不增加节点;交付与重放确定性不绑定为语义世界状态。", "文本“gently”未提供独立力或速度阈值;已用不触地、触架不触桌和联合保持表达给出的验收,不添加柔和程度步骤或臆造冲击力阈值。" ], "conf": 6, "conf_reason_zh": "两盘计数步骤、初态、相同模式物理规则和视频目标外观明确;成功视频支持两步完成,但没有任何状态日志,上游逐盘检查源码缺失,仅能据说明给出精确视频验收,自动绑定及边界实现未核实。", "review": { "status": "agreed", "rounds": 1, "reviewers": [ { "engine": "claude", "model": "claude-sonnet-5-5", "round": 1, "verdict": "agree", "conf": 7, "summary_zh": "两个计数节点(第1盘/第2盘入右架)符合文本与D5当前处理;取出并入放入、共同离手与50步窗口归入success文本、三条约束及后果均合理;无状态日志(P8)故when/never为null,可接受。未发现需改内容的错误。", "n_issues": 0, "n_major": 0 } ], "open_zh": [] }, "evidence_check": { "n_state_runs": 0, "n_faithful": 0, "success_runs_reaching_all": "0/0", "spec_success_eq_grader": "0/0", "diverged_runs": [], "notes_zh": [ "when=null 的节点不在状态日志上检查:['plates_in_right_rack_1', 'plates_in_right_rack_2']" ] }, "provenance": { "prompt": "prompts/01_task_dag.md", "prompt_sha256": "1142cdb56758062bbd0ff585ad99f3fefe07a96377b61614dc4f50b2d16600b7", "engine": "codex", "model": "gpt-6.1-sol", "effort": "high", "created": "2026-10-29T02:00:43Z", "review_prompt": "prompts/02_task_dag_review.md", "review_prompt_sha256": "65812ffc3cb33aaeb10dbe28c925ffe5535a356f8132ec431dcf58b59c2a0b85", "reconcile_prompt": "prompts/05_reconcile.md", "reconcile_prompt_sha256": "b3cd3803cf22f21db996e7ab1145811f823e92c7728c6f21d3b6649cc1ed0e83", "calls": [ { "stage": "dag", "round": 0, "ok": true, "space": "work/tasks/bigym2__move_two_plates/01_dag" }, { "stage": "dag_review", "round": 1, "ok": true, "space": "work/tasks/bigym2__move_two_plates/02_review_r1" } ], "review_engine": "claude", "review_model": "claude-sonnet-5-5", "inputs_manifest_sha256": "844e10d6f5e2d8a3fb803d802e7ffb60a75a0ee86f43fdc98b58e74f0f0eec60", "repo_commit": "4c6d01b9990e3b78ddb620bbd237932b3696f60e" } }