robodojo_subtask / catalog.json
qiaoqy's picture
Complete eight-task RoboDojo subtask dataset: preserve original 400 and append 400 annotated demonstrations
b059b4c verified
Raw History Blame Contribute Delete
64.8 kB
{
"schema_version": 1,
"catalog_version": "20261020.v1",
"status": "semantic_catalog_only_not_frame_annotations",
"created_date_hong_kong": "2026-10-20",
"generator": {
"method": "LLM reasoning in the current Codex session using local task code and inspected demonstration images",
"model_id": "not_recorded_by_session",
"external_api_called": false,
"calvin_images_replayed": false
},
"reference_method": "Tower CALVIN-style observable atomic goals; lift and placement separated; coarse-to-fine boundaries and 5% expansion reserved for later annotation.",
"tasks": [
{
"task_id": "classify_objects",
"name_zh": "物体分类",
"baseline": {
"score": 24.67,
"sr_percent": 12.67
},
"long_instruction": "Sort the objects by category into the three baskets.",
"subtasks": [
{
"id": "lift_sort_object",
"name_zh": "拿起待分类物体",
"instruction_template": "Pick up {object} from the table.",
"parameters": [
"object"
],
"start_evidence": "目标物体仍由桌面支撑。",
"completion_evidence": "目标被夹持且清楚离开桌面;仅闭爪不算完成。",
"optional": false
},
{
"id": "place_sort_object",
"name_zh": "放入对应类别篮筐",
"instruction_template": "Put the held {object} into the {basket} basket for {category}.",
"parameters": [
"object",
"basket",
"category"
],
"start_evidence": "目标已持起,尚未在对应篮筐内释放。",
"completion_evidence": "目标释放并落稳在对应篮筐内,没有放错类别;夹爪不再支撑。",
"optional": false
},
{
"id": "handover_object",
"name_zh": "按需双手交接",
"instruction_template": "Pass the held {object} to the other gripper.",
"parameters": [
"object"
],
"start_evidence": "一手持物,另一手接近接收。",
"completion_evidence": "接收手独立支撑物体,原持握手释放并分离。",
"optional": true
}
],
"sequence": "对实际出现的每个物体重复 lift_sort_object → [handover_object] → place_sort_object;类别、篮筐映射及物体顺序按本条示教绑定。",
"expected_occurrences": "2M + H;配置选3类、每类1–3件,M通常为3–9;H为实际交接次数。不得固定6段或18段。",
"constraints": [
"每条episode独立建立category→basket映射;未归类的新类别参考本条示教/任务元数据,不使用跨episode固定映射。",
"篮筐视觉描述用white/blue/red及视角内位置;cat0等隐藏标签只存审计元数据,不进入视觉模型指令。",
"同类多件也逐件保留拿起/放置;同时操作需保留并行信息,不强制按类别串行。"
]
},
{
"task_id": "fill_pen_holder",
"name_zh": "装满笔筒",
"baseline": {
"score": 23.23,
"sr_percent": 7.33
},
"long_instruction": "Hold the pen holder with one hand, place all pens into it with the other hand, then put it back down.",
"subtasks": [
{
"id": "lift_pen_holder",
"name_zh": "拿起并保持笔筒",
"instruction_template": "Pick up the pen holder and keep it upright.",
"parameters": [],
"start_evidence": "笔筒由桌面支撑。",
"completion_evidence": "持筒手可靠持握并提离桌面,筒口可供插入。",
"optional": false
},
{
"id": "lift_pen",
"name_zh": "拿起一支笔",
"instruction_template": "Pick up {pen} from the table while holding the pen holder steady with the other hand.",
"parameters": [
"pen"
],
"start_evidence": "笔仍由桌面支撑,另一手保持笔筒。",
"completion_evidence": "操作手持笔清楚离开桌面,持筒手仍保持笔筒稳定。",
"optional": false
},
{
"id": "insert_pen",
"name_zh": "将笔插入并释放",
"instruction_template": "Insert the held {pen} into the pen holder and release it while keeping the holder steady.",
"parameters": [
"pen"
],
"start_evidence": "操作手持笔,笔尚未在筒内释放。",
"completion_evidence": "笔尖朝下进入筒内、释放并由筒支撑;持筒手仍保持稳定。",
"optional": false
},
{
"id": "put_down_pen_holder",
"name_zh": "放回装满的笔筒",
"instruction_template": "Put the filled pen holder upright on the table and release it.",
"parameters": [],
"start_evidence": "可见笔已插入,笔筒仍由持筒手保持。",
"completion_evidence": "笔筒直立落稳在桌面,持筒手释放,已装入的笔留在筒内。",
"optional": false
}
],
"sequence": "lift_pen_holder → (lift_pen → insert_pen) × 4 → put_down_pen_holder",
"expected_occurrences": "完整顺序示教通常10段;若双手同时开始拿筒和首支笔,使用联合目标,不编造先后。",
"constraints": [
"配置含2支pen和2支oil_pen,共4件;以可见颜色/形状及实例轨迹区分,不能将target编号当颜色。",
"持筒手可左可右;持续扶筒是每个拿笔/插笔阶段的约束,不重复切出hold阶段。",
"插入前的转笔、对准、运送及释放后退让归入insert_pen。"
]
},
{
"task_id": "organize_table",
"name_zh": "整理桌面",
"baseline": {
"score": 23.33,
"sr_percent": 0.0
},
"long_instruction": "Place the alarm clock on the drawer, put the figurine on the stand, place the mouse on the mouse pad, and push the keyboard into the frame.",
"subtasks": [
{
"id": "lift_alarm",
"name_zh": "拿起闹钟",
"instruction_template": "Pick up the alarm clock from the table.",
"parameters": [],
"start_evidence": "目标仍在原支撑面。",
"completion_evidence": "目标被可靠持握且离开原支撑面。",
"optional": false
},
{
"id": "place_alarm",
"name_zh": "放好闹钟",
"instruction_template": "Place the held alarm clock on top of the drawer and release it.",
"parameters": [],
"start_evidence": "目标已被持起,尚未在指定位置释放。",
"completion_evidence": "目标在指定支撑面稳定释放,保持任务要求的正向/直立姿态。",
"optional": false
},
{
"id": "lift_figurine",
"name_zh": "拿起摆件",
"instruction_template": "Pick up the figurine from the table.",
"parameters": [],
"start_evidence": "目标仍在原支撑面。",
"completion_evidence": "目标被可靠持握且离开原支撑面。",
"optional": false
},
{
"id": "place_figurine",
"name_zh": "放好摆件",
"instruction_template": "Place the held figurine on the stand and release it.",
"parameters": [],
"start_evidence": "目标已被持起,尚未在指定位置释放。",
"completion_evidence": "目标在指定支撑面稳定释放,保持任务要求的正向/直立姿态。",
"optional": false
},
{
"id": "lift_mouse",
"name_zh": "拿起鼠标",
"instruction_template": "Pick up the mouse from the table.",
"parameters": [],
"start_evidence": "目标仍在原支撑面。",
"completion_evidence": "目标被可靠持握且离开原支撑面。",
"optional": false
},
{
"id": "place_mouse",
"name_zh": "放好鼠标",
"instruction_template": "Place the held mouse on the mouse pad and release it.",
"parameters": [],
"start_evidence": "目标已被持起,尚未在指定位置释放。",
"completion_evidence": "目标在指定支撑面稳定释放,保持任务要求的正向/直立姿态。",
"optional": false
},
{
"id": "push_keyboard",
"name_zh": "推齐键盘",
"instruction_template": "Push the keyboard into the frame and align it with the frame.",
"parameters": [],
"start_evidence": "键盘未在框内对齐,仍由桌面支撑。",
"completion_evidence": "键盘位于目标框区域、朝向正确且静止,双手停止推动并离开。",
"optional": false
}
],
"sequence": "(lift_alarm → place_alarm)、(lift_figurine → place_figurine)、(lift_mouse → place_mouse)、push_keyboard;各物体对内部有先后,物体对之间按实际演示排序。",
"expected_occurrences": "标准完整示教7段;键盘属于接触推动,不能人为拆成拿起/放置。",
"constraints": [
"garage为摆件、cube_cushion为底座;不是清理车库。drawer目标是抽屉柜顶部,不是打开抽屉。",
"成功要求包括鼠标在垫内且朝向正确、键盘位置朝向、摆件在底座且直立、闹钟在柜顶且直立。",
"双手协同推键盘用一个语义目标,不把左右手分成两条互斥阶段。"
]
},
{
"task_id": "make_kong",
"name_zh": "Make Kong(麻将杠牌)",
"baseline": {
"score": 26.67,
"sr_percent": 26.67
},
"long_instruction": "Wait for the opponent to discard a tile, then declare a kong with the matching tiles.",
"subtasks": [
{
"id": "wait_for_discard",
"name_zh": "等待对手出牌",
"instruction_template": "Wait for the opponent to discard a tile.",
"parameters": [],
"start_evidence": "对手辅助臂的出牌动作尚未结束。",
"completion_evidence": "弃牌已可辨认且出牌过程结束,可据其牌面确定匹配三张;不能按固定秒数结束。",
"optional": false
},
{
"id": "reveal_matching_tile",
"name_zh": "翻倒一张匹配手牌",
"instruction_template": "Turn the {tile} matching the discarded tile face up on the table.",
"parameters": [
"tile"
],
"start_evidence": "当前匹配牌在己方立牌中,尚未牌面朝上。",
"completion_evidence": "该牌稳定牌面朝上放于桌面,非目标牌保持原排列和姿态。",
"optional": false
},
{
"id": "lift_replacement_tile",
"name_zh": "拿起补牌",
"instruction_template": "Pick up the top replacement tile from the side stack.",
"parameters": [],
"start_evidence": "侧方牌堆顶部补牌仍由牌堆支撑。",
"completion_evidence": "补牌由夹爪持握且清楚离开牌堆。",
"optional": false
},
{
"id": "place_replacement_tile",
"name_zh": "将补牌立入手牌",
"instruction_template": "Place the held replacement tile upright in the designated position in the hand and release it.",
"parameters": [],
"start_evidence": "补牌已持起,尚未在手牌目标处稳定释放。",
"completion_evidence": "补牌在手牌目标位置立稳并释放,已翻倒的三张保持牌面朝上,其余手牌不被扰乱。",
"optional": false
},
{
"id": "handover_object",
"name_zh": "按需双手交接",
"instruction_template": "Pass the held {object} to the other gripper.",
"parameters": [
"object"
],
"start_evidence": "一手持物,另一手接近接收。",
"completion_evidence": "接收手独立支撑物体,原持握手释放并分离。",
"optional": true
}
],
"sequence": "wait_for_discard → reveal_matching_tile × 3 → lift_replacement_tile → [handover_object] → place_replacement_tile",
"expected_occurrences": "基础完整序列6段,加实际交接H段;三张匹配牌按本条示教操作顺序,不绑定固定花色或牌列编号。",
"constraints": [
"这里的Kong是麻将杠牌,不是填充狗玩具;不要写成取四张牌堆叠。",
"弃牌是外部事件/选择依据;演示主要翻倒己方匹配三张,再从侧堆取补牌。不能凭麻将常识添加拾取对手弃牌的步骤。",
"翻倒/揭示牌是接触姿态变化原子任务,不强拆没有独立悬空持握证据的拿起/放置。",
"补牌过程中如有独立接收/释放证据才使用handover_object;稀疏图像中双手靠近不等于交接。",
"外部support_arm0出牌不作为受训左右臂的专家动作;等待阶段监督应使用原始左右臂动作。"
]
},
{
"task_id": "put_bottles_into_dustbin",
"name_zh": "瓶子扔入垃圾桶",
"long_instruction": "Pick up the bottles and throw them into the dustbin, using handover when needed.",
"subtasks": [
{
"id": "lift_bottle",
"name_zh": "拾起瓶子",
"instruction_template": "Pick up the {bottle_description} bottle from the table.",
"parameters": [
"bottle_description"
],
"start_evidence": "The visually identified bottle is supported by the tabletop and is not held clear of it. Reaching, positioning the gripper and closing around the bottle belong to this goal.",
"completion_evidence": "A gripper visibly holds the identified bottle clear of the tabletop, with no remaining table support. Closure or contact alone is insufficient; if clearance is occluded, completion remains uncertain.",
"optional": false
},
{
"id": "handover_bottle",
"name_zh": "双手交接瓶子",
"instruction_template": "Pass the {bottle_description} bottle from the {giver_gripper} gripper to the {receiver_gripper} gripper.",
"parameters": [
"bottle_description",
"giver_gripper",
"receiver_gripper"
],
"start_evidence": "The giver visibly holds the identified bottle clear of the table, and the other gripper is available to receive it. Bringing the bottle and receiver together and establishing a receiving grasp are phases of this goal.",
"completion_evidence": "The receiver independently holds the same bottle after the giver releases and separates. Two grippers touching or enclosing the bottle does not establish a completed handover. An occluded giver release or uncertain independent support must remain uncertain.",
"optional": true
},
{
"id": "deposit_bottle",
"name_zh": "放瓶入桶",
"instruction_template": "Put the held {bottle_description} bottle into the dustbin and release it.",
"parameters": [
"bottle_description"
],
"start_evidence": "The identified bottle is visibly held clear of its previous support, following pickup or a completed handover, and has not yet been released into the bin. Transport to the bin and alignment over its opening belong to this goal.",
"completion_evidence": "The identified bottle is released into the bin and is no longer supported by either gripper. A visible bottle inside the bin together with a released, empty gripper provides strong confirmation. Merely holding it above the opening, its disappearance behind the rim, or a cleared table alone is insufficient. Occluded or failed releases remain uncertain.",
"optional": false
}
],
"sequence": "For each visually bound bottle: lift_bottle → optional handover_bottle → deposit_bottle. Repeat these labels across the four bottles without imposing a universal bottle order. The supplied samples show direct left-gripper disposal throughout episode 0; three direct disposals followed by a right-to-left handover and disposal in episode 49; and two direct disposals followed by two right-to-left handover/disposal chains in episode 99. These are episode-specific patterns, not required global sequences. Across different bottles, impose only visually supported ordering and allow overlap when actually observed; neither concurrency nor strict serialization follows from the task source.",
"expected_occurrences": "Reference pattern for a complete four-bottle run: 4 lift_bottle goals, 4 deposit_bottle goals, and 0–2 optional handover_bottle goals in these demonstrations, yielding 8–10 goal instances. These are reference counts, not annotation quotas: do not force counts, infer missing successes, or turn repeated grasp attempts into completed lifts. Failed, occluded or unfinished goals may reduce the number of confirmed completions.",
"constraints": [
"Bind bottle_description to visible color, shape, markings and, when needed, initial relative location. Preserve that object binding through pickup, any handover and deposit, even when its orientation or location changes.",
"Use descriptions such as green, yellow with a white cap and black label, blue-and-white, white with green/red markings, short white with a screw cap, or cream-to-peach cylindrical only where supported by that episode's images. Do not infer contents, brands or hidden simulator identities.",
"Bind giver_gripper and receiver_gripper to visually verified left/right robot grippers, not image-side position alone.",
"Consult synchronized head, left-wrist and right-wrist views together. A wrist close-up may establish appearance while the head view establishes clearance, ownership or bin-relative position.",
"Keep visible lifting separate from release into the bin. Reaching, grasp closure, alignment and retreat are phases, not additional atomic goals; transport stays within deposit except for motion needed to execute a handover.",
"A real handover requires giver release and independent receiver support. Proximity, dual contact or simultaneous gripping alone is not enough.",
"No bin-lid operation is demonstrated or required by the supplied source; do not add lid-opening or lid-closing goals.",
"The source specifies four bottles and bin containment, with open-gripper scoring and a return-to-origin check. These constraints do not establish visible successes or justify separate gripper-opening or return-home catalog goals.",
"The older catalog supplies reference semantics only. Its object identities, example sequences and frame references must not override these supplied episodes.",
"Sample references are evidence anchors, not exact start or completion boundaries. Preserve uncertainty when support, release, containment or identity cannot be resolved across the available views."
],
"visual_findings": [
{
"episode": "episode_0000000.hdf5",
"evidence": "Frame 0 shows a green bottle, a yellow bottle with a white cap, a slender white bottle initially lying on the table, and an upright cream-to-peach cylindrical bottle; the open black bin is beside the table. The head and left-wrist views show the green bottle held clear toward the bin at frame 31, followed by an empty gripper over the bin at 47 and green material visible inside at 62. The yellow bottle is approached/grasped at 78, held clear at 94, and absent from the holding gripper over the bin at 109; its final resting position is less directly visible. The slender white bottle is approached at 125–156, visibly held clear at 172, and visible inside the bin at 188. The cream-to-peach bottle remains table-supported during the prolonged approach/grasp sequence at 219–298; frame 313 provides clear held-above-table evidence, 329 shows transport over the bin, and 345 shows the gripper empty above it. Its exact final position is obscured by the bin view. Frame 361 shows a cleared table and empty grippers. No handover is visible; all four chains use the left gripper. These samples do not locate exact transitions."
},
{
"episode": "episode_0000049.hdf5",
"evidence": "Frames 0 and 14 distinguish a short white screw-cap bottle, a green bottle, a taller white bottle with green/red markings, and an initially horizontal blue-and-white bottle. The short white bottle is visibly held clear by the left gripper at 43, followed by an empty gripper over the bin at 58, although its settled position is not clearly shown. The green bottle is held clear at 101, with the gripper empty over the bin at 116; containment is partly occluded. The taller marked white bottle is held clear at 160–174 and is visible inside the bin with the gripper empty at 189. The right gripper approaches the blue-and-white bottle at 203–218 and visibly holds it clear at 233–247. The left gripper approaches at 262 and both grippers contact it at 276–291; frame 305 shows the left gripper independently holding it over the bin while the right gripper is open and separated, supporting a completed right-to-left handover. At 320 the left gripper is empty over the bin, and 335 shows a cleared table and empty grippers, but the blue-and-white bottle's final resting position is not clearly visible. Only one handover is supported here, unlike the older reference sequence."
},
{
"episode": "episode_0000099.hdf5",
"evidence": "Frames 0 and 17 show a cream-to-peach cylindrical bottle, a yellow white-capped bottle with a black label, a blue-and-white bottle, and a shorter white screw-cap bottle. The cream-to-peach bottle is grasped at 35 and held clear over the bin at 53; the empty left gripper at 71 supports a release, with the final position partly hidden. The yellow bottle is visibly held clear at 107, remains held at the bin at 125, and is followed by an empty left gripper at 143; frame 125 alone is not release evidence. The right gripper holds the short white bottle clear at 179–197; both grippers converge at 214, and at 232 the left gripper holds it independently over the bin while the right gripper is open and separated. At 250 the left gripper is empty and a pale bottle is visible inside the bin, supporting deposit. The right gripper grasps the blue-and-white bottle at 286 and visibly holds it clear at 304–322. The left gripper approaches at 340 and both contact it at 358; frame 376 shows independent left-gripper holding with the right gripper open and separated, supporting the second handover. At 394 the blue-and-white bottle is visible inside the bin below the empty gripper, and 412 shows a cleared table and empty grippers. Exact lift, handover and release boundaries are not established by these sparse samples."
}
],
"rationale": "The synchronized views support three reusable observable state changes: table-supported to visibly held-clear, giver-held to independently receiver-held, and held to released inside the bin. The demonstrated bottle appearances and differing handover patterns can be expressed through visual parameters without separate color-specific labels or simulator IDs. Keeping these changes distinct avoids treating grasp closure as pickup, dual contact as handover, or transport over the bin as successful disposal. The source clarifies the four-bottle containment objective but does not supply visual evidence or temporal boundaries.",
"catalog_provenance": {
"source_files": [
{
"path": "robodojo/vendor/RoboDojo/task/RoboDojo/tasks/put_bottles_into_dustbin.py",
"sha256": "3ba6e19df0c5ff19ba961092f031ce14e91e72a9a4a5934eece6000a1ffa1c48"
},
{
"path": "robodojo/vendor/RoboDojo/task/RoboDojo/config/put_bottles_into_dustbin.yml",
"sha256": "7eddf43e287b6e8aea686c6d686fb87e13ac1be1cb5aac16a8a2b0d2e9eb976e"
}
],
"request_metadata": "/mnt/dongxu-fs1/data-hdd/qiyuanqiao/robodojo_20261017/annotations/long_horizon_remaining_subtasks_20261026/catalog/put_bottles_into_dustbin/request/metadata.json",
"parsed_response_sha256": "e27397a11c520c6a3f58dad7c6287f8b211225f6ffb6371a53b487e285bb2718",
"sampled_episodes": [
"episode_0000000.hdf5",
"episode_0000049.hdf5",
"episode_0000099.hdf5"
]
}
},
{
"task_id": "play_tic_tac_toe",
"name_zh": "井字棋",
"long_instruction": "Play tic-tac-toe as the first player and fill the board with the opponent.",
"subtasks": [
{
"id": "pick_up_piece",
"name_zh": "拾起棋子",
"instruction_template": "Pick up the {piece_description} with your {arm} arm.",
"parameters": [
"piece_description",
"arm"
],
"start_evidence": "A visually identified loose player piece rests on the table, and the selected controlled arm begins approaching it.",
"completion_evidence": "The selected piece is visibly held by the selected gripper and lifted clear of its original supporting surface. Gripper closure alone, or disappearance behind the hand, does not establish completion.",
"optional": false
},
{
"id": "place_piece_in_cell",
"name_zh": "落子入格",
"instruction_template": "Place the held {piece_description} in the {cell_description} cell of the board using your {arm} arm.",
"parameters": [
"piece_description",
"cell_description",
"arm"
],
"start_evidence": "The selected controlled arm holds the visually tracked piece, and the target board cell is visually identifiable and unoccupied.",
"completion_evidence": "The piece rests within the specified cell on the board without gripper support, with release or subsequent hand separation visible. A piece merely held over or touching the cell is insufficient.",
"optional": false
},
{
"id": "wait_for_opponent",
"name_zh": "等待对手落子",
"instruction_template": "Wait for the external opponent arm to finish placing its next piece on the board.",
"parameters": [],
"start_evidence": "Following a player placement, the controlled robot yields access to the board while an opponent response is pending or the external support arm visibly begins its move.",
"completion_evidence": "A new opponent piece visibly rests on the board without support, and the external arm has disengaged from it. If landing or release remains occluded, completion remains uncertain rather than being inferred from arm withdrawal alone.",
"optional": false
}
],
"sequence": "Repeat pick_up_piece -> place_piece_in_cell for visually bound player pieces. The first player placement precedes the first wait_for_opponent. In the demonstrated turn-taking pattern, each of the first four player placements is followed by an opponent response before the next player placement; the fifth player placement has no required subsequent opponent wait. Piece selection, arm choice and target cells are instance-specific, not a fixed ordering encoded in the labels. Waiting and preparation for the next pickup may overlap if the synchronized views actually show this; represent such intervals jointly rather than forcing the whole wait and pickup into serial segments. Do not infer concurrency from widely separated samples. Approach, grasp closure, transport, alignment and retreat belong inside these goals, not separate catalog entries.",
"expected_occurrences": "Reference for a complete board-filling rollout: five player pickups, five player placements and four external-opponent waits. The source permits up to five player pieces and four opponent responses; partial or uncertain demonstrations may support fewer completed occurrences, and retries may produce additional attempts. Do not force counts, fabricate missing lifts or releases, or promote uncertain completions to success to match the reference.",
"constraints": [
"This task is board filling as the first player, not finding a winning move or demonstrating a winning strategy.",
"Bind piece_description to visible attributes and relations, such as 'gold ring-shaped piece at the left end of the near-side row'. Track that piece through the synchronized views; do not substitute simulator identities or assume identity after an unresolved occlusion.",
"Bind arm to the controlled robot's left or right arm using camera identity and visible continuity. The far-side white support arm is an external opponent, not a third controlled manipulator.",
"Bind cell_description using a fixed head-camera orientation: left and right are head-image left and right; the top row is the far row nearest the external arm, and the bottom row is the near row nearest the controlled robot. Use descriptions such as 'top-left', 'middle-right' or 'center'. Preserve this convention when consulting rotated wrist views.",
"Use the head view to establish board occupancy and spatial relations, and the synchronized wrist views to check piece shape, grasping, local alignment and release. Wrist-image directions must not redefine board directions.",
"Separate visible lifting from placement and release. A later resting piece may support a placement outcome but does not establish the exact earlier lift or release boundary.",
"The visible player pieces are gold rings and the visible opponent pieces are gold crosses. Their roles are supported by the demonstrated manipulation, not merely by private asset categories.",
"Do not encode private cell indices, hidden object IDs, metric simulator tolerances or origin-pose thresholds in language or visual completion criteria.",
"The source's return-to-origin condition explains opponent triggering but does not justify a separate mandatory reset goal. Observable withdrawal is a phase of placement or yielding for the opponent.",
"A waiting goal describes the controlled robot yielding for an external event; it must never be labeled as the controlled robot picking up or placing the opponent's cross.",
"All sampled frame references are evidence anchors, not exact temporal boundaries. Failed, ambiguous or occluded completions must remain uncertain in later annotations."
],
"visual_findings": [
{
"episode": "episode_0000000.hdf5",
"evidence": "Frame 0 shows an empty black board with yellow grid lines, five near-side gold rings, four far-side gold crosses, two controlled side arms and a far-side external arm. Frames 38 and 76 show the left controlled hand at a ring and then carrying/aligning it over the board; frame 115 shows a released ring in the top-middle cell with the controlled arms withdrawn. The external arm is over the board at 153, and frame 191 shows a cross in the middle-right cell. Further left-arm ring manipulations appear at 230/268 and 421/460, with additional resting rings visible at 306 and 498. Right-arm manipulation appears at 613/651 and 805/843; the later head views at 690 and 882 show the corresponding added rings. External-arm board visits at 345, 536 and 766 support repeated opponent waits. Frame 882 shows a filled board: rings at top-middle, top-right, center, bottom-left and bottom-middle, and crosses in the remaining cells. These anchors support repeated pickup, placement and external waiting, but do not resolve every lift/release instant or prove overlap."
},
{
"episode": "episode_0000049.hdf5",
"evidence": "Frame 0 again shows five near-side rings, four far-side crosses and an empty board. The left hand approaches a ring at 37 and is over the board at 75; frame 112 shows a resting ring in bottom-middle. The support arm visits the board at 150, and a cross is visible at middle-left by 187. The next left-arm ring manipulation is visible at 225/262, with rings at bottom-left and bottom-middle by 300. The external arm remains around the board across 337/375, so these sparse views should not be treated as exact opponent-release boundaries or as proof of separate extra turns. A further left-arm placement is visible at 450 and a resting center ring at 487. Right-arm ring manipulation appears at 600/637 and 787/825, followed by resting rings at middle-right by 675 and top-left by 863. External-arm activity at 525/562 and 712/750 supports later waiting goals. Frame 863 shows a filled board, with rings at top-left, center, middle-right, bottom-left and bottom-middle. The sampled views support the goal types without fixing grasp or release boundaries."
},
{
"episode": "episode_0000099.hdf5",
"evidence": "Frame 0 shows the same visible board and separate ring/cross supplies. Frames 37/74 show left-hand ring manipulation, followed by a resting top-left ring at 111. The external arm visibly holds a cross at 148, with a middle-left cross visible at 186. A second left-arm ring manipulation appears at 223/260, and frame 297 shows rings at top-left and top-middle. The external arm holds another cross at 334 and is over the center region at 372; frame 409 shows the added center cross while the right controlled arm approaches a ring. Right-arm ring placements are supported by 409/446 followed by a resting bottom-middle ring at 483, by 595/632 followed by a bottom-left ring at 669, and by 781/818 followed by a bottom-right ring at 856. External-arm activity at 521/558 and 707/744 supports the remaining opponent waits. Frame 856 shows a filled board, with rings at top-left, top-middle and all three bottom cells. The different arm allocation and cell order demonstrate why these must be parameterized reusable goals rather than a fixed trajectory; sparse samples do not establish exact boundaries or concurrency."
}
],
"rationale": "Across the synchronized head and wrist views, loose rings become controlled-hand-held pieces and subsequently remain on specified board cells after hand separation. These are distinct observable outcomes, motivating separate pickup and placement goals. Repeated external-arm cross placements while the controlled robot yields motivate a separate opponent-waiting goal, without attributing those manipulations to the robot. Differences in target cells and arm allocation support parameterized labels with repetitions. The official source clarifies first-player board filling and the five-player/four-opponent reference structure, but visual evidence—not source identities, triggers or tolerances—governs instance binding and completion certainty.",
"catalog_provenance": {
"source_files": [
{
"path": "robodojo/vendor/RoboDojo/task/RoboDojo/tasks/play_tic_tac_toe.py",
"sha256": "8f75ebfe55c997ad9bd3b342db46e843e87397db94356f31cadcbe0cb483bca7"
},
{
"path": "robodojo/vendor/RoboDojo/task/RoboDojo/config/play_tic_tac_toe.yml",
"sha256": "63f863a7fb81a040f65803c64f97c4e9577d4fe32cd0f92563bdc72d1aa336d2"
}
],
"request_metadata": "/mnt/dongxu-fs1/data-hdd/qiyuanqiao/robodojo_20261017/annotations/long_horizon_remaining_subtasks_20261026/catalog/play_tic_tac_toe/request/metadata.json",
"parsed_response_sha256": "91046779868b580252ecd80a61681b7c3b1b418e6c93d19aa9a0b537769ea31f",
"sampled_episodes": [
"episode_0000000.hdf5",
"episode_0000049.hdf5",
"episode_0000099.hdf5"
]
}
},
{
"task_id": "fill_egg_holder",
"name_zh": "装满蛋托",
"long_instruction": "Place the four eggs from the basket into the egg holder, then close the lid.",
"subtasks": [
{
"id": "lift_egg_from_basket",
"name_zh": "从篮中提起蛋",
"instruction_template": "Lift the egg described as {egg_description} out of the basket.",
"parameters": [
"egg_description"
],
"start_evidence": "A visually identified egg is in the basket, and a gripper approaches or contacts that egg to begin picking it up.",
"completion_evidence": "The same egg is visibly held clear of its basket support. Gripper closure alone, or an egg obscured inside the basket, does not establish a completed lift.",
"optional": false
},
{
"id": "place_egg_in_cup",
"name_zh": "将蛋放入蛋杯",
"instruction_template": "Place and release the egg described as {egg_description} into {cup_description} of the egg holder.",
"parameters": [
"egg_description",
"cup_description"
],
"start_evidence": "The egg bound to the preceding lift is visibly held clear of the basket, with the open holder and intended cup visually identifiable.",
"completion_evidence": "The gripper releases the egg, and the egg visibly remains seated in the specified cup independently of the gripper. An egg merely above a cup or hidden by the hand is insufficient; use subsequent unobstructed views when available.",
"optional": false
},
{
"id": "close_holder_lid",
"name_zh": "合上蛋盒盖",
"instruction_template": "Close the lid of {holder_description}.",
"parameters": [
"holder_description"
],
"start_evidence": "The holder lid is open, and a gripper approaches or contacts the lid to rotate it toward the base after the filling actions.",
"completion_evidence": "The articulated lid visibly reaches its closed position over the base and remains closed after the hand moves away. A partially lowered lid is not a completed closure.",
"optional": false
}
],
"sequence": "For each visually bound egg, lift_egg_from_basket(egg_description) precedes place_egg_in_cup(the same egg_description, cup_description). Repeat this pair for the eggs being transferred; the supplied demonstrations show four pairs, serially performed with the right hand in episode 0 and the left hand in episodes 49 and 99. Egg selection order and cup order are not fixed: episode 0 first fills the hinge-side left cup in the head view, whereas episodes 49 and 99 first fill the hinge-side right cup. The filling placements precede close_holder_lid. Do not impose an order between different eggs beyond observed hand/resource dependencies, and do not infer concurrency from sparse samples. No independent lid-opening, holder-holding or handover goal is supported by these demonstrations.",
"expected_occurrences": "Reference pattern per supplied episode: 4 lift_egg_from_basket occurrences, 4 place_egg_in_cup occurrences and 1 close_holder_lid occurrence. These are reference counts, not annotation quotas: do not force counts, fabricate missing completions, or convert an occluded or failed attempt into success. Retries may repeat labels, and only visually supported completions should be marked complete.",
"constraints": [
"Bind egg_description to visible appearance or an explicitly anchored initial basket position, and preserve that same binding across its lift and placement. Similar-looking eggs must not acquire invented identities or fixed simulator IDs.",
"Bind cup_description to a visible cup using an explicit reference convention, such as hinge-side versus front row and left versus right in the head view. Keep the physical cup binding stable across wrist-camera viewpoint changes.",
"A visible lift and a placement/release are separate goals. Reaching, grasp closure, transport, alignment and retreat are phases within these goals, not additional atomic goals.",
"Use synchronized head, left-wrist and right-wrist views together. Visibility in one view can resolve occlusion in another; if all views leave release or support unresolved, completion remains uncertain.",
"Do not infer that an egg is supported by a cup while the hand obscures the contact, or infer the eggs' hidden support state after the lid closes.",
"The holder begins open in all three supplied episodes. Do not add a lid-opening goal merely because opening could be necessary in another initialization.",
"No independent holder stabilization or handover is visibly demonstrated here. If a persistent stabilizing hold is encountered in later demonstrations, treat it as a constraint rather than automatically splitting out a goal.",
"Closing the articulated lid is a contact-driven goal distinct from placing the last egg. Partial downward motion is not sufficient for completion.",
"The official source clarifies intended seating of four eggs and lid closure, but its containment, height, joint-ratio and robot-origin checks are not directly observable RGB measurements and must not be asserted as visual facts.",
"The demonstrated final retreat is not a separate catalog goal. Do not infer exact robot-origin achievement from the final images.",
"Frame references below are sampled evidence of states and transitions, not exact action start or completion boundaries."
],
"visual_findings": [
{
"episode": "episode_0000000.hdf5",
"evidence": "Frame 0 shows an open white four-cup holder to the left of the basket; the right-wrist view at frame 21 shows four eggs in the basket. The right hand approaches an egg at frame 42 and visibly carries it clear of the basket at frame 63. Frames 84 and 105 show transport/alignment above the holder; frame 127 shows the egg with the jaws separated, and frame 148 confirms it remains in the hinge-side left cup while three eggs remain in the basket. Further held-clear eggs appear at frames 190 and 296; frames 232 and 338 show two and three eggs respectively remaining in the holder after those transfers. Frame 359 shows the last basket egg, frame 381 shows it held clear, and frames 402–423 show its approach to the remaining cup. Frame 444 shows four eggs in the holder with the gripper no longer carrying an egg. Frames 444–465 provide evidence of lid-closing interaction and downward lid motion; frame 487 shows the lid closed and the right arm withdrawn. The left arm remains away from the manipulation. Sparse views do not establish exact lift, release or contact boundaries."
},
{
"episode": "episode_0000049.hdf5",
"evidence": "Frame 0 shows the basket on the left and an already open four-cup holder near the center; the left-wrist view at frame 16 exposes four eggs. The left hand encloses an egg at frame 33 and holds it clear of the basket at frame 50, then carries it toward the holder at frames 67–84. Frame 101 shows one egg remaining in the hinge-side right cup after hand separation. A second pickup is supported by frames 118–135, transport by frame 152, and a released egg in the other hinge-side cup by frames 169–186. Frames 202–219 show another pickup and held-clear transport; frame 253 shows three eggs in the holder. Frames 270–287 show the final basket egg and its pickup, frame 304 shows it carried clear, frame 321 shows placement approach, and frame 338 shows four eggs in the holder with the hand separated. Frames 355–372 show the lid-closing interaction and partial lowering; frame 389 shows a closed holder and the left hand withdrawn. The right hand stays away from the task objects. Release instants and exact lid-contact timing remain unresolved between samples."
},
{
"episode": "episode_0000099.hdf5",
"evidence": "Frames 0–16 show a left-side basket containing four eggs and an already open white holder. The left hand approaches the central visible egg at frame 32, holds an egg clear at frame 48 and transports it above the holder at frame 64. Frames 80–97 show the egg near the cup with opening/separation of the jaws; frame 113 confirms one egg remains in the hinge-side right cup. Frames 129–145 show selection and lifting of another egg, frame 161 shows held transport, and frame 177 shows two seated eggs with the jaws separated. Frames 194–210 show the next pickup, frame 226 shows transport toward the holder, and frames 242–258 show three eggs remaining in the holder. Frames 274–291 show the final egg being grasped near the basket, frame 307 shows it carried clear, and frame 323 shows it at the remaining cup with the hand separating. Frame 339 confirms four eggs in the open holder and shows the hand transitioning to lid interaction. Frame 355 shows the lid substantially lowered but not yet unequivocally closed; frame 372 shows the closed holder and withdrawn left hand. The right arm does not independently manipulate the holder or receive an egg. These samples support the reusable goals without establishing exact temporal boundaries."
}
],
"rationale": "Across all three synchronized camera sets, eggs are visibly carried clear of the basket before being lowered and released into individual cups, supporting separate lift and placement goals rather than a single pick-and-place label. Later views of eggs remaining in cups provide stronger placement evidence than hand-occluded contact alone. After the four transfers, the manipulating hand interacts with the articulated lid and a later view shows the holder closed, supporting a distinct lid-contact goal. The changing active hand and first target cup demonstrate why the catalog should reuse parameterized labels rather than prescribe egg identities, cup order or arm identity. The source supplies intended semantics, while the demonstrated images supply the evidence for the catalog.",
"catalog_provenance": {
"source_files": [
{
"path": "robodojo/vendor/RoboDojo/task/RoboDojo/tasks/fill_egg_holder.py",
"sha256": "506dfb772f8882ae53966e3580affde19c3c0f11591c78cc5f5b1d997ba04f7e"
},
{
"path": "robodojo/vendor/RoboDojo/task/RoboDojo/config/fill_egg_holder.yml",
"sha256": "d79129278c9a5488d63bb1d3865f2c6b39f21f85957ec94ace091af9ae29e31f"
}
],
"request_metadata": "/mnt/dongxu-fs1/data-hdd/qiyuanqiao/robodojo_20261017/annotations/long_horizon_remaining_subtasks_20261026/catalog/fill_egg_holder/request/metadata.json",
"parsed_response_sha256": "e1f7ea0c1c60aa2d0962f3beebfcc7a10a2debd39dbd8ac21e7f7ba86f4f8e64",
"sampled_episodes": [
"episode_0000000.hdf5",
"episode_0000049.hdf5",
"episode_0000099.hdf5"
]
}
},
{
"task_id": "play_stacking_toy",
"name_zh": "套柱积木",
"long_instruction": "Place all stacking toy pieces onto the correct pegs.",
"subtasks": [
{
"id": "lift_piece",
"name_zh": "提起积木",
"instruction_template": "Lift the {color} {shape} piece initially at {initial_location} off its support with the {arm} arm.",
"parameters": [
"color",
"shape",
"initial_location",
"arm"
],
"start_evidence": "The visually identified piece rests on the table or another visible support, and the selected gripper begins interacting with it.",
"completion_evidence": "The piece is visibly clear of its previous support and controlled by the selected gripper. Gripper closure, contact, or a wrist close-up without evidence of clearance is insufficient.",
"optional": false
},
{
"id": "place_piece_on_matching_peg",
"name_zh": "套柱稳放",
"instruction_template": "Place the {color} {shape} piece initially at {initial_location} onto {matching_peg} with the {arm} arm, and release it so it remains stably seated.",
"parameters": [
"color",
"shape",
"initial_location",
"matching_peg",
"arm"
],
"start_evidence": "The identified piece is held clear of its previous support, and the arm begins moving or positioning it for the visually identified matching peg.",
"completion_evidence": "The peg passes through the piece's hole, the piece has settled onto the base or the existing stack on that peg, and the gripper no longer supports it. Visible separation or subsequent views must establish that the piece remains seated without gripper support; alignment, peg entry, or gripper opening alone does not establish completion.",
"optional": false
},
{
"id": "handover_piece",
"name_zh": "双臂交接",
"instruction_template": "Pass the {color} {shape} piece initially at {initial_location} from the {giving_arm} arm to the {receiving_arm} arm.",
"parameters": [
"color",
"shape",
"initial_location",
"giving_arm",
"receiving_arm"
],
"start_evidence": "One gripper visibly holds the identified piece and the other approaches that same piece for an actual transfer.",
"completion_evidence": "The receiving gripper visibly controls the piece after the giving gripper releases and separates, without the piece becoming unsupported.",
"optional": true
}
],
"sequence": "For each visually tracked piece, lift_piece precedes place_piece_on_matching_peg. Optional handover_piece may intervene only when a genuine same-piece transfer is visible; placement then uses the receiving arm. Repeat across actual pieces without a universal shape, color, piece or arm order. Preserve simultaneous independent piece chains and continuing instance identities. Ordinary withdrawal, including return toward the initial resting pose after the last release, remains inside the adjacent placement goal; do not create a return-to-rest goal.",
"expected_occurrences": "Reference for a complete unretried run: 10 lift_piece goals and 10 place_piece_on_matching_peg goals, with actual optional handovers added only when visually supported. Four source-defined groups have sizes four, three, two and one; shape/color assignment and manipulation order vary. Concurrent intervals may combine goals, so interval count need not equal 20. Do not force counts or infer successful completions from the source; failed, unfinished or ambiguous attempts remain explicit.",
"constraints": [
"Use the synchronized head, left-wrist, and right-wrist views together. Head views establish scene layout and support relationships; wrist views help identify held pieces, holes, and peg engagement.",
"Bind color, shape, and initial_location to visible attributes of the particular piece. Preserve that binding as the piece moves; initial_location is an episode-relative visual description, not its changing current position.",
"Bind matching_peg using a visible base-relative description and demonstrated matching relationship, such as the peg already carrying pieces of the same visible group. Do not use hidden block or pole identifiers.",
"The source specifies ten pieces in groups of four, three, two, and one. It does not establish a universal manipulation order or a universal color/shape assignment to those group sizes.",
"Separate visible clearance from the support in lift_piece from stable release in place_piece_on_matching_peg. Reaching and grasp closure belong inside the lift; transport, hole alignment, peg entry, sliding down, release, and local withdrawal belong inside placement.",
"Preserve simultaneous independent goals when supported by the synchronized views. Nearby grippers, arm crossings, or different arms handling different pieces are not evidence of a handover.",
"Judge seating relative to the base or existing stack: the piece must be threaded onto the matching peg and remain supported after the gripper ceases supporting it. A partially inserted or still-held piece is not a confirmed placement.",
"Keep failed, ambiguous, or occluded completions uncertain in subsequent annotations. Later clear views may support a stable outcome but do not reveal the exact release instant.",
"Source distance, depth, gripper-opening, and origin thresholds explain benchmark semantics but are not directly measurable from these RGB samples. Do not claim those numerical conditions have been verified visually.",
"Sampled frame references below are evidence anchors, not exact subtask start or end boundaries.",
"Initial reaching and final retreat, including visible return toward the resting pose, stay in the adjacent manipulation goal according to the shared frozen boundary protocol; no independent return-to-rest or gripper-opening label.",
"The source return-to-origin check does not create an observable semantic subtask or establish an exact origin pose from RGB."
],
"visual_findings": [
{
"episode": "episode_0000000.hdf5",
"evidence": "Frame 0 shows a white four-peg base and scattered red circular, yellow square, purple star-shaped, and blue rectangular pieces, with some pieces partly obscured by grippers. Left-wrist views at frames 38 and 77 show a red circular piece controlled by the left gripper and then threaded over a peg; frame 115's head view shows a red piece remaining on the base while the arm handles another. Repeated red-piece engagement is visible at 154 and 231, followed by right-arm red handling at 269–308. Yellow-square handling and peg engagement recur at 346–385, 423–500, and 539–577; the head view at 423 shows the earlier yellow piece remaining seated while the arm moves to another. Purple-star handling and engagement appear at 616–654 and 693–731, with the star stack visible at 770. The left wrist shows the blue rectangle held at 808 and threaded onto the remaining peg at 847. At 886 both arms are back near their initial resting poses and all four visible groups remain on the base. The sampled progression is consistent with the four-red, three-yellow, two-purple, one-blue arrangement, but individual releases and exact boundaries are not all directly sampled."
},
{
"episode": "episode_0000049.hdf5",
"evidence": "Frame 0 shows blue rectangular pieces on both sides of the base, yellow squares, red circles, and a purple star, unlike episode 0's group assignment. Blue-piece handling and peg engagement are visible in the left wrist at 40–80, 121–161, and 202–242, followed by right-wrist blue handling and engagement at 283–364. The existing blue stack remains on the base as a yellow square is handled at 404 and is visibly seated at 445. Further yellow handling and insertion appear at 485–526 and 566–607. Red circles are handled by the left arm at 647–688 and the right arm at 728–809. At 850 the left wrist shows the purple star held; at 890 it is threaded onto the last peg. At 931 the arms have returned to resting poses and the four groups remain on the base. This supports repeated lift/placement chains with a blue-four, yellow-three, red-two, purple-one arrangement, rather than a fixed order inherited from episode 0. Peg engagement frames alone do not prove release; later unsupported persistence supplies stronger outcome evidence."
},
{
"episode": "episode_0000099.hdf5",
"evidence": "Frame 0 shows scattered yellow squares, blue rectangles, red circles, and one visible purple star around the four-peg base. The left wrist shows a yellow square controlled at 38 and threaded onto the left-end peg at 77; another is handled at 116 and engaged at 155. The right arm handles yellow pieces at 194–233 and 311–350; frame 272 includes nearby blue pieces and a yellow piece beside the right gripper, so that sample alone should not be assigned an unambiguous successful lift. Blue rectangular handling and engagement recur at 389–428, 466–505, and 544–583, while the yellow stack persists. Frame 622 shows both arms away from rest in different work regions, not an established same-piece handover. Red circles are handled by the left arm at 661–700 and 739–778. The right wrist shows the purple star held at 817 and threaded onto the final peg at 856. At 895 both arms are near their initial resting poses and the grouped pieces remain on the base. The progression is consistent with yellow-four, blue-three, red-two, purple-one groups. Occlusion and sampling prevent exact release times or unconditional success claims for every intermediate attempt."
}
],
"rationale": "Across all three demonstrations, wrist views distinguish gripper-controlled pieces from later peg engagement, and head views show previously handled pieces persisting on the base while manipulation continues elsewhere. This supports reusable lift and stable-placement goals rather than separate reach, transport, alignment, or gripper-opening labels. Different visible group assignments and arm use show why labels must repeat with visual parameters instead of hidden identities or a fixed ten-step order. Final head views support an observable per-arm return goal. Handover is explicitly optional because these samples do not clearly demonstrate one. The source supplies group structure and intended seating semantics, while the RGB evidence supplies the actual behavior and its uncertainty.",
"catalog_provenance": {
"source_files": [
{
"path": "robodojo/vendor/RoboDojo/task/RoboDojo/tasks/play_stacking_toy.py",
"sha256": "23361329ee7efe330da1f1c241415205bafed2b6c082a913659d868a80d338c3"
},
{
"path": "robodojo/vendor/RoboDojo/task/RoboDojo/config/play_stacking_toy.yml",
"sha256": "a09afc106161064a2ea6a40882d717f7872bd4875309a97431d6863a9b7ffa1c"
}
],
"request_metadata": "/mnt/dongxu-fs1/data-hdd/qiyuanqiao/robodojo_20261017/annotations/long_horizon_remaining_subtasks_20261026/catalog/play_stacking_toy/request/metadata.json",
"parsed_response_sha256": "0287e14aa03c77439b2765c83e0f4165393aa4976fb26172a528142d34f36fb0",
"sampled_episodes": [
"episode_0000000.hdf5",
"episode_0000049.hdf5",
"episode_0000099.hdf5"
],
"protocol_curation": {
"method": "Codex LLM review of generated catalog against shared 20261020 annotation protocol and official source",
"removed_subtask_ids": [
"return_arm_to_rest"
],
"reason": "Final retreat belongs to the last manipulation goal in the existing protocol. Keep lifting, placement and optional actual handover at matching granularity.",
"pre_curation_catalog_sha256": "82010eb15c2d5e52c027ebcdc9ff615a15ab648e2ee93ec9a68bdffa0186f538",
"visual_findings_preserved": true,
"raw_response_preserved": true
}
}
}
],
"sources": [
{
"path": "pi05/tasks/tower_1014_lh_tower_three_blocks/README.md",
"sha256": "2d5ba49de400993a38ac5b205d1cb4720b8e96a0481f3a69c09d3eb6671e8281"
},
{
"path": "pi05/tasks/tower_1014_lh_tower_three_blocks/annotate.py",
"sha256": "9c85842eb781ccc8d8ca55b12e63ca7a8a4d6b2e19a5bbd769f4676be151f11b"
},
{
"path": "pi05/tasks/tower_1014_lh_tower_three_blocks/common.py",
"sha256": "1ed31505987bbb7796a326850f004928a048be2d3119cbbd0c18f0fe62b59c11"
},
{
"path": "robodojo/vendor/RoboDojo/task/RoboDojo/tasks/classify_objects.py",
"sha256": "aef4333c088766375e21a928d53a9a93aaee285fecae41e22385acdc74dd11ac"
},
{
"path": "robodojo/vendor/RoboDojo/task/RoboDojo/config/classify_objects.yml",
"sha256": "80a754d2aedd6a89f58185eae36f5b2eb1a28250f40a67325f576fde34650384"
},
{
"path": "robodojo/vendor/RoboDojo/task/RoboDojo/tasks/fill_pen_holder.py",
"sha256": "8706dce239487e078b21af0b95bcc5cd4eabdcdadcdcbc6400d6b138d982f1e6"
},
{
"path": "robodojo/vendor/RoboDojo/task/RoboDojo/config/fill_pen_holder.yml",
"sha256": "ac72e5582495e8757c547997b317e2fe990370009527b1faa6bfbf4e1275537c"
},
{
"path": "robodojo/vendor/RoboDojo/task/RoboDojo/tasks/organize_table.py",
"sha256": "9e3057e70b8ad0b8cb553be6dec00d2b71fada6f4087edcb38fc56508f572dbd"
},
{
"path": "robodojo/vendor/RoboDojo/task/RoboDojo/config/organize_table.yml",
"sha256": "9ffa15dd279060d4f042285a2c53a341c7208e2f89a2f3dd36ee08b22aa453fa"
},
{
"path": "robodojo/vendor/RoboDojo/task/RoboDojo/tasks/make_kong.py",
"sha256": "b9266689d65f9213b05219ad952344f9fd63be8892097ec0103af99791b245d5"
},
{
"path": "robodojo/vendor/RoboDojo/task/RoboDojo/config/make_kong.yml",
"sha256": "8a33c9e57b6e6c719c4800911c37272ba055ad0f4cda2f1de9d531ee2910351a"
}
],
"visual_evidence": "visual_evidence.json",
"append_batches": [
{
"id": "remaining_subtasks_20261026",
"catalog_sha256": "5e5d6fe25b0522b3ecf8dd1a5dbfc005763ff278368921e38a18f72cfc93e15c",
"original_catalog_sha256": "6eaa01dc772799ab24651a14e07ecd5f849de62a607672aaccfb7efae7fc24aa",
"metadata": {
"schema_version": 1,
"catalog_version": "20261026.v1",
"status": "frozen_semantic_catalog",
"created_date_hong_kong": "2026-10-26",
"generator": {
"method": "Visual LLM catalog reasoning from official task source/config and 3 synchronized RGB demonstrations per task",
"model_id": "gpt-6.1-sol",
"reasoning_effort": "medium",
"external_api_called": true,
"calvin_images_replayed": false,
"protocol_review": "Codex LLM checked all four task catalogs; removed a generated standalone return-to-rest goal for stacking to preserve the existing atomic-goal protocol. Raw response and curation provenance preserved."
},
"reference_method": "Same CALVIN-style observable goals and lift/placement distinction as the 20261020 four-task catalog; same coarse16/refine3/fine1+5% boundary protocol.",
"sources": [
{
"path": "robodojo/vendor/RoboDojo/task/RoboDojo/tasks/put_bottles_into_dustbin.py",
"sha256": "3ba6e19df0c5ff19ba961092f031ce14e91e72a9a4a5934eece6000a1ffa1c48"
},
{
"path": "robodojo/vendor/RoboDojo/task/RoboDojo/config/put_bottles_into_dustbin.yml",
"sha256": "7eddf43e287b6e8aea686c6d686fb87e13ac1be1cb5aac16a8a2b0d2e9eb976e"
},
{
"path": "robodojo/vendor/RoboDojo/task/RoboDojo/tasks/play_tic_tac_toe.py",
"sha256": "8f75ebfe55c997ad9bd3b342db46e843e87397db94356f31cadcbe0cb483bca7"
},
{
"path": "robodojo/vendor/RoboDojo/task/RoboDojo/config/play_tic_tac_toe.yml",
"sha256": "63f863a7fb81a040f65803c64f97c4e9577d4fe32cd0f92563bdc72d1aa336d2"
},
{
"path": "robodojo/vendor/RoboDojo/task/RoboDojo/tasks/fill_egg_holder.py",
"sha256": "506dfb772f8882ae53966e3580affde19c3c0f11591c78cc5f5b1d997ba04f7e"
},
{
"path": "robodojo/vendor/RoboDojo/task/RoboDojo/config/fill_egg_holder.yml",
"sha256": "d79129278c9a5488d63bb1d3865f2c6b39f21f85957ec94ace091af9ae29e31f"
},
{
"path": "robodojo/vendor/RoboDojo/task/RoboDojo/tasks/play_stacking_toy.py",
"sha256": "23361329ee7efe330da1f1c241415205bafed2b6c082a913659d868a80d338c3"
},
{
"path": "robodojo/vendor/RoboDojo/task/RoboDojo/config/play_stacking_toy.yml",
"sha256": "a09afc106161064a2ea6a40882d717f7872bd4875309a97431d6863a9b7ffa1c"
}
],
"visual_evidence": "visual_evidence.json"
},
"provenance_path": "provenance/remaining_subtasks_20261026"
}
]
}