diff --git a/README.md b/README.md index 635fcc8d99c988f9a248ffbcac9bb3bf01f81a88..26405d9f8b27bd59cb46296a7d73946730a62c0f 100644 --- a/README.md +++ b/README.md @@ -385,26 +385,61 @@ reads only the highest `phaseN/` folder of a task (its DAG with its own trajecto - Layout as in phase 1: `phase2///task_dag.json` and `phase2///trajectories/.json`, provenance work-directory paths relative, `dag_sha256` = sha256 of `task_dag.json` as published here. A task is uploaded whole (its DAG together with all its phase-2 trajectories). -- Work in progress: tasks are added as soon as their DAG and all their trajectories pass validation. Status: 50 of 434 tasks, 100 of 868 trajectories. +- Work in progress: tasks are added as soon as their DAG and all their trajectories pass validation. Status: 86 of 434 tasks, 172 of 868 trajectories. ### Phase 2 tasks | folder | trajectories | also in phase 1 | | --- | --- | --- | +| `phase2/behavior-1k/carrying_out_garden_furniture/` | 2 | no | | `phase2/behavior-1k/installing_a_fax_machine/` | 2 | no | | `phase2/behavior-1k/store_batteries/` | 2 | no | | `phase2/behavior-1k/store_produce/` | 2 | no | +| `phase2/dexhand/book_page_turn/` | 2 | no | +| `phase2/dexhand/grinder_crank/` | 2 | no | +| `phase2/dexhand/jar_lid_open/` | 2 | no | +| `phase2/dexhand/pancake_serve/` | 2 | no | +| `phase2/dexhand/remote_battery_swap/` | 2 | no | | `phase2/dexjoco/fold_glasses/` | 2 | no | | `phase2/dextoolbench/hammer_mallet_hammer_swing_down/` | 2 | no | | `phase2/dextoolbench/screwdriver_short_screwdriver_spin_horizontal/` | 2 | no | +| `phase2/embodied-olympics-fun/alpine_slalom/` | 2 | no | | `phase2/embodied-olympics-fun/backflip/` | 2 | no | +| `phase2/embodied-olympics-fun/badminton_serve/` | 2 | yes | | `phase2/embodied-olympics-fun/balance_beam/` | 2 | no | | `phase2/embodied-olympics-fun/balance_board/` | 2 | yes | +| `phase2/embodied-olympics-fun/bowling/` | 2 | no | +| `phase2/embodied-olympics-fun/box_jump/` | 2 | yes | +| `phase2/embodied-olympics-fun/boxing/` | 2 | no | +| `phase2/embodied-olympics-fun/catch/` | 2 | no | +| `phase2/embodied-olympics-fun/catch_and_shoot/` | 2 | no | +| `phase2/embodied-olympics-fun/clean_and_jerk/` | 2 | yes | +| `phase2/embodied-olympics-fun/discus/` | 2 | no | | `phase2/embodied-olympics-fun/dunk/` | 2 | no | +| `phase2/embodied-olympics-fun/dunk_between_legs/` | 2 | yes | +| `phase2/embodied-olympics-fun/free_throw/` | 2 | no | +| `phase2/embodied-olympics-fun/freeski_360/` | 2 | no | +| `phase2/embodied-olympics-fun/golf_putt/` | 2 | no | +| `phase2/embodied-olympics-fun/high_bar_swing/` | 2 | no | +| `phase2/embodied-olympics-fun/high_jump/` | 2 | yes | | `phase2/embodied-olympics-fun/javelin/` | 2 | no | +| `phase2/embodied-olympics-fun/juggling/` | 2 | no | +| `phase2/embodied-olympics-fun/juggling_feet/` | 2 | no | +| `phase2/embodied-olympics-fun/jump_rope/` | 2 | no | | `phase2/embodied-olympics-fun/layup/` | 2 | no | +| `phase2/embodied-olympics-fun/pole_vault/` | 2 | yes | | `phase2/embodied-olympics-fun/shot_put/` | 2 | no | +| `phase2/embodied-olympics-fun/springboard_3m/` | 2 | no | | `phase2/embodied-olympics-fun/sprint/` | 2 | no | +| `phase2/embodied-olympics-fun/steeplechase/` | 2 | yes | +| `phase2/embodied-olympics-fun/table_tennis_return/` | 2 | no | +| `phase2/embodied-olympics-fun/taekwondo_kick/` | 2 | no | +| `phase2/embodied-olympics-fun/tennis_return/` | 2 | no | +| `phase2/embodied-olympics-fun/tennis_serve/` | 2 | no | +| `phase2/embodied-olympics-fun/toes_to_bar/` | 2 | no | +| `phase2/embodied-olympics-fun/triple_jump/` | 2 | yes | +| `phase2/embodied-olympics-fun/volleyball_spike/` | 2 | yes | +| `phase2/embodied-olympics-match/tabletennis_match_i04/` | 2 | no | | `phase2/embodied-safety-exam/a29/` | 2 | no | | `phase2/embodied-safety-exam/c02_i01/` | 2 | no | | `phase2/ispatialgym/mental_rotation/` | 2 | no | @@ -415,6 +450,7 @@ reads only the highest `phaseN/` folder of a task (its DAG with its own trajecto | `phase2/mujoco-playground/go1_joystick_rough_terrain/` | 2 | no | | `phase2/myosuite/dex_hammer_use/` | 2 | no | | `phase2/pandadyn/d05_moving_ramp_relay/` | 2 | no | +| `phase2/roboarc/blanket_flatten_04/` | 2 | no | | `phase2/roboarc/cast_coil_07/` | 2 | no | | `phase2/roboarc/ikea_ikea365/` | 2 | no | | `phase2/roboarc/ikea_lack/` | 2 | no | diff --git a/phase2/behavior-1k/carrying_out_garden_furniture/task_dag.json b/phase2/behavior-1k/carrying_out_garden_furniture/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..e914e66eb0f3c6cfe7d252afd845a92b8241cc3b --- /dev/null +++ b/phase2/behavior-1k/carrying_out_garden_furniture/task_dag.json @@ -0,0 +1,280 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "behavior-1k/carrying_out_garden_furniture", + "suite": "behavior-1k", + "base": "behavior-carrying-out-garden-furniture-i00", + "title_zh": "把花园家具搬出屋", + "summary_zh": "草坪椅放在花园地面、独轮车放在任一草坪上，两处关系同时成立且全部目标物体离手、线速度不超过 0.05 m/s，连续保持 30 个控制步。", + "embodiment_zh": "特权与标准模式均为 R1Pro 轮式移动双臂机器人；左右臂各七个关节，各配二指夹爪并使用辅助抓取，躯干可调，底盘可平移与转向；头部及双腕相机用于观察。", + "task_class": "compositional", + "analysis_level": "step", + "spec": { + "version": "1.2", + "task": "behavior-carrying-out-garden-furniture-i00", + "bench": "behavior-1k", + "progress": "live", + "entities": {}, + "milestones": [ + { + "id": "chair_on_garden_floor", + "zh": "草坪椅放花园地面", + "level": "semantic", + "mode": "state", + "when": null, + "after": [], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "头部或腕部相机中，草坪椅已从客厅移到户外花园地面，椅体接触该地面，且地面在椅体正下方而非上方；仍在客厅、悬空或仅靠近花园地面均未达成，离开该支承面即失去此关系。整项成功还须两处目标关系同时成立，全部目标物体均未附着夹爪且线速度不超过 0.05 m/s，连续保持 30 个控制步（30 Hz，标称 1 秒）。", + "step": { + "cat": "关系", + "item": "放置" + }, + "contact": [ + { + "who": "手指/夹爪", + "how": "抓握", + "plan": 1, + "note_zh": "一侧二指夹爪夹住草坪椅的框架，随底盘搬运到目标支承面。" + }, + { + "who": "底盘", + "how": "支撑", + "plan": 1, + "note_zh": "轮式底盘接触地面并承载机器人，持有草坪椅时从客厅移动到花园。" + }, + { + "who": "双臂协作", + "how": "抓握", + "plan": 2, + "note_zh": "左右二指夹爪分别抓住草坪椅两侧的框架，共同搬运。" + }, + { + "who": "底盘", + "how": "支撑", + "plan": 2, + "note_zh": "轮式底盘接触地面并承载双臂持有的草坪椅，从客厅移动到目标位置。" + }, + { + "who": "手指/夹爪", + "how": "推/扫", + "plan": 3, + "note_zh": "手掌或张开的夹爪接触草坪椅的框架，将其推移到目标支承面。" + }, + { + "who": "底盘", + "how": "支撑", + "plan": 3, + "note_zh": "轮式底盘接触地面，随对草坪椅的推移从客厅移动到花园。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/standard/instruction.md:5", + "inputs/grader/efe_behavior-files/rcb_behavior.py:103-105", + "inputs/grader/efe_behavior-files/rcb_behavior.py:1414-1434", + "inputs/bddl/carrying_out_garden_furniture.bddl:13-18", + "inputs/bddl/carrying_out_garden_furniture.bddl:23", + "inputs/bddl/carrying_out_garden_furniture.literals.json:3" + ], + "conf": 6, + "note_zh": "仅一个放置步骤；抓取、抬起、搬出房屋、经过门口和松手不另拆。节点描述 ontop 关系，关系本身可在持物时成立，故 freeze_held=false；全场离手、静止及共同保持窗口属于 success 的联合验收，因无状态日志只能以文字表达，不把两个节点各自改为全场成功检查。未要求椅子特定朝向或每条椅腿均触地。" + }, + { + "id": "wheelbarrow_on_lawn", + "zh": "独轮车放草坪", + "level": "semantic", + "mode": "state", + "when": null, + "after": [], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "头部或腕部相机中，独轮车已从客厅移到花园的草坪，车体与任一草坪接触且草坪在车体正下方而非上方；参考运行文字记录中的车斗为红色，草坪为绿色，放在露台硬地、悬在草坪上方或仅靠近草坪均未达成，离开草坪支承即失去此关系。整项成功还须两处目标关系同时成立，全部目标物体均未附着夹爪且线速度不超过 0.05 m/s，连续保持 30 个控制步（30 Hz，标称 1 秒）。", + "step": { + "cat": "关系", + "item": "放置" + }, + "contact": [ + { + "who": "手指/夹爪", + "how": "抓握", + "plan": 1, + "note_zh": "一侧二指夹爪夹住独轮车的把手或框架，随底盘搬运到目标支承面。" + }, + { + "who": "底盘", + "how": "支撑", + "plan": 1, + "note_zh": "轮式底盘接触地面并承载机器人，持有独轮车时从客厅移动到花园。" + }, + { + "who": "双臂协作", + "how": "抓握", + "plan": 2, + "note_zh": "左右二指夹爪分别抓住独轮车两侧的把手或框架，共同搬运。" + }, + { + "who": "底盘", + "how": "支撑", + "plan": 2, + "note_zh": "轮式底盘接触地面并承载双臂持有的独轮车，从客厅移动到目标位置。" + }, + { + "who": "手指/夹爪", + "how": "推/扫", + "plan": 3, + "note_zh": "手掌或张开的夹爪接触独轮车的框架，将其推移到目标支承面。" + }, + { + "who": "底盘", + "how": "支撑", + "plan": 3, + "note_zh": "轮式底盘接触地面，随对独轮车的推移从客厅移动到花园。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/standard/instruction.md:5", + "inputs/grader/efe_behavior-files/rcb_behavior.py:103-105", + "inputs/grader/efe_behavior-files/rcb_behavior.py:1414-1434", + "inputs/bddl/carrying_out_garden_furniture.bddl:14-18", + "inputs/bddl/carrying_out_garden_furniture.bddl:24-27", + "inputs/bddl/carrying_out_garden_furniture.literals.json:4" + ], + "conf": 6, + "note_zh": "存在量词允许任一草坪；本实例的目标文字只列 lawn.n.01_1。独轮车与草坪椅不同，各成一步，两步可以任意顺序完成。单爪搬运、双爪抓把手与推移作为替代接触方案，不强制参考运行的左手抓取、沿泳池行走或北侧草坪落点。全局离手、静止和共同保持窗口放入整体成功的文字说明；因此本关系节点不冻结持物帧。" + } + ], + "constraints": [ + { + "id": "robot_fallen", + "zh": "机器人跌倒", + "never": null, + "criterion_zh": "机器人底座低于场景地面超过 0.03 m，或底座向上轴的世界 z 分量小于 0.75（倾斜约超过 41.4°），该析取条件连续满足 300 个控制步（30 Hz，标称 10 秒）后，跌倒监视器锁存并终止；不足窗口的短暂倾倒不触发。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_behavior-files/rcb_behavior.py:92-99", + "inputs/grader/efe_behavior-files/rcb_behavior.py:124-144", + "inputs/grader/efe_behavior-files/rcb_behavior.py:1483-1489" + ] + }, + { + "id": "simulator_breakdown", + "zh": "仿真状态崩溃", + "never": null, + "criterion_zh": "机器人位置、四元数、线速度或关节位置出现非有限值，机器人任一位置坐标绝对值超过 10000 m，或底座线速度范数超过 100 m/s，立即判为崩溃；每 30 个控制步另查全部目标物体，其位置出现非有限值或任一坐标绝对值超过 10000 m 也触发。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_behavior-files/rcb_behavior.py:106-113", + "inputs/grader/efe_behavior-files/rcb_behavior.py:1437-1471" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "chair_on_garden_floor", + "level": "new_code", + "source": "Log flags.lit:(ontop lawn_chair.n.01_1 floor.n.01_2); inputs/bddl/carrying_out_garden_furniture.literals.json:3" + }, + { + "atom": "wheelbarrow_on_any_lawn", + "level": "new_code", + "source": "Log each candidate flags.lit:(ontop barrow.n.03_1 lawn.n.01_1); inputs/bddl/carrying_out_garden_furniture.literals.json:4" + }, + { + "atom": "joint_success_rule", + "level": "new_code", + "source": "Log flags.success_rule from Session.succeeded / RestWatch.done, with release and rest diagnostics; inputs/grader/efe_behavior-files/rcb_behavior.py:1401-1434" + }, + { + "atom": "robot_fallen", + "level": "new_code", + "source": "Log Session._fall.fallen after FallWatch.update; inputs/grader/efe_behavior-files/rcb_behavior.py:1483-1489" + }, + { + "atom": "simulator_breakdown", + "level": "new_code", + "source": "Log Session.breakdown is not None; inputs/grader/efe_behavior-files/rcb_behavior.py:1437-1471" + } + ] + }, + "native_metrics": [ + { + "name": "q_score", + "unit": "比例", + "threshold": 1.0, + "direction": "higher_better" + } + ], + "open_questions_zh": [ + "状态证据缺口：inputs/states/index.json 的 logs 为空，按 P8 所有节点 when、约束 never 和整体 success 均为 null；atoms 列出未来需记录的字段，当前不能核验时间线或 success 与评分器的一致性。q_score=1 仅表示裸目标全部满足，不替代离手、静止和共同保持窗口。", + "参考证据缺口：成功参考运行 pc43j4z 无视频、无状态日志（oracle.json），仅能读取 run.json 的动作与文字输出；其中步骤 168 自述椅子已落在户外露台，步骤 248 释放独轮车到绿色草坪，步骤 251 输出 reason: success。视觉判据据文本和这些记录编写，不能声称已由参考帧核实。", + "门与路线：scene_objects.json 列出客厅至花园的 sliding_door_lrvyvc_0，以及走廊、门厅、车库等至花园的其他门，但不含门初态及无门通路的完整拓扑；参考运行步骤 64-73 打开客厅露台门只证明这一条路线。依 S2 对文本未点名前置步骤可省略的规则，本文件仅保留文本两项放置，不设某扇门或某一路线的强制节点；任务无关门恢复目标。", + "粒度采用 D5 当前读法：搬出房屋的目的由两个最终放置关系覆盖，离开室内不是另受评的目标，因此不再拆出取出或导航步骤。", + "模式差异：标准模式由服务对同一现场 episode 判定；特权模式要求两个新进程回放均成功、终态散列一致且初始实例匹配（privileged/grade.py:168-174）。两者共用放置、离手和静止的物理规则；重放一致性、交付及基础设施有效性不是世界状态步骤，不能用未记录的评分结论绑定语义节点。" + ], + "conf": 6, + "conf_reason_zh": "两个目标及其无序结构由文本和 BDDL 明确支持，初始两个目标均不成立；评分代码给出了共同保持、速度和失败阈值。缺少全部状态日志和参考视频，无法核验实际视觉判据、通路初态及动态绑定，故置信度为 6。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "文本只描述两项放置（草坪椅放花园地面、独轮车放草坪），文件两个目标节点、无序、无多余节点；无状态日志故 when/success 为 null 符合 P8；约束为评分器失败条件（立即终止）。标签与接触方案合理。", + "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：['chair_on_garden_floor', 'wheelbarrow_on_lawn']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:37:21Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/behavior-1k__carrying_out_garden_furniture/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/behavior-1k__carrying_out_garden_furniture/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "b424f28fe8f92beab70840d2d990f646065e6cbed9f2eb00c278e09022ca7e48", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/behavior-1k/carrying_out_garden_furniture/trajectories/behavior-carrying-out-garden-furniture-i00-privileged--rcrje99.json b/phase2/behavior-1k/carrying_out_garden_furniture/trajectories/behavior-carrying-out-garden-furniture-i00-privileged--rcrje99.json new file mode 100644 index 0000000000000000000000000000000000000000..f6dfcbadf740f8c788d9138db45d2535a3c0b997 --- /dev/null +++ b/phase2/behavior-1k/carrying_out_garden_furniture/trajectories/behavior-carrying-out-garden-furniture-i00-privileged--rcrje99.json @@ -0,0 +1,258 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "behavior-carrying-out-garden-furniture-i00-privileged--rcrje99", + "taskKey": "behavior-1k/carrying_out_garden_furniture", + "dag_sha256": "1d06f3b5a0c891414123553d44f860000675991a757c0131c143640e6ba130b3", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "all_success": 1, + "deterministic": 1, + "instance_ok": 1, + "infra_error": 0, + "instance_exact": 0, + "n_actions": 4235, + "sim_steps": 4235, + "fallen": 0, + "final_reward": 0.0, + "q_score": 1.0, + "replay_wall_s": 1875.6, + "replay_timeout_s": 4023.3, + "grader_error": 0, + "video_rendered": 1, + "agent_sessions": 3, + "agent_resets": 4, + "agent_sim_steps": 8470, + "agent_records": 2, + "agent_replays": 1, + "agent_cost_self_reported": 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, + "regrade_0154": { + "tag": "behavior-254-0154", + "old": 1, + "new": 0, + "why": "stricter rule", + "at": "2026-10-18 02:08:19" + } + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/behavior-carrying-out-garden-furniture-i00-privileged--rcrje99/replay.mp4", + "duration_s": 141.167, + "fps": 30.0, + "frames": 4235, + "sha256": "047685dc756b11672ba6a4fa4c2577c963f19afae62bc549b12d784f8f9bde33" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 2, + "final": 2, + "nodes": { + "chair_on_garden_floor": { + "method": "visual", + "intervals": [ + [ + 52.5, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察：视频 0.0 s 草坪椅仍在客厅，尚未位于户外花园地面。" + }, + { + "t": 52.5, + "log_step": null, + "text_zh": "观察：视频 49.0 s 和视频 51.0 s 右腕视角仍可见草坪椅靠背编织框紧贴夹爪；视频 52.0–53.0 s 椅背逐渐离开右腕视野，机器人夹爪撤离；视频 57.0 s 头部视角可见椅子独立留在户外木质露台地面。推断：结合松手撤离及随后椅子留在地面的画面，将该支承关系起点估为视频 52.5 s，误差约 1 s；此前持握时的椅地接触点被遮挡，不能仅凭仍在持握判定关系未成立。" + }, + { + "t": 141.13, + "log_step": null, + "text_zh": "推断：评分器 reward.q_score=1.0 表明两项新增目标关系在终态均成立，与视频中椅子留在户外的观察一致；后续录像未见机器人再次搬动椅子，故该关系保持到视频结束。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "抓握", + "note_zh": "观察：视频 28.0–51.0 s 右夹爪夹住草坪椅靠背框架，椅子随右臂和底盘一起搬出客厅；视频 51.0 s 仍可见椅背紧贴夹爪，视频 52.0–53.0 s 松手撤离，视频 57.0 s 椅子留在露台地面。" + }, + { + "who": "底盘", + "how": "支撑", + "note_zh": "观察：机器人底盘承载持椅的右臂，从客厅经门口行驶到户外露台后放椅（视频 28.0–53.0 s）。" + } + ] + }, + "wheelbarrow_on_lawn": { + "method": "visual", + "intervals": [ + [ + 141.0, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 77.0, + "log_step": null, + "text_zh": "观察：视频 74.0–77.0 s 红色独轮车仍在客厅；视频 78.0–85.0 s 右夹爪合在白色把手处，车体离开原位置，随后随机器人搬出门口。" + }, + { + "t": 141.13, + "log_step": null, + "text_zh": "观察：视频 130.0–139.5 s 机器人仍沿户外硬质步道拖动独轮车，视频 140.0–141.13 s 车斗前部移至绿色草坪上方，右夹爪仍抓住把手，车体继续移动。" + }, + { + "t": 141.0, + "log_step": null, + "text_zh": "推断：结合视频 140.0–141.13 s 车体进入草坪区域和评分器 reward.q_score=1.0（两项新增目标关系均满足），独轮车与草坪的支承关系约在视频 141.0 s 成立，时刻误差约 0.5 s；轮地接触点被车斗和画面边缘遮挡。节点 freeze_held=false，因此持车不否定此关系，但不代表全局离手静置验收通过。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "抓握", + "note_zh": "观察：视频 78.0–85.0 s 右夹爪合拢夹住独轮车一侧白色把手，抬起把手并拖动车体；视频 140.0–141.13 s 仍保持该抓握。" + }, + { + "who": "底盘", + "how": "支撑", + "note_zh": "观察：机器人底盘在右手持独轮车期间从客厅驶出，沿露台和花园步道行驶至草坪边（视频 85.0–141.13 s）。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "q_score", + "value": 1.0, + "unit": "比例", + "threshold": 1.0, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "wheelbarrow_on_lawn", + "t": 141.0, + "log_step": null, + "where": "终态验收", + "evidence_zh": "观察：视频 140.0–141.13 s 独轮车进入草坪区域时，机器人右夹爪仍夹住把手并随底盘继续拖行，直到最后一帧均未见释放或静置；评分器 reward.q_score=1.0，但 success=0，regrade_0154 记载 old=1、new=0、why=\"stricter rule\"。推断：两处目标关系已成立，最后处理的独轮车尚未满足全局离手、静止并共同保持的终态验收。" + }, + "phenomena": [ + { + "cat": "接触", + "item": "没松开", + "t": 141.0, + "log_step": null, + "evidence_zh": "观察：视频 140.0–141.13 s 机器人右夹爪持续夹在独轮车白色把手上，红色车斗相对右臂保持固定而随机器人移动；录像结束前没有张爪撤离或独轮车独立静置的动作。" + } + ], + "root_causes": [], + "gap_zh": "观察：inputs/ 中有完整视频和运行记录，但没有状态日志，也没有参考视频。草坪椅落地和独轮车轮地接触点局部遮挡，草坪椅关系起点约有 1 s 不确定性，独轮车关系起点约有 0.5 s 不确定性；终态关系由视频和 q_score=1.0 合并判断。log step 228 的交付脚本安排在后续导航完成后张开右夹爪并保持，而视频结束时仍在拖车；log step 248 的旧服务回放结果报告 success=true、4235 个控制步，与视频帧数一致，复评字段仅给 stricter rule，未给离手或速度的细分结果。推断：无法据现有证据区分旧服务提前结束、脚本后段未执行及其他执行细节，故不把未释放强归于规划、控制或交付根因。", + "unmet": { + "chair_on_garden_floor": true, + "wheelbarrow_on_lawn": true + } + }, + "summary_zh": "机器人将草坪椅搬到户外露台并松手，随后抓住独轮车把手拖到草坪，两项目标支承关系最终均成立（q_score=1.0）。视频约 141.0 s 的独轮车终态验收未完成：最后一帧仍持车移动，未见离手静置；旧服务报告成功，但更严格的复评将 success 改为 0。没有状态日志，轮地接触时刻需结合评分总值推断，且日志不能解释为何交付录像结束于脚本预定松手之前，因此根因留空。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 2, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "fix", + "conf": 7, + "summary_zh": "终局判断（两节点 q_score=1.0 达成、独轮车终帧仍持握、根因留空）与证据一致；但草坪椅放置/松手时刻标成视频 50.0 s 偏早约 2.5 s：右腕视角 51 s 仍夹着椅背编织框，约 52–53 s 才松开。", + "n_issues": 1, + "n_major": 1 + }, + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 2, + "verdict": "agree", + "conf": 7, + "summary_zh": "两节点（椅子约52.5 s、独轮车约141 s）、终态验收断点、没松开现象、根因留空并附gap_zh均与帧和评分字段一致；第1轮的椅子时刻已修正。", + "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": "2026-10-29T10:07:19Z", + "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/behavior-carrying-out-garden-furniture-i00-privileged--rcrje99/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/behavior-carrying-out-garden-furniture-i00-privileged--rcrje99/04_review_r1" + }, + { + "stage": "reconcile", + "round": 1, + "ok": true, + "space": "work/runs/behavior-carrying-out-garden-furniture-i00-privileged--rcrje99/05_reconcile_r1" + }, + { + "stage": "traj_review", + "round": 2, + "ok": true, + "space": "work/runs/behavior-carrying-out-garden-furniture-i00-privileged--rcrje99/04_review_r2" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "851a385149534227eee010242cbcf13a2bb2e34e396c16a6c4e95a93822a84d8", + "video_sha256": "047685dc756b11672ba6a4fa4c2577c963f19afae62bc549b12d784f8f9bde33", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/behavior-1k/carrying_out_garden_furniture/trajectories/behavior-carrying-out-garden-furniture-i00-privileged--wu84vbd.json b/phase2/behavior-1k/carrying_out_garden_furniture/trajectories/behavior-carrying-out-garden-furniture-i00-privileged--wu84vbd.json new file mode 100644 index 0000000000000000000000000000000000000000..ca4824fbe76af5f1444730725a493cbd2dad1c22 --- /dev/null +++ b/phase2/behavior-1k/carrying_out_garden_furniture/trajectories/behavior-carrying-out-garden-furniture-i00-privileged--wu84vbd.json @@ -0,0 +1,226 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "behavior-carrying-out-garden-furniture-i00-privileged--wu84vbd", + "taskKey": "behavior-1k/carrying_out_garden_furniture", + "dag_sha256": "1d06f3b5a0c891414123553d44f860000675991a757c0131c143640e6ba130b3", + "outcome": { + "success": 1, + "reward": 1.0, + "source": "catalog", + "raw": { + "success": 1, + "all_success": 1, + "deterministic": 1, + "instance_ok": 1, + "infra_error": 0, + "instance_exact": 0, + "n_actions": 6460, + "sim_steps": 6377, + "fallen": 0, + "final_reward": 1.0, + "q_score": 1.0, + "replay_wall_s": 5078.0, + "replay_timeout_s": 4430.0, + "grader_error": 0, + "video_rendered": 1, + "agent_sessions": 2, + "agent_resets": 3, + "agent_sim_steps": 6250, + "agent_records": 0, + "agent_replays": 0, + "agent_cost_self_reported": 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/behavior-carrying-out-garden-furniture-i00-privileged--wu84vbd/replay.mp4", + "duration_s": 212.567, + "fps": 30.0, + "frames": 6377, + "sha256": "79a379daf9ad9592b0e6e8ab454cb7c43c758344b9299ac57e4ed43e866edce4" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 2, + "final": 2, + "nodes": { + "chair_on_garden_floor": { + "method": "visual", + "intervals": [ + [ + 71.5, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0, + "log_step": null, + "text_zh": "观察：视频 0 s，带白色坐垫的藤编草坪椅在客厅木地板上，尚未放到花园地面。" + }, + { + "t": 25, + "log_step": null, + "text_zh": "观察：视频 19–27 s，机器人右夹爪靠近并夹住椅框，右腕视角中的椅框随手臂抬起；视频 48–65 s，椅子随机器人穿过门口来到泳池旁的户外木露台。" + }, + { + "t": 71.5, + "log_step": null, + "text_zh": "观察：视频 70–73 s，机器人右夹爪松开椅框，椅子向下落放；视频 74–76 s，机器人转身后可见白色坐垫和椅腿留在露台上，椅子未随夹爪离开。" + }, + { + "t": 71.5, + "log_step": 139, + "text_zh": "推断：取视频约 71.5 s 的落放过程作为花园地面支承关系的达成锚点，误差约 0.5 s；搬运时椅脚受遮挡，不能排除持物期间已较早接触该地面。log step 139 的原始连续录制输出为 chair ontop True、2189 行；它与此处可见落放一致，但未用 agent 工作时间换算视频时间。" + }, + { + "t": 212.53, + "log_step": null, + "text_zh": "推断：落放后未见机器人重新提起椅子；评分器 reward.success=1、all_success=1、q_score=1.0，确认终态包含草坪椅在花园地面的目标关系。终帧不再朝向椅子，终态判断由可见落放和评分字段共同支持。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "抓握", + "note_zh": "机器人右侧二指夹爪夹持草坪椅框架，将椅子带出客厅后松手落放；右腕视角可见椅框随夹爪移动。" + }, + { + "who": "底盘", + "how": "支撑", + "note_zh": "轮式底盘承载机器人，在持有草坪椅时从客厅穿过滑门驶到户外露台。" + } + ] + }, + "wheelbarrow_on_lawn": { + "method": "visual", + "intervals": [ + [ + 212.5, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 90, + "log_step": null, + "text_zh": "观察：视频 90–92 s，红色车斗的独轮车仍在客厅，机器人向它靠近；视频 92–98 s，右夹爪贴近车斗后缘并夹持，车斗随后相对室内地板移动并保持在右腕相机旁。" + }, + { + "t": 132.5, + "log_step": null, + "text_zh": "观察：视频约 123–133 s，机器人拖着红色独轮车从客厅通过门口来到露台；视频 138–203 s，机器人继续带着它沿户外露台移动，红色车斗一直可在右腕视角辨认。" + }, + { + "t": 212.5, + "log_step": null, + "text_zh": "观察：视频 204–209 s，独轮车随机器人沿木露台的草坪边缘移动；视频 210–212.53 s，机器人向草坪方向移动，灰色车把、红色车斗与草地同框，末段右夹爪张开，车斗相对夹爪下移。车轮和支腿的具体落地点被车斗及视野边缘遮挡。" + }, + { + "t": 212.5, + "log_step": null, + "text_zh": "推断：结合视频末段车斗到达草坪并在夹爪旁下移，以及评分器 success=1、all_success=1、q_score=1.0，将独轮车放草坪记为最终成立。区间起点取视频约 212.5 s 的末段落放锚点；精确的首次支承接触不可见，可能早于该锚点约 1–2 s，不能由画面独立确认。" + }, + { + "t": 212.53, + "log_step": 167, + "text_zh": "观察：log step 164、166–167 显示交付文件由原始连续录制的前 6260 行加向草坪平移、开右夹爪的 200 行组成；视频末段与此平移和开爪方向一致。log step 151 的跌倒以及 log step 163 的重置回放偏离属于其他会话测试，评分视频末段没有相同跌倒；未把这些测试状态移入本节点。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "抓握", + "note_zh": "机器人右侧二指夹爪夹住红色独轮车车斗后缘，随底盘移动拖运，末段张开右夹爪；左手未共同夹持车体。" + }, + { + "who": "底盘", + "how": "支撑", + "note_zh": "轮式底盘承载机器人并在右手牵引独轮车时穿门、沿露台绕泳池移动到草坪边缘，最后驶向草坪。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "q_score", + "value": 1.0, + "unit": "比例", + "threshold": 1.0, + "direction": "higher_better" + } + ], + "failure": null, + "summary_zh": "评分回放完成两项目标：草坪椅在视频约 71.5 s 落放于户外露台，独轮车在视频约 212.5 s 到达草坪并落放，评分器 success=1、q_score=1.0，终态为 2/2。log step 168 的悲观自述及会话中的跌倒、重置回放偏离不代表这次评分回放失败，故没有失败断点或原因标签。无状态日志且参考运行无视频，家具底部接触点受遮挡，首次达成时刻只能近似定位；DAG 所述离手、速度阈值和连续 30 控制步窗口不能由末段画面独立核验，按本次实际评分规则保留成功判定。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "评分回放 success=1、q_score=1.0，两个目标节点均为 true，无失败断点，与帧一致；椅子约 71.5 s 落放、独轮车 212.5 s 在草坪（207–209 s 腕视角下仍为木露台，212 s 为草地），时间与接触标注合理。", + "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": "2026-10-29T09:51:23Z", + "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/behavior-carrying-out-garden-furniture-i00-privileged--wu84vbd/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/behavior-carrying-out-garden-furniture-i00-privileged--wu84vbd/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "8705631a189b0c0d3396b092018686ccc93276b467a87da3a363a4ab623aab4f", + "video_sha256": "79a379daf9ad9592b0e6e8ab454cb7c43c758344b9299ac57e4ed43e866edce4", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/dexhand/book_page_turn/task_dag.json b/phase2/dexhand/book_page_turn/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..1339e50e9ea5a9a77828ef3eaa1689d15f9fce8d --- /dev/null +++ b/phase2/dexhand/book_page_turn/task_dag.json @@ -0,0 +1,276 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "dexhand/book_page_turn", + "suite": "dexhand", + "base": "dexhand-book-page-turn-i00", + "title_zh": "翻过一页硬纸书", + "summary_zh": "仅把右侧顶层第2号叶片翻到左侧页堆，使书显示第6、7页，并在书体和其他页未越过失败阈值的情况下离手稳定50次控制。", + "embodiment_zh": "两种模式均为固定天机Marvin支架、两条七关节机械臂，每臂腕部安装一只灵巧手；默认Sharpa Wave（每手22个手指执行器），可选Wuji Hand 2.2（每手20个）；双手均可参与，无移动底盘。", + "task_class": "atomic", + "analysis_level": "step", + "spec": { + "version": "1.2", + "task": "dexhand-book-page-turn-i00", + "bench": "dexhand", + "progress": "live", + "entities": {}, + "milestones": [ + { + "id": "target_page_turned", + "zh": "翻过指定页", + "mode": "state", + "when": null, + "after": [], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "总览中，原在右侧顶层的第2号硬纸页绕书脊翻到左侧页堆，摊开面由第4、5页变为第6、7页，翻动时须依次经过铰链角大于30度和大于90度；完成时铰链角大于170度、与左侧第0号叶片／第1号叶片／左封面之一实际接触、铰链角速度绝对值不大于0.2弧度每秒、所有其他叶片及两封面相对静置后初态的角度变化绝对值均不大于5度，且两手、两臂和支架都不再接触书的任何部位，这组条件连续成立50次控制（50赫兹，名义1秒，录像叠字hold达到50/50）。纸页仍在右侧、悬起未落到左堆、手还碰书或稳定计数未满均未完成；达到后若纸页离开左堆、重新被碰触或上述联合条件失效，就失去该状态，录像末尾追加的冻结画面不作为新的保持时长。", + "step": { + "cat": "位姿", + "item": "翻转" + }, + "contact": [ + { + "who": "手指/夹爪", + "how": "推/扫", + "plan": 1, + "note_zh": "一只灵巧手（如右手）的食指指尖拨动第2号叶片外缘，将其带过书脊；成功录像用右食指拨页，没有合手夹持。" + }, + { + "who": "手指/夹爪", + "how": "抓握", + "plan": 2, + "note_zh": "一只灵巧手的拇指与其余手指捏住第2号叶片边缘，将其翻到左侧页堆。" + }, + { + "who": "双臂协作", + "how": "抓握", + "plan": 3, + "note_zh": "一只灵巧手抓住书的固定部分稳住书体，另一只灵巧手捏住第2号叶片翻页；两只手最终都须离开书。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:3", + "inputs/task/standard/instruction.md:3", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/feaa1d1d97/checker.py:10-11,44-49,116-146,149-173", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/0e164430d9/task.json:11-15,37-49", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/13c6999f14/runtime.py:62-74,127-138", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/1afea4c64a/scene.py:140-177,253-261,288-291", + "inputs/oracle/replay.mp4:0.0-6.0", + "inputs/oracle/run.json:steps[86]" + ], + "conf": 6, + "note_zh": "一个指定叶片的一次翻页为一个语义步骤（位姿·翻转的定义明确包含翻页）。分离顶页、抬起、过竖直是途经状态，原生四阶段0→1→2→3→4只绑定本步骤验收；松手和稳定不另建节点，其余页及封面在5度范围内属于翻页质量条件。hold_s为0是预留使用原生hold_steps>=50的计数验收，不再叠加保持窗口；当前没有状态日志，when为null。显式关闭持物冻结以便操作铰接页时不掩盖状态丢失；接触方案可用任一只手，方案3表示可选的双手做法而非额外步骤。" + } + ], + "constraints": [ + { + "id": "book_displaced", + "zh": "书脊移位", + "never": null, + "criterion_zh": "书脊主体相对被动静置后的初始位置，三维位移范数大于0.03米；每个2毫秒物理子步检查，首次越界立即失败，之后移回也不能补救。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/feaa1d1d97/checker.py:99-101", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/13c6999f14/runtime.py:111-118", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/0e164430d9/task.json:45" + ] + }, + { + "id": "book_rotated", + "zh": "书脊转动", + "never": null, + "criterion_zh": "书脊主体相对被动静置后的初始四元数，最短旋转角大于10度；不是只检查倾斜角，每个2毫秒物理子步检查，首次越界立即失败。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/feaa1d1d97/checker.py:87-90,102-104", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/13c6999f14/runtime.py:111-118", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/0e164430d9/task.json:46" + ] + }, + { + "id": "second_leaf_turned", + "zh": "误翻其他页", + "never": null, + "criterion_zh": "除第2号叶片外，任一叶片的绑定铰链角相对静置后初态变化绝对值大于30度；每个2毫秒物理子步检查，首次越界立即失败。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/feaa1d1d97/checker.py:108-113", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/13c6999f14/runtime.py:111-118", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/0e164430d9/task.json:44" + ] + }, + { + "id": "book_closing", + "zh": "封面合拢", + "never": null, + "criterion_zh": "左封面或右封面的绑定铰链角相对静置后初态变化绝对值大于30度；每个2毫秒物理子步检查，首次越界立即失败。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/feaa1d1d97/checker.py:108-113", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/13c6999f14/runtime.py:111-118", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/0e164430d9/task.json:44" + ] + }, + { + "id": "book_below_table", + "zh": "书体穿过桌面", + "never": null, + "criterion_zh": "书脊子树内任一书体的质心高度小于0.778米，即低于0.78米桌面超过2毫米；每个2毫秒物理子步检查，立即失败。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/feaa1d1d97/checker.py:105-107", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/13c6999f14/runtime.py:111-118", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/0e164430d9/task.json:15,47" + ] + }, + { + "id": "unstable_physics", + "zh": "物理状态失稳", + "never": null, + "criterion_zh": "任一广义位置或速度非有限，或出现坏位置、坏速度、坏加速度物理警告；每个2毫秒物理子步检查，立即失败。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/feaa1d1d97/checker.py:94-98", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/13c6999f14/runtime.py:111-118" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "target_page_turned", + "level": "new_code", + "source": "inputs/grader/efe_dexhand-files/privileged/environment/variants/feaa1d1d97/checker.py:149-173；需逐控制记录原生stage、target_angle_deg、target_on_left_stack、target_speed_rad_s、max_other_leaf_change_deg、hand_book_contacts及hold_steps（inputs/grader/efe_dexhand-files/privileged/environment/variants/13c6999f14/runtime.py:127-138），验收用有序阶段及hold_steps>=50；现无状态日志，不能绑定。" + }, + { + "atom": "book_displaced", + "level": "new_code", + "source": "inputs/grader/efe_dexhand-files/privileged/environment/variants/feaa1d1d97/checker.py:99-101; inputs/grader/efe_dexhand-files/privileged/environment/variants/13c6999f14/runtime.py:111-118; inputs/grader/efe_dexhand-files/privileged/environment/variants/0e164430d9/task.json:45；需逐物理子步记录invariant首个失败原因及相关原生量，现无状态日志。" + }, + { + "atom": "book_rotated", + "level": "new_code", + "source": "inputs/grader/efe_dexhand-files/privileged/environment/variants/feaa1d1d97/checker.py:87-90,102-104; inputs/grader/efe_dexhand-files/privileged/environment/variants/13c6999f14/runtime.py:111-118; inputs/grader/efe_dexhand-files/privileged/environment/variants/0e164430d9/task.json:46；需逐物理子步记录invariant首个失败原因及相关原生量，现无状态日志。" + }, + { + "atom": "second_leaf_turned", + "level": "new_code", + "source": "inputs/grader/efe_dexhand-files/privileged/environment/variants/feaa1d1d97/checker.py:108-113; inputs/grader/efe_dexhand-files/privileged/environment/variants/13c6999f14/runtime.py:111-118; inputs/grader/efe_dexhand-files/privileged/environment/variants/0e164430d9/task.json:44；需逐物理子步记录invariant首个失败原因及相关原生量，现无状态日志。" + }, + { + "atom": "book_closing", + "level": "new_code", + "source": "inputs/grader/efe_dexhand-files/privileged/environment/variants/feaa1d1d97/checker.py:108-113; inputs/grader/efe_dexhand-files/privileged/environment/variants/13c6999f14/runtime.py:111-118; inputs/grader/efe_dexhand-files/privileged/environment/variants/0e164430d9/task.json:44；需逐物理子步记录invariant首个失败原因及相关原生量，现无状态日志。" + }, + { + "atom": "book_below_table", + "level": "new_code", + "source": "inputs/grader/efe_dexhand-files/privileged/environment/variants/feaa1d1d97/checker.py:105-107; inputs/grader/efe_dexhand-files/privileged/environment/variants/13c6999f14/runtime.py:111-118; inputs/grader/efe_dexhand-files/privileged/environment/variants/0e164430d9/task.json:15,47；需逐物理子步记录invariant首个失败原因及相关原生量，现无状态日志。" + }, + { + "atom": "unstable_physics", + "level": "new_code", + "source": "inputs/grader/efe_dexhand-files/privileged/environment/variants/feaa1d1d97/checker.py:94-98; inputs/grader/efe_dexhand-files/privileged/environment/variants/13c6999f14/runtime.py:111-118；需逐物理子步记录invariant首个失败原因及相关原生量，现无状态日志。" + } + ] + }, + "native_metrics": [ + { + "name": "target_angle_deg", + "unit": "deg", + "threshold": 170, + "direction": "higher_better" + }, + { + "name": "target_speed_rad_s", + "unit": "rad/s", + "threshold": 0.2, + "direction": "lower_better" + }, + { + "name": "max_other_leaf_change_deg", + "unit": "deg", + "threshold": 5, + "direction": "lower_better" + }, + { + "name": "book_shift_m", + "unit": "m", + "threshold": 0.03, + "direction": "lower_better" + }, + { + "name": "book_rotation_deg", + "unit": "deg", + "threshold": 10, + "direction": "lower_better" + } + ], + "open_questions_zh": [ + "inputs/states/index.json的logs为空：节点when、约束never及success均为null（P8），现有录像可看见翻页、离手及原生叠字，但无法独立核验全部接触、瞬时角速度和2毫秒子步失败条件；需增加原生特征、保持计数及失败原因的状态记录。", + "模式差异：物理任务的checker.py、runtime.py与task.json两模式逐字相同；特权模式评分器还要求两个独立进程回放均成功且最终状态哈希一致，标准模式直接判定最终在线回合。上述交付、确定性与基础设施判定不建物理步骤，也不冒充已记录世界状态（inputs/grader/efe_dexhand-files/privileged/tests/grade.py:103-146；inputs/grader/efe_dexhand-files/standard/tests/grade.py:23-48）。" + ], + "conf": 6, + "conf_reason_zh": "单步粒度来自两模式一致的任务文本，全部验收及失败阈值直接取自相同检查器和配置，初态由场景与静置重置代码核实，成功录像展示完整翻页及hold达到50/50；没有任何状态日志，联合条件和子步约束不能程序核验，因此置信度为6。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "单步（翻第2号叶片）符合任务文本：抬起/过竖直/落放为途经阶段，离手并入验收；约束均为评分器失败条件且立即终止，阈值与来源一致；无状态日志故when/success为null合理且已列入开放问题。接触方案与成功录像（右手拨页）一致。", + "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：['target_page_turned']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:48:34Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/dexhand__book_page_turn/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/dexhand__book_page_turn/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "f3ee22385400828cf1678e6adec47813418ee530a553917f7f4ef20872572780", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/dexhand/book_page_turn/trajectories/dexhand-book-page-turn-i00-standard--68s3td9.json b/phase2/dexhand/book_page_turn/trajectories/dexhand-book-page-turn-i00-standard--68s3td9.json new file mode 100644 index 0000000000000000000000000000000000000000..62f59ab72d4a405310e74042c0f9f7628a4ca62f --- /dev/null +++ b/phase2/dexhand/book_page_turn/trajectories/dexhand-book-page-turn-i00-standard--68s3td9.json @@ -0,0 +1,213 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "dexhand-book-page-turn-i00-standard--68s3td9", + "taskKey": "dexhand/book_page_turn", + "dag_sha256": "b1614d794a716fb93ce7267d3c538874fe875ffd93b9add17265ecb5aed9cd91", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "calls": 68, + "deterministic": 1, + "finalized": 1, + "hand_wuji": 0, + "infra_error": 0, + "live_matches_replay": 1, + "live_success": 0, + "n_actions": 318, + "no_verdict": 0, + "replay_success": 0, + "reset_budget": 20, + "resets": 9, + "resume_state_mismatch": 0, + "resumes": 0, + "sim_error": 0, + "sim_startups": 1, + "sim_steps": 318, + "success": 0, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.549, + "peak_processes": 11, + "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/dexhand-book-page-turn-i00-standard--68s3td9/replay.mp4", + "duration_s": 8.6, + "fps": 5.0, + "frames": 43, + "sha256": "b93ad3313a33671e1acc493cb3c632e5e8e9c23d7330fe92f9fb99aacf399cc3" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "target_page_turned": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察：视频 0.0 s，书平摊在桌面，目标星形图案页仍在原侧，叠字为叶片2角度−0.3度、stage 0/4、hold 0/50。" + }, + { + "t": 6.4, + "log_step": 164, + "text_zh": "观察：逐帧检查视频 0.0–6.4 s，机器人左手从桌后侧抬起并伸食指靠近星形页外缘，目标页未出现抬起、越过竖直或落到对侧的过程；视频 6.2–6.4 s 整本书向画面左下方移位。视频 6.4 s 叠字为 FAIL: book_displaced，叶片2角度3.3度、stage 0/4、其他页最大变化9.4度、hold 0/50、书体移位33.3毫米、旋转4.49度；log step 164 的服务状态报告末次 episode 10 在318次控制后 task_failure。" + }, + { + "t": 8.4, + "log_step": null, + "text_zh": "观察：视频 6.6–8.4 s 为失败终态冻结画面，目标页仍在原侧，stage 0/4、hold 0/50未变，翻过指定页的联合验收从未成立。成功参考视频 2.0–6.0 s 可见页板立起、翻到对侧及hold 50/50，这些完成特征在本次视频均未出现。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "推/扫", + "note_zh": "观察：视频 6.0–6.2 s，机器人左手伸直的食指抵近星形目标页外缘，指尖出现接触色标；视频 6.4 s 整本书被推移。推断：左食指拨推页边时把作用传给整书，没有形成对向夹持；右手停在桌旁，未参与稳书。末帧hand-book contacts为0表示该帧接触已解除，不代表此前没有接触。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "target_angle_deg", + "unit": "deg", + "threshold": 170, + "direction": "higher_better", + "value": null + }, + { + "name": "target_speed_rad_s", + "unit": "rad/s", + "threshold": 0.2, + "direction": "lower_better", + "value": null + }, + { + "name": "max_other_leaf_change_deg", + "unit": "deg", + "threshold": 5, + "direction": "lower_better", + "value": null + }, + { + "name": "book_shift_m", + "unit": "m", + "threshold": 0.03, + "direction": "lower_better", + "value": null + }, + { + "name": "book_rotation_deg", + "unit": "deg", + "threshold": 10, + "direction": "lower_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "target_page_turned", + "t": 6.4, + "log_step": 163, + "where": "操作中", + "evidence_zh": "观察：视频 6.0–6.4 s，机器人左食指靠近并拨压目标页外缘，页板未翻起而整本书移动；视频 6.4 s 评分叠字显示 FAIL: book_displaced、book shift 33.3 mm，stage仍为0/4。log step 163启动最后一次重置后的控制脚本，log step 164的episode 10、318次控制及task_failure与本次回放结局一致；断点属于翻指定页这一步。" + }, + "unmet": { + "target_page_turned": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "带动本体", + "t": 6.4, + "log_step": null, + "evidence_zh": "观察：视频 6.2–6.4 s，机器人左手食指在星形目标页外缘动作后，书的两侧页堆和书脊一起向画面左下方移动，而目标页未相对书脊翻起；视频 6.4 s 叠字报告书体移位33.3毫米并以book_displaced结束，叶片2角度仅3.3度、stage 0/4。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "反馈控制", + "t": 6.4, + "log_step": 161, + "evidence_zh": "推断：末次脚本的末端偏差校正存在腕部与指腹目标换算错误。log step 156定义pad_rows把输入指腹目标减去R@OFF得到腕目标；log step 161新增校正先算腕目标w_corr，却把w_corr−R_PRESS@OFF传给pad_rows，导致腕目标再次减去同一偏置，成为w_corr−2R_PRESS@OFF。log step 163启动这一修改版本，log step 164报告其失败；视频 6.0–6.4 s实际呈现伸指逼近页边并推移整书，未实现分离抬页。该代码证据支持末端校正控制有误，末次实际校正误差大小未记录，不能仅凭agent在log step 165的下垂解释确定全部动力学原因。", + "source_zh": "日志" + } + ], + "gap_zh": "没有状态日志；视频仅5帧/秒，视频 6.2–6.4 s之间的瞬时接触力和具体碰撞子步不能恢复。末次episode的校正过程打印未出现在现有日志中，因而无法量化该次末端误差或确认每一轮校正的执行结果；先前回合的2.7厘米误差不作为本次测量。DAG列出的5项原生指标均未在run.json评分字段中报告，native_metrics按规则填null，视频叠字的可见数值仅用于观察证据。" + }, + "summary_zh": "观察：本次评分回合中，机器人左食指靠近星形目标页外缘，但未完成任何抬页或翻页阶段，翻过指定页节点始终未达成，最终为0/1。观察：视频 6.4 s整本书被带动，叠字以book_displaced失败结束；推断：日志中末次脚本的指腹／腕目标偏置重复扣除支持反馈控制错误。没有状态日志及末次校正过程测量，无法还原瞬时接触力或量化该次末端偏差。", + "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（翻指定页，操作中，视频 6.4 s book_displaced）、接触（左食指推/扫）与视频叠字一致；反馈控制根因明确标为推断，基于log中pad_rows与w_corr的重复扣除代码，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": "2026-10-29T10:06:16Z", + "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/dexhand-book-page-turn-i00-standard--68s3td9/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/dexhand-book-page-turn-i00-standard--68s3td9/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "13f19807faaf605cd3293ac9fbed9e66cb7664ababdc705092829f643cf3ea7e", + "video_sha256": "b93ad3313a33671e1acc493cb3c632e5e8e9c23d7330fe92f9fb99aacf399cc3", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/dexhand/book_page_turn/trajectories/dexhand-book-page-turn-i00-standard--p9y4nub.json b/phase2/dexhand/book_page_turn/trajectories/dexhand-book-page-turn-i00-standard--p9y4nub.json new file mode 100644 index 0000000000000000000000000000000000000000..25737b175ee3cc846c328f960592c1cd71cedeb3 --- /dev/null +++ b/phase2/dexhand/book_page_turn/trajectories/dexhand-book-page-turn-i00-standard--p9y4nub.json @@ -0,0 +1,202 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "dexhand-book-page-turn-i00-standard--p9y4nub", + "taskKey": "dexhand/book_page_turn", + "dag_sha256": "b1614d794a716fb93ce7267d3c538874fe875ffd93b9add17265ecb5aed9cd91", + "outcome": { + "success": 1, + "reward": 1.0, + "source": "catalog", + "raw": { + "calls": 178, + "deterministic": 1, + "finalized": 1, + "hand_wuji": 0, + "infra_error": 0, + "live_matches_replay": 1, + "live_success": 1, + "n_actions": 2508, + "no_verdict": 0, + "replay_success": 1, + "reset_budget": 20, + "resets": 3, + "resume_state_mismatch": 0, + "resumes": 0, + "sim_error": 0, + "sim_startups": 1, + "sim_steps": 2508, + "success": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.328, + "peak_processes": 10, + "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/dexhand-book-page-turn-i00-standard--p9y4nub/replay.mp4", + "duration_s": 52.4, + "fps": 5.0, + "frames": 262, + "sha256": "c43a93eb1cb88f3af6b51610a80d149e933537e5b287ef5d4179f8d1534be212" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "target_page_turned": { + "method": "visual", + "intervals": [ + [ + 50.2, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察：视频 0.0 s 书摊开，第2号页仍平放在右侧，原生叠字为 stage 0/4、hold 0/50；本节点初始未完成。" + }, + { + "t": 36.0, + "log_step": null, + "text_zh": "观察：视频 34.8–39.6 s 右手伸出的食指在指定页外缘拨动，页片逐渐抬离下层页；视频 36.0 s 叠字给出页角1.8度、hand-book contacts 2。右手没有以拇指和食指对夹页片，另一只手未参与。" + }, + { + "t": 40.0, + "log_step": null, + "text_zh": "观察：视频 40.0 s 第2号页角31.7度、stage 1/4（lifted）；视频 44.0 s 页片越过竖直，页角92.8度、stage 2/4（past_vertical），其余页仍平放。" + }, + { + "t": 49.2, + "log_step": null, + "text_zh": "观察：视频 46.0–48.8 s 右食指在竖起页片的页面一侧移动，页片继续朝左堆倾倒，原生 hand-book contacts 为0，右手随后抬离；视频 49.2 s 页片已落到左侧，摊开面为左侧绿色图案、右侧黄色图案，与成功参考终态一致，叠字为页角181.6度、stage 3/4（laid）、hold 2/50、hand-book contacts 0。此时保持计数未满，尚不计完成。" + }, + { + "t": 50.2, + "log_step": 49, + "text_zh": "观察：视频 49.4、49.6、49.8、50.0 s 的原生保持计数依次为12/50、22/50、32/50、42/50；视频 50.2 s 首次可见 PASS、stage 4/4（held）、hold 50/50，页角181.7度、其余页最大变化0.0度、hand-book contacts 0，双手均离书。log step 49 的机器人服务返回 done=True、reason=success、control_steps=2508，评分字段 live_success=1、replay_success=1、live_matches_replay=1。节点时间取视频帧，精度约0.2 s。" + }, + { + "t": 52.2, + "log_step": null, + "text_zh": "观察：视频 50.4–52.2 s 为标有 final state 的冻结终态，指定页留在左堆且右手悬空，PASS 与 hold 50/50 保持显示；已逐帧核对视频 51.4–52.2 s。冻结画面未计入新的保持时长，节点完成依据为视频 50.2 s 之前的实际保持过程。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "推/扫", + "note_zh": "观察：视频 34.8–43.6 s 右手食指指尖沿第2号硬纸页外缘拨动并将其抬过接近竖直的位置，手没有对向捏夹，左手未参与；视频 44.0 s 已越过竖直且原生 hand-book contacts 为0，之后右手指在页片旁移动并撤离，页片继续向左落下。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "target_angle_deg", + "unit": "deg", + "threshold": 170, + "direction": "higher_better", + "value": null + }, + { + "name": "target_speed_rad_s", + "unit": "rad/s", + "threshold": 0.2, + "direction": "lower_better", + "value": null + }, + { + "name": "max_other_leaf_change_deg", + "unit": "deg", + "threshold": 5, + "direction": "lower_better", + "value": null + }, + { + "name": "book_shift_m", + "unit": "m", + "threshold": 0.03, + "direction": "lower_better", + "value": null + }, + { + "name": "book_rotation_deg", + "unit": "deg", + "threshold": 10, + "direction": "lower_better", + "value": null + } + ], + "failure": null, + "summary_zh": "观察：受评视频中，机器人用右食指拨起第2号硬纸页，依次抬过30度和90度，再使其落到左侧页堆并离手；视频 50.2 s 首次显示 PASS、stage 4/4 和 hold 50/50，完成1/1节点。观察：log step 49–50 的机器人服务确认第4回合成功，评分字段也确认在线与回放均成功且一致，前三次重置涉及的会话尝试不计入本次节点进度。观察：没有状态日志可独立复核瞬时角速度和全部细分接触，联合验收由原生叠字及成功判定佐证；评分字段未报告DAG所列五项指标的数值，因此 native_metrics 的 value 均保留null，末尾冻结画面不计保持时长。", + "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 target_page_turned 在视频 50.2 s 首次出现 PASS、stage 4/4、hold 50/50，与抽帧一致；触点为右食指推/扫，无失败层，标注可接受。", + "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": "2026-10-29T10:06:44Z", + "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/dexhand-book-page-turn-i00-standard--p9y4nub/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/dexhand-book-page-turn-i00-standard--p9y4nub/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "a7612cae7d8fc3c09e7f79251e9899850162b8879583a1be5d353ec43e97dd55", + "video_sha256": "c43a93eb1cb88f3af6b51610a80d149e933537e5b287ef5d4179f8d1534be212", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/dexhand/grinder_crank/task_dag.json b/phase2/dexhand/grinder_crank/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..6173954c35ffbb2130d8af31f9ff1961f69ffa64 --- /dev/null +++ b/phase2/dexhand/grinder_crank/task_dag.json @@ -0,0 +1,397 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "dexhand/grinder_crank", + "suite": "dexhand", + "base": "dexhand-grinder-crank-i00", + "title_zh": "手摇磨豆机", + "summary_zh": "曲柄相对机身净顺时针转过至少720°后，双手脱离，磨豆机底座落台、直立且低速静止的联合条件连续保持50个控制步，并且全过程未触发失败约束。", + "embodiment_zh": "两模式均为固定在支架上的天机Marvin双臂机器人，每臂7个关节，左右腕各安装一只同型号灵巧手；默认两只Sharpa Wave（每手22个手指执行器），也支持两只Wuji Hand 2.2（每手20个手指执行器），底座固定、无移动底盘。", + "task_class": "atomic", + "analysis_level": "step", + "spec": { + "version": "1.2", + "task": "dexhand-grinder-crank-i00", + "bench": "dexhand", + "progress": "live", + "entities": {}, + "milestones": [ + { + "id": "crank_turned", + "zh": "曲柄转满两圈", + "level": "semantic", + "mode": "state", + "when": null, + "after": [], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "总览画面中，黑色圆旋钮随金属曲柄绕磨豆机顶轴从上方看净顺时针转至少两整圈（相对机身的展开转角≥720°，不是旋钮自转），随后双手与机身、曲柄、旋钮均无接触，底座接触台面、机身竖轴倾角≤10°，质心线速度≤0.02米/秒、机身角速度≤0.2弧度/秒、曲柄速度绝对值≤0.5弧度/秒，这些条件同时连续满足50个50赫兹控制步才完成。回放顶部的Crank至少为2.00/2.00、hand contacts为R 0 L 0且released hold为50/50可辅助确认；仅转够角度但仍握住、仍在晃动或保持不足均未完成，接触恢复、转角降回阈值以下或任一验收条件失去即不再满足。", + "step": { + "cat": "机构", + "item": "旋转" + }, + "contact": [ + { + "who": "双臂协作", + "how": "抓握", + "plan": 1, + "note_zh": "一只灵巧手抓握磨豆机身抵抗转矩，另一只灵巧手抓握黑色旋钮带动曲柄；典型方案为左手扶机身、右手转旋钮，左右角色不作为必经条件。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:3", + "inputs/task/standard/instruction.md:3", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/f0caa6842d/checker.py:64,97-99,178-201", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/f0caa6842d/checker.py:113-137,203-224", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/1d21e930d4/task.json:139-150", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/2b6719073f/runtime.py:45-69,100-115", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/49286a9a86/scene.py:101-124", + "inputs/grader/efe_dexhand-files/privileged/environment/assets/source/hario_mechanism.py:77-105", + "inputs/oracle/replay.mp4 @ 0.0 s, 12.0 s, 21.0 s, 29.0 s", + "inputs/oracle/run.json:steps[id=94].out.text" + ], + "conf": 7, + "note_zh": "两个整圈由同一个展开净转角阈值验收，按暂行D6只建一个步骤；评分器stage=1只表示转角曾达标，不能替代最终仍达标且释放静置的验收。抓握、必要的换握和最后松手均不另建节点；机身的开局直立和位置范围属于失败约束。hold_s为0表示不在原生released_hold_steps达到50之外再次计时；没有状态日志，完整保持窗由视频判据承担。曲柄作为操作机构显式禁用持物冻结；转角判定实现使用720°减1e-9°的浮点容差。" + } + ], + "constraints": [ + { + "id": "grinder_tipped", + "zh": "磨豆机倾倒", + "never": null, + "criterion_zh": "磨豆机自身竖轴偏离世界竖直方向超过10°；每个2毫秒物理子步检查，首次越限立即结束，扶正不能补救。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/f0caa6842d/checker.py:101-111,166-167", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/1d21e930d4/task.json:141" + ] + }, + { + "id": "grinder_slid", + "zh": "磨豆机滑移", + "never": null, + "criterion_zh": "磨豆机底座中心相对开局位置的水平位移大于0.02米；首次越限立即结束，移回不能补救。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/f0caa6842d/checker.py:104-106,168-169", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/1d21e930d4/task.json:142" + ] + }, + { + "id": "grinder_lifted", + "zh": "磨豆机抬离", + "never": null, + "criterion_zh": "磨豆机底座中心相对开局高度上升大于0.02米；首次越限立即结束。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/f0caa6842d/checker.py:104-106,170-171", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/1d21e930d4/task.json:143" + ] + }, + { + "id": "grinder_below_counter", + "zh": "底座跌穿台面", + "never": null, + "criterion_zh": "磨豆机底座中心世界高度小于0.695米，即台面0.700米减去0.005米容差；首次越限立即结束。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/f0caa6842d/checker.py:34,164-165", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/1d21e930d4/task.json:16-26,144" + ] + }, + { + "id": "grinder_spun", + "zh": "磨豆机身自转", + "never": null, + "criterion_zh": "磨豆机身局部横轴的水平投影相对开局方向的有符号偏航角绝对值大于45°；首次越限立即结束，并非限制曲柄自身转角。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/f0caa6842d/checker.py:107-110,172-173", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/1d21e930d4/task.json:145" + ] + }, + { + "id": "crank_reversed", + "zh": "曲柄反向超限", + "never": null, + "criterion_zh": "相对机身的曲柄展开净顺时针角小于负90°，即从开局净逆时针转过90°以上；不把转动中的任意短暂回退都判作失败。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/f0caa6842d/checker.py:97-99,174-175", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/1d21e930d4/task.json:146" + ] + }, + { + "id": "robot_collision", + "zh": "非手部碰撞", + "never": null, + "criterion_zh": "任一机械臂、躯干或支架几何体与任何几何体发生接触，即使碰到另一只手也失败；只有手部允许接触，每个物理子步检查。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/f0caa6842d/checker.py:40-61,153-163" + ] + }, + { + "id": "unstable_physics", + "zh": "物理状态失稳", + "never": null, + "criterion_zh": "物理引擎报告BADQPOS、BADQVEL或BADQACC任一种警告，首次出现即失败；该数值异常须由引擎诊断判断，不能仅凭画面猜测。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/f0caa6842d/checker.py:140-146" + ] + }, + { + "id": "direct_object_manipulation", + "zh": "绕过执行器操控", + "never": null, + "criterion_zh": "任务禁止直接传送物体、建立抓取焊接、脚本驱动物体或向磨豆机直接施力；合法交付只能包含机器人14个臂关节和双手手指的执行器目标。此为文本禁止的操控方式，不能从单张画面判断，也不声称评分器具有独立作弊检测字段。", + "consequence": "历史否决", + "sources": [ + "inputs/task/privileged/instruction.md:3", + "inputs/task/standard/instruction.md:3", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/2b6719073f/runtime.py:92-108" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "net_clockwise_deg", + "level": "new_code", + "source": "inputs/grader/efe_dexhand-files/privileged/environment/variants/f0caa6842d/checker.py:97-99,183,209；需逐控制步记录Checker.features(data)[\"net_clockwise_deg\"]，本任务未提供状态日志。" + }, + { + "atom": "hand_contacts", + "level": "new_code", + "source": "inputs/grader/efe_dexhand-files/privileged/environment/variants/f0caa6842d/checker.py:113-131,190-191,218；需逐控制步记录Checker.features(data)[\"hand_contacts\"]，本任务未提供状态日志。" + }, + { + "atom": "base_on_counter", + "level": "new_code", + "source": "inputs/grader/efe_dexhand-files/privileged/environment/variants/f0caa6842d/checker.py:113-131,192,219；需逐控制步记录Checker.features(data)[\"base_on_counter\"]，本任务未提供状态日志。" + }, + { + "atom": "tilt_deg", + "level": "new_code", + "source": "inputs/grader/efe_dexhand-files/privileged/environment/variants/f0caa6842d/checker.py:101-111,193,213；需逐控制步记录Checker.features(data)[\"tilt_deg\"]，本任务未提供状态日志。" + }, + { + "atom": "linear_speed_m_s", + "level": "new_code", + "source": "inputs/grader/efe_dexhand-files/privileged/environment/variants/f0caa6842d/checker.py:133-137,194,220；需逐控制步记录Checker.features(data)[\"linear_speed_m_s\"]，本任务未提供状态日志。" + }, + { + "atom": "angular_speed_rad_s", + "level": "new_code", + "source": "inputs/grader/efe_dexhand-files/privileged/environment/variants/f0caa6842d/checker.py:133-137,195,221；需逐控制步记录Checker.features(data)[\"angular_speed_rad_s\"]，本任务未提供状态日志。" + }, + { + "atom": "crank_speed_rad_s", + "level": "new_code", + "source": "inputs/grader/efe_dexhand-files/privileged/environment/variants/f0caa6842d/checker.py:133-137,196,222；需逐控制步记录Checker.features(data)[\"crank_speed_rad_s\"]，本任务未提供状态日志。" + }, + { + "atom": "released_hold_steps", + "level": "new_code", + "source": "inputs/grader/efe_dexhand-files/privileged/environment/variants/f0caa6842d/checker.py:178-201,223；需逐控制步记录Checker.features(data)[\"released_hold_steps\"]，本任务未提供状态日志。" + }, + { + "atom": "slide_m", + "level": "new_code", + "source": "inputs/grader/efe_dexhand-files/privileged/environment/variants/f0caa6842d/checker.py:101-111,168-169,214；需逐控制步记录Checker.features(data)[\"slide_m\"]，本任务未提供状态日志。" + }, + { + "atom": "lift_m", + "level": "new_code", + "source": "inputs/grader/efe_dexhand-files/privileged/environment/variants/f0caa6842d/checker.py:101-111,170-171,215；需逐控制步记录Checker.features(data)[\"lift_m\"]，本任务未提供状态日志。" + }, + { + "atom": "base_z_m", + "level": "new_code", + "source": "inputs/grader/efe_dexhand-files/privileged/environment/variants/f0caa6842d/checker.py:101-111,164-165,216；需逐控制步记录Checker.features(data)[\"base_z_m\"]，本任务未提供状态日志。" + }, + { + "atom": "body_yaw_deg", + "level": "new_code", + "source": "inputs/grader/efe_dexhand-files/privileged/environment/variants/f0caa6842d/checker.py:107-110,172-173,217；需逐控制步记录Checker.features(data)[\"body_yaw_deg\"]，本任务未提供状态日志。" + }, + { + "atom": "robot_collision", + "level": "new_code", + "source": "inputs/grader/efe_dexhand-files/privileged/environment/variants/f0caa6842d/checker.py:153-163；需在每个子步记录非手部几何接触或原生failure与failure_detail。" + }, + { + "atom": "unstable_physics", + "level": "new_code", + "source": "inputs/grader/efe_dexhand-files/privileged/environment/variants/f0caa6842d/checker.py:144-146；需记录物理引擎BADQPOS/BADQVEL/BADQACC警告及首次失败。" + }, + { + "atom": "direct_object_manipulation", + "level": "infeasible", + "source": "inputs/task/privileged/instruction.md:3；仅靠机器人与物体的状态不能可靠识别外部直接施力、传送或焊接，需独立的执行接口审计。" + } + ] + }, + "native_metrics": [ + { + "name": "net_clockwise_deg", + "unit": "deg", + "threshold": 720, + "direction": "higher_better", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/f0caa6842d/checker.py:183" + ] + }, + { + "name": "tilt_deg", + "unit": "deg", + "threshold": 10, + "direction": "lower_better", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/1d21e930d4/task.json:141" + ] + }, + { + "name": "slide_m", + "unit": "m", + "threshold": 0.02, + "direction": "lower_better", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/1d21e930d4/task.json:142" + ] + }, + { + "name": "lift_m", + "unit": "m", + "threshold": 0.02, + "direction": "lower_better", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/1d21e930d4/task.json:143" + ] + }, + { + "name": "base_z_m", + "unit": "m", + "threshold": 0.695, + "direction": "higher_better", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/f0caa6842d/checker.py:34,164-165" + ] + }, + { + "name": "abs_body_yaw_deg", + "unit": "deg", + "threshold": 45, + "direction": "lower_better", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/1d21e930d4/task.json:145" + ] + }, + { + "name": "linear_speed_m_s", + "unit": "m/s", + "threshold": 0.02, + "direction": "lower_better", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/1d21e930d4/task.json:147" + ] + }, + { + "name": "angular_speed_rad_s", + "unit": "rad/s", + "threshold": 0.2, + "direction": "lower_better", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/1d21e930d4/task.json:148" + ] + }, + { + "name": "crank_speed_rad_s", + "unit": "rad/s", + "threshold": 0.5, + "direction": "lower_better", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/1d21e930d4/task.json:149" + ] + }, + { + "name": "released_hold_steps", + "unit": "control", + "threshold": 50, + "direction": "higher_better", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/1d21e930d4/task.json:150" + ] + } + ], + "open_questions_zh": [ + "inputs/states/index.json的logs为空，没有可安全绑定的逐帧字段；依P8，唯一节点when、全部约束never及success均为null，不能进行状态日志的达标与评分一致性验证。原生特征虽在评分器代码和成功运行记录中存在，仍需由回放适配器逐帧记录后才能绑定。", + "暂行DECISIONS.md D6、D10：两整圈按单个720°累计转角目标建一个旋转步骤；不拆逐圈计数，也不把评分器crank_turned与released_hold两个阶段拆成语义节点。", + "文本与评分器的方式核查范围不同：文本要求一手扶机身、另一手通过旋钮转曲柄，Checker.control只核查相对转角与最终释放静止，未独立核查哪只手扶机身或是否始终抓旋钮；这些是接触方案，不补成未受评的动作节点。", + "两模式共享完全一致的checker.py、runtime.py和初态实例；privileged评分入口要求两次新进程重放都成功且终态哈希相同（inputs/grader/efe_dexhand-files/privileged/tests/grade.py:113-118），standard评分入口读取可信服务的实时终结记录（inputs/grader/efe_dexhand-files/standard/tests/grade.py:23-29）。本DAG表示共同的物理任务核心，不把交付有效性、重放一致性或服务终结状态绑定为语义条件。", + "没有状态日志时，速度和严格无接触难以仅从静帧精确量测；成功参考视频顶部显示Crank +2.09/2.00、R 0 L 0、released hold 50/50，run.json的step 94中两次重放输出净转角753.373°及保持50步，支持成功终态，但不以此代替日志覆盖率。回放标明2倍速，50个控制步的保持窗应按控制步或模拟时间判断，不能直接要求播放时长1秒。" + ], + "conf": 7, + "conf_reason_zh": "单节点粒度直接来自文本且符合D6；阈值、联合保持窗、初态与失败规则均有源码证据，两模式物理核心相同，成功视频及两次重放记录与目标一致；无逐帧状态日志，精确速度和接触验收尚不能自动核对。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "单节点（曲柄转满两圈）符合文本与暂行D6/D10：抓握、松手静置均折入该节点的判据，评分器全部失败条件均作为立即终止约束，success因无日志为null合规。无状态日志，无法做回放一致性验证，已在open_questions_zh说明。接触标签双臂协作×抓握合理。", + "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：['crank_turned']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:49:41Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/dexhand__grinder_crank/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/dexhand__grinder_crank/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "8499b9e9a61418a5c37ec88c964607a0275a1aad5082831163414661eb6f6b9e", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/dexhand/grinder_crank/trajectories/dexhand-grinder-crank-i00-privileged--7n7fpdp.json b/phase2/dexhand/grinder_crank/trajectories/dexhand-grinder-crank-i00-privileged--7n7fpdp.json new file mode 100644 index 0000000000000000000000000000000000000000..23c23652a7f56b4d7f896fd5dc1365ed85bf6f6d --- /dev/null +++ b/phase2/dexhand/grinder_crank/trajectories/dexhand-grinder-crank-i00-privileged--7n7fpdp.json @@ -0,0 +1,221 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "dexhand-grinder-crank-i00-privileged--7n7fpdp", + "taskKey": "dexhand/grinder_crank", + "dag_sha256": "70aeb6508b690fe8f013c97c5b13c4633c4dfa3bf5e280aa5f5507c412b2a337", + "outcome": { + "success": 1, + "reward": 1.0, + "source": "catalog", + "raw": { + "deterministic": 1, + "hand_wuji": 0, + "infra_error": 0, + "n_actions": 876, + "sim_steps": 875, + "stages_completed": 2, + "success": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 2.45, + "peak_processes": 9, + "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/dexhand-grinder-crank-i00-privileged--7n7fpdp/replay.mp4", + "duration_s": 19.6, + "fps": 10.0, + "frames": 196, + "sha256": "c5e408a176fc1eb021680ff8c6aa642d7270742481c6b11a914cf7d52c5db86f" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "crank_turned": { + "method": "visual", + "intervals": [ + [ + 17.5, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 6.0, + "log_step": null, + "text_zh": "观察：视频 6.0 s，机器人左手四指弯曲环抱磨豆机身，右手食指向下伸直抵住黑色旋钮，其余手指收起；视频 8.0、10.0、12.0、14.0 s 的旋钮位置随曲柄绕顶轴变化，左手持续固定机身。右手是单指推动，没有夹住旋钮。" + }, + { + "t": 14.4, + "log_step": null, + "text_zh": "观察：视频 14.0 s 顶部基准回放叠字显示 Crank +1.97/2.00，视频 14.4 s 为 +2.03/2.00，hand contacts R 1 L 26、released hold 0/50；该叠字报告评分器的曲柄净顺时针圈数、双手接触数及联合释放静止保持步数。此时转角达标，但双手尚未脱离，节点尚未完成。" + }, + { + "t": 16.5, + "log_step": null, + "text_zh": "观察：视频 16.0–16.5 s，机器人左手逐渐张开，右手已抬离；视频 16.5 s 叠字显示 R 0 L 0、released hold 0/50、Crank +2.04/2.00，机身底座仍在台面上。视频 17.4 s 左手已后撤，保持计数为45/50，仍未完成验收。" + }, + { + "t": 17.5, + "log_step": 33, + "text_zh": "观察：视频 17.5 s 首次显示 PASS，Crank +2.04/2.00、hand contacts R 0 L 0、released hold 50/50、step 875；磨豆机直立落台，双手均已脱离。视频 17.6–19.5 s 逐帧检查均维持相同成功终态；视频 19.5 s 仍为 PASS。log step 33 对实际交付文件的两次独立重放均报告 success=True、failure=None、steps=875，且终态哈希一致；inputs/run.json 的 reward.success=1、stages_completed=2、deterministic=1 与视频相符。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "观察：视频 6.0–14.4 s，机器人左手掌与弯曲四指环抱磨豆机身，固定机身并配合右手转动曲柄；此项仅描述左手与机身的抓握接触。" + }, + { + "who": "双臂协作", + "how": "推/扫", + "note_zh": "观察：视频 6.0、8.0、10.0、12.0、14.0、14.4 s，机器人右手以伸直食指抵住黑色旋钮并沿圆周推动，其他手指收起，没有形成对向夹持；同时左手抓握机身。随后右手先抬离，左手再张开后撤，视频 16.5 s 双手接触数均为0。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "net_clockwise_deg", + "unit": "deg", + "threshold": 720, + "direction": "higher_better", + "value": null + }, + { + "name": "tilt_deg", + "unit": "deg", + "threshold": 10, + "direction": "lower_better", + "value": null + }, + { + "name": "slide_m", + "unit": "m", + "threshold": 0.02, + "direction": "lower_better", + "value": null + }, + { + "name": "lift_m", + "unit": "m", + "threshold": 0.02, + "direction": "lower_better", + "value": null + }, + { + "name": "base_z_m", + "unit": "m", + "threshold": 0.695, + "direction": "higher_better", + "value": null + }, + { + "name": "abs_body_yaw_deg", + "unit": "deg", + "threshold": 45, + "direction": "lower_better", + "value": null + }, + { + "name": "linear_speed_m_s", + "unit": "m/s", + "threshold": 0.02, + "direction": "lower_better", + "value": null + }, + { + "name": "angular_speed_rad_s", + "unit": "rad/s", + "threshold": 0.2, + "direction": "lower_better", + "value": null + }, + { + "name": "crank_speed_rad_s", + "unit": "rad/s", + "threshold": 0.5, + "direction": "lower_better", + "value": null + }, + { + "name": "released_hold_steps", + "unit": "control", + "threshold": 50, + "direction": "higher_better", + "value": null + } + ], + "failure": null, + "summary_zh": "观察：机器人左手抓握磨豆机身，右手用食指推动旋钮，曲柄净顺时针转过两圈后双手依次脱离；视频 17.5 s 首次显示 PASS、R 0 L 0、released hold 50/50，唯一节点完成，成功状态保持至视频 19.5 s。观察：评分器 success=1、stages_completed=2，log step 33 对交付文件的两次独立重放均成功且终态哈希相同，与视频一致。观察：本次没有状态日志，精确速度依靠回放的联合验收计数确认；评分器 reward 未报告 DAG 所列原生指标的数值，故 native_metrics 的 value 均留为 null。", + "conf": 9, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "成功运行：crank_turned 在视频 17.5 s（step 875）首次 PASS，与帧、log step 33 及评分器一致；接触标注（左手抓握机身、右手食指推）与帧相符，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": "2026-10-29T10:07:30Z", + "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/dexhand-grinder-crank-i00-privileged--7n7fpdp/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/dexhand-grinder-crank-i00-privileged--7n7fpdp/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "7337829cdb8e3c83734fcaf92c7c9de89f0a5651df7806621c205f2f8cd64fb8", + "video_sha256": "c5e408a176fc1eb021680ff8c6aa642d7270742481c6b11a914cf7d52c5db86f", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/dexhand/grinder_crank/trajectories/dexhand-grinder-crank-i00-standard--2ormf7y.json b/phase2/dexhand/grinder_crank/trajectories/dexhand-grinder-crank-i00-standard--2ormf7y.json new file mode 100644 index 0000000000000000000000000000000000000000..a071f2d39c911e881a58e81a96ebc2d2dc0e3814 --- /dev/null +++ b/phase2/dexhand/grinder_crank/trajectories/dexhand-grinder-crank-i00-standard--2ormf7y.json @@ -0,0 +1,257 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "dexhand-grinder-crank-i00-standard--2ormf7y", + "taskKey": "dexhand/grinder_crank", + "dag_sha256": "70aeb6508b690fe8f013c97c5b13c4633c4dfa3bf5e280aa5f5507c412b2a337", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "calls": 148, + "control_steps_total": 1828, + "elapsed_s": 3644.084, + "finalized": 1, + "gave_up": 0, + "graded_by_live": 1, + "grader_error": 0, + "hand_wuji": 0, + "infra_error": 0, + "instance_mismatch": 0, + "live_success": 0, + "missing_log": 0, + "n_actions": 241, + "no_verdict": 0, + "replays": 0, + "reset_budget": 20, + "resets": 10, + "resume_state_mismatch": 0, + "resumes": 0, + "sim_error": 0, + "sim_startups": 1, + "sim_steps": 241, + "success": 0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.659, + "peak_processes": 11, + "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/dexhand-grinder-crank-i00-standard--2ormf7y/replay.mp4", + "duration_s": 7.2, + "fps": 5.0, + "frames": 36, + "sha256": "5bfd68bc6f6d33c3883c594391bf70d27477e832a3e0d7c74112963bc4e82f42" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "crank_turned": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察：视频 0.0 s 至 4.4 s，机器人左手靠近并逐渐合拢包住磨豆机身，机身发生偏转；黑色旋钮没有绕顶轴完成整圈，画面原生叠字 Crank 始终约为 0.00/2.00 圈。" + }, + { + "t": 5.0, + "log_step": 470, + "text_zh": "观察：视频 4.6 s 至 5.0 s，机器人右手双指向旋钮下探并接触，左手仍扶握机身；机身倾斜加大。视频 5.0 s 叠字显示 FAIL: grinder_tipped、tilt 10.2 deg、Crank -0.00/2.00、hand contacts R 2 L 1、released hold 0/50；视频 5.2 s 至 7.0 s 保持同一失败终态，未出现两圈或松手保持达标。log step 470 的服务终结记录为 Episode 11、control_steps 241、done=True、reason=task_failure，与视频末端 step 241 相符。" + }, + { + "t": 5.0, + "log_step": 464, + "text_zh": "观察：log step 464 重置到 Episode 11，log step 465 确认该局 control_steps=0，log step 466 启动控制器；评分字段 n_actions=241、sim_steps=241。此前十次试验不计本次节点进度。成功参考视频 29.0 s 显示 PASS、Crank +2.09/2.00、R 0 L 0、released hold 50/50，仅用于对照完成判据，不是本次进度证据。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "观察：视频约 3.2 s 至 5.0 s，机器人左手手指合拢包住磨豆机身进行扶握，右手随后同时接近并作用于旋钮区域；此项记录左手对机身的抓握。" + }, + { + "who": "双臂协作", + "how": "压/擦", + "note_zh": "观察：视频 4.6 s 至 5.0 s，机器人右手食指、中指向下接触黑色旋钮附近的曲柄组件，左手同时扶握机身；右手呈双指下探压碰，画面未显示其对旋钮形成对向夹紧。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "net_clockwise_deg", + "unit": "deg", + "threshold": 720, + "direction": "higher_better", + "value": null + }, + { + "name": "tilt_deg", + "unit": "deg", + "threshold": 10, + "direction": "lower_better", + "value": null + }, + { + "name": "slide_m", + "unit": "m", + "threshold": 0.02, + "direction": "lower_better", + "value": null + }, + { + "name": "lift_m", + "unit": "m", + "threshold": 0.02, + "direction": "lower_better", + "value": null + }, + { + "name": "base_z_m", + "unit": "m", + "threshold": 0.695, + "direction": "higher_better", + "value": null + }, + { + "name": "abs_body_yaw_deg", + "unit": "deg", + "threshold": 45, + "direction": "lower_better", + "value": null + }, + { + "name": "linear_speed_m_s", + "unit": "m/s", + "threshold": 0.02, + "direction": "lower_better", + "value": null + }, + { + "name": "angular_speed_rad_s", + "unit": "rad/s", + "threshold": 0.2, + "direction": "lower_better", + "value": null + }, + { + "name": "crank_speed_rad_s", + "unit": "rad/s", + "threshold": 0.5, + "direction": "lower_better", + "value": null + }, + { + "name": "released_hold_steps", + "unit": "control", + "threshold": 50, + "direction": "higher_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "crank_turned", + "t": 5.0, + "log_step": 470, + "where": "操作中", + "evidence_zh": "观察：视频 4.6 s 至 5.0 s，机器人左手扶握机身、右手下探接触旋钮区域时，机身倾角由 7.5° 增至 10.2°，视频 5.0 s 显示 FAIL: grinder_tipped；曲柄仍约零圈，尚未进入转满两圈后的释放验收。log step 470 确认最后评分局 Episode 11 在 241 个控制步后 task_failure，此后视频至 7.0 s 均为冻结终态。" + }, + "unmet": { + "crank_turned": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "带动本体", + "t": 5.0, + "log_step": null, + "evidence_zh": "观察：视频 3.2 s 左手扶握时机身已偏转，视频 4.6 s 至 5.0 s 右手双指下探触碰曲柄组件时，磨豆机身继续被带斜，叠字倾角由 7.5° 增至 10.2° 并显示 grinder_tipped；旋钮没有产生所需的相对机身两圈转动。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "视频", + "evidence_zh": "推断：视频 4.6 s 至 5.0 s，机器人右手接触旋钮区域时，左手的扶握未能保持机身稳定，接触引起的机身倾斜先于有效曲柄旋转触发失败；因此归为双手接触与稳定控制失败。该判断依据实际接触后的机身运动及原生倾角叠字，不依据 agent 在 log step 473 所称控制器已验证的总结。" + } + ], + "gap_zh": "未提供逐帧状态日志，无法分离左右手的接触力、精确接触部位及腕部跟踪误差；log step 462 的控制器主体在输入日志中已截断，因此不进一步断言错误握法、特定高度参数或逆解误差。inputs/run.json 的 reward 未报告 DAG 所列原生数值指标，native_metrics 的 value 均留 null；视频叠字中的倾角等读数只作为视觉证据。" + }, + "summary_zh": "本次评分局为最后重置后的 Episode 11：机器人左手扶握机身、右手下探接触旋钮，唯一节点“曲柄转满两圈”未达成，最终为 0/1。视频约 5.0 s 机身被带斜至叠字 10.2°，触发 grinder_tipped，曲柄仍约零圈、释放保持 0/50；根因保守归为控制。没有逐帧接触力或状态日志，无法细分接触设计与执行误差。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "节点未达成、断点在 crank_turned 操作中、倾倒终态（视频 5.0 s 叠字 FAIL: grinder_tipped、tilt 10.2°）、接触与根因（控制，已标推断）与帧一致，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": "2026-10-29T10:07:47Z", + "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/dexhand-grinder-crank-i00-standard--2ormf7y/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/dexhand-grinder-crank-i00-standard--2ormf7y/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "049b406b79245d81a8585407efa38dca8ae15a26eb6c07d87b1aa67ba38e2b37", + "video_sha256": "5bfd68bc6f6d33c3883c594391bf70d27477e832a3e0d7c74112963bc4e82f42", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/dexhand/jar_lid_open/task_dag.json b/phase2/dexhand/jar_lid_open/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..d9d5c5fc2033cfddd4ecf74ea8e1653672ab373a --- /dev/null +++ b/phase2/dexhand/jar_lid_open/task_dag.json @@ -0,0 +1,385 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "dexhand/jar_lid_open", + "suite": "dexhand", + "base": "dexhand-jar-lid-open-i00", + "title_zh": "拧开罐盖并放到桌面", + "summary_zh": "右手依次拧开红色罐盖、抬离罐口并将盖直立放在罐旁台面，盖与罐均离手稳定保持 50 个控制步，且全程没有触发失败条件。", + "embodiment_zh": "两个模式均为固定底座的天机 Marvin 双臂机器人，每臂七关节；只用右臂及右侧灵巧手操作，左臂停放且左手张开。每回合选 Sharpa Wave（默认，22 个手指执行器）或 Wuji Hand 2.2（20 个手指执行器），无移动底盘；两种手型共用场景与成功判据。", + "task_class": "compositional", + "analysis_level": "step", + "spec": { + "version": "1.2", + "task": "dexhand-jar-lid-open-i00", + "bench": "dexhand", + "progress": "live", + "entities": {}, + "milestones": [ + { + "id": "lid_unscrewed", + "zh": "拧开罐盖", + "mode": "event", + "when": null, + "after": [], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "红盖仍套在玻璃罐口的螺纹区时，相对罐身从上方看逆时针累计转过至少 1.049342105263158 rad；盖原点在罐坐标系中比闭合高度高出小于 6 mm、距罐轴小于 4 mm，评分器 stage 至少为 1。视频中可看红盖顶面白色标记相对罐身转动，顶栏由 Stage 0/3 变为 1/3；仅握住盖、整罐一起转或未到解锁角不算，达到后即保留完成记录。", + "step": { + "cat": "关系", + "item": "拧" + }, + "contact": [ + { + "who": "手指/夹爪", + "how": "抓握", + "plan": 1, + "note_zh": "右手拇指与其余手指夹住红色罐盖外缘，相对罐身逆时针拧转。" + }, + { + "who": "手指/夹爪", + "how": "压/擦", + "plan": 2, + "note_zh": "右手掌面或指腹压住红色罐盖顶面，以摩擦接触带动盖逆时针转动；随后抬盖所需的抓握是后续接触。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:3", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/2009ba31aa/checker.py:39-40", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/2009ba31aa/checker.py:179-182", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/2009ba31aa/checker.py:222-238", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/a4f4d5a733/scene.py:69-86", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/a4f4d5a733/scene.py:219-220", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/0c73de31b5/runtime.py:58-64", + "inputs/oracle/replay.mp4:约 8-14 秒，红盖相对转动并进入 Stage 1/3" + ], + "conf": 6, + "note_zh": "初始红盖闭合且累计相对转角为零。逆时针拧开是独立语义步骤，原生 stage 1 是其验收；不拆成接近、抓取或旋转若干小段。螺纹由被动接触几何构成，两种手型共享条件。" + }, + { + "id": "lid_on_worktop", + "zh": "罐盖放到桌面", + "mode": "state", + "when": null, + "after": [ + "lid_unscrewed" + ], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "拧开后，盖裙最低点曾比罐口高出严格大于 30 mm（stage 2），然后红盖正面朝上、开口朝下放在罐旁浅色台面，中心距罐轴至少 90 mm，接触台面且不接触任何机器人部件或罐身，倾角不大于 0.15 rad、线速度不大于 0.02 m/s、角速度不大于 0.3 rad/s，同时玻璃罐接触台面、不接触机器人、倾角不大于 0.10 rad、线速度不大于 0.02 m/s，全部连续满足 50 个控制步（50 Hz，约 1 秒）。视频中罐口已露出，红盖位于罐旁桌面并停止移动，手指虽可在附近但须离盖，顶栏 hold 达 50/50、Stage 3/3；仍悬在手中、贴罐、倾斜或未保持满窗口均未完成，之后若被拿离、碰触或失去这些终态条件即不再满足。", + "step": { + "cat": "关系", + "item": "放置" + }, + "contact": [ + { + "who": "手指/夹爪", + "how": "抓握", + "plan": 1, + "note_zh": "右手手指抓住已拧开的红色罐盖，抬离罐口并移到罐旁台面，放稳后松开。" + }, + { + "who": "手指/夹爪", + "how": "托/舀", + "plan": 2, + "note_zh": "右手指腹从下方托住已脱离螺纹的罐盖，移至罐旁台面，盖得到台面承托后抽离手指。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:3", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/2009ba31aa/checker.py:33", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/2009ba31aa/checker.py:41-43", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/2009ba31aa/checker.py:235-238", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/2009ba31aa/checker.py:241-277", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/0c73de31b5/runtime.py:45-46", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/a4f4d5a733/scene.py:315-322", + "inputs/oracle/replay.mp4:约 14-18 秒抬离，约 24-26.2 秒盖落桌并保持至 PASS", + "inputs/oracle/run.json:step 56，stage 3、hold_steps 50、lid_distance_m 0.1394785507；step 67，两次重放成功" + ], + "conf": 6, + "note_zh": "抬离是放置过程中的原生有序验收阶段，不另建节点。罐身原本直立在桌上，保持稳定是此步验收及全程失败约束，不另计保留初态。连续窗口由评分器 hold_steps >= 50 表达；未来记录该计数后无需再施加时间保持，因此 hold_s 为 0，避免重复等待。freeze_held=false 使重新触碰或持盖时不会隐藏放置条件丢失。" + } + ], + "constraints": [ + { + "id": "jar_tipped", + "zh": "罐身倾倒", + "never": null, + "criterion_zh": "任一 2 ms 物理子步中，玻璃罐相对世界竖直方向的倾角大于 15°；首次违反即终止，之后扶正不能撤销。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/2009ba31aa/checker.py:34", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/2009ba31aa/checker.py:185-187" + ] + }, + { + "id": "jar_slid", + "zh": "罐身滑移", + "never": null, + "criterion_zh": "玻璃罐原点相对重置位置的水平位移大于 0.03 m，立即终止。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/2009ba31aa/checker.py:35", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/2009ba31aa/checker.py:188-190" + ] + }, + { + "id": "jar_lifted", + "zh": "罐身抬起", + "never": null, + "criterion_zh": "玻璃罐原点比重置时高出大于 0.015 m，立即终止。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/2009ba31aa/checker.py:36", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/2009ba31aa/checker.py:191-192" + ] + }, + { + "id": "lid_knocked_off", + "zh": "罐盖提前落桌", + "never": null, + "criterion_zh": "尚未达到抬离阶段（stage < 2）时，红盖与工作台发生有效接触，立即终止。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/2009ba31aa/checker.py:211-212" + ] + }, + { + "id": "lid_dropped", + "zh": "罐盖掉落", + "never": null, + "criterion_zh": "任一物理子步，盖质心向下速度大于 1.0 m/s；或盖没有任何其他接触类别的有效接触时，向下速度大于 0.3 m/s，即立即终止，无重新接住的宽限期。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/2009ba31aa/checker.py:37-38", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/2009ba31aa/checker.py:150-161", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/2009ba31aa/checker.py:210-221" + ] + }, + { + "id": "lid_off_table", + "zh": "罐盖离开台面", + "never": null, + "criterion_zh": "盖原点 z < 0.745 m，或 |x| > 0.45 m、|y| > 0.58 m，或盖接触地面，立即终止；评分器也把罐身接触地面记为 lid_off_table。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/2009ba31aa/checker.py:193-209", + "inputs/task/privileged/environment/instance.json:147-151" + ] + }, + { + "id": "jar_off_table", + "zh": "罐身离开台面", + "never": null, + "criterion_zh": "罐原点 z < 0.745 m，或 |x| > 0.45 m、|y| > 0.58 m，立即终止；罐身接触地面同样终止，原生原因名为 lid_off_table。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/2009ba31aa/checker.py:193-209", + "inputs/task/privileged/environment/instance.json:147-151" + ] + }, + { + "id": "object_left_workspace", + "zh": "物体越出工作区", + "never": null, + "criterion_zh": "罐或盖的原点离开闭区间工作空间：x 为 [-0.45, 0.45] m、y 为 [-0.40, 0.40] m、z 为 [0.73, 1.25] m，立即终止。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/2009ba31aa/checker.py:91-92", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/2009ba31aa/checker.py:193-195", + "inputs/task/privileged/environment/instance.json:153-164" + ] + }, + { + "id": "lid_forced_through_thread", + "zh": "罐盖强穿螺纹", + "never": null, + "criterion_zh": "盖从螺纹区向上离开且仍距罐轴小于 0.004 m 时，相对累计转角模 pi/2 不在 [RELEASE_ANGLE - 0.03, STOP_ANGLE + 0.03] rad 内即失败；端点由场景给出，约为 [1.0193421053, 1.2914542215] rad，不能直接向上硬拔。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/2009ba31aa/checker.py:39-44", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/2009ba31aa/checker.py:222-230", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/a4f4d5a733/scene.py:73-86" + ] + }, + { + "id": "left_arm_contact", + "zh": "左臂接触物体", + "never": null, + "criterion_zh": "停放的左臂任一连杆或左手与罐或盖发生有效接触，立即终止；只移动左臂而未接触不触发此检查。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/2009ba31aa/checker.py:81-88", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/2009ba31aa/checker.py:199-207" + ] + }, + { + "id": "unstable_physics", + "zh": "物理失稳", + "never": null, + "criterion_zh": "任一物理子步出现 MuJoCo BADQPOS、BADQVEL 或 BADQACC 警告，即记为物理失稳并立即终止。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/0c73de31b5/runtime.py:23", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/0c73de31b5/runtime.py:101-108" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "lid_unscrewed", + "level": "new_code", + "source": "inputs/grader/efe_dexhand-files/privileged/environment/variants/2009ba31aa/checker.py:111-115,279-295 Checker.state().stage / Checker.features().stage；需逐帧记录 stage，取 stage >= 1；当前无状态日志。" + }, + { + "atom": "lid_on_worktop", + "level": "new_code", + "source": "inputs/grader/efe_dexhand-files/privileged/environment/variants/2009ba31aa/checker.py:241-273 Checker.final_features() 的全部接触、距离、倾角、速度字段，以及 stage >= 2；需逐帧记录，当前无状态日志。" + }, + { + "atom": "placement_hold", + "level": "new_code", + "source": "inputs/grader/efe_dexhand-files/privileged/environment/variants/2009ba31aa/checker.py:274-284 Checker.features().hold_steps；需记录 hold_steps >= 50，原生计数已含连续保持窗口。" + }, + { + "atom": "task_failure", + "level": "new_code", + "source": "inputs/grader/efe_dexhand-files/privileged/environment/variants/2009ba31aa/checker.py:169-239,292-294 Checker.features().failure / failure_detail；inputs/grader/efe_dexhand-files/privileged/environment/variants/0c73de31b5/runtime.py:101-108 记录物理失稳；需逐物理子步或永久失败原因日志，当前无状态日志。" + }, + { + "atom": "task_success", + "level": "new_code", + "source": "inputs/grader/efe_dexhand-files/privileged/environment/variants/2009ba31aa/checker.py:275-277,292 Checker.features().success；需记录原生逐帧成功标志，当前无状态日志。" + } + ] + }, + "native_metrics": [ + { + "name": "max_seated_turn_rad", + "unit": "rad", + "threshold": 1.049342105263158, + "direction": "higher_better" + }, + { + "name": "max_clearance_m", + "unit": "m", + "threshold": 0.03, + "direction": "higher_better" + }, + { + "name": "lid_distance_m", + "unit": "m", + "threshold": 0.09, + "direction": "higher_better" + }, + { + "name": "lid_tilt_rad", + "unit": "rad", + "threshold": 0.15, + "direction": "lower_better" + }, + { + "name": "lid_speed_m_s", + "unit": "m/s", + "threshold": 0.02, + "direction": "lower_better" + }, + { + "name": "lid_spin_rad_s", + "unit": "rad/s", + "threshold": 0.3, + "direction": "lower_better" + }, + { + "name": "jar_tilt_rad", + "unit": "rad", + "threshold": 0.1, + "direction": "lower_better" + }, + { + "name": "jar_speed_m_s", + "unit": "m/s", + "threshold": 0.02, + "direction": "lower_better" + }, + { + "name": "jar_slide_m", + "unit": "m", + "threshold": 0.03, + "direction": "lower_better" + } + ], + "open_questions_zh": [ + "输入 inputs/states/index.json 的 logs 为空，成功示例也没有状态日志；按 P8，两个节点的 when、全部约束的 never 以及 success 均为 null。现有视频和运行记录可支持语义及验收解读，但无法逐物理子步核验接触、瞬时速度、解锁角和历史失败，atoms 仅列需要增加的记录，不冒充已存在的绑定。", + "模式差异：物理 checker.py、runtime.py、scene.py 两个模式逐字节相同；privileged 的 grade.py:114-123 要求两次独立重放均成功且最终哈希一致，standard 的 grade.py:23-32 采用已终结实时回合的成功值。DAG 表达共同物理任务，归档合法性、确定性和服务交付不作为语义步骤或世界状态。", + "文本与评分器差异：instruction.md:3 要求转盖时不要带转罐身，且左臂保持停放；checker.py:185-207 只对罐倾斜、水平位移、抬起及左臂接触作失败检查，没有罐身纯偏航旋转或左臂单纯离开停放姿态的独立阈值。按 N1 记录差异，不臆造转角阈值或阶段。", + "P14：接触方案 1 为右手抓握；方案 2 分别为压住盖面借摩擦拧转、从下方托盖搬放，作为任务几何允许的明显替代接触，未由当前这一条成功示例验证；不要求示例使用的特定手指组合。", + "连续保持由原生 hold_steps >= 50 验收，hold_s=0 避免再次计时；若将来只记录逐控制步原始放置条件而不记录计数，需按 D31 改为 49 × 0.02 = 0.98 秒并核对采样频率。max_clearance_m 的 0.03 m 阈值为严格大于，其余终态最大值为小于等于、最小距离为大于等于。" + ], + "conf": 6, + "conf_reason_zh": "两个语义步骤与文本及 S1 的明确案例一致，必要顺序、所有验收阈值、起态和失败规则均有代码依据，成功视频及记录覆盖两步；缺少全部状态日志，不能执行世界状态绑定和重放一致性自检，替代接触未实证，因此置信度为 6。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "两个节点（拧开罐盖→放到罐旁台面）与指令的开罐、放置两步对应，「Lift clear」已折入第二节点的 criterion，未单列；success 为 null（无状态日志，P8），11 个约束与评分器失败条件一致且均为立即终止，边、标签、open_questions 合理。无需改动。", + "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：['lid_unscrewed', 'lid_on_worktop']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:50:53Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/dexhand__jar_lid_open/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/dexhand__jar_lid_open/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "1bf88932f7cdbf41c1108214bf7838bac0bc06bae6bf438f5dbc9105b4340137", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/dexhand/jar_lid_open/trajectories/dexhand-jar-lid-open-i00-standard--s96sbbw.json b/phase2/dexhand/jar_lid_open/trajectories/dexhand-jar-lid-open-i00-standard--s96sbbw.json new file mode 100644 index 0000000000000000000000000000000000000000..eff15921e0180334fb14a9b397e187d60fe393d5 --- /dev/null +++ b/phase2/dexhand/jar_lid_open/trajectories/dexhand-jar-lid-open-i00-standard--s96sbbw.json @@ -0,0 +1,269 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "dexhand-jar-lid-open-i00-standard--s96sbbw", + "taskKey": "dexhand/jar_lid_open", + "dag_sha256": "b2060ab237a25379dd35ea85a182cf91836d35380185b84c082a6e96311fd522", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "calls": 706, + "deterministic": 1, + "finalized": 1, + "hand_wuji": 0, + "infra_error": 0, + "live_matches_replay": 1, + "live_success": 0, + "n_actions": 2497, + "no_verdict": 0, + "replay_success": 0, + "reset_budget": 20, + "resets": 10, + "resume_state_mismatch": 0, + "resumes": 0, + "sim_error": 0, + "sim_startups": 1, + "sim_steps": 2497, + "success": 0, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.65, + "peak_processes": 11, + "peak_disk_gb": 0.993, + "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/dexhand-jar-lid-open-i00-standard--s96sbbw/replay.mp4", + "duration_s": 52.2, + "fps": 5.0, + "frames": 261, + "sha256": "6ca7b310b4c1f7572b3c505e8ffa081acbb9e36f3f3c7ec4fe59d4303e4d5113" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "lid_unscrewed": { + "method": "visual", + "intervals": [ + [ + 33.2, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察：视频 0.0 s 红盖闭合在玻璃罐口，顶栏 Stage 0/3、lid turn 0.00 rad。" + }, + { + "t": 33.2, + "log_step": null, + "text_zh": "观察：视频 30.0–33.2 s 机器人右手夹住红盖拧转，白色标记随盖相对罐身转动；视频 33.0 s 顶栏仍为 Stage 0/3、lid turn 1.03 rad，视频 33.2 s 首次显示 Stage 1/3、lid turn 1.06 rad，红盖仍在罐口。此后该事件保持已完成。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "抓握", + "note_zh": "观察：视频 21.0–33.2 s，机器人右手拇指与弯曲的其余手指从红盖外缘相向夹住，随后拧转红盖；此前视频 12.0–20.0 s 手指在盖缘逐渐接近、调整接触。" + } + ] + }, + "lid_on_worktop": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 38.2, + "log_step": null, + "text_zh": "观察：视频 36.0–38.2 s 机器人右手将红盖向上抬离罐口，盖与罐口之间出现清楚间隙；视频 38.2 s 顶栏 Stage 2/3、clearance 36 mm、hold 0/50，抬离阶段已达到，但盖仍由手指夹持。" + }, + { + "t": 41.2, + "log_step": null, + "text_zh": "观察：视频 40.6 s 红盖尚大致水平；视频 41.2–42.0 s 右手搬运时，红盖在指间明显滑转，由水平转为接近竖直、悬垂在腕部下方，罐留在原处。" + }, + { + "t": 50.0, + "log_step": 209, + "text_zh": "观察：视频 44.0–49.8 s 机器人右臂向罐旁桌面下降，盖逐渐被腕臂遮挡，顶栏始终 Stage 2/3、hold 0/50；视频 50.0 s 顶栏显示 FAIL: lid_dropped，并保持到视频 52.0 s 最后一帧。log step 209 的服务终结状态为 episode 11、control_steps 2497、task_failure；评分器 live_matches_replay=1、replay_success=0。没有达到放置所需的 Stage 3/3 和连续保持窗口。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "抓握", + "note_zh": "观察：视频 36.0–40.6 s，机器人右手以拇指和其余弯曲手指对向夹持红盖外缘，抬离罐口并开始搬运；视频 41.2–42.0 s 盖在这只手中滑转、下垂，仍局部挂在手指处。末段腕臂遮挡了具体脱离接触的瞬间。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "max_seated_turn_rad", + "unit": "rad", + "threshold": 1.049342105263158, + "direction": "higher_better", + "value": null + }, + { + "name": "max_clearance_m", + "unit": "m", + "threshold": 0.03, + "direction": "higher_better", + "value": null + }, + { + "name": "lid_distance_m", + "unit": "m", + "threshold": 0.09, + "direction": "higher_better", + "value": null + }, + { + "name": "lid_tilt_rad", + "unit": "rad", + "threshold": 0.15, + "direction": "lower_better", + "value": null + }, + { + "name": "lid_speed_m_s", + "unit": "m/s", + "threshold": 0.02, + "direction": "lower_better", + "value": null + }, + { + "name": "lid_spin_rad_s", + "unit": "rad/s", + "threshold": 0.3, + "direction": "lower_better", + "value": null + }, + { + "name": "jar_tilt_rad", + "unit": "rad", + "threshold": 0.1, + "direction": "lower_better", + "value": null + }, + { + "name": "jar_speed_m_s", + "unit": "m/s", + "threshold": 0.02, + "direction": "lower_better", + "value": null + }, + { + "name": "jar_slide_m", + "unit": "m", + "threshold": 0.03, + "direction": "lower_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "lid_on_worktop", + "t": 41.2, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察：视频 38.2 s 机器人已将红盖抬离罐口，正在持盖搬向桌面；视频 41.2–42.0 s 盖开始在指间滑转并下垂，之后未见恢复稳定水平夹持或完成放置。视频 50.0 s 评分顶栏显示 FAIL: lid_dropped、Stage 2/3、hold 0/50；log step 209 的第 11 回合以 task_failure 终结。断点取搬运中首次未修复的夹持失稳。" + }, + "unmet": { + "lid_unscrewed": true, + "lid_on_worktop": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "没保持住", + "t": 41.2, + "log_step": null, + "evidence_zh": "观察：视频 38.2–40.6 s 红盖随机器人右手被平稳抬起；视频 41.2–42.0 s 在右腕姿态没有相应大幅翻转时，盖自行在指间滑转至接近竖直并向下悬垂。视频 50.0 s 顶栏最终报 lid_dropped。这里是已建立夹持后未保持住盖的稳定承载，末段具体掉落接触被腕臂遮挡。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "视频", + "evidence_zh": "推断：视频 40.6–42.0 s 机器人右腕姿态基本不变而盖相对手指滑转、下垂，说明实际夹持控制没有继续约束盖的姿态。log step 195 的 pressure 脚本由指尖雅可比和预设刚度计算指关节偏置，log step 202 将名义压力参数由 2 改为 8 后运行最后回合，log step 203 的后续脚本依次下降和张指释放；log step 205 打印的是脚本计算的目标力，并非触觉实测。结合视频中的失稳及视频 50.0 s 的掉盖判定，归为保持和释放控制不足；现有证据不能可靠细分为接触布局、实际夹持力或关节跟踪问题。" + } + ], + "gap_zh": "没有状态日志或逐帧接触、实测指尖力及盖倾角数据。视频约 44.0 s 后红盖被腕臂遮挡，无法直接看到松指时盖与台面的全部接触关系，不能断言是悬空提前释放、夹持残留后甩落还是其他具体脱离机制；掉盖终结由本次视频顶栏及 log step 209 的服务状态确认。评分器 reward 未报告 DAG 的九项原生连续指标，因此其 value 均为 null；未将视频顶栏瞬时读数或 agent 自行估计填作评分器指标。" + }, + "summary_zh": "评分对应最后的第 11 回合（log step 203 重置，log step 209 终结）；机器人右手在视频 33.2 s 拧开红盖，并在视频 38.2 s 抬离罐口，最终完成 1/2 个语义节点。视频 41.2–42.0 s 搬运时盖在指间滑转下垂，此后未完成稳定放置，视频 50.0 s 以 lid_dropped 判负，现象为没保持住，根因归为控制。末段腕臂遮挡及缺少接触、实测力日志，使掉盖的具体释放机制和控制子类仍不确定。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "节点值、时间、接触、断点和现象与视频顶栏及帧一致：视频 33.2 s 进入 Stage 1/3，38.2 s Stage 2/3 抬离，41.2–42.0 s 盖在指间滑转下垂，50.0 s 显示 FAIL: lid_dropped。根因为推断且已标注，不确定性已写入 gap_zh，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": "2026-10-29T10:06:36Z", + "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/dexhand-jar-lid-open-i00-standard--s96sbbw/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/dexhand-jar-lid-open-i00-standard--s96sbbw/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "53f5bf408bf45f438d86da3ecd0954282d0480c04374a000a6f79e7178583567", + "video_sha256": "6ca7b310b4c1f7572b3c505e8ffa081acbb9e36f3f3c7ec4fe59d4303e4d5113", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/dexhand/jar_lid_open/trajectories/dexhand-jar-lid-open-i00-standard--vdr3ybj.json b/phase2/dexhand/jar_lid_open/trajectories/dexhand-jar-lid-open-i00-standard--vdr3ybj.json new file mode 100644 index 0000000000000000000000000000000000000000..8e67471719328fe9ec23b5a507e227a135c66178 --- /dev/null +++ b/phase2/dexhand/jar_lid_open/trajectories/dexhand-jar-lid-open-i00-standard--vdr3ybj.json @@ -0,0 +1,244 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "dexhand-jar-lid-open-i00-standard--vdr3ybj", + "taskKey": "dexhand/jar_lid_open", + "dag_sha256": "b2060ab237a25379dd35ea85a182cf91836d35380185b84c082a6e96311fd522", + "outcome": { + "success": 1, + "reward": 1.0, + "source": "catalog", + "raw": { + "calls": 9, + "deterministic": 1, + "finalized": 1, + "hand_wuji": 0, + "infra_error": 0, + "live_matches_replay": 1, + "live_success": 1, + "n_actions": 580, + "no_verdict": 0, + "replay_success": 1, + "reset_budget": 20, + "resets": 0, + "resume_state_mismatch": 0, + "resumes": 0, + "sim_error": 0, + "sim_startups": 1, + "sim_steps": 580, + "success": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.439, + "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/dexhand-jar-lid-open-i00-standard--vdr3ybj/replay.mp4", + "duration_s": 13.8, + "fps": 5.0, + "frames": 69, + "sha256": "b7f4354dc7ec0484f43f78a03dcae4a01b6b988817cd5809dd0c35e9d3a5cae0" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 2, + "final": 2, + "nodes": { + "lid_unscrewed": { + "method": "visual", + "intervals": [ + [ + 5.6, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 5.6, + "log_step": null, + "text_zh": "观察：视频 5.4 s 红盖仍在罐口，机器人右手以拇指与对侧手指夹住盖缘，顶栏为 Stage 0/3、lid turn 1.03 rad；视频 5.6 s 顶栏变为 Stage 1/3、lid turn 1.08 rad，白色盖面标记随右手转动，罐身仍直立在原处。首次达到时间定位在视频 5.4–5.6 s 之间，取首个确认帧视频 5.6 s。" + }, + { + "t": 6.6, + "log_step": null, + "text_zh": "观察：视频 6.4 s 盖与罐口已出现间隙，视频 6.6 s 顶栏为 Stage 2/3、clearance 46 mm，右手继续夹持红盖，罐留在桌上。拧开节点为已完成即保留的事件，随后搬放时转角变化不撤销该事件。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "抓握", + "note_zh": "观察：视频约 3.0–5.6 s，机器人右手拇指与对侧食指、中指在红盖外缘形成对向夹持，随腕部转动拧开盖；没有使用左手固定罐身。" + } + ] + }, + "lid_on_worktop": { + "method": "visual", + "intervals": [ + [ + 11.6, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 6.8, + "log_step": null, + "text_zh": "观察：视频 6.8 s，机器人右手夹持红盖悬在罐口上方，罐口已露出，顶栏 Stage 2/3、clearance 63 mm；随后视频约 7.0–10.4 s，右手把盖移到罐旁并降低至浅色台面。" + }, + { + "t": 11.6, + "log_step": null, + "text_zh": "观察：视频 10.6–11.4 s 右手手指逐渐张开并抬离，红盖正面朝上留在罐旁台面；视频 11.4 s 顶栏为 Stage 2/3、hold 40/50，视频 11.6 s 首次显示 Stage 3/3、hold 50/50 和 PASS。此时盖与罐分离，罐直立，右手离盖，按评分器完整保持窗口的确认帧取视频 11.6 s 为完成时刻。" + }, + { + "t": 13.6, + "log_step": null, + "text_zh": "观察：视频 12.8、13.0、13.2、13.4、13.6 s 的末秒画面均保持同一终态，顶栏持续 PASS、Stage 3/3、hold 50/50；红盖在桌面，右手悬于其上，玻璃罐直立。inputs/run.json 的 reward 同时报告 live_success=1、replay_success=1、live_matches_replay=1，且 resets=0、resumes=0。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "抓握", + "note_zh": "观察：视频约 6.2–10.6 s，机器人右手拇指与食指、中指对向夹住红盖外缘，抬离罐口并搬到罐旁桌面；随后张指松开并抬手，盖留在台面。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "max_seated_turn_rad", + "unit": "rad", + "threshold": 1.049342105263158, + "direction": "higher_better", + "value": null + }, + { + "name": "max_clearance_m", + "unit": "m", + "threshold": 0.03, + "direction": "higher_better", + "value": null + }, + { + "name": "lid_distance_m", + "unit": "m", + "threshold": 0.09, + "direction": "higher_better", + "value": null + }, + { + "name": "lid_tilt_rad", + "unit": "rad", + "threshold": 0.15, + "direction": "lower_better", + "value": null + }, + { + "name": "lid_speed_m_s", + "unit": "m/s", + "threshold": 0.02, + "direction": "lower_better", + "value": null + }, + { + "name": "lid_spin_rad_s", + "unit": "rad/s", + "threshold": 0.3, + "direction": "lower_better", + "value": null + }, + { + "name": "jar_tilt_rad", + "unit": "rad", + "threshold": 0.1, + "direction": "lower_better", + "value": null + }, + { + "name": "jar_speed_m_s", + "unit": "m/s", + "threshold": 0.02, + "direction": "lower_better", + "value": null + }, + { + "name": "jar_slide_m", + "unit": "m", + "threshold": 0.03, + "direction": "lower_better", + "value": null + } + ], + "failure": null, + "summary_zh": "观察：机器人右手夹持红盖相对罐身拧转，视频约 5.6 s 达到拧开节点，继而抬离并把盖平放在罐旁台面，视频约 11.6 s 达到放置节点，顶栏显示 Stage 3/3、hold 50/50、PASS，最终完成 2/2 个节点。观察：末秒画面持续保持盖已放稳且离手，评分记录实时与重放均成功，没有失败断点。证据缺口：没有状态日志，节点时刻取自视频相邻帧及顶栏，精度约 0.2 s；评分字段未报告 DAG 所列九项原生连续指标，均保留 null，无法独立逐物理子步核验接触和瞬时速度。", + "conf": 9, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "成功运行，两个节点时刻（视频 5.6 s Stage 1/3；11.6 s Stage 3/3、hold 50/50、PASS）与顶栏一致，接触标注合理，无失败断点，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": "2026-10-29T10:04:13Z", + "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/dexhand-jar-lid-open-i00-standard--vdr3ybj/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/dexhand-jar-lid-open-i00-standard--vdr3ybj/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "12a5a500f1a5352a0cce7548b687176795c1aea7b2d01b6b7c03d2a7adccd26e", + "video_sha256": "b7f4354dc7ec0484f43f78a03dcae4a01b6b988817cd5809dd0c35e9d3a5cae0", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/dexhand/pancake_serve/task_dag.json b/phase2/dexhand/pancake_serve/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..cfa200e3dbb61a1dcc2fba041e7b2b8836b44e27 --- /dev/null +++ b/phase2/dexhand/pancake_serve/task_dag.json @@ -0,0 +1,388 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "dexhand/pancake_serve", + "suite": "dexhand", + "base": "dexhand-pancake-serve-i00", + "title_zh": "用锅铲盛煎饼", + "summary_zh": "用手持锅铲从煎锅托起煎饼，依序移至餐盘上方并平稳装盘撤铲，静置达标且全程不掉落、不碰饼、不扰动锅盘或越界。", + "embodiment_zh": "两模式均为固定底座的天机 Marvin 双臂机器人，每臂7关节并装一只灵巧手；默认 Sharpa Wave（每手22指关节执行器），也支持 Wuji Hand 2.2（每手20执行器），右臂和右手操作锅铲，左臂停在躯干旁，无移动底盘。锅铲为自由工具，铲片有三个被动弹簧铰链，靠手指及掌面接触握持。", + "task_class": "compositional", + "analysis_level": "step", + "spec": { + "version": "1.2", + "task": "dexhand-pancake-serve-i00", + "bench": "dexhand", + "progress": "live", + "entities": {}, + "milestones": [ + { + "id": "turner_in_hand", + "zh": "拿起锅铲", + "mode": "event", + "when": null, + "after": [], + "hold_s": 0, + "freeze_held": false, + "goal": false, + "criterion_zh": "总览画面中，灵巧手手指合拢握住黑色锅铲柄，锅铲离开其开局位于白盘前方的台面，并随手移动、由手承持。手仅靠近或压住仍躺在台面上的锅铲，不算完成。", + "step": { + "cat": "位姿", + "item": "拿起" + }, + "contact": [ + { + "who": "手指/夹爪", + "how": "抓握", + "plan": 1, + "note_zh": "右侧灵巧手用拇指和其余手指握住锅铲柄并提起锅铲。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "task/privileged/instruction.md:3 先拿起台面上的锅铲，只有锅铲可移动煎饼", + "task/privileged/instruction.md:21-23 锅铲初始未在手中，仅由接触握持", + "task/privileged/environment/instance.json:65-244 两种手型的冻结初始位姿", + "grader/efe_dexhand-files/privileged/environment/variants/24a9fdfcda/runtime.py:57-66 按冻结位姿复位，没有抓取焊接", + "grader/efe_dexhand-files/privileged/environment/variants/282560417a/checker.py:190-198 必须使用铲片独立承托煎饼", + "grader/efe_dexhand-files/privileged/environment/variants/282560417a/checker.py:214 手与锅铲的原生接触特征", + "oracle/replay.mp4 @ 0.0-3.5 s 台面上的锅铲被右手握住并提离" + ], + "conf": 6, + "note_zh": "按 Y17 与 DECISIONS.md D2 保留必需工具拿起步骤：文本明确要求先拿起，锅铲初始自由放在台面且未安装在机器人上，评分器强制铲片独立承托煎饼；该工具必须用于所有合法路线。手柄握法采用典型方案，不限定手指数或特定指尖。此步为过程事件，不要求结束时仍把锅铲举在空中；评分器没有单独的拿起阶段。" + }, + { + "id": "pancake_on_blade", + "zh": "煎饼托上铲", + "mode": "event", + "when": null, + "after": [ + "turner_in_hand" + ], + "hold_s": 0.18, + "freeze_held": false, + "goal": true, + "criterion_zh": "总览中，黑色铲片从下方托住金黄色煎饼，煎饼完全离开黑色煎锅且只接触铲片所属的锅铲，不接触锅、盘、机器人、台面或地面，饼原点比复位高度升高至少0.015 m、相对水平倾角不超过30°，连续10个50 Hz控制步成立，画面顶部的 on blade 勾选后即通过。煎饼仍贴锅底或锅沿、只是铲片伸到饼边下方，均未完成；这是通过后锁存的事件，之后煎饼装盘不撤销本步。", + "step": { + "cat": "位姿", + "item": "拿起" + }, + "contact": [ + { + "who": "手持工具", + "how": "托/舀", + "plan": 1, + "note_zh": "右手持锅铲，以铲片从下方托住并抬起煎饼；手指不直接接触煎饼。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "task/privileged/instruction.md:3 滑入铲片并将煎饼从锅中托起", + "grader/efe_dexhand-files/privileged/environment/variants/282560417a/checker.py:100-118 有效接触与铲片识别", + "grader/efe_dexhand-files/privileged/environment/variants/282560417a/checker.py:180-198 独立铲片接触、升高和倾角", + "grader/efe_dexhand-files/privileged/environment/variants/282560417a/checker.py:227-230 连续10控制步后阶段锁存", + "task/privileged/environment/instance.json:56-58 搬运倾角30度、控制间隔0.02秒", + "task/privileged/environment/instance.json:364 升高至少0.015米", + "task/privileged/environment/instance.json:376 连续10控制步", + "oracle/replay.mp4 @ 6.0-6.5 s 煎饼随铲片离锅且 on blade 勾选" + ], + "conf": 6, + "note_zh": "文本独立要求煎饼托上铲，托稳后可暂停再搬运，故与装盘分开；直接产生变化的是用工具将煎饼拿起，接触为铲片托/舀，不另列手握工具。hold_s按 DECISIONS.md D31 取(10-1)×0.02=0.18秒；无状态日志，实际验收仍由 criterion_zh 的10步原生窗口或视频阶段标记判定。" + }, + { + "id": "pancake_on_plate", + "zh": "煎饼装盘", + "mode": "state", + "when": null, + "after": [ + "pancake_on_blade" + ], + "hold_s": 0.98, + "freeze_held": false, + "goal": true, + "criterion_zh": "在托上铲阶段通过之后，先让煎饼仍只接触锅铲、饼心距白盘心水平不超过0.058 m且饼原点高于盘原点加0.006 m，再让煎饼只接触白盘、饼心仍在该水平半径内、相对水平倾角不超过20°、质心线速度不超过0.02 m/s且角速度不超过0.3 rad/s，同时锅铲不接触饼且评分器采样的铲片轮廓点距饼心水平距离全部至少0.08 m，整组装盘条件连续50个50 Hz控制步成立才完成。画面中金黄色煎饼平落在右侧白盘内、黑色铲片完全撤到饼旁并保持安静，顶部 hold 到50/50且 PASS；饼仍搭在铲片上、超出盘内允许位置或铲片未撤开都不算完成，达标后若这些状态丢失则本步失去。", + "step": { + "cat": "关系", + "item": "放置" + }, + "contact": [ + { + "who": "手持工具", + "how": "托/舀", + "plan": 1, + "note_zh": "右手持锅铲，以铲片托住煎饼送到白色陶瓷盘并降低，再从煎饼下抽出铲片。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "task/privileged/instruction.md:3 盘上方、装盘、撤铲与静置验收", + "grader/efe_dexhand-files/privileged/environment/variants/282560417a/checker.py:65-72 铲片轮廓实际采样点", + "grader/efe_dexhand-files/privileged/environment/variants/282560417a/checker.py:185-202 距离、盘上方、倾角、速度、撤铲及仅盘接触", + "grader/efe_dexhand-files/privileged/environment/variants/282560417a/checker.py:227-239 原生顺序与50控制步成功窗口", + "task/privileged/environment/instance.json:58 控制间隔0.02秒", + "task/privileged/environment/instance.json:63 静置50控制步", + "task/privileged/environment/instance.json:377-386 饼半径0.07米、盘半径0.128米、速度与装盘阈值", + "task/privileged/environment/instance.json:402 撤铲边缘间隙0.01米，与饼半径合为0.08米", + "oracle/replay.mp4 @ 7.5-10.0 s 煎饼进入白盘，撤铲，hold到50/50" + ], + "conf": 6, + "note_zh": "原生 over_plate 只标记装盘过程中的必经中间状态，不单列步骤；该时刻要求 z_pancake > z_plate+0.006，装盘时不再要求悬空。装盘的50步静置与撤铲均并入验收，hold_s按 D31 取(50-1)×0.02=0.98秒；原生计数器一旦单独记录，绑定计数器时无需重复时间窗。freeze_held=false避免手仍握锅铲时遮蔽装盘达成或丢失，机器人不得直接持饼。" + } + ], + "constraints": [ + { + "id": "pancake_dropped", + "zh": "煎饼掉落", + "never": null, + "criterion_zh": "煎饼接触台面或地面，或没有任何有效接触且质心竖直速度小于-0.35 m/s，或煎饼原点低于z=0.85 m，即立即失败且不可撤销。", + "consequence": "立即终止", + "sources": [ + "grader/efe_dexhand-files/privileged/environment/variants/282560417a/checker.py:140-158", + "task/privileged/environment/instance.json:58-63 控制步长、台面高度与掉落阈值", + "task/privileged/environment/instance.json:387-419 支承物阈值和工作区", + "grader/efe_dexhand-files/privileged/environment/variants/24a9fdfcda/runtime.py:103-106 每1毫秒物理子步检查，无宽限期", + "task/privileged/environment/instance.json:378 物理步长0.001秒" + ] + }, + { + "id": "robot_touches_pancake", + "zh": "机器人碰饼", + "never": null, + "criterion_zh": "煎饼与任意机器人几何体发生有效接触即立即失败，包括左右手、臂连杆和支架；手不能直接扶住或夹住煎饼。", + "consequence": "立即终止", + "sources": [ + "grader/efe_dexhand-files/privileged/environment/variants/282560417a/checker.py:43-60,153-154", + "task/privileged/environment/instance.json:58-63 控制步长、台面高度与掉落阈值", + "task/privileged/environment/instance.json:387-419 支承物阈值和工作区", + "grader/efe_dexhand-files/privileged/environment/variants/24a9fdfcda/runtime.py:103-106 每1毫秒物理子步检查，无宽限期", + "task/privileged/environment/instance.json:378 物理步长0.001秒" + ] + }, + { + "id": "turner_dropped", + "zh": "锅铲掉落", + "never": null, + "criterion_zh": "锅铲碰地，或锅铲与机器人没有有效接触时质心竖直速度小于-0.35 m/s，或锅铲原点低于z=0.85 m，即立即失败；初始静放台面不构成掉落。", + "consequence": "立即终止", + "sources": [ + "grader/efe_dexhand-files/privileged/environment/variants/282560417a/checker.py:147-160", + "task/privileged/environment/instance.json:58-63 控制步长、台面高度与掉落阈值", + "task/privileged/environment/instance.json:387-419 支承物阈值和工作区", + "grader/efe_dexhand-files/privileged/environment/variants/24a9fdfcda/runtime.py:103-106 每1毫秒物理子步检查，无宽限期", + "task/privileged/environment/instance.json:378 物理步长0.001秒" + ] + }, + { + "id": "pan_disturbed", + "zh": "煎锅移位倾倒", + "never": null, + "criterion_zh": "黑色煎锅原点相对复位位置的三维位移超过0.03 m，或其上轴相对复位上轴倾斜超过15°，即立即失败；绕竖直轴的纯自转不等于倾倒。", + "consequence": "立即终止", + "sources": [ + "grader/efe_dexhand-files/privileged/environment/variants/282560417a/checker.py:124-127,161-167", + "task/privileged/environment/instance.json:58-63 控制步长、台面高度与掉落阈值", + "task/privileged/environment/instance.json:387-419 支承物阈值和工作区", + "grader/efe_dexhand-files/privileged/environment/variants/24a9fdfcda/runtime.py:103-106 每1毫秒物理子步检查，无宽限期", + "task/privileged/environment/instance.json:378 物理步长0.001秒" + ] + }, + { + "id": "plate_disturbed", + "zh": "餐盘移位倾倒", + "never": null, + "criterion_zh": "白色陶瓷盘原点相对复位位置的三维位移超过0.03 m，或其上轴相对复位上轴倾斜超过15°，即立即失败；盘中煎饼的倾角阈值另属装盘验收。", + "consequence": "立即终止", + "sources": [ + "grader/efe_dexhand-files/privileged/environment/variants/282560417a/checker.py:124-127,161-167", + "task/privileged/environment/instance.json:58-63 控制步长、台面高度与掉落阈值", + "task/privileged/environment/instance.json:387-419 支承物阈值和工作区", + "grader/efe_dexhand-files/privileged/environment/variants/24a9fdfcda/runtime.py:103-106 每1毫秒物理子步检查，无宽限期", + "task/privileged/environment/instance.json:378 物理步长0.001秒" + ] + }, + { + "id": "object_outside_workspace", + "zh": "物体越界", + "never": null, + "criterion_zh": "煎饼、煎锅、餐盘或锅铲任一主体原点离开闭工作区x∈[0.23,1.13] m、y∈[-0.65,0.45] m、z∈[0.80,1.50] m即立即失败，检查的是主体原点而非所有网格顶点。", + "consequence": "立即终止", + "sources": [ + "grader/efe_dexhand-files/privileged/environment/variants/282560417a/checker.py:63-74,168-171", + "task/privileged/environment/instance.json:58-63 控制步长、台面高度与掉落阈值", + "task/privileged/environment/instance.json:387-419 支承物阈值和工作区", + "grader/efe_dexhand-files/privileged/environment/variants/24a9fdfcda/runtime.py:103-106 每1毫秒物理子步检查，无宽限期", + "task/privileged/environment/instance.json:378 物理步长0.001秒" + ] + }, + { + "id": "unstable_physics", + "zh": "物理失稳", + "never": null, + "criterion_zh": "物理引擎报告位置、速度或加速度异常警告，或者任意广义位置或速度出现非有限值，即立即失败；仅凭视频抖动不能精确确定此项。", + "consequence": "立即终止", + "sources": [ + "grader/efe_dexhand-files/privileged/environment/variants/282560417a/checker.py:134-139", + "task/privileged/environment/instance.json:58-63 控制步长、台面高度与掉落阈值", + "task/privileged/environment/instance.json:387-419 支承物阈值和工作区", + "grader/efe_dexhand-files/privileged/environment/variants/24a9fdfcda/runtime.py:103-106 每1毫秒物理子步检查，无宽限期", + "task/privileged/environment/instance.json:378 物理步长0.001秒" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "turner_in_hand", + "level": "new_code", + "source": "grader/efe_dexhand-files/privileged/environment/variants/282560417a/checker.py:214 需记录 hand_holds_spatula，并记录锅铲离开台面且随手承持的状态；本任务没有状态日志" + }, + { + "atom": "pancake_on_blade", + "level": "new_code", + "source": "grader/efe_dexhand-files/privileged/environment/variants/282560417a/checker.py:198,227-230 需记录 on_blade、on_blade_run 与 stage；本任务没有状态日志" + }, + { + "atom": "pancake_on_plate", + "level": "new_code", + "source": "grader/efe_dexhand-files/privileged/environment/variants/282560417a/checker.py:199-202,231-239 需记录 over_plate、served、stage、hold，包含完整顺序与撤铲、静止验收；本任务没有状态日志" + }, + { + "atom": "task_failure", + "level": "new_code", + "source": "grader/efe_dexhand-files/privileged/environment/variants/282560417a/checker.py:130-177 需逐物理子步记录 failure、failure_detail 和失败时刻；覆盖全部七项约束" + }, + { + "atom": "task_success", + "level": "new_code", + "source": "grader/efe_dexhand-files/privileged/environment/variants/282560417a/checker.py:222-239 需记录原生 success 与 failure；成功为无既往失败且依次通过 on_blade、over_plate、served的50步窗口" + } + ] + }, + "native_metrics": [ + { + "name": "pancake_lift_m", + "unit": "m", + "threshold": 0.015, + "direction": "higher_better" + }, + { + "name": "pancake_tilt_deg", + "unit": "deg", + "threshold": 30, + "direction": "lower_better" + }, + { + "name": "pancake_plate_dist_m", + "unit": "m", + "threshold": 0.058, + "direction": "lower_better" + }, + { + "name": "served_pancake_tilt_deg", + "unit": "deg", + "threshold": 20, + "direction": "lower_better" + }, + { + "name": "pancake_speed_m_s", + "unit": "m/s", + "threshold": 0.02, + "direction": "lower_better" + }, + { + "name": "pancake_ang_speed_rad_s", + "unit": "rad/s", + "threshold": 0.3, + "direction": "lower_better" + }, + { + "name": "blade_clearance_m", + "unit": "m", + "threshold": 0.08, + "direction": "higher_better" + }, + { + "name": "pan_shift_m", + "unit": "m", + "threshold": 0.03, + "direction": "lower_better" + }, + { + "name": "plate_shift_m", + "unit": "m", + "threshold": 0.03, + "direction": "lower_better" + } + ], + "open_questions_zh": [ + "inputs/states/index.json 的 logs 为空；按 DECISIONS.md P8，全部节点 when、全部约束 never 和 success 均为 null，不能声称已用状态日志验证达到全部步骤或评分一致。原生字段只列为待新增记录的原子，精确接触、毫米级阈值和速度需依靠原生画面计数及后续状态记录核实。", + "DECISIONS.md D2/Y17 的本任务应用：保留拿起锅铲；指令明确要求先从台面拿起，评分器 on_blade 必须接触真实铲片且只接触锅铲（checker.py:190-198），禁止机器人直接碰饼（153-154），初态与视频均显示工具未在手中，因此所有遵守文本的成功路线都需要该工具。checker.py:214 虽计算手铲接触，但成功谓词未独立要求锅铲离台并在手，故本节点 goal=false，不把典型握法变成额外成功门槛。", + "文本步骤未受评：拿起锅铲没有独立的评分阶段或离台高度判据；按 N1 保留为视频判断的非目标过程步骤。这里既不把一般物体抓取扩成节点，也不把仅接触锅铲等同于完成拿起。", + "DECISIONS.md D10/S2 的粒度读法：煎饼托上铲与煎饼装盘分开，可托稳后暂停；原生 over_plate 是装盘的中间验收而不另建节点，撤铲与50步静止也并入装盘。煎饼托上铲按其直接产生的离锅升高状态标位姿·拿起，工具接触标手持工具×托/舀。", + "文本与实现的精度差异：指令称铲片轮廓每一点均须撤出0.08 m，而 checker.py:65-72,192-197 只检查每条铲片的四个角点和一个前缘中点，并非整条轮廓的连续最小距离；装盘标准采用实现的采样点检查，0.08 m来自0.07 m饼半径加0.01 m间隙。", + "两模式场景 checker.py 逐字相同；privileged/tests/grade.py:114-119 还要求两次新进程回放均成功且终态哈希相同，standard/tests/grade.py:22-37 评分服务最后一次实时回合并检查终结与基础设施状态。这些交付有效性条件不拆成世界状态节点；success 无日志可绑定而置 null。", + "成功视频长10.6秒，画面终态显示模拟时间13.66秒，不能把视频秒直接当作模拟秒；10步和50步窗口按控制步及画面原生计数判定，hold_s分别用 D31 的0.18和0.98秒，不按该视频播放速度估计。" + ], + "conf": 6, + "conf_reason_zh": "步骤来自文本，初态、阶段顺序、阈值和即时失败均有源代码依据，成功视频可见全部三步；但没有任何状态日志可核实条件绑定和细微接触，工具拿起的Y17应用及轮廓采样与文字差异需明确保留。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "三步（拿起锅铲/煎饼托上铲/煎饼装盘）与指令三阶段一致，over_plate、撤铲和静置并入装盘，阈值（0.015 m、30°、10步、0.058 m、20°、0.02 m/s、0.3 rad/s、0.08 m、50步）与checker一致；七项约束为立即终止；无状态日志，when/success为null并已在open_questions_zh说明。", + "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：['turner_in_hand', 'pancake_on_blade', 'pancake_on_plate']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:50:30Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/dexhand__pancake_serve/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/dexhand__pancake_serve/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "bf734c9818b5ead55baa27eccc2ca9b41453cb631fd2a908092ac14007d3fac4", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/dexhand/pancake_serve/trajectories/dexhand-pancake-serve-i00-privileged--wxe6lrk.json b/phase2/dexhand/pancake_serve/trajectories/dexhand-pancake-serve-i00-privileged--wxe6lrk.json new file mode 100644 index 0000000000000000000000000000000000000000..caa5d8c61f3cad11409e94bd639c251f376b5190 --- /dev/null +++ b/phase2/dexhand/pancake_serve/trajectories/dexhand-pancake-serve-i00-privileged--wxe6lrk.json @@ -0,0 +1,267 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "dexhand-pancake-serve-i00-privileged--wxe6lrk", + "taskKey": "dexhand/pancake_serve", + "dag_sha256": "88a62bb839c4f72417c49b00bde10e86ccf1f60461bb3d1dbe0d25557e9b5275", + "outcome": { + "success": 1, + "reward": 1.0, + "source": "catalog", + "raw": { + "deterministic": 1, + "hand_wuji": 0, + "infra_error": 0, + "n_actions": 2699, + "sim_steps": 2699, + "stages_completed": 3, + "success": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 6.225, + "peak_processes": 13, + "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/dexhand-pancake-serve-i00-privileged--wxe6lrk/replay.mp4", + "duration_s": 13.48, + "fps": 25.0, + "frames": 337, + "sha256": "32034b651695c042aa6e497d7d9587ce66f50b3111d438466e7b5a5b13e3bbcd" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 3, + "base": 0, + "peak": 3, + "final": 3, + "nodes": { + "turner_in_hand": { + "method": "visual", + "intervals": [ + [ + 2.48, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察：视频 0.00 s，黑色锅铲平放在白盘前方台面，机器人右手张开，煎饼仍在黑锅内。" + }, + { + "t": 2.48, + "log_step": null, + "text_zh": "观察：视频 2.24–2.72 s，机器人右手手指围拢黑色铲柄并抬手，锅铲离开台面、随右手升起；视频 2.48 s 已可见铲片与台面分离，随后右手持铲移向黑锅。拿起时刻约为视频 2.5 s。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "抓握", + "note_zh": "观察：机器人右侧灵巧手手指合拢包住黑色锅铲柄，握持并提离台面；左手未参与。" + } + ] + }, + "pancake_on_blade": { + "method": "visual", + "intervals": [ + [ + 8.56, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 8.48, + "log_step": null, + "text_zh": "观察：视频 8.08–8.48 s，机器人右手持铲从锅内向上托起煎饼，顶部 on blade 尚未勾选；仅铲片进入饼下不提前记作完成。" + }, + { + "t": 8.56, + "log_step": null, + "text_zh": "观察：视频 8.56 s，煎饼由黑色铲片承托并离开锅底，顶部 on blade 首次在已检查帧中勾选（视频 8.48 s 尚未勾选）；视频 8.64–9.50 s，煎饼继续随铲抬高、移向白盘，黑锅原处已无煎饼。原生勾选确认独立铲片承托、升高、倾角及连续10控制步验收；本节点通过后锁存。" + } + ], + "contact_actual": [ + { + "who": "手持工具", + "how": "托/舀", + "note_zh": "观察：机器人右手持黑色锅铲，铲片滑到煎饼下方并从下方托起煎饼；未见手直接扶饼。" + } + ] + }, + "pancake_on_plate": { + "method": "visual", + "intervals": [ + [ + 12.44, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 10.0, + "log_step": null, + "text_zh": "观察：视频 10.00–11.50 s，机器人右手持铲把已离锅的煎饼送到白盘上方并降低，顶部 over plate 已勾选；铲片仍在饼下时 hold 为0/50，装盘尚未验收。" + }, + { + "t": 12.44, + "log_step": null, + "text_zh": "观察：视频 12.08–12.40 s，机器人向左后方抽出铲片，煎饼留在白盘内，hold 从0/50增至42/50；视频 12.44 s 顶部显示 PASS、served 勾选和 hold 50/50，饼平放盘内、铲片已撤开，盘心距离显示2.4 cm。此处采用视频内评分器原生计数验收，不把加速视频的播放时长当作仿真静置时长。" + }, + { + "t": 13.44, + "log_step": null, + "text_zh": "观察：视频 12.48、12.80、13.00、13.20、13.44 s 均保持相同成功终态，煎饼仍在盘内且铲片在旁，末尾未见装盘关系丢失。" + }, + { + "t": null, + "log_step": 113, + "text_zh": "观察：log step 106、113 的两次独立新进程交付文件回放报告均为 success=true、stage=3、2699个控制步，终态哈希相同；inputs/run.json 评分器字段 success=1、stages_completed=3、deterministic=1，与视频 PASS 一致。" + } + ], + "contact_actual": [ + { + "who": "手持工具", + "how": "托/舀", + "note_zh": "观察：机器人右手以锅铲从下方承托煎饼送至白盘并降低，随后抽出铲片，让煎饼留在白盘上静置。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "pancake_lift_m", + "unit": "m", + "threshold": 0.015, + "direction": "higher_better", + "value": null + }, + { + "name": "pancake_tilt_deg", + "unit": "deg", + "threshold": 30, + "direction": "lower_better", + "value": null + }, + { + "name": "pancake_plate_dist_m", + "unit": "m", + "threshold": 0.058, + "direction": "lower_better", + "value": null + }, + { + "name": "served_pancake_tilt_deg", + "unit": "deg", + "threshold": 20, + "direction": "lower_better", + "value": null + }, + { + "name": "pancake_speed_m_s", + "unit": "m/s", + "threshold": 0.02, + "direction": "lower_better", + "value": null + }, + { + "name": "pancake_ang_speed_rad_s", + "unit": "rad/s", + "threshold": 0.3, + "direction": "lower_better", + "value": null + }, + { + "name": "blade_clearance_m", + "unit": "m", + "threshold": 0.08, + "direction": "higher_better", + "value": null + }, + { + "name": "pan_shift_m", + "unit": "m", + "threshold": 0.03, + "direction": "lower_better", + "value": null + }, + { + "name": "plate_shift_m", + "unit": "m", + "threshold": 0.03, + "direction": "lower_better", + "value": null + } + ], + "failure": null, + "summary_zh": "机器人依序完成拿起锅铲（约视频 2.5 s）、煎饼托上铲（视频 8.56 s 原生标记通过）及装盘撤铲静置（视频 12.44 s PASS、hold 50/50），末尾视频 13.44 s 保持成功终态，全部3个节点最终成立。交付文件的两次新进程回放均成功且终态哈希相同（log step 106、113），与评分器 success=1 一致；会话中恢复快照后的失败调试不计入交付视频的失败。无状态日志，细微接触和速度阈值依赖画面原生验收标记，评分器 reward 未报告的9项原生数值指标均保留为null；画面仿真时钟与视频时钟不同，节点时刻均取自视频帧。", + "conf": 9, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "三个节点（拿起锅铲约视频 2.5 s、托饼上铲约 8.6 s、装盘 12.44 s PASS）与帧和画面原生勾选一致，成功运行，无失败层，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": "2026-10-29T10:08:43Z", + "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/dexhand-pancake-serve-i00-privileged--wxe6lrk/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/dexhand-pancake-serve-i00-privileged--wxe6lrk/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "b85109a49620f0f6197ba6e2da838794f0744b2acfa218f50946af47abce4b48", + "video_sha256": "32034b651695c042aa6e497d7d9587ce66f50b3111d438466e7b5a5b13e3bbcd", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/dexhand/pancake_serve/trajectories/dexhand-pancake-serve-i00-standard--mhaxhhl.json b/phase2/dexhand/pancake_serve/trajectories/dexhand-pancake-serve-i00-standard--mhaxhhl.json new file mode 100644 index 0000000000000000000000000000000000000000..a8b5f71fe03adde9f02e6595d1a42bb48edae999 --- /dev/null +++ b/phase2/dexhand/pancake_serve/trajectories/dexhand-pancake-serve-i00-standard--mhaxhhl.json @@ -0,0 +1,260 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "dexhand-pancake-serve-i00-standard--mhaxhhl", + "taskKey": "dexhand/pancake_serve", + "dag_sha256": "88a62bb839c4f72417c49b00bde10e86ccf1f60461bb3d1dbe0d25557e9b5275", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "calls": 99, + "deterministic": 1, + "finalized": 1, + "hand_wuji": 0, + "infra_error": 0, + "live_matches_replay": 1, + "live_success": 0, + "n_actions": 500, + "no_verdict": 0, + "replay_success": 0, + "reset_budget": 20, + "resets": 7, + "resume_state_mismatch": 0, + "resumes": 0, + "sim_error": 0, + "sim_startups": 1, + "sim_steps": 500, + "success": 0, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.533, + "peak_processes": 12, + "peak_disk_gb": 1.026, + "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/dexhand-pancake-serve-i00-standard--mhaxhhl/replay.mp4", + "duration_s": 11.4, + "fps": 5.0, + "frames": 57, + "sha256": "bb9280c5551759edd745331843991e76b8b9a9c9b809bef26c87579f2314f8a2" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 3, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "turner_in_hand": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 2.4, + "log_step": null, + "text_zh": "观察：视频 1.6–2.2 s，机器人右手屈指靠到黑色铲柄旁；视频 2.4–2.8 s，锅铲在台面上转向、向右滑移，未见整件锅铲离台由手承持；视频 3.0–3.4 s，右手空手抬离，锅铲留在台面。" + }, + { + "t": 11.2, + "log_step": 180, + "text_zh": "观察：视频 3.4–11.2 s，锅铲一直留在台面前部，没有再次被拿起。agent 在 log step 180 自述“成功拿起并抬高12 cm”，但该说法不符合本评分视频，不能据此判定本节点达成。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "推/扫", + "note_zh": "观察：视频 2.2–2.8 s，机器人右手屈曲手指接触、拨动黑色锅铲柄，使锅铲在台面转向滑移；未见手指从相对两侧夹紧铲柄并提离台面，不标抓握。" + } + ] + }, + "pancake_on_blade": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 5.8, + "log_step": null, + "text_zh": "观察：逐帧检查视频 0.0–11.2 s，金黄色煎饼始终留在黑色煎锅内；视频 3.4–6.4 s，机器人右手空手移向锅边，黑色锅铲仍在台面前部，铲片未进入煎饼下方，未实际开始工具托饼。顶部 on blade 始终未勾选。" + } + ] + }, + "pancake_on_plate": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 11.2, + "log_step": 175, + "text_zh": "观察：视频 6.6–8.0 s，机器人空手向白盘一侧移动后撤回；视频 10.2–11.2 s 的全部尾帧中，白盘仍空、煎饼仍在锅内、锅铲仍在台面，顶部 hold 为0/50且没有PASS。log step 175 的服务返回显示最后评分回合为 episode 8、执行500个控制步、done=False、reason=None；没有煎饼实际搬运装盘。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "pancake_lift_m", + "unit": "m", + "threshold": 0.015, + "direction": "higher_better", + "value": null + }, + { + "name": "pancake_tilt_deg", + "unit": "deg", + "threshold": 30, + "direction": "lower_better", + "value": null + }, + { + "name": "pancake_plate_dist_m", + "unit": "m", + "threshold": 0.058, + "direction": "lower_better", + "value": null + }, + { + "name": "served_pancake_tilt_deg", + "unit": "deg", + "threshold": 20, + "direction": "lower_better", + "value": null + }, + { + "name": "pancake_speed_m_s", + "unit": "m/s", + "threshold": 0.02, + "direction": "lower_better", + "value": null + }, + { + "name": "pancake_ang_speed_rad_s", + "unit": "rad/s", + "threshold": 0.3, + "direction": "lower_better", + "value": null + }, + { + "name": "blade_clearance_m", + "unit": "m", + "threshold": 0.08, + "direction": "higher_better", + "value": null + }, + { + "name": "pan_shift_m", + "unit": "m", + "threshold": 0.03, + "direction": "lower_better", + "value": null + }, + { + "name": "plate_shift_m", + "unit": "m", + "threshold": 0.03, + "direction": "lower_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "turner_in_hand", + "t": 2.8, + "log_step": 175, + "where": "操作中", + "evidence_zh": "观察：视频 2.2–2.8 s，机器人右手使锅铲在台面转动，但未将其抓起；视频 3.0 s 起右手空手上抬，锅铲留在原工作台前部直到视频 11.2 s，没有补做成功。最早未修复的是拿起锅铲，而非后续托饼或装盘。log step 175 记录执行最后的 episode 8；评分器字段 n_actions=500、live_matches_replay=1，与该回合相符。" + }, + "unmet": { + "pancake_on_blade": false, + "pancake_on_plate": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "没建立接触", + "t": 2.8, + "log_step": null, + "evidence_zh": "观察：视频 1.6–3.2 s，机器人右手屈指靠在铲柄一侧，锅铲在台面滑转，随后右手空手抬起，锅铲没有随手离台；所需的夹持承重接触未建立。这里确有拨动接触，但不等于建立拿起所需的抓持，未见先拿稳再滑落的过程。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "抓取和接触设计", + "alt_feasible": "已验证有", + "source_zh": "视频 / 日志 / agent 自述", + "evidence_zh": "推断：右手选定的屈指笼式接触未把铲柄夹在相对接触面之间，产生拨转而不能承重；依据视频 2.2–3.2 s 的台面滑转及空手抬离。log step 146 的脚本预设俯掌笼式握法，log step 175 只微调抓取腕位后执行最后回合；agent 在 log step 180 也承认该接触笼不能可靠夹持。可行替代仅由成功参考确认：同任务 oracle 视频 2.0–3.5 s 显示同型右手拿起并持续承持锅铲，参考不是本回合的进度证据。" + } + ], + "gap_zh": "观察：本任务没有状态日志；仅有5 fps总览视频，无法核实手柄背面的全部接触点及接触力。agent 在 log step 180 给出的握隙与柄尺寸均属自述、待核实，且其“成功拿起后侧移滑落”的叙述与本评分视频不符；本标签以逐帧可见的台面拨转和空手抬离为准，未采用该叙述或把早期测试算入进度。评分器未报告DAG所列九项原生连续指标，故其值均为null。" + }, + "summary_zh": "观察：评分回合中，机器人右手只将锅铲在台面拨转，随后空手完成后续移动；未达到拿起锅铲、煎饼托上铲或煎饼装盘，终态为0/3。推断：断在拿起锅铲，屈指笼式接触没有形成承重抓持；煎饼始终留锅、白盘始终空。观察：没有状态日志，agent关于已经拿起锅铲的自述不符合本回合视频，精确接触力与握隙无法独立核实。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "三个节点均未达成、终态0/3与帧一致：锅铲始终留在台面，煎饼留锅，白盘空；断点在拿起锅铲，接触/抓取设计的归因有视频支持。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": "2026-10-29T10:07:44Z", + "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/dexhand-pancake-serve-i00-standard--mhaxhhl/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/dexhand-pancake-serve-i00-standard--mhaxhhl/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "349aab808e497af3ff3460fb2aa431554a29b2d983ebf6d970d1015ef205bd32", + "video_sha256": "bb9280c5551759edd745331843991e76b8b9a9c9b809bef26c87579f2314f8a2", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/dexhand/remote_battery_swap/task_dag.json b/phase2/dexhand/remote_battery_swap/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..8da12de25966b3f4216c6170c32dfab31e144985 --- /dev/null +++ b/phase2/dexhand/remote_battery_swap/task_dag.json @@ -0,0 +1,575 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "dexhand/remote_battery_swap", + "suite": "dexhand", + "base": "dexhand-remote-battery-swap-i00", + "title_zh": "更换遥控器电池", + "summary_zh": "依序开盖、将灰色旧电池放入托盘、把绿色新电池按正确极性装入槽并关盖扣锁，最终所有任务物体离手、所需位置和静止条件共同连续保持 50 个控制步且全程无失败。", + "embodiment_zh": "两模式均为固定底座的天机 Marvin 机器人，左右各七关节机械臂并装灵巧手；仅右臂和右手操作，左臂及左手停放且不得触碰任务物体，底座不移动；可用 Sharpa Wave（每手 22 个手指执行器，默认）或 Wuji Hand 2.2（每手 20 个），任务几何及判据相同。", + "task_class": "compositional", + "analysis_level": "step", + "spec": { + "version": "1.2", + "task": "dexhand-remote-battery-swap-i00", + "bench": "dexhand", + "progress": "live", + "entities": {}, + "milestones": [ + { + "id": "battery_cover_opened", + "zh": "打开电池盖", + "mode": "event", + "when": null, + "after": [], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "总览中黑色滑盖向遥控器尾端伸出，原先被遮住的灰色电池及仓口露出；盖的滑动开度至少 0.060 m（全行程 0.085 m）才算打开，锁片仅被按下或小幅露缝均不算。此步骤一经通过即保留，随后关盖不撤销。", + "step": { + "cat": "机构", + "item": "打开" + }, + "contact": [ + { + "who": "手指/夹爪", + "how": "压/擦", + "note_zh": "右手指腹压住滑盖弹性锁片，同时靠摩擦将滑盖向尾端滑开", + "plan": 1 + }, + { + "who": "手指/夹爪", + "how": "抓握", + "note_zh": "右手拇指压住锁片，其他手指夹住滑盖边缘向尾端拉开", + "plan": 2 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:3", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:221-226", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/e30cc06716/task.json:272", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/cc9b3e00d4/scene.py:188-205", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/93b12f9846/runtime.py:63-72", + "inputs/oracle/replay.mp4 @ 0-6 s", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:256-271", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/e30cc06716/task.json:49" + ], + "conf": 6, + "note_zh": "文本明确要求的开盖步骤；按锁片解锁和滑开是同一机构开启动作，不拆出按压节点。评分器最终要求 stages_latched.cover_open 为真，因此列为 goal 事件；关盖后无需继续保持开度。" + }, + { + "id": "old_cell_in_tray", + "zh": "旧电池入托盘", + "mode": "state", + "when": null, + "after": [ + "battery_cover_opened" + ], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "总览可见灰色旧电池从黑色仓内转到右侧白色托盘底面，右手已不接触它；中心在托盘坐标系内满足 |x|<=0.037 m、|y|<=0.032 m、z<=0.014 m，并确实接触托盘底板，仅悬在盘上或搁在盘沿不算完成，离开上述位置或重新被手接触即失去。最终联合验收还要求四阶段已依序锁存、两节电池及关盖条件仍成立，所有任务物体无机器人接触，遥控器、托盘和两节电池的线速度均不超过 0.02 m/s、角速度均不超过 0.5 rad/s，连续 50 个控制步；该跨步骤窗口归入整体成功，不单列节点。", + "step": { + "cat": "关系", + "item": "放入" + }, + "contact": [ + { + "who": "手指/夹爪", + "how": "抓握", + "note_zh": "右手两根或多根手指捏住灰色旧电池，移至白色托盘内后放开", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:3", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:184-206", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:256-271", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/e30cc06716/task.json:59-62", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/e30cc06716/task.json:275", + "inputs/oracle/replay.mp4 @ 8-17 s" + ], + "conf": 6, + "note_zh": "从仓中取出与移入托盘合为一次转移，以托盘内落底并离手为验收；取出没有独立评分目标（DECISIONS.md D5）。视频中盘沿上的电池还需拨入，后续右手对灰色电池的推拨是同一步的修正；不把特定捏取手指或修正动作写成必经步骤。" + }, + { + "id": "new_cell_seated", + "zh": "新电池入槽", + "mode": "state", + "when": null, + "after": [ + "old_cell_in_tray" + ], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "开盖时可见绿色电池沿遥控器长轴卧在原灰色电池的位置，正极朝金色触点，右手离开电池；相对槽中心沿槽、跨槽、高度误差分别不超过 0.005、0.003、0.002 m，正极方向相对遥控器 +x 夹角不超过 10 度，且同时接触两条支承肋，反向、斜搭、未落到底或仍被手接触均不算，随后移位或重新被手接触会失去。最终联合验收还要求四阶段已依序锁存、两节电池及关盖条件仍成立，所有任务物体无机器人接触，遥控器、托盘和两节电池的线速度均不超过 0.02 m/s、角速度均不超过 0.5 rad/s，连续 50 个控制步；该跨步骤窗口归入整体成功，不单列节点。", + "step": { + "cat": "关系", + "item": "插入" + }, + "contact": [ + { + "who": "手指/夹爪", + "how": "抓握", + "note_zh": "右手手指捏住绿色新电池，校正朝向并放入电池槽后放开", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:3", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:191-217", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:256-271", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/e30cc06716/task.json:276-279", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/cc9b3e00d4/scene.py:148-173", + "inputs/oracle/replay.mp4 @ 18-24 s" + ], + "conf": 6, + "note_zh": "电池是进入有严格配合公差的槽，标关系·插入；极性、双肋接触、离手为验收，不另建对准、翻转或释放节点。盖合上后会遮住电池，可结合盖合前的清晰画面及原生阶段叠字判定是否持续成立。" + }, + { + "id": "battery_cover_closed", + "zh": "关上电池盖", + "mode": "state", + "when": null, + "after": [ + "new_cell_seated" + ], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "黑色滑盖滑回覆盖整个电池仓、绿色电池不再露出，弹性锁片回弹；开度<=0.002 m且锁片按深<=0.0015 m才算关上并锁住，留缝、滑盖伸出或锁片仍压下超过阈值即不成立。此状态开局已成立，开盖后失去，最终关盖恢复；全部工作完成后还须满足共同离手静止窗口：最终联合验收还要求四阶段已依序锁存、两节电池及关盖条件仍成立，所有任务物体无机器人接触，遥控器、托盘和两节电池的线速度均不超过 0.02 m/s、角速度均不超过 0.5 rad/s，连续 50 个控制步；该跨步骤窗口归入整体成功，不单列节点。", + "step": { + "cat": "机构", + "item": "关上" + }, + "contact": [ + { + "who": "手指/夹爪", + "how": "压/擦", + "note_zh": "右手指腹按住滑盖锁片并沿滑轨向仓口方向擦推，随后离开使锁片回弹", + "plan": 1 + }, + { + "who": "手指/夹爪", + "how": "抓握", + "note_zh": "右手拇指压下锁片、其余手指夹持滑盖边缘推回，随后放开滑盖和锁片", + "plan": 2 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:3", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:221-226", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:256-271", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/e30cc06716/task.json:273-274", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/cc9b3e00d4/scene.py:188-205", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/93b12f9846/runtime.py:63-72", + "inputs/oracle/replay.mp4 @ 0 s, 24-29 s" + ], + "conf": 6, + "note_zh": "初态 cover_slide=0、tab_depth=0，按 S4(b) 是从起点计分的恢复步骤，不能改成 final。state 不受 after 门控；after 仅表达第三阶段后最终关盖的原生顺序，阶段锁存历史及联合保持归整体成功；hold_s=0 不重复计入 50 步窗口。" + } + ], + "constraints": [ + { + "id": "object_below_table", + "zh": "物体低于桌面", + "never": null, + "criterion_zh": "每个 0.002 s 物理子步检查：旧电池、新电池、滑盖、遥控器或托盘任一刚体原点 z<0.30 m 即失败。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/instruction.md:3", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:113-120" + ] + }, + { + "id": "object_outside_workspace", + "zh": "物体越界", + "never": null, + "criterion_zh": "旧电池、新电池、滑盖、遥控器或托盘任一刚体原点离开 x∈[-0.37,0.47] m、y∈[-0.34,0.34] m、z∈[0.30,0.80] m 的闭区间工作箱即失败。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:117-122", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/e30cc06716/task.json:254-265" + ] + }, + { + "id": "object_touches_floor", + "zh": "物体触地", + "never": null, + "criterion_zh": "遥控器（含滑盖及锁片子体）、托盘或任一电池与地面产生有效接触即失败。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:47-61", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:138-145" + ] + }, + { + "id": "remote_displaced", + "zh": "遥控器移位", + "never": null, + "criterion_zh": "遥控器刚体原点相对重置位置的水平位移超过 0.03 m 即失败，不是沿某一个轴分别比较。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:92-95", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:123-125", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/e30cc06716/task.json:266" + ] + }, + { + "id": "remote_tipped", + "zh": "遥控器倾斜", + "never": null, + "criterion_zh": "遥控器局部上轴相对世界竖直的倾角超过 15 度即失败。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:109-110", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:126-128", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/e30cc06716/task.json:267" + ] + }, + { + "id": "tray_displaced", + "zh": "托盘移位", + "never": null, + "criterion_zh": "托盘刚体原点相对重置位置的水平位移超过 0.03 m 即失败。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:92-95", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:129-131", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/e30cc06716/task.json:268" + ] + }, + { + "id": "tray_tipped", + "zh": "托盘倾斜", + "never": null, + "criterion_zh": "托盘局部上轴相对世界竖直的倾角超过 15 度即失败。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:109-110", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:132-134", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/e30cc06716/task.json:269" + ] + }, + { + "id": "old_cell_touches_table", + "zh": "旧电池触桌", + "never": null, + "criterion_zh": "灰色旧电池一旦与桌面产生有效接触即失败，即使还被手抓住也不豁免。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:135-153" + ] + }, + { + "id": "new_cell_returns_to_table", + "zh": "新电池再触桌", + "never": null, + "criterion_zh": "绿色新电池在某控制步满足中心比初始高度高出超过 0.01 m 且不接触桌面后，任何后续物理子步再次触桌即失败。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:154-155", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:250-255", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/e30cc06716/task.json:271" + ] + }, + { + "id": "cell_falls_without_hand", + "zh": "电池无手快落", + "never": null, + "criterion_zh": "任一电池没有机器人接触且竖直速度小于 -sqrt(2*9.81*0.04) m/s（约 -0.886 m/s）即失败，即便接触其他表面也不豁免。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:77", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:156-160", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/e30cc06716/task.json:270" + ] + }, + { + "id": "cell_falls_unsupported", + "zh": "电池悬空快落", + "never": null, + "criterion_zh": "任一电池完全没有有效接触且竖直速度小于 -0.5 m/s 即失败；没有等待或补接宽限期，后来接住不能撤销失败。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:138-151", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:161-164", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/e30cc06716/task.json:283" + ] + }, + { + "id": "left_arm_contact", + "zh": "左臂接触任务物", + "never": null, + "criterion_zh": "停放的左肩至腕的臂链或左手接触遥控器、滑盖、锁片、托盘或任一电池即失败；移动左臂本身未被此检查禁止。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:17-28", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:63-65", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:138-142" + ] + }, + { + "id": "unstable_physics", + "zh": "物理失稳", + "never": null, + "criterion_zh": "物理子步出现 MuJoCo BADQPOS、BADQVEL 或 BADQACC 警告，运行时立即以 unstable_physics 结束；仅凭画面中的正常摆动不能断定该约束被违反。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/93b12f9846/runtime.py:125-134" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "battery_cover_open", + "level": "new_code", + "source": "待在逐帧/子步适配器记录 Checker.features.conditions.cover_open；来源 inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:221-226；当前没有状态日志，未假设已有字段路径。" + }, + { + "atom": "old_cell_in_tray", + "level": "new_code", + "source": "待在逐帧/子步适配器记录 Checker.features.conditions.old_cell_in_tray；来源 inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:199-206；当前没有状态日志，未假设已有字段路径。" + }, + { + "atom": "new_cell_seated", + "level": "new_code", + "source": "待在逐帧/子步适配器记录 Checker.features.conditions.new_cell_seated；来源 inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:207-217；当前没有状态日志，未假设已有字段路径。" + }, + { + "atom": "battery_cover_closed", + "level": "new_code", + "source": "待在逐帧/子步适配器记录 Checker.features.conditions.cover_closed；来源 inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:225-226；当前没有状态日志，未假设已有字段路径。" + }, + { + "atom": "ordered_stages", + "level": "new_code", + "source": "待在逐帧/子步适配器记录 Checker.features.stages_latched and stage_steps；来源 inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:256-275；当前没有状态日志，未假设已有字段路径。" + }, + { + "atom": "released_rest_hold", + "level": "new_code", + "source": "待在逐帧/子步适配器记录 Checker.features.hand_contacts, max_speed_m_s, max_spin_rad_s, hold_steps；来源 inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:263-276；当前没有状态日志，未假设已有字段路径。" + }, + { + "atom": "task_success", + "level": "new_code", + "source": "待在逐帧/子步适配器记录 Session.features.success；来源 inputs/grader/efe_dexhand-files/privileged/environment/variants/93b12f9846/runtime.py:97-99；当前没有状态日志，未假设已有字段路径。" + }, + { + "atom": "task_failure", + "level": "new_code", + "source": "待在逐帧/子步适配器记录 Session.failure, failure_detail, new_cell_lifted and substep violation record；来源 inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:113-165; inputs/grader/efe_dexhand-files/privileged/environment/variants/93b12f9846/runtime.py:125-139；当前没有状态日志，未假设已有字段路径。" + } + ] + }, + "native_metrics": [ + { + "name": "cover_slide_m_open", + "unit": "m", + "threshold": 0.06, + "direction": "higher_better", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:222; inputs/grader/efe_dexhand-files/privileged/environment/variants/e30cc06716/task.json:272" + ] + }, + { + "name": "cover_slide_m_closed", + "unit": "m", + "threshold": 0.002, + "direction": "lower_better", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:226; inputs/grader/efe_dexhand-files/privileged/environment/variants/e30cc06716/task.json:273" + ] + }, + { + "name": "tab_depth_m_closed", + "unit": "m", + "threshold": 0.0015, + "direction": "lower_better", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:226; inputs/grader/efe_dexhand-files/privileged/environment/variants/e30cc06716/task.json:274" + ] + }, + { + "name": "old_cell_tray_local_x_abs", + "unit": "m", + "threshold": 0.037, + "direction": "lower_better", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:202-205", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/e30cc06716/task.json:59-62" + ] + }, + { + "name": "old_cell_tray_local_y_abs", + "unit": "m", + "threshold": 0.032, + "direction": "lower_better", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:202-205", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/e30cc06716/task.json:59-62" + ] + }, + { + "name": "old_cell_tray_local_z", + "unit": "m", + "threshold": 0.014, + "direction": "lower_better", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:203-205", + "inputs/grader/efe_dexhand-files/privileged/environment/variants/e30cc06716/task.json:275" + ] + }, + { + "name": "new_cell_slot_error_x_abs", + "unit": "m", + "threshold": 0.005, + "direction": "lower_better", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:214-216; inputs/grader/efe_dexhand-files/privileged/environment/variants/e30cc06716/task.json:276" + ] + }, + { + "name": "new_cell_slot_error_y_abs", + "unit": "m", + "threshold": 0.003, + "direction": "lower_better", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:214-216; inputs/grader/efe_dexhand-files/privileged/environment/variants/e30cc06716/task.json:277" + ] + }, + { + "name": "new_cell_slot_error_z_abs", + "unit": "m", + "threshold": 0.002, + "direction": "lower_better", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:214-216; inputs/grader/efe_dexhand-files/privileged/environment/variants/e30cc06716/task.json:278" + ] + }, + { + "name": "new_cell_axis_angle_deg", + "unit": "deg", + "threshold": 10, + "direction": "lower_better", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:207-216; inputs/grader/efe_dexhand-files/privileged/environment/variants/e30cc06716/task.json:279" + ] + }, + { + "name": "max_speed_m_s", + "unit": "m/s", + "threshold": 0.02, + "direction": "lower_better", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:219; inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:267; inputs/grader/efe_dexhand-files/privileged/environment/variants/e30cc06716/task.json:280" + ] + }, + { + "name": "max_spin_rad_s", + "unit": "rad/s", + "threshold": 0.5, + "direction": "lower_better", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:220; inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:268; inputs/grader/efe_dexhand-files/privileged/environment/variants/e30cc06716/task.json:281" + ] + }, + { + "name": "remote_displacement_m", + "unit": "m", + "threshold": 0.03, + "direction": "lower_better", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:123-125" + ] + }, + { + "name": "remote_tilt_deg", + "unit": "deg", + "threshold": 15, + "direction": "lower_better", + "sources": [ + "inputs/grader/efe_dexhand-files/privileged/environment/variants/5655757645/checker.py:126-128" + ] + } + ], + "open_questions_zh": [ + "无状态日志：inputs/states/index.json 的 logs 为空，成功样例也没有状态日志。依 DECISIONS.md P8，全部 when、never 和 success 均为 null，需据视频判定；毫米级公差、真实接触、极性角度和子步跌落无法仅凭总览精确复算，原生特征列为待记录的 new_code，而非虚构现有字段。", + "按 DECISIONS.md D5／CASES.md OPEN-5，将“旧电池出仓并入托盘”合为旧电池入托盘：评分器只独立检查托盘内落底离手，没有独立出仓目标；取出、搬运和末端拨正均在这一步内。", + "评分器要求，文本未提：第三阶段 new_cell_seated 已要求绿色电池无机器人接触（checker.py:214-216），任务段落在第四阶段后才明确要求全局松手；将此作为新电池入槽的验收，不添加释放节点（DECISIONS.md N1）。", + "任务设计问题：文本称每个任务物体静止，但 checker.py:13、219-220、263-268 仅计算遥控器、托盘、旧电池和新电池四个自由体原点速度，未单独约束滑盖和锁片的关节速度；DAG 按代码描述实际静止验收，关盖几何条件仍须持续满足。", + "两模式的 checker.py 与 task.json 内容相同，物理成功条件相同；privileged 评分还要求两次新进程回放均成功且终态哈希一致（privileged/tests/grade.py:113-123），standard 取服务最终现场回合结果（standard/tests/grade.py:22-35）。交付格式、确定性、服务终结和基础设施有效性不作为世界状态节点或约束；无可用日志时 success 保持 null。", + "共同成功条件精确为：四阶段按序锁存，old_cell_in_tray、new_cell_seated、cover_closed 当前均真，hand_contacts 总和为零，四自由体最大线速度<=0.02 m/s、最大角速度<=0.5 rad/s，连续 50 次控制后 hold_steps>=50（checker.py:256-271）。窗口跨越三个末态步骤，归整体 success；节点 hold_s=0，避免关盖在初态的计分被全局松手条件掩盖；未来有日志应优先记录并比较原生 hold_steps，而非另估时长。", + "原视频叠字为 3.2 倍速（inputs/oracle/replay.mp4），物理连续 50 控制步的一秒不能用一秒视频长度直接替代。现有视频末帧显示 PASS、Stages 4/4、released hold 50/50；仅用这次成功说明可达状态和视觉线索，不将其手指选择、拨正动作或时间点作为必经路线。" + ], + "conf": 6, + "conf_reason_zh": "四个步骤来自两模式相同的任务文本，阶段顺序、阈值、初态及失败规则都有源码依据，成功视频显示各阶段及最终通过；没有任何状态日志，精细接触和阈值只能使用视频与原生叠字判断，不能进行逐帧绑定和回放一致性验证。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "四个步骤（开盖、旧电池入托盘、新电池入槽、关盖锁定）与任务文本一致，顺序边符合评分器的阶段锁存，最终 50 步松手静止窗口归入整体成功，失败条件均为立即终止。无状态日志，when/never/success 为 null 符合 P8。旧电池出仓并入托盘按 D5 合并，已列入 open_questions_zh。未发现需要修改内容的错误。小提醒：前三个节点的 criterion_zh 重复了整段最终联合验收文字，略冗长，属措辞问题，不要求修改。", + "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：['battery_cover_opened', 'old_cell_in_tray', 'new_cell_seated', 'battery_cover_closed']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:51:40Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/dexhand__remote_battery_swap/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/dexhand__remote_battery_swap/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "46b7545e632de525f598881c9d7d0254516796cc8af90b0f2a88e4f311c958db", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/dexhand/remote_battery_swap/trajectories/dexhand-remote-battery-swap-i00-privileged--dncwpct.json b/phase2/dexhand/remote_battery_swap/trajectories/dexhand-remote-battery-swap-i00-privileged--dncwpct.json new file mode 100644 index 0000000000000000000000000000000000000000..a1e9f14ab6cb59c8d297281747b2c646b2fb50c5 --- /dev/null +++ b/phase2/dexhand/remote_battery_swap/trajectories/dexhand-remote-battery-swap-i00-privileged--dncwpct.json @@ -0,0 +1,345 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "dexhand-remote-battery-swap-i00-privileged--dncwpct", + "taskKey": "dexhand/remote_battery_swap", + "dag_sha256": "448d38d5df99705f24cec0642f2338ce7100127136b831930f8d1fee7f80f537", + "outcome": { + "success": 1, + "reward": 1.0, + "source": "catalog", + "raw": { + "deterministic": 1, + "hand_wuji": 0, + "infra_error": 0, + "n_actions": 1015, + "sim_steps": 1015, + "stages_completed": 4, + "success": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 2.231, + "peak_processes": 8, + "peak_disk_gb": 0.25, + "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/dexhand-remote-battery-swap-i00-privileged--dncwpct/replay.mp4", + "duration_s": 22.4, + "fps": 10.0, + "frames": 224, + "sha256": "698e0a5bb9f58572e4e50d223dd756e4fb78b28b46d10cb1e0334093d93c8c75" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 4, + "base": 1, + "peak": 4, + "final": 4, + "nodes": { + "battery_cover_opened": { + "method": "visual", + "intervals": [ + [ + 2.7, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 2.7, + "log_step": null, + "text_zh": "观察：视频 1.2–3.6 s，机器人右手食指下压滑盖锁片并向尾端擦推，黑色滑盖伸出、灰色旧电池及仓口露出；视频 2.6 s 原生叠字为 door 56 mm、Stages 0/4，视频 2.7 s 变为 door 64 mm、Stages 1/4，达到 60 mm 开盖判据。" + }, + { + "t": 22.3, + "log_step": null, + "text_zh": "观察：视频 22.3 s 显示 PASS、Stages 4/4；开盖为锁存事件，最终盖子关回不撤销该节点。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "压/擦", + "note_zh": "机器人右手食指指端按住滑盖的弹性锁片，并保持压接向遥控器尾端擦推滑盖；没有夹住盖子。" + } + ] + }, + "old_cell_in_tray": { + "method": "visual", + "intervals": [ + [ + 9.1, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 6.8, + "log_step": null, + "text_zh": "观察：视频 5.0–6.8 s，机器人右手拇指与食指从两侧夹住灰色电池并抬起；视频 6.8 s 可见灰色电池随指端悬在原仓口之外，原仓内不再有灰色电池。" + }, + { + "t": 9.1, + "log_step": null, + "text_zh": "观察：视频 7.1–9.5 s，机器人将灰色电池移到右侧白色托盘并降下、张指；视频 9.0 s 叠字仍为 Stages 1/4，视频 9.1 s 变为 Stages 2/4，随后右手抬离，灰色电池留在盘底。" + }, + { + "t": 22.3, + "log_step": null, + "text_zh": "观察：视频 9.5–22.3 s 灰色电池持续留在白色托盘内，机器人未再操作它；视频 21.3–22.3 s 逐帧显示其位置不变，叠字 PASS、Stages 4/4、released hold 50/50。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "抓握", + "note_zh": "机器人右手拇指和食指对向捏夹灰色旧电池，提离遥控器电池仓并搬入白色托盘，降下后张指释放。" + } + ] + }, + "new_cell_seated": { + "method": "visual", + "intervals": [ + [ + 15.7, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 13.9, + "log_step": null, + "text_zh": "观察：视频 10.5–12.6 s，机器人右手下降到桌上的绿色电池处，掌部遮住夹取瞬间；视频 13.1 s 原桌面位置已无绿色电池，视频 13.9 s 可见绿色电池夹在右手指端并随手移至遥控器仓口上方。" + }, + { + "t": 15.7, + "log_step": null, + "text_zh": "观察：视频 14.1–16.5 s，机器人将绿色电池降到灰色电池原来的槽内并张指抬手；视频 15.6 s 叠字仍为 Stages 2/4，视频 15.7 s 变为 Stages 3/4；视频 16.2 s 可见绿色电池沿槽长轴平卧，金色端朝向槽的金色触点，手指已抬开。" + }, + { + "t": 22.3, + "log_step": null, + "text_zh": "推断：视频 16.2–17.7 s 绿色电池稳定留在槽内，随后被机器人滑回的盖子遮住；视频 20.3 s 及视频 22.3 s 原生叠字均显示 PASS、Stages 4/4、released hold 50/50，且 inputs/run.json 评分器字段 success=1、stages_completed=4，支持最终仍满足极性、落底及离手判据。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "抓握", + "note_zh": "机器人右手拇指和食指捏夹绿色新电池，提离桌面、调整方向后搬到电池槽，降入并张指释放；掌部遮住最初合指瞬间，搬运至仓口时可见电池在对向指端之间。" + } + ] + }, + "battery_cover_closed": { + "method": "visual", + "intervals": [ + [ + 0, + 1.6 + ], + [ + 19.3, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0, + "log_step": null, + "text_zh": "观察：视频 0.0 s 初始黑色滑盖覆盖电池仓，没有电池露出，原生叠字 door 0 mm；该 state 节点从初态成立，不受 after 顺序边门控。" + }, + { + "t": 1.6, + "log_step": null, + "text_zh": "推断：视频 1.2–1.8 s 机器人右手食指下降按压锁片，随后滑盖移出；视频 2.0 s 叠字 door 5 mm 已超过关盖阈值。锁片按深未在叠字中给出，结合下降压接动作，将初始锁闭状态失去时间估为视频 1.6 s，约有 0.3 s 不确定性。" + }, + { + "t": 19.3, + "log_step": null, + "text_zh": "观察：视频 17.5–19.4 s，机器人右手食指压住盖上锁片并向仓口擦推滑盖，绿色电池被覆盖；视频 19.2 s 仍为 Stages 3/4，视频 19.3 s 变为 Stages 4/4，随后指端抬离、released hold 开始累计。" + }, + { + "t": 22.3, + "log_step": null, + "text_zh": "观察：视频 20.3 s 叠字显示 PASS、door -0 mm、released hold 50/50；视频 21.3–22.3 s 逐帧显示右手离开滑盖、黑盖保持闭合，灰色电池仍在托盘。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "压/擦", + "note_zh": "机器人右手食指指端按下滑盖锁片，沿滑轨向仓口方向擦推盖子，随后抬起食指使锁片回弹；开盖阶段的同类按压使初始锁闭状态失去。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "cover_slide_m_open", + "unit": "m", + "threshold": 0.06, + "direction": "higher_better", + "value": null + }, + { + "name": "cover_slide_m_closed", + "unit": "m", + "threshold": 0.002, + "direction": "lower_better", + "value": null + }, + { + "name": "tab_depth_m_closed", + "unit": "m", + "threshold": 0.0015, + "direction": "lower_better", + "value": null + }, + { + "name": "old_cell_tray_local_x_abs", + "unit": "m", + "threshold": 0.037, + "direction": "lower_better", + "value": null + }, + { + "name": "old_cell_tray_local_y_abs", + "unit": "m", + "threshold": 0.032, + "direction": "lower_better", + "value": null + }, + { + "name": "old_cell_tray_local_z", + "unit": "m", + "threshold": 0.014, + "direction": "lower_better", + "value": null + }, + { + "name": "new_cell_slot_error_x_abs", + "unit": "m", + "threshold": 0.005, + "direction": "lower_better", + "value": null + }, + { + "name": "new_cell_slot_error_y_abs", + "unit": "m", + "threshold": 0.003, + "direction": "lower_better", + "value": null + }, + { + "name": "new_cell_slot_error_z_abs", + "unit": "m", + "threshold": 0.002, + "direction": "lower_better", + "value": null + }, + { + "name": "new_cell_axis_angle_deg", + "unit": "deg", + "threshold": 10, + "direction": "lower_better", + "value": null + }, + { + "name": "max_speed_m_s", + "unit": "m/s", + "threshold": 0.02, + "direction": "lower_better", + "value": null + }, + { + "name": "max_spin_rad_s", + "unit": "rad/s", + "threshold": 0.5, + "direction": "lower_better", + "value": null + }, + { + "name": "remote_displacement_m", + "unit": "m", + "threshold": 0.03, + "direction": "lower_better", + "value": null + }, + { + "name": "remote_tilt_deg", + "unit": "deg", + "threshold": 15, + "direction": "lower_better", + "value": null + } + ], + "failure": null, + "summary_zh": "机器人依序开盖、将灰色旧电池夹送入托盘、将绿色新电池按正确极性放入槽并关盖锁住，最终四个节点全部成立；关盖节点从初态计入，开盖时失去，最终恢复。视频 20.3 s 及视频 21.3–22.3 s 的画面显示 PASS、Stages 4/4、released hold 50/50，与评分器 success=1、deterministic=1 一致，无失败断点。没有状态日志，首次按压导致锁闭失去的时间约有 0.3 s 不确定性，槽内微小公差及接触验收结合原生阶段叠字与评分结果判断；评分器未报告 DAG 所列原生连续指标数值，均保留为 null。", + "conf": 9, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "成功运行，四个节点的时间与叠字（视频 2.7 s Stages 1/4、9.1 s 2/4、15.7 s 3/4、19.3 s 4/4）吻合，contact_actual 与画面一致，无失败断点，未发现需要修改之处。", + "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": "2026-10-29T10:05:35Z", + "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/dexhand-remote-battery-swap-i00-privileged--dncwpct/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/dexhand-remote-battery-swap-i00-privileged--dncwpct/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "094d2d9b53b609ccf0c545ab63c5783a3f37411078f1e94fd268198ef7364fd9", + "video_sha256": "698e0a5bb9f58572e4e50d223dd756e4fb78b28b46d10cb1e0334093d93c8c75", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/dexhand/remote_battery_swap/trajectories/dexhand-remote-battery-swap-i00-standard--9s6thft.json b/phase2/dexhand/remote_battery_swap/trajectories/dexhand-remote-battery-swap-i00-standard--9s6thft.json new file mode 100644 index 0000000000000000000000000000000000000000..e00af170fb87993df0240947c16d25ad15dbd17a --- /dev/null +++ b/phase2/dexhand/remote_battery_swap/trajectories/dexhand-remote-battery-swap-i00-standard--9s6thft.json @@ -0,0 +1,324 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "dexhand-remote-battery-swap-i00-standard--9s6thft", + "taskKey": "dexhand/remote_battery_swap", + "dag_sha256": "448d38d5df99705f24cec0642f2338ce7100127136b831930f8d1fee7f80f537", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "calls": 806, + "control_steps_total": 5986, + "elapsed_s": 2522.632, + "finalized": 1, + "gave_up": 0, + "graded_by_live": 1, + "grader_error": 0, + "hand_wuji": 0, + "infra_error": 0, + "instance_mismatch": 0, + "live_success": 0, + "missing_log": 0, + "n_actions": 422, + "no_verdict": 0, + "replays": 0, + "reset_budget": 20, + "resets": 10, + "resume_state_mismatch": 0, + "resumes": 0, + "sim_error": 0, + "sim_startups": 1, + "sim_steps": 422, + "success": 0, + "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": 0.252, + "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/dexhand-remote-battery-swap-i00-standard--9s6thft/replay.mp4", + "duration_s": 10.8, + "fps": 5.0, + "frames": 54, + "sha256": "e601b95e6817963ae1f178a0589e66b632361a0c5f7c64322e374f54c6d7e574" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 4, + "base": null, + "peak": null, + "final": null, + "nodes": { + "battery_cover_opened": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 5.8, + "log_step": null, + "text_zh": "观察：视频 0.0–8.6 s，机器人右手伸出食指，在黑色滑盖尾端按压并横移；盖未向尾端伸出到露出灰色电池的程度，视频叠字始终为 Stages 0/4，未达到 0.060 m 的开盖节点。" + }, + { + "t": 8.6, + "log_step": null, + "text_zh": "观察：视频 8.6 s 叠字变为 FAIL: remote_tipped；视频 8.8–10.6 s 为相同终态，电池仓仍被盖遮住。成功参考视频 4.0–6.0 s 可清楚看到滑盖向尾端伸出及灰色电池露出，本次没有该线索。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "压/擦", + "note_zh": "观察：视频约 4.6–8.6 s，机器人右手食指指端压住黑色滑盖尾部并沿盖面横向擦移；没有形成相向夹持。" + } + ] + }, + "old_cell_in_tray": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.6, + "log_step": null, + "text_zh": "观察：逐帧检查视频 0.0–10.6 s，机器人右手始终在遥控器滑盖附近，没有从仓内取出并搬运灰色电池；右侧白色托盘从初态到视频 10.6 s 始终为空。" + } + ] + }, + "new_cell_seated": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 10.6, + "log_step": null, + "text_zh": "观察：视频 0.0–10.6 s，绿色新电池始终卧在遥控器前方桌面原位，机器人右手未接近、夹取或搬运它；最终仍在桌面，没有进入电池槽。" + } + ] + }, + "battery_cover_closed": { + "method": "unknown", + "intervals": null, + "final": null, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察：视频 0.0 s 标为 initial state，黑色滑盖覆盖电池仓，叠字 door 0 mm；任务初态规定盖已关上且锁片未按下，故关盖条件在初态成立。" + }, + { + "t": 8.6, + "log_step": null, + "text_zh": "观察：视频约 4.6–8.6 s 机器人右手食指压擦滑盖尾部；视频 8.6–10.6 s 盖仍覆盖仓口且叠字 door -0 mm，但锁片所在接触区持续被食指遮住，叠字没有报告锁片按深。" + }, + { + "t": 10.6, + "log_step": null, + "text_zh": "推断：本次没有主动执行换电后的关盖步骤。初态关盖成立，但仅凭最终盖的位置不能确定锁片按深是否恢复到 <=0.0015 m；缺少状态日志和可见锁片侧视，无法可靠重建关盖条件的完整区间及终值，保留未知。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "cover_slide_m_open", + "unit": "m", + "threshold": 0.06, + "direction": "higher_better", + "value": null + }, + { + "name": "cover_slide_m_closed", + "unit": "m", + "threshold": 0.002, + "direction": "lower_better", + "value": null + }, + { + "name": "tab_depth_m_closed", + "unit": "m", + "threshold": 0.0015, + "direction": "lower_better", + "value": null + }, + { + "name": "old_cell_tray_local_x_abs", + "unit": "m", + "threshold": 0.037, + "direction": "lower_better", + "value": null + }, + { + "name": "old_cell_tray_local_y_abs", + "unit": "m", + "threshold": 0.032, + "direction": "lower_better", + "value": null + }, + { + "name": "old_cell_tray_local_z", + "unit": "m", + "threshold": 0.014, + "direction": "lower_better", + "value": null + }, + { + "name": "new_cell_slot_error_x_abs", + "unit": "m", + "threshold": 0.005, + "direction": "lower_better", + "value": null + }, + { + "name": "new_cell_slot_error_y_abs", + "unit": "m", + "threshold": 0.003, + "direction": "lower_better", + "value": null + }, + { + "name": "new_cell_slot_error_z_abs", + "unit": "m", + "threshold": 0.002, + "direction": "lower_better", + "value": null + }, + { + "name": "new_cell_axis_angle_deg", + "unit": "deg", + "threshold": 10, + "direction": "lower_better", + "value": null + }, + { + "name": "max_speed_m_s", + "unit": "m/s", + "threshold": 0.02, + "direction": "lower_better", + "value": null + }, + { + "name": "max_spin_rad_s", + "unit": "rad/s", + "threshold": 0.5, + "direction": "lower_better", + "value": null + }, + { + "name": "remote_displacement_m", + "unit": "m", + "threshold": 0.03, + "direction": "lower_better", + "value": null + }, + { + "name": "remote_tilt_deg", + "unit": "deg", + "threshold": 15, + "direction": "lower_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "battery_cover_opened", + "t": 8.4, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察：视频约 5.8–8.4 s，机器人右手食指继续压擦滑盖，遥控器整件随之改变朝向并翘起，滑盖未露出灰色电池；视频 8.6 s 首次显示 FAIL: remote_tipped，随后冻结至视频 10.6 s。机器人一直在尝试开盖，该步骤没有随后修复。" + }, + "unmet": { + "battery_cover_opened": false, + "old_cell_in_tray": false, + "new_cell_seated": false, + "battery_cover_closed": null + }, + "phenomena": [ + { + "cat": "接触", + "item": "带动本体", + "t": 8.4, + "log_step": null, + "evidence_zh": "观察：视频约 5.8–8.6 s，机器人右手食指在滑盖尾部施力时，遥控器本体跟随手指转动、倾起，盖未相对本体滑开到所需位置；视频 8.6 s 的原生叠字为 FAIL: remote_tipped。这不是机器人本体失去平衡，而是操作盖时带动了遥控器基体。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "视频 / 日志", + "evidence_zh": "推断：视频约 4.6–8.6 s，机器人已经把右手食指放在滑盖尾部并执行按压、横移，但所需的解锁和相对滑动没有实现，作用力反而带动遥控器倾起，支持开盖接触控制失败。log step 60 中 agent 的控制脚本输出 cavity True，随后又输出 BLOCKED、S1 done，最终服务返回 EpisodeOver: task_failure；log step 61 中 agent 承认灰度仓口检测误把锁片当作开仓。这些日志只作为控制过程的补充证据，不把脚本自报的 S1 done 当作节点完成，也不将其中工作耗时换算为视频时间。" + } + ], + "gap_zh": "没有状态日志；视频提供总览及阶段、盖位移、失败原因叠字，但没有锁片按深或接触力。锁片在受操作及终态时被右手食指遮挡，因此关盖节点虽初态已成立，后续完整区间和终值仍未知；也不能据现有证据把控制根因进一步确定为具体按压参数、接触几何或反馈失稳。日志包含十次重置，未给出最后现场回合各输出与视频帧的逐控制步对应，本文件所有视频锚点均直接取自已检查画面。" + }, + "summary_zh": "本次评分视频中，机器人右手只尝试打开电池盖，未达到开度要求，旧电池未入托盘，绿色新电池始终留在桌面。视频约 8.4 s 开盖操作带动遥控器本体倾起，视频 8.6 s 显示 remote_tipped 判负，支持控制失败。盖在初态已关闭，但锁片受操作后持续被手指遮挡且无按深记录，关盖节点的完整时间区间及终值保留未知。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "视频显示右手食指压擦滑盖，遥控器被带动倾起，8.6 s 叠字 FAIL: remote_tipped，盖开度未达节点；开盖未达成、旧电池未入托盘、新电池始终在桌面，断点 battery_cover_opened、现象带动本体、根因控制均有证据支持；关盖未知的理由成立（锁片被遮、无状态日志）。", + "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": "2026-10-29T10:06:14Z", + "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/dexhand-remote-battery-swap-i00-standard--9s6thft/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/dexhand-remote-battery-swap-i00-standard--9s6thft/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "3a2dc72ffda5ab1f7f518b507fc2889c72d19041eae68cc06bae0e9680275042", + "video_sha256": "e601b95e6817963ae1f178a0589e66b632361a0c5f7c64322e374f54c6d7e574", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/alpine_slalom/task_dag.json b/phase2/embodied-olympics-fun/alpine_slalom/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..b5181dd993f79823f68c6ffaa0b55235635809dc --- /dev/null +++ b/phase2/embodied-olympics-fun/alpine_slalom/task_dag.json @@ -0,0 +1,1575 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "embodied-olympics-fun/alpine_slalom", + "suite": "embodied-olympics-fun", + "base": "embodied-olympics-fun-alpine-slalom-i00", + "title_zh": "高山滑雪回转", + "summary_zh": "机器人自行起滑，以双脚及两只雪板尖依次通过全部32道旗门，并在规则和总时限内以双脚越过两柱间的终点。", + "embodiment_zh": "两种模式均为MuJoCo中的Unitree G1人形机器人，29个关节、高1.32米、重33千克；双臂末端是无可动手指的橡胶拳垫，不是二指夹爪或灵巧手；双脚各刚性绑定一只1.28米回转雪板，无雪杖、无轮式底盘，靠腿和躯干控制滑行。特权模式可读仿真状态，标准模式通过机器人服务获取相机及本体感知，受评本体相同。", + "task_class": "compositional", + "analysis_level": "step", + "spec": { + "version": "1.2", + "task": "embodied-olympics-fun-alpine-slalom-i00", + "bench": "embodied-olympics-fun", + "progress": "live", + "entities": {}, + "milestones": [ + { + "id": "gate_1_passed", + "zh": "通过第1门", + "mode": "event", + "when": null, + "after": [], + "hold_s": 0, + "goal": true, + "criterion_zh": "先在回合开始后20秒内自行起滑，双脚跨过起点红线；随后两只雪板尖和双脚四个受检点均沿下坡方向穿过第1道红色旗门的两杆间门线（坡面里程12米，位于起始朝向的左侧，横向范围[0.8, 4.8]米，含端点），才记为过门。画面能分辨时，应见双脚连同两板尖从该对红杆上坡侧全部进入下坡侧；仅躯干越线、跨杆两侧或仍有一点未过均不算，尚未记过门时逆向退回的点须重过，碰倒杆本身不扣分。", + "step": { + "cat": "本体", + "item": "行进" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚经固定雪板支承在雪面上，双腿和躯干调节重心完成过门。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:719-731", + "inputs/task/privileged/environment/rfo_task.py:866-881", + "inputs/task/privileged/environment/rfo_task.py:899-930", + "inputs/task/privileged/environment/rfo_task.py:1034-1037", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/instruction.md:34-38" + ], + "conf": 6, + "note_zh": "每门仅一个步骤，同时接受两板尖与双脚；按裁判已通过门数锁存，过门后不要求停稳，也不随机器人远离而失去。固定远景中后段旗门和机器人过小或离开画面，不能从不可见推断未完成；需结合该次裁判阶段记录。起点阶段仅为首门之前的途中事件及计时验收，不另建起滑节点（S8）；雪板已在初态固定于双脚。" + }, + { + "id": "gate_2_passed", + "zh": "通过第2门", + "mode": "event", + "when": null, + "after": [ + "gate_1_passed" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "第1道门已通过后，两只雪板尖和双脚四个受检点均沿下坡方向穿过第2道蓝色旗门的两杆间门线（坡面里程21.6米，位于起始朝向的右侧，横向范围[-4.8, -0.8]米，含端点），才记为过门。画面能分辨时，应见双脚连同两板尖从该对蓝杆上坡侧全部进入下坡侧；仅躯干越线、跨杆两侧或仍有一点未过均不算，尚未记过门时逆向退回的点须重过，碰倒杆本身不扣分。", + "step": { + "cat": "本体", + "item": "行进" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚经固定雪板支承在雪面上，双腿和躯干调节重心完成过门。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:719-731", + "inputs/task/privileged/environment/rfo_task.py:866-881", + "inputs/task/privileged/environment/rfo_task.py:899-930", + "inputs/task/privileged/environment/rfo_task.py:1034-1037", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/instruction.md:34-38" + ], + "conf": 6, + "note_zh": "每门仅一个步骤，同时接受两板尖与双脚；按裁判已通过门数锁存，过门后不要求停稳，也不随机器人远离而失去。固定远景中后段旗门和机器人过小或离开画面，不能从不可见推断未完成；需结合该次裁判阶段记录。" + }, + { + "id": "gate_3_passed", + "zh": "通过第3门", + "mode": "event", + "when": null, + "after": [ + "gate_2_passed" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "第2道门已通过后，两只雪板尖和双脚四个受检点均沿下坡方向穿过第3道红色旗门的两杆间门线（坡面里程31.2米，位于起始朝向的左侧，横向范围[0.8, 4.8]米，含端点），才记为过门。画面能分辨时，应见双脚连同两板尖从该对红杆上坡侧全部进入下坡侧；仅躯干越线、跨杆两侧或仍有一点未过均不算，尚未记过门时逆向退回的点须重过，碰倒杆本身不扣分。", + "step": { + "cat": "本体", + "item": "行进" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚经固定雪板支承在雪面上，双腿和躯干调节重心完成过门。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:719-731", + "inputs/task/privileged/environment/rfo_task.py:866-881", + "inputs/task/privileged/environment/rfo_task.py:899-930", + "inputs/task/privileged/environment/rfo_task.py:1034-1037", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/instruction.md:34-38" + ], + "conf": 6, + "note_zh": "每门仅一个步骤，同时接受两板尖与双脚；按裁判已通过门数锁存，过门后不要求停稳，也不随机器人远离而失去。固定远景中后段旗门和机器人过小或离开画面，不能从不可见推断未完成；需结合该次裁判阶段记录。" + }, + { + "id": "gate_4_passed", + "zh": "通过第4门", + "mode": "event", + "when": null, + "after": [ + "gate_3_passed" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "第3道门已通过后，两只雪板尖和双脚四个受检点均沿下坡方向穿过第4道蓝色旗门的两杆间门线（坡面里程40.8米，位于起始朝向的右侧，横向范围[-4.8, -0.8]米，含端点），才记为过门。画面能分辨时，应见双脚连同两板尖从该对蓝杆上坡侧全部进入下坡侧；仅躯干越线、跨杆两侧或仍有一点未过均不算，尚未记过门时逆向退回的点须重过，碰倒杆本身不扣分。", + "step": { + "cat": "本体", + "item": "行进" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚经固定雪板支承在雪面上，双腿和躯干调节重心完成过门。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:719-731", + "inputs/task/privileged/environment/rfo_task.py:866-881", + "inputs/task/privileged/environment/rfo_task.py:899-930", + "inputs/task/privileged/environment/rfo_task.py:1034-1037", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/instruction.md:34-38" + ], + "conf": 5, + "note_zh": "每门仅一个步骤，同时接受两板尖与双脚；按裁判已通过门数锁存，过门后不要求停稳，也不随机器人远离而失去。固定远景中后段旗门和机器人过小或离开画面，不能从不可见推断未完成；需结合该次裁判阶段记录。" + }, + { + "id": "gate_5_passed", + "zh": "通过第5门", + "mode": "event", + "when": null, + "after": [ + "gate_4_passed" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "第4道门已通过后，两只雪板尖和双脚四个受检点均沿下坡方向穿过第5道红色旗门的两杆间门线（坡面里程50.4米，位于起始朝向的左侧，横向范围[0.8, 4.8]米，含端点），才记为过门。画面能分辨时，应见双脚连同两板尖从该对红杆上坡侧全部进入下坡侧；仅躯干越线、跨杆两侧或仍有一点未过均不算，尚未记过门时逆向退回的点须重过，碰倒杆本身不扣分。", + "step": { + "cat": "本体", + "item": "行进" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚经固定雪板支承在雪面上，双腿和躯干调节重心完成过门。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:719-731", + "inputs/task/privileged/environment/rfo_task.py:866-881", + "inputs/task/privileged/environment/rfo_task.py:899-930", + "inputs/task/privileged/environment/rfo_task.py:1034-1037", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/instruction.md:34-38" + ], + "conf": 5, + "note_zh": "每门仅一个步骤，同时接受两板尖与双脚；按裁判已通过门数锁存，过门后不要求停稳，也不随机器人远离而失去。固定远景中后段旗门和机器人过小或离开画面，不能从不可见推断未完成；需结合该次裁判阶段记录。" + }, + { + "id": "gate_6_passed", + "zh": "通过第6门", + "mode": "event", + "when": null, + "after": [ + "gate_5_passed" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "第5道门已通过后，两只雪板尖和双脚四个受检点均沿下坡方向穿过第6道蓝色旗门的两杆间门线（坡面里程60米，位于起始朝向的右侧，横向范围[-4.8, -0.8]米，含端点），才记为过门。画面能分辨时，应见双脚连同两板尖从该对蓝杆上坡侧全部进入下坡侧；仅躯干越线、跨杆两侧或仍有一点未过均不算，尚未记过门时逆向退回的点须重过，碰倒杆本身不扣分。", + "step": { + "cat": "本体", + "item": "行进" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚经固定雪板支承在雪面上，双腿和躯干调节重心完成过门。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:719-731", + "inputs/task/privileged/environment/rfo_task.py:866-881", + "inputs/task/privileged/environment/rfo_task.py:899-930", + "inputs/task/privileged/environment/rfo_task.py:1034-1037", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/instruction.md:34-38" + ], + "conf": 5, + "note_zh": "每门仅一个步骤，同时接受两板尖与双脚；按裁判已通过门数锁存，过门后不要求停稳，也不随机器人远离而失去。固定远景中后段旗门和机器人过小或离开画面，不能从不可见推断未完成；需结合该次裁判阶段记录。" + }, + { + "id": "gate_7_passed", + "zh": "通过第7门", + "mode": "event", + "when": null, + "after": [ + "gate_6_passed" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "第6道门已通过后，两只雪板尖和双脚四个受检点均沿下坡方向穿过第7道红色旗门的两杆间门线（坡面里程69.6米，位于起始朝向的左侧，横向范围[0.8, 4.8]米，含端点），才记为过门。画面能分辨时，应见双脚连同两板尖从该对红杆上坡侧全部进入下坡侧；仅躯干越线、跨杆两侧或仍有一点未过均不算，尚未记过门时逆向退回的点须重过，碰倒杆本身不扣分。", + "step": { + "cat": "本体", + "item": "行进" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚经固定雪板支承在雪面上，双腿和躯干调节重心完成过门。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:719-731", + "inputs/task/privileged/environment/rfo_task.py:866-881", + "inputs/task/privileged/environment/rfo_task.py:899-930", + "inputs/task/privileged/environment/rfo_task.py:1034-1037", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/instruction.md:34-38" + ], + "conf": 5, + "note_zh": "每门仅一个步骤，同时接受两板尖与双脚；按裁判已通过门数锁存，过门后不要求停稳，也不随机器人远离而失去。固定远景中后段旗门和机器人过小或离开画面，不能从不可见推断未完成；需结合该次裁判阶段记录。" + }, + { + "id": "gate_8_passed", + "zh": "通过第8门", + "mode": "event", + "when": null, + "after": [ + "gate_7_passed" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "第7道门已通过后，两只雪板尖和双脚四个受检点均沿下坡方向穿过第8道蓝色旗门的两杆间门线（坡面里程79.2米，位于起始朝向的右侧，横向范围[-4.8, -0.8]米，含端点），才记为过门。画面能分辨时，应见双脚连同两板尖从该对蓝杆上坡侧全部进入下坡侧；仅躯干越线、跨杆两侧或仍有一点未过均不算，尚未记过门时逆向退回的点须重过，碰倒杆本身不扣分。", + "step": { + "cat": "本体", + "item": "行进" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚经固定雪板支承在雪面上，双腿和躯干调节重心完成过门。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:719-731", + "inputs/task/privileged/environment/rfo_task.py:866-881", + "inputs/task/privileged/environment/rfo_task.py:899-930", + "inputs/task/privileged/environment/rfo_task.py:1034-1037", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/instruction.md:34-38" + ], + "conf": 5, + "note_zh": "每门仅一个步骤，同时接受两板尖与双脚；按裁判已通过门数锁存，过门后不要求停稳，也不随机器人远离而失去。固定远景中后段旗门和机器人过小或离开画面，不能从不可见推断未完成；需结合该次裁判阶段记录。" + }, + { + "id": "gate_9_passed", + "zh": "通过第9门", + "mode": "event", + "when": null, + "after": [ + "gate_8_passed" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "第8道门已通过后，两只雪板尖和双脚四个受检点均沿下坡方向穿过第9道红色旗门的两杆间门线（坡面里程88.8米，位于起始朝向的左侧，横向范围[0.8, 4.8]米，含端点），才记为过门。画面能分辨时，应见双脚连同两板尖从该对红杆上坡侧全部进入下坡侧；仅躯干越线、跨杆两侧或仍有一点未过均不算，尚未记过门时逆向退回的点须重过，碰倒杆本身不扣分。", + "step": { + "cat": "本体", + "item": "行进" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚经固定雪板支承在雪面上，双腿和躯干调节重心完成过门。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:719-731", + "inputs/task/privileged/environment/rfo_task.py:866-881", + "inputs/task/privileged/environment/rfo_task.py:899-930", + "inputs/task/privileged/environment/rfo_task.py:1034-1037", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/instruction.md:34-38" + ], + "conf": 5, + "note_zh": "每门仅一个步骤，同时接受两板尖与双脚；按裁判已通过门数锁存，过门后不要求停稳，也不随机器人远离而失去。固定远景中后段旗门和机器人过小或离开画面，不能从不可见推断未完成；需结合该次裁判阶段记录。" + }, + { + "id": "gate_10_passed", + "zh": "通过第10门", + "mode": "event", + "when": null, + "after": [ + "gate_9_passed" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "第9道门已通过后，两只雪板尖和双脚四个受检点均沿下坡方向穿过第10道蓝色旗门的两杆间门线（坡面里程98.4米，位于起始朝向的右侧，横向范围[-4.8, -0.8]米，含端点），才记为过门。画面能分辨时，应见双脚连同两板尖从该对蓝杆上坡侧全部进入下坡侧；仅躯干越线、跨杆两侧或仍有一点未过均不算，尚未记过门时逆向退回的点须重过，碰倒杆本身不扣分。", + "step": { + "cat": "本体", + "item": "行进" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚经固定雪板支承在雪面上，双腿和躯干调节重心完成过门。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:719-731", + "inputs/task/privileged/environment/rfo_task.py:866-881", + "inputs/task/privileged/environment/rfo_task.py:899-930", + "inputs/task/privileged/environment/rfo_task.py:1034-1037", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/instruction.md:34-38" + ], + "conf": 5, + "note_zh": "每门仅一个步骤，同时接受两板尖与双脚；按裁判已通过门数锁存，过门后不要求停稳，也不随机器人远离而失去。固定远景中后段旗门和机器人过小或离开画面，不能从不可见推断未完成；需结合该次裁判阶段记录。" + }, + { + "id": "gate_11_passed", + "zh": "通过第11门", + "mode": "event", + "when": null, + "after": [ + "gate_10_passed" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "第10道门已通过后，两只雪板尖和双脚四个受检点均沿下坡方向穿过第11道红色旗门的两杆间门线（坡面里程108米，位于起始朝向的左侧，横向范围[0.8, 4.8]米，含端点），才记为过门。画面能分辨时，应见双脚连同两板尖从该对红杆上坡侧全部进入下坡侧；仅躯干越线、跨杆两侧或仍有一点未过均不算，尚未记过门时逆向退回的点须重过，碰倒杆本身不扣分。", + "step": { + "cat": "本体", + "item": "行进" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚经固定雪板支承在雪面上，双腿和躯干调节重心完成过门。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:719-731", + "inputs/task/privileged/environment/rfo_task.py:866-881", + "inputs/task/privileged/environment/rfo_task.py:899-930", + "inputs/task/privileged/environment/rfo_task.py:1034-1037", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/instruction.md:34-38" + ], + "conf": 5, + "note_zh": "每门仅一个步骤，同时接受两板尖与双脚；按裁判已通过门数锁存，过门后不要求停稳，也不随机器人远离而失去。固定远景中后段旗门和机器人过小或离开画面，不能从不可见推断未完成；需结合该次裁判阶段记录。" + }, + { + "id": "gate_12_passed", + "zh": "通过第12门", + "mode": "event", + "when": null, + "after": [ + "gate_11_passed" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "第11道门已通过后，两只雪板尖和双脚四个受检点均沿下坡方向穿过第12道蓝色旗门的两杆间门线（坡面里程117.6米，位于起始朝向的右侧，横向范围[-4.8, -0.8]米，含端点），才记为过门。画面能分辨时，应见双脚连同两板尖从该对蓝杆上坡侧全部进入下坡侧；仅躯干越线、跨杆两侧或仍有一点未过均不算，尚未记过门时逆向退回的点须重过，碰倒杆本身不扣分。", + "step": { + "cat": "本体", + "item": "行进" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚经固定雪板支承在雪面上，双腿和躯干调节重心完成过门。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:719-731", + "inputs/task/privileged/environment/rfo_task.py:866-881", + "inputs/task/privileged/environment/rfo_task.py:899-930", + "inputs/task/privileged/environment/rfo_task.py:1034-1037", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/instruction.md:34-38" + ], + "conf": 5, + "note_zh": "每门仅一个步骤，同时接受两板尖与双脚；按裁判已通过门数锁存，过门后不要求停稳，也不随机器人远离而失去。固定远景中后段旗门和机器人过小或离开画面，不能从不可见推断未完成；需结合该次裁判阶段记录。" + }, + { + "id": "gate_13_passed", + "zh": "通过第13门", + "mode": "event", + "when": null, + "after": [ + "gate_12_passed" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "第12道门已通过后，两只雪板尖和双脚四个受检点均沿下坡方向穿过第13道红色旗门的两杆间门线（坡面里程127.2米，位于起始朝向的左侧，横向范围[0.8, 4.8]米，含端点），才记为过门。画面能分辨时，应见双脚连同两板尖从该对红杆上坡侧全部进入下坡侧；仅躯干越线、跨杆两侧或仍有一点未过均不算，尚未记过门时逆向退回的点须重过，碰倒杆本身不扣分。", + "step": { + "cat": "本体", + "item": "行进" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚经固定雪板支承在雪面上，双腿和躯干调节重心完成过门。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:719-731", + "inputs/task/privileged/environment/rfo_task.py:866-881", + "inputs/task/privileged/environment/rfo_task.py:899-930", + "inputs/task/privileged/environment/rfo_task.py:1034-1037", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/instruction.md:34-38" + ], + "conf": 5, + "note_zh": "每门仅一个步骤，同时接受两板尖与双脚；按裁判已通过门数锁存，过门后不要求停稳，也不随机器人远离而失去。固定远景中后段旗门和机器人过小或离开画面，不能从不可见推断未完成；需结合该次裁判阶段记录。" + }, + { + "id": "gate_14_passed", + "zh": "通过第14门", + "mode": "event", + "when": null, + "after": [ + "gate_13_passed" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "第13道门已通过后，两只雪板尖和双脚四个受检点均沿下坡方向穿过第14道蓝色旗门的两杆间门线（坡面里程136.8米，位于起始朝向的右侧，横向范围[-4.8, -0.8]米，含端点），才记为过门。画面能分辨时，应见双脚连同两板尖从该对蓝杆上坡侧全部进入下坡侧；仅躯干越线、跨杆两侧或仍有一点未过均不算，尚未记过门时逆向退回的点须重过，碰倒杆本身不扣分。", + "step": { + "cat": "本体", + "item": "行进" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚经固定雪板支承在雪面上，双腿和躯干调节重心完成过门。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:719-731", + "inputs/task/privileged/environment/rfo_task.py:866-881", + "inputs/task/privileged/environment/rfo_task.py:899-930", + "inputs/task/privileged/environment/rfo_task.py:1034-1037", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/instruction.md:34-38" + ], + "conf": 5, + "note_zh": "每门仅一个步骤，同时接受两板尖与双脚；按裁判已通过门数锁存，过门后不要求停稳，也不随机器人远离而失去。固定远景中后段旗门和机器人过小或离开画面，不能从不可见推断未完成；需结合该次裁判阶段记录。" + }, + { + "id": "gate_15_passed", + "zh": "通过第15门", + "mode": "event", + "when": null, + "after": [ + "gate_14_passed" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "第14道门已通过后，两只雪板尖和双脚四个受检点均沿下坡方向穿过第15道红色旗门的两杆间门线（坡面里程146.4米，位于起始朝向的左侧，横向范围[0.8, 4.8]米，含端点），才记为过门。画面能分辨时，应见双脚连同两板尖从该对红杆上坡侧全部进入下坡侧；仅躯干越线、跨杆两侧或仍有一点未过均不算，尚未记过门时逆向退回的点须重过，碰倒杆本身不扣分。", + "step": { + "cat": "本体", + "item": "行进" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚经固定雪板支承在雪面上，双腿和躯干调节重心完成过门。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:719-731", + "inputs/task/privileged/environment/rfo_task.py:866-881", + "inputs/task/privileged/environment/rfo_task.py:899-930", + "inputs/task/privileged/environment/rfo_task.py:1034-1037", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/instruction.md:34-38" + ], + "conf": 5, + "note_zh": "每门仅一个步骤，同时接受两板尖与双脚；按裁判已通过门数锁存，过门后不要求停稳，也不随机器人远离而失去。固定远景中后段旗门和机器人过小或离开画面，不能从不可见推断未完成；需结合该次裁判阶段记录。" + }, + { + "id": "gate_16_passed", + "zh": "通过第16门", + "mode": "event", + "when": null, + "after": [ + "gate_15_passed" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "第15道门已通过后，两只雪板尖和双脚四个受检点均沿下坡方向穿过第16道蓝色旗门的两杆间门线（坡面里程156米，位于起始朝向的右侧，横向范围[-4.8, -0.8]米，含端点），才记为过门。画面能分辨时，应见双脚连同两板尖从该对蓝杆上坡侧全部进入下坡侧；仅躯干越线、跨杆两侧或仍有一点未过均不算，尚未记过门时逆向退回的点须重过，碰倒杆本身不扣分。", + "step": { + "cat": "本体", + "item": "行进" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚经固定雪板支承在雪面上，双腿和躯干调节重心完成过门。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:719-731", + "inputs/task/privileged/environment/rfo_task.py:866-881", + "inputs/task/privileged/environment/rfo_task.py:899-930", + "inputs/task/privileged/environment/rfo_task.py:1034-1037", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/instruction.md:34-38" + ], + "conf": 5, + "note_zh": "每门仅一个步骤，同时接受两板尖与双脚；按裁判已通过门数锁存，过门后不要求停稳，也不随机器人远离而失去。固定远景中后段旗门和机器人过小或离开画面，不能从不可见推断未完成；需结合该次裁判阶段记录。" + }, + { + "id": "gate_17_passed", + "zh": "通过第17门", + "mode": "event", + "when": null, + "after": [ + "gate_16_passed" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "第16道门已通过后，两只雪板尖和双脚四个受检点均沿下坡方向穿过第17道红色旗门的两杆间门线（坡面里程165.6米，位于起始朝向的左侧，横向范围[0.8, 4.8]米，含端点），才记为过门。画面能分辨时，应见双脚连同两板尖从该对红杆上坡侧全部进入下坡侧；仅躯干越线、跨杆两侧或仍有一点未过均不算，尚未记过门时逆向退回的点须重过，碰倒杆本身不扣分。", + "step": { + "cat": "本体", + "item": "行进" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚经固定雪板支承在雪面上，双腿和躯干调节重心完成过门。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:719-731", + "inputs/task/privileged/environment/rfo_task.py:866-881", + "inputs/task/privileged/environment/rfo_task.py:899-930", + "inputs/task/privileged/environment/rfo_task.py:1034-1037", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/instruction.md:34-38" + ], + "conf": 5, + "note_zh": "每门仅一个步骤，同时接受两板尖与双脚；按裁判已通过门数锁存，过门后不要求停稳，也不随机器人远离而失去。固定远景中后段旗门和机器人过小或离开画面，不能从不可见推断未完成；需结合该次裁判阶段记录。" + }, + { + "id": "gate_18_passed", + "zh": "通过第18门", + "mode": "event", + "when": null, + "after": [ + "gate_17_passed" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "第17道门已通过后，两只雪板尖和双脚四个受检点均沿下坡方向穿过第18道蓝色旗门的两杆间门线（坡面里程175.2米，位于起始朝向的右侧，横向范围[-4.8, -0.8]米，含端点），才记为过门。画面能分辨时，应见双脚连同两板尖从该对蓝杆上坡侧全部进入下坡侧；仅躯干越线、跨杆两侧或仍有一点未过均不算，尚未记过门时逆向退回的点须重过，碰倒杆本身不扣分。", + "step": { + "cat": "本体", + "item": "行进" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚经固定雪板支承在雪面上，双腿和躯干调节重心完成过门。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:719-731", + "inputs/task/privileged/environment/rfo_task.py:866-881", + "inputs/task/privileged/environment/rfo_task.py:899-930", + "inputs/task/privileged/environment/rfo_task.py:1034-1037", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/instruction.md:34-38" + ], + "conf": 5, + "note_zh": "每门仅一个步骤，同时接受两板尖与双脚；按裁判已通过门数锁存，过门后不要求停稳，也不随机器人远离而失去。固定远景中后段旗门和机器人过小或离开画面，不能从不可见推断未完成；需结合该次裁判阶段记录。" + }, + { + "id": "gate_19_passed", + "zh": "通过第19门", + "mode": "event", + "when": null, + "after": [ + "gate_18_passed" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "第18道门已通过后，两只雪板尖和双脚四个受检点均沿下坡方向穿过第19道红色旗门的两杆间门线（坡面里程184.8米，位于起始朝向的左侧，横向范围[0.8, 4.8]米，含端点），才记为过门。画面能分辨时，应见双脚连同两板尖从该对红杆上坡侧全部进入下坡侧；仅躯干越线、跨杆两侧或仍有一点未过均不算，尚未记过门时逆向退回的点须重过，碰倒杆本身不扣分。", + "step": { + "cat": "本体", + "item": "行进" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚经固定雪板支承在雪面上，双腿和躯干调节重心完成过门。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:719-731", + "inputs/task/privileged/environment/rfo_task.py:866-881", + "inputs/task/privileged/environment/rfo_task.py:899-930", + "inputs/task/privileged/environment/rfo_task.py:1034-1037", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/instruction.md:34-38" + ], + "conf": 5, + "note_zh": "每门仅一个步骤，同时接受两板尖与双脚；按裁判已通过门数锁存，过门后不要求停稳，也不随机器人远离而失去。固定远景中后段旗门和机器人过小或离开画面，不能从不可见推断未完成；需结合该次裁判阶段记录。" + }, + { + "id": "gate_20_passed", + "zh": "通过第20门", + "mode": "event", + "when": null, + "after": [ + "gate_19_passed" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "第19道门已通过后，两只雪板尖和双脚四个受检点均沿下坡方向穿过第20道蓝色旗门的两杆间门线（坡面里程194.4米，位于起始朝向的右侧，横向范围[-4.8, -0.8]米，含端点），才记为过门。画面能分辨时，应见双脚连同两板尖从该对蓝杆上坡侧全部进入下坡侧；仅躯干越线、跨杆两侧或仍有一点未过均不算，尚未记过门时逆向退回的点须重过，碰倒杆本身不扣分。", + "step": { + "cat": "本体", + "item": "行进" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚经固定雪板支承在雪面上，双腿和躯干调节重心完成过门。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:719-731", + "inputs/task/privileged/environment/rfo_task.py:866-881", + "inputs/task/privileged/environment/rfo_task.py:899-930", + "inputs/task/privileged/environment/rfo_task.py:1034-1037", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/instruction.md:34-38" + ], + "conf": 5, + "note_zh": "每门仅一个步骤，同时接受两板尖与双脚；按裁判已通过门数锁存，过门后不要求停稳，也不随机器人远离而失去。固定远景中后段旗门和机器人过小或离开画面，不能从不可见推断未完成；需结合该次裁判阶段记录。" + }, + { + "id": "gate_21_passed", + "zh": "通过第21门", + "mode": "event", + "when": null, + "after": [ + "gate_20_passed" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "第20道门已通过后，两只雪板尖和双脚四个受检点均沿下坡方向穿过第21道红色旗门的两杆间门线（坡面里程204米，位于起始朝向的左侧，横向范围[0.8, 4.8]米，含端点），才记为过门。画面能分辨时，应见双脚连同两板尖从该对红杆上坡侧全部进入下坡侧；仅躯干越线、跨杆两侧或仍有一点未过均不算，尚未记过门时逆向退回的点须重过，碰倒杆本身不扣分。", + "step": { + "cat": "本体", + "item": "行进" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚经固定雪板支承在雪面上，双腿和躯干调节重心完成过门。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:719-731", + "inputs/task/privileged/environment/rfo_task.py:866-881", + "inputs/task/privileged/environment/rfo_task.py:899-930", + "inputs/task/privileged/environment/rfo_task.py:1034-1037", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/instruction.md:34-38" + ], + "conf": 5, + "note_zh": "每门仅一个步骤，同时接受两板尖与双脚；按裁判已通过门数锁存，过门后不要求停稳，也不随机器人远离而失去。固定远景中后段旗门和机器人过小或离开画面，不能从不可见推断未完成；需结合该次裁判阶段记录。" + }, + { + "id": "gate_22_passed", + "zh": "通过第22门", + "mode": "event", + "when": null, + "after": [ + "gate_21_passed" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "第21道门已通过后，两只雪板尖和双脚四个受检点均沿下坡方向穿过第22道蓝色旗门的两杆间门线（坡面里程213.6米，位于起始朝向的右侧，横向范围[-4.8, -0.8]米，含端点），才记为过门。画面能分辨时，应见双脚连同两板尖从该对蓝杆上坡侧全部进入下坡侧；仅躯干越线、跨杆两侧或仍有一点未过均不算，尚未记过门时逆向退回的点须重过，碰倒杆本身不扣分。", + "step": { + "cat": "本体", + "item": "行进" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚经固定雪板支承在雪面上，双腿和躯干调节重心完成过门。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:719-731", + "inputs/task/privileged/environment/rfo_task.py:866-881", + "inputs/task/privileged/environment/rfo_task.py:899-930", + "inputs/task/privileged/environment/rfo_task.py:1034-1037", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/instruction.md:34-38" + ], + "conf": 5, + "note_zh": "每门仅一个步骤，同时接受两板尖与双脚；按裁判已通过门数锁存，过门后不要求停稳，也不随机器人远离而失去。固定远景中后段旗门和机器人过小或离开画面，不能从不可见推断未完成；需结合该次裁判阶段记录。" + }, + { + "id": "gate_23_passed", + "zh": "通过第23门", + "mode": "event", + "when": null, + "after": [ + "gate_22_passed" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "第22道门已通过后，两只雪板尖和双脚四个受检点均沿下坡方向穿过第23道红色旗门的两杆间门线（坡面里程223.2米，位于起始朝向的左侧，横向范围[0.8, 4.8]米，含端点），才记为过门。画面能分辨时，应见双脚连同两板尖从该对红杆上坡侧全部进入下坡侧；仅躯干越线、跨杆两侧或仍有一点未过均不算，尚未记过门时逆向退回的点须重过，碰倒杆本身不扣分。", + "step": { + "cat": "本体", + "item": "行进" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚经固定雪板支承在雪面上，双腿和躯干调节重心完成过门。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:719-731", + "inputs/task/privileged/environment/rfo_task.py:866-881", + "inputs/task/privileged/environment/rfo_task.py:899-930", + "inputs/task/privileged/environment/rfo_task.py:1034-1037", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/instruction.md:34-38" + ], + "conf": 5, + "note_zh": "每门仅一个步骤，同时接受两板尖与双脚；按裁判已通过门数锁存，过门后不要求停稳，也不随机器人远离而失去。固定远景中后段旗门和机器人过小或离开画面，不能从不可见推断未完成；需结合该次裁判阶段记录。" + }, + { + "id": "gate_24_passed", + "zh": "通过第24门", + "mode": "event", + "when": null, + "after": [ + "gate_23_passed" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "第23道门已通过后，两只雪板尖和双脚四个受检点均沿下坡方向穿过第24道蓝色旗门的两杆间门线（坡面里程232.8米，位于起始朝向的右侧，横向范围[-4.8, -0.8]米，含端点），才记为过门。画面能分辨时，应见双脚连同两板尖从该对蓝杆上坡侧全部进入下坡侧；仅躯干越线、跨杆两侧或仍有一点未过均不算，尚未记过门时逆向退回的点须重过，碰倒杆本身不扣分。", + "step": { + "cat": "本体", + "item": "行进" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚经固定雪板支承在雪面上，双腿和躯干调节重心完成过门。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:719-731", + "inputs/task/privileged/environment/rfo_task.py:866-881", + "inputs/task/privileged/environment/rfo_task.py:899-930", + "inputs/task/privileged/environment/rfo_task.py:1034-1037", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/instruction.md:34-38" + ], + "conf": 5, + "note_zh": "每门仅一个步骤，同时接受两板尖与双脚；按裁判已通过门数锁存，过门后不要求停稳，也不随机器人远离而失去。固定远景中后段旗门和机器人过小或离开画面，不能从不可见推断未完成；需结合该次裁判阶段记录。" + }, + { + "id": "gate_25_passed", + "zh": "通过第25门", + "mode": "event", + "when": null, + "after": [ + "gate_24_passed" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "第24道门已通过后，两只雪板尖和双脚四个受检点均沿下坡方向穿过第25道红色旗门的两杆间门线（坡面里程242.4米，位于起始朝向的左侧，横向范围[0.8, 4.8]米，含端点），才记为过门。画面能分辨时，应见双脚连同两板尖从该对红杆上坡侧全部进入下坡侧；仅躯干越线、跨杆两侧或仍有一点未过均不算，尚未记过门时逆向退回的点须重过，碰倒杆本身不扣分。", + "step": { + "cat": "本体", + "item": "行进" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚经固定雪板支承在雪面上，双腿和躯干调节重心完成过门。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:719-731", + "inputs/task/privileged/environment/rfo_task.py:866-881", + "inputs/task/privileged/environment/rfo_task.py:899-930", + "inputs/task/privileged/environment/rfo_task.py:1034-1037", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/instruction.md:34-38" + ], + "conf": 5, + "note_zh": "每门仅一个步骤，同时接受两板尖与双脚；按裁判已通过门数锁存，过门后不要求停稳，也不随机器人远离而失去。固定远景中后段旗门和机器人过小或离开画面，不能从不可见推断未完成；需结合该次裁判阶段记录。" + }, + { + "id": "gate_26_passed", + "zh": "通过第26门", + "mode": "event", + "when": null, + "after": [ + "gate_25_passed" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "第25道门已通过后，两只雪板尖和双脚四个受检点均沿下坡方向穿过第26道蓝色旗门的两杆间门线（坡面里程252米，位于起始朝向的右侧，横向范围[-4.8, -0.8]米，含端点），才记为过门。画面能分辨时，应见双脚连同两板尖从该对蓝杆上坡侧全部进入下坡侧；仅躯干越线、跨杆两侧或仍有一点未过均不算，尚未记过门时逆向退回的点须重过，碰倒杆本身不扣分。", + "step": { + "cat": "本体", + "item": "行进" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚经固定雪板支承在雪面上，双腿和躯干调节重心完成过门。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:719-731", + "inputs/task/privileged/environment/rfo_task.py:866-881", + "inputs/task/privileged/environment/rfo_task.py:899-930", + "inputs/task/privileged/environment/rfo_task.py:1034-1037", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/instruction.md:34-38" + ], + "conf": 5, + "note_zh": "每门仅一个步骤，同时接受两板尖与双脚；按裁判已通过门数锁存，过门后不要求停稳，也不随机器人远离而失去。固定远景中后段旗门和机器人过小或离开画面，不能从不可见推断未完成；需结合该次裁判阶段记录。" + }, + { + "id": "gate_27_passed", + "zh": "通过第27门", + "mode": "event", + "when": null, + "after": [ + "gate_26_passed" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "第26道门已通过后，两只雪板尖和双脚四个受检点均沿下坡方向穿过第27道红色旗门的两杆间门线（坡面里程261.6米，位于起始朝向的左侧，横向范围[0.8, 4.8]米，含端点），才记为过门。画面能分辨时，应见双脚连同两板尖从该对红杆上坡侧全部进入下坡侧；仅躯干越线、跨杆两侧或仍有一点未过均不算，尚未记过门时逆向退回的点须重过，碰倒杆本身不扣分。", + "step": { + "cat": "本体", + "item": "行进" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚经固定雪板支承在雪面上，双腿和躯干调节重心完成过门。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:719-731", + "inputs/task/privileged/environment/rfo_task.py:866-881", + "inputs/task/privileged/environment/rfo_task.py:899-930", + "inputs/task/privileged/environment/rfo_task.py:1034-1037", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/instruction.md:34-38" + ], + "conf": 5, + "note_zh": "每门仅一个步骤，同时接受两板尖与双脚；按裁判已通过门数锁存，过门后不要求停稳，也不随机器人远离而失去。固定远景中后段旗门和机器人过小或离开画面，不能从不可见推断未完成；需结合该次裁判阶段记录。" + }, + { + "id": "gate_28_passed", + "zh": "通过第28门", + "mode": "event", + "when": null, + "after": [ + "gate_27_passed" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "第27道门已通过后，两只雪板尖和双脚四个受检点均沿下坡方向穿过第28道蓝色旗门的两杆间门线（坡面里程271.2米，位于起始朝向的右侧，横向范围[-4.8, -0.8]米，含端点），才记为过门。画面能分辨时，应见双脚连同两板尖从该对蓝杆上坡侧全部进入下坡侧；仅躯干越线、跨杆两侧或仍有一点未过均不算，尚未记过门时逆向退回的点须重过，碰倒杆本身不扣分。", + "step": { + "cat": "本体", + "item": "行进" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚经固定雪板支承在雪面上，双腿和躯干调节重心完成过门。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:719-731", + "inputs/task/privileged/environment/rfo_task.py:866-881", + "inputs/task/privileged/environment/rfo_task.py:899-930", + "inputs/task/privileged/environment/rfo_task.py:1034-1037", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/instruction.md:34-38" + ], + "conf": 5, + "note_zh": "每门仅一个步骤，同时接受两板尖与双脚；按裁判已通过门数锁存，过门后不要求停稳，也不随机器人远离而失去。固定远景中后段旗门和机器人过小或离开画面，不能从不可见推断未完成；需结合该次裁判阶段记录。" + }, + { + "id": "gate_29_passed", + "zh": "通过第29门", + "mode": "event", + "when": null, + "after": [ + "gate_28_passed" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "第28道门已通过后，两只雪板尖和双脚四个受检点均沿下坡方向穿过第29道红色旗门的两杆间门线（坡面里程280.8米，位于起始朝向的左侧，横向范围[0.8, 4.8]米，含端点），才记为过门。画面能分辨时，应见双脚连同两板尖从该对红杆上坡侧全部进入下坡侧；仅躯干越线、跨杆两侧或仍有一点未过均不算，尚未记过门时逆向退回的点须重过，碰倒杆本身不扣分。", + "step": { + "cat": "本体", + "item": "行进" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚经固定雪板支承在雪面上，双腿和躯干调节重心完成过门。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:719-731", + "inputs/task/privileged/environment/rfo_task.py:866-881", + "inputs/task/privileged/environment/rfo_task.py:899-930", + "inputs/task/privileged/environment/rfo_task.py:1034-1037", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/instruction.md:34-38" + ], + "conf": 5, + "note_zh": "每门仅一个步骤，同时接受两板尖与双脚；按裁判已通过门数锁存，过门后不要求停稳，也不随机器人远离而失去。固定远景中后段旗门和机器人过小或离开画面，不能从不可见推断未完成；需结合该次裁判阶段记录。" + }, + { + "id": "gate_30_passed", + "zh": "通过第30门", + "mode": "event", + "when": null, + "after": [ + "gate_29_passed" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "第29道门已通过后，两只雪板尖和双脚四个受检点均沿下坡方向穿过第30道蓝色旗门的两杆间门线（坡面里程290.4米，位于起始朝向的右侧，横向范围[-4.8, -0.8]米，含端点），才记为过门。画面能分辨时，应见双脚连同两板尖从该对蓝杆上坡侧全部进入下坡侧；仅躯干越线、跨杆两侧或仍有一点未过均不算，尚未记过门时逆向退回的点须重过，碰倒杆本身不扣分。", + "step": { + "cat": "本体", + "item": "行进" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚经固定雪板支承在雪面上，双腿和躯干调节重心完成过门。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:719-731", + "inputs/task/privileged/environment/rfo_task.py:866-881", + "inputs/task/privileged/environment/rfo_task.py:899-930", + "inputs/task/privileged/environment/rfo_task.py:1034-1037", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/instruction.md:34-38" + ], + "conf": 5, + "note_zh": "每门仅一个步骤，同时接受两板尖与双脚；按裁判已通过门数锁存，过门后不要求停稳，也不随机器人远离而失去。固定远景中后段旗门和机器人过小或离开画面，不能从不可见推断未完成；需结合该次裁判阶段记录。" + }, + { + "id": "gate_31_passed", + "zh": "通过第31门", + "mode": "event", + "when": null, + "after": [ + "gate_30_passed" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "第30道门已通过后，两只雪板尖和双脚四个受检点均沿下坡方向穿过第31道红色旗门的两杆间门线（坡面里程300米，位于起始朝向的左侧，横向范围[0.8, 4.8]米，含端点），才记为过门。画面能分辨时，应见双脚连同两板尖从该对红杆上坡侧全部进入下坡侧；仅躯干越线、跨杆两侧或仍有一点未过均不算，尚未记过门时逆向退回的点须重过，碰倒杆本身不扣分。", + "step": { + "cat": "本体", + "item": "行进" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚经固定雪板支承在雪面上，双腿和躯干调节重心完成过门。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:719-731", + "inputs/task/privileged/environment/rfo_task.py:866-881", + "inputs/task/privileged/environment/rfo_task.py:899-930", + "inputs/task/privileged/environment/rfo_task.py:1034-1037", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/instruction.md:34-38" + ], + "conf": 5, + "note_zh": "每门仅一个步骤，同时接受两板尖与双脚；按裁判已通过门数锁存，过门后不要求停稳，也不随机器人远离而失去。固定远景中后段旗门和机器人过小或离开画面，不能从不可见推断未完成；需结合该次裁判阶段记录。" + }, + { + "id": "gate_32_passed", + "zh": "通过第32门", + "mode": "event", + "when": null, + "after": [ + "gate_31_passed" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "第31道门已通过后，两只雪板尖和双脚四个受检点均沿下坡方向穿过第32道蓝色旗门的两杆间门线（坡面里程309.6米，位于起始朝向的右侧，横向范围[-4.8, -0.8]米，含端点），才记为过门。画面能分辨时，应见双脚连同两板尖从该对蓝杆上坡侧全部进入下坡侧；仅躯干越线、跨杆两侧或仍有一点未过均不算，尚未记过门时逆向退回的点须重过，碰倒杆本身不扣分。", + "step": { + "cat": "本体", + "item": "行进" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚经固定雪板支承在雪面上，双腿和躯干调节重心完成过门。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:719-731", + "inputs/task/privileged/environment/rfo_task.py:866-881", + "inputs/task/privileged/environment/rfo_task.py:899-930", + "inputs/task/privileged/environment/rfo_task.py:1034-1037", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/instruction.md:34-38" + ], + "conf": 5, + "note_zh": "每门仅一个步骤，同时接受两板尖与双脚；按裁判已通过门数锁存，过门后不要求停稳，也不随机器人远离而失去。固定远景中后段旗门和机器人过小或离开画面，不能从不可见推断未完成；需结合该次裁判阶段记录。" + }, + { + "id": "finish_crossed", + "zh": "越过终点", + "mode": "event", + "when": null, + "after": [ + "gate_32_passed" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "32道旗门均依次通过后，双脚都到达坡面里程320米的终点线下坡侧，且检查当时双脚横向坐标都在[-6, 6]米内；须在从回合初态起90秒的总时限内由裁判接受，无需停下或静置。画面能分辨时，应见机器人双脚滑至终点横幅下两立柱之间并越过地面终点线；只看见起点黑色拱架或机器人消失在远景中不足以确认，双脚未全部过线或检查时有脚在柱外即未完成。", + "step": { + "cat": "本体", + "item": "行进" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚经固定雪板支承在雪面上，双腿和躯干控制滑行至终点两柱之间。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:734-736", + "inputs/task/privileged/environment/rfo_task.py:780-787", + "inputs/task/privileged/environment/rfo_task.py:931-942", + "inputs/task/privileged/environment/rfo_task.py:1033-1037", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:174-178", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:283-289", + "inputs/oracle/run.json: reward.metric_gates, reward.metric_time_s, reward.stages_reached" + ], + "conf": 5, + "note_zh": "终点按代码在双脚均位于终点线下坡侧时检查其当前横坐标，非逐脚越线插值；与文字差异见疑问。成功立即锁存，不另设站稳节点。成功录像后段机器人及终点离开固定远景的可辨范围，最终达成由该次裁判记录支持。" + } + ], + "constraints": [ + { + "id": "fell", + "zh": "摔倒", + "never": null, + "criterion_zh": "成功前，脚和手以外任一机器人部位实际接触雪面（每1毫秒物理步检查），或骨盆沿当地雪面法向的高度小于0.35米（每10毫秒控制步检查），裁判判摔倒并终止；手触雪及雪板正常支承不构成摔倒。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:672-673", + "inputs/task/privileged/environment/rfo_task.py:883-893", + "inputs/task/privileged/environment/rfo_task.py:944-951", + "inputs/task/privileged/environment/rfo_task.py:1037" + ] + }, + { + "id": "missed_gate", + "zh": "漏过旗门", + "never": null, + "criterion_zh": "当前旗门尚未记为通过时，任一受检板尖或脚向下跨过后续旗门线，立即判漏门；抵达终点的代码分支则在双脚都到终点线下坡侧而仍有未过旗门时判漏门。当前门仍有点未过、但尚未越过后续门线，仅为尚未完成，不立即构成漏门。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:907-935" + ] + }, + { + "id": "stopped_after_start", + "zh": "起滑后停滞", + "never": null, + "criterion_zh": "起点阶段已成立后，骨盆平移速度模长连续150个控制步低于0.1米/秒（150×0.01秒=1.5秒），裁判判停滞并终止；起点前不应用此规则，任一步达到或超过0.1米/秒会重置停滞计数。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:952-961", + "inputs/task/privileged/environment/rfo_task.py:1037" + ] + }, + { + "id": "did_not_start", + "zh": "未按时起滑", + "never": null, + "criterion_zh": "回合时间达到20秒的控制步检查时，双脚尚未都跨过起点线，裁判判未起滑并终止；这是未出发规则，区别于90秒总回合上限。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:899-902", + "inputs/task/privileged/environment/rfo_task.py:954-958", + "inputs/task/privileged/environment/rfo_task.py:1035" + ] + }, + { + "id": "missed_finish", + "zh": "错过终点", + "never": null, + "criterion_zh": "所有旗门已过且双脚都位于终点线下坡侧时，任一脚当时的横向坐标绝对值大于6米，裁判判错过终点并终止；代码不在第一只脚单独越线时立即判罚。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:931-942", + "inputs/task/privileged/environment/rfo_task.py:1037" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "gate_1_passed", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:926-930; log task.trackers.passed >= 1 or referee.stages membership gate 1; no state log supplied" + }, + { + "atom": "gate_2_passed", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:926-930; log task.trackers.passed >= 2 or referee.stages membership gate 2; no state log supplied" + }, + { + "atom": "gate_3_passed", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:926-930; log task.trackers.passed >= 3 or referee.stages membership gate 3; no state log supplied" + }, + { + "atom": "gate_4_passed", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:926-930; log task.trackers.passed >= 4 or referee.stages membership gate 4; no state log supplied" + }, + { + "atom": "gate_5_passed", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:926-930; log task.trackers.passed >= 5 or referee.stages membership gate 5; no state log supplied" + }, + { + "atom": "gate_6_passed", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:926-930; log task.trackers.passed >= 6 or referee.stages membership gate 6; no state log supplied" + }, + { + "atom": "gate_7_passed", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:926-930; log task.trackers.passed >= 7 or referee.stages membership gate 7; no state log supplied" + }, + { + "atom": "gate_8_passed", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:926-930; log task.trackers.passed >= 8 or referee.stages membership gate 8; no state log supplied" + }, + { + "atom": "gate_9_passed", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:926-930; log task.trackers.passed >= 9 or referee.stages membership gate 9; no state log supplied" + }, + { + "atom": "gate_10_passed", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:926-930; log task.trackers.passed >= 10 or referee.stages membership gate 10; no state log supplied" + }, + { + "atom": "gate_11_passed", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:926-930; log task.trackers.passed >= 11 or referee.stages membership gate 11; no state log supplied" + }, + { + "atom": "gate_12_passed", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:926-930; log task.trackers.passed >= 12 or referee.stages membership gate 12; no state log supplied" + }, + { + "atom": "gate_13_passed", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:926-930; log task.trackers.passed >= 13 or referee.stages membership gate 13; no state log supplied" + }, + { + "atom": "gate_14_passed", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:926-930; log task.trackers.passed >= 14 or referee.stages membership gate 14; no state log supplied" + }, + { + "atom": "gate_15_passed", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:926-930; log task.trackers.passed >= 15 or referee.stages membership gate 15; no state log supplied" + }, + { + "atom": "gate_16_passed", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:926-930; log task.trackers.passed >= 16 or referee.stages membership gate 16; no state log supplied" + }, + { + "atom": "gate_17_passed", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:926-930; log task.trackers.passed >= 17 or referee.stages membership gate 17; no state log supplied" + }, + { + "atom": "gate_18_passed", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:926-930; log task.trackers.passed >= 18 or referee.stages membership gate 18; no state log supplied" + }, + { + "atom": "gate_19_passed", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:926-930; log task.trackers.passed >= 19 or referee.stages membership gate 19; no state log supplied" + }, + { + "atom": "gate_20_passed", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:926-930; log task.trackers.passed >= 20 or referee.stages membership gate 20; no state log supplied" + }, + { + "atom": "gate_21_passed", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:926-930; log task.trackers.passed >= 21 or referee.stages membership gate 21; no state log supplied" + }, + { + "atom": "gate_22_passed", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:926-930; log task.trackers.passed >= 22 or referee.stages membership gate 22; no state log supplied" + }, + { + "atom": "gate_23_passed", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:926-930; log task.trackers.passed >= 23 or referee.stages membership gate 23; no state log supplied" + }, + { + "atom": "gate_24_passed", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:926-930; log task.trackers.passed >= 24 or referee.stages membership gate 24; no state log supplied" + }, + { + "atom": "gate_25_passed", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:926-930; log task.trackers.passed >= 25 or referee.stages membership gate 25; no state log supplied" + }, + { + "atom": "gate_26_passed", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:926-930; log task.trackers.passed >= 26 or referee.stages membership gate 26; no state log supplied" + }, + { + "atom": "gate_27_passed", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:926-930; log task.trackers.passed >= 27 or referee.stages membership gate 27; no state log supplied" + }, + { + "atom": "gate_28_passed", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:926-930; log task.trackers.passed >= 28 or referee.stages membership gate 28; no state log supplied" + }, + { + "atom": "gate_29_passed", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:926-930; log task.trackers.passed >= 29 or referee.stages membership gate 29; no state log supplied" + }, + { + "atom": "gate_30_passed", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:926-930; log task.trackers.passed >= 30 or referee.stages membership gate 30; no state log supplied" + }, + { + "atom": "gate_31_passed", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:926-930; log task.trackers.passed >= 31 or referee.stages membership gate 31; no state log supplied" + }, + { + "atom": "gate_32_passed", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:926-930; log task.trackers.passed >= 32 or referee.stages membership gate 32; no state log supplied" + }, + { + "atom": "finish_crossed", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:931-940; log referee.stages membership finish; no state log supplied" + }, + { + "atom": "start_crossed", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:899-902; log task.trackers.started and t_start for first-gate acceptance; no state log supplied" + }, + { + "atom": "fell", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:672-673; log the corresponding referee failure event and its measured world-state inputs; no state log supplied" + }, + { + "atom": "missed_gate", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:907-935; log the corresponding referee failure event and its measured world-state inputs; no state log supplied" + }, + { + "atom": "stopped_after_start", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:952-961; log the corresponding referee failure event and its measured world-state inputs; no state log supplied" + }, + { + "atom": "did_not_start", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:899-902; log the corresponding referee failure event and its measured world-state inputs; no state log supplied" + }, + { + "atom": "missed_finish", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:931-942; log the corresponding referee failure event and its measured world-state inputs; no state log supplied" + }, + { + "atom": "referee_success", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:174-178; log per-frame benchmark referee.success, referee.stages, referee.failed; no state log supplied" + } + ] + }, + "native_metrics": [ + { + "name": "gates", + "unit": "count", + "threshold": 32, + "direction": "higher_better", + "note_zh": "裁判已通过旗门计数；达到32之后还必须越过终点。", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:926-930", + "inputs/task/privileged/environment/rfo_task.py:1036" + ] + }, + { + "name": "speed_m_s", + "unit": "m/s", + "threshold": 0.1, + "direction": "higher_better", + "note_zh": "仅表示起滑后停滞判定的速度下界，须连续低于此值1.5秒才失败；不是越快得分越高的评分指标。", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:952-961", + "inputs/task/privileged/environment/rfo_task.py:1037" + ] + }, + { + "name": "time_s", + "unit": "s", + "threshold": null, + "direction": "lower_better", + "note_zh": "终点记录的比赛用时为终点回合时间减起滑时间；90秒限制约束整个回合，不是此净用时的独立阈值；成功为二值，不按用时给分。", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:938", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:284-287" + ] + } + ], + "open_questions_zh": [ + "输入没有任何状态日志（inputs/states/index.json 的 logs 为空）；依P8，全部节点的 when、约束的 never 及 success 均为 null。atoms 的 new_code 仅列出将来需要记录的裁判世界状态，未将运行记录中的最终评分冒充逐帧状态。", + "录像证据限制：已查看成功录像全程每5秒抽帧及0、10、77.35秒原帧。固定远景早段可见黑色起点拱架、红色起点线及机器人向远处红蓝杆滑行，后段无法逐一辨认双脚、板尖、门序或终点；成功记录报告32门、34个原生阶段、7738控制步、比赛用时73.028秒，两次回放终态哈希相同（inputs/oracle/run.json 的 reward 及步骤80）。这不能替代缺失的逐帧过门证据。", + "颗粒度采用32个过门步骤加终点，起滑阶段并入首门验收（S2、S5、S8）；文本的 start 原生阶段不单列，禁止停滞也不把整场32次计数合为一项，逐门计数按S5执行。", + "文本与代码不一致：文字称任一脚在柱外越过终点线即失败，也称未过门时任一板尖或脚越过终点线即漏门；实际代码只在双脚均到终点线下坡侧时检查是否漏门及当时双脚横坐标（rfo_task.py:916-917,931-942），不保留各脚越线瞬间的横坐标。终点节点与相应约束按实际评分器接受规则书写，不新增步骤（N1）。", + "文本与代码的过门细节：两板尖受检位置由雪板中段刚体位置及其前向乘tip_x推算（rfo_task.py:870-873），并非柔性前段真实端点；未通过门的点逆向退回会清除记录，但已通过门的 passed 计数不减少（919-929）。因此过门节点用event，精确验收应以该次裁判几何和阶段为准。", + "场地文字近似坐标有误：instruction.md:10称起点边缘约在x=-1.6米、z=0.14米；按场景profile公式及本实例18.3度、6米圆角计算，边缘约x=-2.833米、z=0.617米（rfo_task.py:689-694）。判据以可见起点红线和场景几何为准，不据文字近似坐标另设阈值。", + "模式交付判定不同：特权模式要求两次新进程回放均成功且终态哈希一致（grader/efe_embodied-olympics-fun-files/grade.py:105-115）；标准模式取服务的实时裁判结果并要求有效终结且无基础设施错误（grader/efe_standard-files/v2_1/grade.py:93-99）。两者任务模块及实例完全相同，DAG建模共同世界状态；交付有效性和回放确定性不作为语义步骤或世界状态原子。", + "20秒未出发是明确的DNS规则，列为立即终止约束；90秒为从初态计起的总回合上限，依P21不列为约束。骨盆速度和净比赛用时虽列原生指标，评分为二值；速度只用于停滞判定，用时不单独决定得分。", + "任务说明还规定数值发散失败，world.py:872-883确实会撤销同一控制步新触发的成功；这是仿真有效性而非可由普通视频独立判断的语义世界步骤，未增设节点。无状态日志时也不能用最终视频位置穷尽该有效性判定。" + ], + "conf": 6, + "conf_reason_zh": "32道门及终点直接来自两模式文本，几何、门序、阈值和终止规则均有代码依据；初态与模式共同核心已核对。没有状态日志，固定远景也无法逐门观察精细接受条件，故不可自动检验且节点置信度受限；文字与代码差异已显式列出。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "33个节点（32道旗门+终点）与文本及裁判一致：门里程12+9.6k、横向区间[0.8,4.8]/[-4.8,-0.8]、终点x=320与|y|≤6、成功条件（全部阶段）均与rfo_task.py一致；链式after即裁判阶段顺序；约束(fell/missed_gate/stopped/did_not_start/missed_finish)均为立即终止且阈值正确（0.35 m、0.1 m/s×1.5 s、20 s、6 m）；无状态日志故when/success为null符合规则，文字与代码差异已列入open_questions_zh。起滑并入首门属S2允许的两种读法之一，不提出修改。", + "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：['gate_1_passed', 'gate_2_passed', 'gate_3_passed', 'gate_4_passed', 'gate_5_passed', 'gate_6_passed', 'gate_7_passed', 'gate_8_passed', 'gate_9_passed', 'gate_10_passed', 'gate_11_passed', 'gate_12_passed', 'gate_13_passed', 'gate_14_passed', 'gate_15_passed', 'gate_16_passed', 'gate_17_passed', 'gate_18_passed', 'gate_19_passed', 'gate_20_passed', 'gate_21_passed', 'gate_22_passed', 'gate_23_passed', 'gate_24_passed', 'gate_25_passed', 'gate_26_passed', 'gate_27_passed', 'gate_28_passed', 'gate_29_passed', 'gate_30_passed', 'gate_31_passed', 'gate_32_passed', 'finish_crossed']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:46:13Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__alpine_slalom/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__alpine_slalom/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "036c99c3647b3a34d64e12757f2a1b735f80c82046e20841cc0b227e2e20a9d7", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/embodied-olympics-fun/alpine_slalom/trajectories/embodied-olympics-fun-alpine-slalom-i00-privileged--ev2gd7x.json b/phase2/embodied-olympics-fun/alpine_slalom/trajectories/embodied-olympics-fun-alpine-slalom-i00-privileged--ev2gd7x.json new file mode 100644 index 0000000000000000000000000000000000000000..cedfe494102246e53fbec0dc7e92057e9892ff78 --- /dev/null +++ b/phase2/embodied-olympics-fun/alpine_slalom/trajectories/embodied-olympics-fun-alpine-slalom-i00-privileged--ev2gd7x.json @@ -0,0 +1,1265 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-alpine-slalom-i00-privileged--ev2gd7x", + "taskKey": "embodied-olympics-fun/alpine_slalom", + "dag_sha256": "0382103c21732bd9a43317e1385d02d1f88b24ff2896d1d6062c6a895ff37a48", + "outcome": { + "success": 1, + "reward": 1.0, + "source": "catalog", + "raw": { + "success": 1, + "replay_success": 1, + "deterministic": 1, + "n_actions": 8341, + "sim_steps": 8341, + "horizon": 9000, + "episode_over": 1, + "progress": 1.0, + "stages_reached": 34, + "stages_total": 34, + "fell": 0, + "replay_s": 73.19, + "bad_trajectory": 0, + "metric_gates": 32.0, + "metric_speed_m_s": 5.35, + "metric_time_s": 79.058, + "grader_error": 0, + "video_rendered": 1, + "agent_sessions": 85, + "agent_resets": 234, + "agent_replays": 4, + "agent_cost_self_reported": 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/embodied-olympics-fun-alpine-slalom-i00-privileged--ev2gd7x/replay.mp4", + "duration_s": 83.44, + "fps": 25.0, + "frames": 2086, + "sha256": "f4624b71b383accd420070c7fd58f1b23db22676944ec17fd4f5ba3a29187250" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 33, + "base": 0, + "peak": 33, + "final": 33, + "nodes": { + "gate_1_passed": { + "method": "visual", + "intervals": [ + [ + 11.7, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 11.7, + "log_step": 224, + "text_zh": "观察：log step 224 的交付文件两次新进程回放报告均为 success=true、8341个控制步、仿真 83.41 s、相同终态哈希；其裁判事件包含 gate 1，事件时刻为仿真 11.684 s。最终评分器也报告 metric_gates=32、stages_reached=34、replay_success=1、deterministic=1。" + }, + { + "t": 11.6, + "log_step": null, + "text_zh": "观察：视频 11.6 s、视频 11.8 s，机器人以屈膝姿态随双脚固定雪板滑行，在旗杆间交替转弯；双脚和两板尖的精确过线先后在远景中不能逐点分辨。" + }, + { + "t": 11.7, + "log_step": 224, + "text_zh": "推断：这是交付文件回放的裁判阶段，而非搜索候选的结果。由8341控制步×0.01 s=仿真83.41 s与视频83.44 s时长相符，且已在视频4.24–4.60 s起点越红线及视频11.6–11.8 s首门附近滑行处核对，按同一回合时钟将该门事件定位为约视频 11.7 s（N6，精度约0.2 s）；成功裁判确认四受检点及前序门均被接受，事件锁存至片尾。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：视频 11.6 s、视频 11.8 s，机器人双脚固定于雪板，腿和躯干在板上承重并随雪板沿雪面滑行转弯；未据此声称是否触碰旗杆。" + } + ] + }, + "gate_2_passed": { + "method": "visual", + "intervals": [ + [ + 14.0, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 14.0, + "log_step": 224, + "text_zh": "观察：log step 224 的交付文件两次新进程回放报告均为 success=true、8341个控制步、仿真 83.41 s、相同终态哈希；其裁判事件包含 gate 2，事件时刻为仿真 13.981 s。最终评分器也报告 metric_gates=32、stages_reached=34、replay_success=1、deterministic=1。" + }, + { + "t": 13.8, + "log_step": null, + "text_zh": "观察：视频 13.8 s、视频 14.0 s、视频 14.2 s，机器人以屈膝姿态随双脚固定雪板滑行，在旗杆间交替转弯；双脚和两板尖的精确过线先后在远景中不能逐点分辨。" + }, + { + "t": 14.0, + "log_step": 224, + "text_zh": "推断：这是交付文件回放的裁判阶段，而非搜索候选的结果。由8341控制步×0.01 s=仿真83.41 s与视频83.44 s时长相符，且已在视频4.24–4.60 s起点越红线及视频11.6–11.8 s首门附近滑行处核对，按同一回合时钟将该门事件定位为约视频 14.0 s（N6，精度约0.2 s）；成功裁判确认四受检点及前序门均被接受，事件锁存至片尾。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：视频 13.8 s、视频 14.0 s、视频 14.2 s，机器人双脚固定于雪板，腿和躯干在板上承重并随雪板沿雪面滑行转弯；未据此声称是否触碰旗杆。" + } + ] + }, + "gate_3_passed": { + "method": "visual", + "intervals": [ + [ + 16.2, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 16.2, + "log_step": 224, + "text_zh": "观察：log step 224 的交付文件两次新进程回放报告均为 success=true、8341个控制步、仿真 83.41 s、相同终态哈希；其裁判事件包含 gate 3，事件时刻为仿真 16.170 s。最终评分器也报告 metric_gates=32、stages_reached=34、replay_success=1、deterministic=1。" + }, + { + "t": 16.0, + "log_step": null, + "text_zh": "观察：视频 16.0 s、视频 16.2 s、视频 16.4 s，机器人以屈膝姿态随双脚固定雪板滑行，在旗杆间交替转弯；双脚和两板尖的精确过线先后在远景中不能逐点分辨。" + }, + { + "t": 16.2, + "log_step": 224, + "text_zh": "推断：这是交付文件回放的裁判阶段，而非搜索候选的结果。由8341控制步×0.01 s=仿真83.41 s与视频83.44 s时长相符，且已在视频4.24–4.60 s起点越红线及视频11.6–11.8 s首门附近滑行处核对，按同一回合时钟将该门事件定位为约视频 16.2 s（N6，精度约0.2 s）；成功裁判确认四受检点及前序门均被接受，事件锁存至片尾。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：视频 16.0 s、视频 16.2 s、视频 16.4 s，机器人双脚固定于雪板，腿和躯干在板上承重并随雪板沿雪面滑行转弯；未据此声称是否触碰旗杆。" + } + ] + }, + "gate_4_passed": { + "method": "visual", + "intervals": [ + [ + 18.6, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 18.6, + "log_step": 224, + "text_zh": "观察：log step 224 的交付文件两次新进程回放报告均为 success=true、8341个控制步、仿真 83.41 s、相同终态哈希；其裁判事件包含 gate 4，事件时刻为仿真 18.560 s。最终评分器也报告 metric_gates=32、stages_reached=34、replay_success=1、deterministic=1。" + }, + { + "t": 18.4, + "log_step": null, + "text_zh": "观察：视频 18.4 s、视频 18.6 s、视频 18.8 s，机器人以屈膝姿态随双脚固定雪板滑行，在旗杆间交替转弯；双脚和两板尖的精确过线先后在远景中不能逐点分辨。" + }, + { + "t": 18.6, + "log_step": 224, + "text_zh": "推断：这是交付文件回放的裁判阶段，而非搜索候选的结果。由8341控制步×0.01 s=仿真83.41 s与视频83.44 s时长相符，且已在视频4.24–4.60 s起点越红线及视频11.6–11.8 s首门附近滑行处核对，按同一回合时钟将该门事件定位为约视频 18.6 s（N6，精度约0.2 s）；成功裁判确认四受检点及前序门均被接受，事件锁存至片尾。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：视频 18.4 s、视频 18.6 s、视频 18.8 s，机器人双脚固定于雪板，腿和躯干在板上承重并随雪板沿雪面滑行转弯；未据此声称是否触碰旗杆。" + } + ] + }, + "gate_5_passed": { + "method": "visual", + "intervals": [ + [ + 21.1, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 21.1, + "log_step": 224, + "text_zh": "观察：log step 224 的交付文件两次新进程回放报告均为 success=true、8341个控制步、仿真 83.41 s、相同终态哈希；其裁判事件包含 gate 5，事件时刻为仿真 21.090 s。最终评分器也报告 metric_gates=32、stages_reached=34、replay_success=1、deterministic=1。" + }, + { + "t": 21.0, + "log_step": null, + "text_zh": "观察：视频 21.0 s、视频 21.2 s，机器人已缩小为远处旗门间的小轮廓，不能分辨足部支承、两板尖及四点过线。" + }, + { + "t": 21.1, + "log_step": 224, + "text_zh": "推断：这是交付文件回放的裁判阶段，而非搜索候选的结果。由8341控制步×0.01 s=仿真83.41 s与视频83.44 s时长相符，且已在视频4.24–4.60 s起点越红线及视频11.6–11.8 s首门附近滑行处核对，按同一回合时钟将该门事件定位为约视频 21.1 s（N6，精度约0.2 s）；成功裁判确认四受检点及前序门均被接受，事件锁存至片尾。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察：视频 21.0 s、视频 21.2 s，该门附近的机器人足部和雪板接触区域太小或已超出可辨范围；无状态接触日志，不能逐门确认实际支承及其他接触。" + } + ] + }, + "gate_6_passed": { + "method": "visual", + "intervals": [ + [ + 23.5, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 23.5, + "log_step": 224, + "text_zh": "观察：log step 224 的交付文件两次新进程回放报告均为 success=true、8341个控制步、仿真 83.41 s、相同终态哈希；其裁判事件包含 gate 6，事件时刻为仿真 23.495 s。最终评分器也报告 metric_gates=32、stages_reached=34、replay_success=1、deterministic=1。" + }, + { + "t": 23.4, + "log_step": null, + "text_zh": "观察：视频 23.4 s、视频 23.6 s，机器人已缩小为远处旗门间的小轮廓，不能分辨足部支承、两板尖及四点过线。" + }, + { + "t": 23.5, + "log_step": 224, + "text_zh": "推断：这是交付文件回放的裁判阶段，而非搜索候选的结果。由8341控制步×0.01 s=仿真83.41 s与视频83.44 s时长相符，且已在视频4.24–4.60 s起点越红线及视频11.6–11.8 s首门附近滑行处核对，按同一回合时钟将该门事件定位为约视频 23.5 s（N6，精度约0.2 s）；成功裁判确认四受检点及前序门均被接受，事件锁存至片尾。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察：视频 23.4 s、视频 23.6 s，该门附近的机器人足部和雪板接触区域太小或已超出可辨范围；无状态接触日志，不能逐门确认实际支承及其他接触。" + } + ] + }, + "gate_7_passed": { + "method": "visual", + "intervals": [ + [ + 25.5, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 25.5, + "log_step": 224, + "text_zh": "观察：log step 224 的交付文件两次新进程回放报告均为 success=true、8341个控制步、仿真 83.41 s、相同终态哈希；其裁判事件包含 gate 7，事件时刻为仿真 25.514 s。最终评分器也报告 metric_gates=32、stages_reached=34、replay_success=1、deterministic=1。" + }, + { + "t": 25.4, + "log_step": null, + "text_zh": "观察：视频 25.4 s、视频 25.6 s，固定相机仍朝向起点和近段旗门，机器人及当前门的足部、板尖已不可辨；与成功参考录像相同的远景限制，画面不提供逐点过门证据。" + }, + { + "t": 25.5, + "log_step": 224, + "text_zh": "推断：这是交付文件回放的裁判阶段，而非搜索候选的结果。由8341控制步×0.01 s=仿真83.41 s与视频83.44 s时长相符，且已在视频4.24–4.60 s起点越红线及视频11.6–11.8 s首门附近滑行处核对，按同一回合时钟将该门事件定位为约视频 25.5 s（N6，精度约0.2 s）；成功裁判确认四受检点及前序门均被接受，事件锁存至片尾。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察：视频 25.4 s、视频 25.6 s，该门附近的机器人足部和雪板接触区域太小或已超出可辨范围；无状态接触日志，不能逐门确认实际支承及其他接触。" + } + ] + }, + "gate_8_passed": { + "method": "visual", + "intervals": [ + [ + 27.9, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 27.9, + "log_step": 224, + "text_zh": "观察：log step 224 的交付文件两次新进程回放报告均为 success=true、8341个控制步、仿真 83.41 s、相同终态哈希；其裁判事件包含 gate 8，事件时刻为仿真 27.862 s。最终评分器也报告 metric_gates=32、stages_reached=34、replay_success=1、deterministic=1。" + }, + { + "t": 27.8, + "log_step": null, + "text_zh": "观察：视频 27.8 s、视频 28.0 s，固定相机仍朝向起点和近段旗门，机器人及当前门的足部、板尖已不可辨；与成功参考录像相同的远景限制，画面不提供逐点过门证据。" + }, + { + "t": 27.9, + "log_step": 224, + "text_zh": "推断：这是交付文件回放的裁判阶段，而非搜索候选的结果。由8341控制步×0.01 s=仿真83.41 s与视频83.44 s时长相符，且已在视频4.24–4.60 s起点越红线及视频11.6–11.8 s首门附近滑行处核对，按同一回合时钟将该门事件定位为约视频 27.9 s（N6，精度约0.2 s）；成功裁判确认四受检点及前序门均被接受，事件锁存至片尾。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察：视频 27.8 s、视频 28.0 s，该门附近的机器人足部和雪板接触区域太小或已超出可辨范围；无状态接触日志，不能逐门确认实际支承及其他接触。" + } + ] + }, + "gate_9_passed": { + "method": "visual", + "intervals": [ + [ + 30.1, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 30.1, + "log_step": 224, + "text_zh": "观察：log step 224 的交付文件两次新进程回放报告均为 success=true、8341个控制步、仿真 83.41 s、相同终态哈希；其裁判事件包含 gate 9，事件时刻为仿真 30.073 s。最终评分器也报告 metric_gates=32、stages_reached=34、replay_success=1、deterministic=1。" + }, + { + "t": 30.0, + "log_step": null, + "text_zh": "观察：视频 30.0 s、视频 30.2 s，固定相机仍朝向起点和近段旗门，机器人及当前门的足部、板尖已不可辨；与成功参考录像相同的远景限制，画面不提供逐点过门证据。" + }, + { + "t": 30.1, + "log_step": 224, + "text_zh": "推断：这是交付文件回放的裁判阶段，而非搜索候选的结果。由8341控制步×0.01 s=仿真83.41 s与视频83.44 s时长相符，且已在视频4.24–4.60 s起点越红线及视频11.6–11.8 s首门附近滑行处核对，按同一回合时钟将该门事件定位为约视频 30.1 s（N6，精度约0.2 s）；成功裁判确认四受检点及前序门均被接受，事件锁存至片尾。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察：视频 30.0 s、视频 30.2 s，该门附近的机器人足部和雪板接触区域太小或已超出可辨范围；无状态接触日志，不能逐门确认实际支承及其他接触。" + } + ] + }, + "gate_10_passed": { + "method": "visual", + "intervals": [ + [ + 32.4, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 32.4, + "log_step": 224, + "text_zh": "观察：log step 224 的交付文件两次新进程回放报告均为 success=true、8341个控制步、仿真 83.41 s、相同终态哈希；其裁判事件包含 gate 10，事件时刻为仿真 32.434 s。最终评分器也报告 metric_gates=32、stages_reached=34、replay_success=1、deterministic=1。" + }, + { + "t": 32.2, + "log_step": null, + "text_zh": "观察：视频 32.2 s、视频 32.4 s、视频 32.6 s，固定相机仍朝向起点和近段旗门，机器人及当前门的足部、板尖已不可辨；与成功参考录像相同的远景限制，画面不提供逐点过门证据。" + }, + { + "t": 32.4, + "log_step": 224, + "text_zh": "推断：这是交付文件回放的裁判阶段，而非搜索候选的结果。由8341控制步×0.01 s=仿真83.41 s与视频83.44 s时长相符，且已在视频4.24–4.60 s起点越红线及视频11.6–11.8 s首门附近滑行处核对，按同一回合时钟将该门事件定位为约视频 32.4 s（N6，精度约0.2 s）；成功裁判确认四受检点及前序门均被接受，事件锁存至片尾。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察：视频 32.2 s、视频 32.4 s、视频 32.6 s，该门附近的机器人足部和雪板接触区域太小或已超出可辨范围；无状态接触日志，不能逐门确认实际支承及其他接触。" + } + ] + }, + "gate_11_passed": { + "method": "visual", + "intervals": [ + [ + 34.7, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 34.7, + "log_step": 224, + "text_zh": "观察：log step 224 的交付文件两次新进程回放报告均为 success=true、8341个控制步、仿真 83.41 s、相同终态哈希；其裁判事件包含 gate 11，事件时刻为仿真 34.676 s。最终评分器也报告 metric_gates=32、stages_reached=34、replay_success=1、deterministic=1。" + }, + { + "t": 34.6, + "log_step": null, + "text_zh": "观察：视频 34.6 s、视频 34.8 s，固定相机仍朝向起点和近段旗门，机器人及当前门的足部、板尖已不可辨；与成功参考录像相同的远景限制，画面不提供逐点过门证据。" + }, + { + "t": 34.7, + "log_step": 224, + "text_zh": "推断：这是交付文件回放的裁判阶段，而非搜索候选的结果。由8341控制步×0.01 s=仿真83.41 s与视频83.44 s时长相符，且已在视频4.24–4.60 s起点越红线及视频11.6–11.8 s首门附近滑行处核对，按同一回合时钟将该门事件定位为约视频 34.7 s（N6，精度约0.2 s）；成功裁判确认四受检点及前序门均被接受，事件锁存至片尾。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察：视频 34.6 s、视频 34.8 s，该门附近的机器人足部和雪板接触区域太小或已超出可辨范围；无状态接触日志，不能逐门确认实际支承及其他接触。" + } + ] + }, + "gate_12_passed": { + "method": "visual", + "intervals": [ + [ + 36.8, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 36.8, + "log_step": 224, + "text_zh": "观察：log step 224 的交付文件两次新进程回放报告均为 success=true、8341个控制步、仿真 83.41 s、相同终态哈希；其裁判事件包含 gate 12，事件时刻为仿真 36.792 s。最终评分器也报告 metric_gates=32、stages_reached=34、replay_success=1、deterministic=1。" + }, + { + "t": 36.6, + "log_step": null, + "text_zh": "观察：视频 36.6 s、视频 36.8 s、视频 37.0 s，固定相机仍朝向起点和近段旗门，机器人及当前门的足部、板尖已不可辨；与成功参考录像相同的远景限制，画面不提供逐点过门证据。" + }, + { + "t": 36.8, + "log_step": 224, + "text_zh": "推断：这是交付文件回放的裁判阶段，而非搜索候选的结果。由8341控制步×0.01 s=仿真83.41 s与视频83.44 s时长相符，且已在视频4.24–4.60 s起点越红线及视频11.6–11.8 s首门附近滑行处核对，按同一回合时钟将该门事件定位为约视频 36.8 s（N6，精度约0.2 s）；成功裁判确认四受检点及前序门均被接受，事件锁存至片尾。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察：视频 36.6 s、视频 36.8 s、视频 37.0 s，该门附近的机器人足部和雪板接触区域太小或已超出可辨范围；无状态接触日志，不能逐门确认实际支承及其他接触。" + } + ] + }, + "gate_13_passed": { + "method": "visual", + "intervals": [ + [ + 39.0, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 39.0, + "log_step": 224, + "text_zh": "观察：log step 224 的交付文件两次新进程回放报告均为 success=true、8341个控制步、仿真 83.41 s、相同终态哈希；其裁判事件包含 gate 13，事件时刻为仿真 39.029 s。最终评分器也报告 metric_gates=32、stages_reached=34、replay_success=1、deterministic=1。" + }, + { + "t": 38.8, + "log_step": null, + "text_zh": "观察：视频 38.8 s、视频 39.0 s、视频 39.2 s，固定相机仍朝向起点和近段旗门，机器人及当前门的足部、板尖已不可辨；与成功参考录像相同的远景限制，画面不提供逐点过门证据。" + }, + { + "t": 39.0, + "log_step": 224, + "text_zh": "推断：这是交付文件回放的裁判阶段，而非搜索候选的结果。由8341控制步×0.01 s=仿真83.41 s与视频83.44 s时长相符，且已在视频4.24–4.60 s起点越红线及视频11.6–11.8 s首门附近滑行处核对，按同一回合时钟将该门事件定位为约视频 39.0 s（N6，精度约0.2 s）；成功裁判确认四受检点及前序门均被接受，事件锁存至片尾。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察：视频 38.8 s、视频 39.0 s、视频 39.2 s，该门附近的机器人足部和雪板接触区域太小或已超出可辨范围；无状态接触日志，不能逐门确认实际支承及其他接触。" + } + ] + }, + "gate_14_passed": { + "method": "visual", + "intervals": [ + [ + 41.2, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 41.2, + "log_step": 224, + "text_zh": "观察：log step 224 的交付文件两次新进程回放报告均为 success=true、8341个控制步、仿真 83.41 s、相同终态哈希；其裁判事件包含 gate 14，事件时刻为仿真 41.213 s。最终评分器也报告 metric_gates=32、stages_reached=34、replay_success=1、deterministic=1。" + }, + { + "t": 41.0, + "log_step": null, + "text_zh": "观察：视频 41.0 s、视频 41.2 s、视频 41.4 s，固定相机仍朝向起点和近段旗门，机器人及当前门的足部、板尖已不可辨；与成功参考录像相同的远景限制，画面不提供逐点过门证据。" + }, + { + "t": 41.2, + "log_step": 224, + "text_zh": "推断：这是交付文件回放的裁判阶段，而非搜索候选的结果。由8341控制步×0.01 s=仿真83.41 s与视频83.44 s时长相符，且已在视频4.24–4.60 s起点越红线及视频11.6–11.8 s首门附近滑行处核对，按同一回合时钟将该门事件定位为约视频 41.2 s（N6，精度约0.2 s）；成功裁判确认四受检点及前序门均被接受，事件锁存至片尾。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察：视频 41.0 s、视频 41.2 s、视频 41.4 s，该门附近的机器人足部和雪板接触区域太小或已超出可辨范围；无状态接触日志，不能逐门确认实际支承及其他接触。" + } + ] + }, + "gate_15_passed": { + "method": "visual", + "intervals": [ + [ + 43.5, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 43.5, + "log_step": 224, + "text_zh": "观察：log step 224 的交付文件两次新进程回放报告均为 success=true、8341个控制步、仿真 83.41 s、相同终态哈希；其裁判事件包含 gate 15，事件时刻为仿真 43.506 s。最终评分器也报告 metric_gates=32、stages_reached=34、replay_success=1、deterministic=1。" + }, + { + "t": 43.4, + "log_step": null, + "text_zh": "观察：视频 43.4 s、视频 43.6 s，固定相机仍朝向起点和近段旗门，机器人及当前门的足部、板尖已不可辨；与成功参考录像相同的远景限制，画面不提供逐点过门证据。" + }, + { + "t": 43.5, + "log_step": 224, + "text_zh": "推断：这是交付文件回放的裁判阶段，而非搜索候选的结果。由8341控制步×0.01 s=仿真83.41 s与视频83.44 s时长相符，且已在视频4.24–4.60 s起点越红线及视频11.6–11.8 s首门附近滑行处核对，按同一回合时钟将该门事件定位为约视频 43.5 s（N6，精度约0.2 s）；成功裁判确认四受检点及前序门均被接受，事件锁存至片尾。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察：视频 43.4 s、视频 43.6 s，该门附近的机器人足部和雪板接触区域太小或已超出可辨范围；无状态接触日志，不能逐门确认实际支承及其他接触。" + } + ] + }, + "gate_16_passed": { + "method": "visual", + "intervals": [ + [ + 45.6, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 45.6, + "log_step": 224, + "text_zh": "观察：log step 224 的交付文件两次新进程回放报告均为 success=true、8341个控制步、仿真 83.41 s、相同终态哈希；其裁判事件包含 gate 16，事件时刻为仿真 45.561 s。最终评分器也报告 metric_gates=32、stages_reached=34、replay_success=1、deterministic=1。" + }, + { + "t": 45.4, + "log_step": null, + "text_zh": "观察：视频 45.4 s、视频 45.6 s、视频 45.8 s，固定相机仍朝向起点和近段旗门，机器人及当前门的足部、板尖已不可辨；与成功参考录像相同的远景限制，画面不提供逐点过门证据。" + }, + { + "t": 45.6, + "log_step": 224, + "text_zh": "推断：这是交付文件回放的裁判阶段，而非搜索候选的结果。由8341控制步×0.01 s=仿真83.41 s与视频83.44 s时长相符，且已在视频4.24–4.60 s起点越红线及视频11.6–11.8 s首门附近滑行处核对，按同一回合时钟将该门事件定位为约视频 45.6 s（N6，精度约0.2 s）；成功裁判确认四受检点及前序门均被接受，事件锁存至片尾。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察：视频 45.4 s、视频 45.6 s、视频 45.8 s，该门附近的机器人足部和雪板接触区域太小或已超出可辨范围；无状态接触日志，不能逐门确认实际支承及其他接触。" + } + ] + }, + "gate_17_passed": { + "method": "visual", + "intervals": [ + [ + 47.9, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 47.9, + "log_step": 224, + "text_zh": "观察：log step 224 的交付文件两次新进程回放报告均为 success=true、8341个控制步、仿真 83.41 s、相同终态哈希；其裁判事件包含 gate 17，事件时刻为仿真 47.862 s。最终评分器也报告 metric_gates=32、stages_reached=34、replay_success=1、deterministic=1。" + }, + { + "t": 47.8, + "log_step": null, + "text_zh": "观察：视频 47.8 s、视频 48.0 s，固定相机仍朝向起点和近段旗门，机器人及当前门的足部、板尖已不可辨；与成功参考录像相同的远景限制，画面不提供逐点过门证据。" + }, + { + "t": 47.9, + "log_step": 224, + "text_zh": "推断：这是交付文件回放的裁判阶段，而非搜索候选的结果。由8341控制步×0.01 s=仿真83.41 s与视频83.44 s时长相符，且已在视频4.24–4.60 s起点越红线及视频11.6–11.8 s首门附近滑行处核对，按同一回合时钟将该门事件定位为约视频 47.9 s（N6，精度约0.2 s）；成功裁判确认四受检点及前序门均被接受，事件锁存至片尾。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察：视频 47.8 s、视频 48.0 s，该门附近的机器人足部和雪板接触区域太小或已超出可辨范围；无状态接触日志，不能逐门确认实际支承及其他接触。" + } + ] + }, + "gate_18_passed": { + "method": "visual", + "intervals": [ + [ + 49.9, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 49.9, + "log_step": 224, + "text_zh": "观察：log step 224 的交付文件两次新进程回放报告均为 success=true、8341个控制步、仿真 83.41 s、相同终态哈希；其裁判事件包含 gate 18，事件时刻为仿真 49.879 s。最终评分器也报告 metric_gates=32、stages_reached=34、replay_success=1、deterministic=1。" + }, + { + "t": 49.8, + "log_step": null, + "text_zh": "观察：视频 49.8 s、视频 50.0 s，固定相机仍朝向起点和近段旗门，机器人及当前门的足部、板尖已不可辨；与成功参考录像相同的远景限制，画面不提供逐点过门证据。" + }, + { + "t": 49.9, + "log_step": 224, + "text_zh": "推断：这是交付文件回放的裁判阶段，而非搜索候选的结果。由8341控制步×0.01 s=仿真83.41 s与视频83.44 s时长相符，且已在视频4.24–4.60 s起点越红线及视频11.6–11.8 s首门附近滑行处核对，按同一回合时钟将该门事件定位为约视频 49.9 s（N6，精度约0.2 s）；成功裁判确认四受检点及前序门均被接受，事件锁存至片尾。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察：视频 49.8 s、视频 50.0 s，该门附近的机器人足部和雪板接触区域太小或已超出可辨范围；无状态接触日志，不能逐门确认实际支承及其他接触。" + } + ] + }, + "gate_19_passed": { + "method": "visual", + "intervals": [ + [ + 52.1, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 52.1, + "log_step": 224, + "text_zh": "观察：log step 224 的交付文件两次新进程回放报告均为 success=true、8341个控制步、仿真 83.41 s、相同终态哈希；其裁判事件包含 gate 19，事件时刻为仿真 52.065 s。最终评分器也报告 metric_gates=32、stages_reached=34、replay_success=1、deterministic=1。" + }, + { + "t": 52.0, + "log_step": null, + "text_zh": "观察：视频 52.0 s、视频 52.2 s，固定相机仍朝向起点和近段旗门，机器人及当前门的足部、板尖已不可辨；与成功参考录像相同的远景限制，画面不提供逐点过门证据。" + }, + { + "t": 52.1, + "log_step": 224, + "text_zh": "推断：这是交付文件回放的裁判阶段，而非搜索候选的结果。由8341控制步×0.01 s=仿真83.41 s与视频83.44 s时长相符，且已在视频4.24–4.60 s起点越红线及视频11.6–11.8 s首门附近滑行处核对，按同一回合时钟将该门事件定位为约视频 52.1 s（N6，精度约0.2 s）；成功裁判确认四受检点及前序门均被接受，事件锁存至片尾。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察：视频 52.0 s、视频 52.2 s，该门附近的机器人足部和雪板接触区域太小或已超出可辨范围；无状态接触日志，不能逐门确认实际支承及其他接触。" + } + ] + }, + "gate_20_passed": { + "method": "visual", + "intervals": [ + [ + 54.4, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 54.4, + "log_step": 224, + "text_zh": "观察：log step 224 的交付文件两次新进程回放报告均为 success=true、8341个控制步、仿真 83.41 s、相同终态哈希；其裁判事件包含 gate 20，事件时刻为仿真 54.368 s。最终评分器也报告 metric_gates=32、stages_reached=34、replay_success=1、deterministic=1。" + }, + { + "t": 54.2, + "log_step": null, + "text_zh": "观察：视频 54.2 s、视频 54.4 s、视频 54.6 s，固定相机仍朝向起点和近段旗门，机器人及当前门的足部、板尖已不可辨；与成功参考录像相同的远景限制，画面不提供逐点过门证据。" + }, + { + "t": 54.4, + "log_step": 224, + "text_zh": "推断：这是交付文件回放的裁判阶段，而非搜索候选的结果。由8341控制步×0.01 s=仿真83.41 s与视频83.44 s时长相符，且已在视频4.24–4.60 s起点越红线及视频11.6–11.8 s首门附近滑行处核对，按同一回合时钟将该门事件定位为约视频 54.4 s（N6，精度约0.2 s）；成功裁判确认四受检点及前序门均被接受，事件锁存至片尾。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察：视频 54.2 s、视频 54.4 s、视频 54.6 s，该门附近的机器人足部和雪板接触区域太小或已超出可辨范围；无状态接触日志，不能逐门确认实际支承及其他接触。" + } + ] + }, + "gate_21_passed": { + "method": "visual", + "intervals": [ + [ + 57.9, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 57.9, + "log_step": 224, + "text_zh": "观察：log step 224 的交付文件两次新进程回放报告均为 success=true、8341个控制步、仿真 83.41 s、相同终态哈希；其裁判事件包含 gate 21，事件时刻为仿真 57.851 s。最终评分器也报告 metric_gates=32、stages_reached=34、replay_success=1、deterministic=1。" + }, + { + "t": 57.8, + "log_step": null, + "text_zh": "观察：视频 57.8 s、视频 58.0 s，固定相机仍朝向起点和近段旗门，机器人及当前门的足部、板尖已不可辨；与成功参考录像相同的远景限制，画面不提供逐点过门证据。" + }, + { + "t": 57.9, + "log_step": 224, + "text_zh": "推断：这是交付文件回放的裁判阶段，而非搜索候选的结果。由8341控制步×0.01 s=仿真83.41 s与视频83.44 s时长相符，且已在视频4.24–4.60 s起点越红线及视频11.6–11.8 s首门附近滑行处核对，按同一回合时钟将该门事件定位为约视频 57.9 s（N6，精度约0.2 s）；成功裁判确认四受检点及前序门均被接受，事件锁存至片尾。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察：视频 57.8 s、视频 58.0 s，该门附近的机器人足部和雪板接触区域太小或已超出可辨范围；无状态接触日志，不能逐门确认实际支承及其他接触。" + } + ] + }, + "gate_22_passed": { + "method": "visual", + "intervals": [ + [ + 60.7, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 60.7, + "log_step": 224, + "text_zh": "观察：log step 224 的交付文件两次新进程回放报告均为 success=true、8341个控制步、仿真 83.41 s、相同终态哈希；其裁判事件包含 gate 22，事件时刻为仿真 60.722 s。最终评分器也报告 metric_gates=32、stages_reached=34、replay_success=1、deterministic=1。" + }, + { + "t": 60.6, + "log_step": null, + "text_zh": "观察：视频 60.6 s、视频 60.8 s，固定相机仍朝向起点和近段旗门，机器人及当前门的足部、板尖已不可辨；与成功参考录像相同的远景限制，画面不提供逐点过门证据。" + }, + { + "t": 60.7, + "log_step": 224, + "text_zh": "推断：这是交付文件回放的裁判阶段，而非搜索候选的结果。由8341控制步×0.01 s=仿真83.41 s与视频83.44 s时长相符，且已在视频4.24–4.60 s起点越红线及视频11.6–11.8 s首门附近滑行处核对，按同一回合时钟将该门事件定位为约视频 60.7 s（N6，精度约0.2 s）；成功裁判确认四受检点及前序门均被接受，事件锁存至片尾。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察：视频 60.6 s、视频 60.8 s，该门附近的机器人足部和雪板接触区域太小或已超出可辨范围；无状态接触日志，不能逐门确认实际支承及其他接触。" + } + ] + }, + "gate_23_passed": { + "method": "visual", + "intervals": [ + [ + 63.0, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 63.0, + "log_step": 224, + "text_zh": "观察：log step 224 的交付文件两次新进程回放报告均为 success=true、8341个控制步、仿真 83.41 s、相同终态哈希；其裁判事件包含 gate 23，事件时刻为仿真 62.954 s。最终评分器也报告 metric_gates=32、stages_reached=34、replay_success=1、deterministic=1。" + }, + { + "t": 62.8, + "log_step": null, + "text_zh": "观察：视频 62.8 s、视频 63.0 s，固定相机仍朝向起点和近段旗门，机器人及当前门的足部、板尖已不可辨；与成功参考录像相同的远景限制，画面不提供逐点过门证据。" + }, + { + "t": 63.0, + "log_step": 224, + "text_zh": "推断：这是交付文件回放的裁判阶段，而非搜索候选的结果。由8341控制步×0.01 s=仿真83.41 s与视频83.44 s时长相符，且已在视频4.24–4.60 s起点越红线及视频11.6–11.8 s首门附近滑行处核对，按同一回合时钟将该门事件定位为约视频 63.0 s（N6，精度约0.2 s）；成功裁判确认四受检点及前序门均被接受，事件锁存至片尾。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察：视频 62.8 s、视频 63.0 s，该门附近的机器人足部和雪板接触区域太小或已超出可辨范围；无状态接触日志，不能逐门确认实际支承及其他接触。" + } + ] + }, + "gate_24_passed": { + "method": "visual", + "intervals": [ + [ + 65.0, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 65.0, + "log_step": 224, + "text_zh": "观察：log step 224 的交付文件两次新进程回放报告均为 success=true、8341个控制步、仿真 83.41 s、相同终态哈希；其裁判事件包含 gate 24，事件时刻为仿真 64.999 s。最终评分器也报告 metric_gates=32、stages_reached=34、replay_success=1、deterministic=1。" + }, + { + "t": 64.8, + "log_step": null, + "text_zh": "观察：视频 64.8 s、视频 65.0 s、视频 65.2 s，固定相机仍朝向起点和近段旗门，机器人及当前门的足部、板尖已不可辨；与成功参考录像相同的远景限制，画面不提供逐点过门证据。" + }, + { + "t": 65.0, + "log_step": 224, + "text_zh": "推断：这是交付文件回放的裁判阶段，而非搜索候选的结果。由8341控制步×0.01 s=仿真83.41 s与视频83.44 s时长相符，且已在视频4.24–4.60 s起点越红线及视频11.6–11.8 s首门附近滑行处核对，按同一回合时钟将该门事件定位为约视频 65.0 s（N6，精度约0.2 s）；成功裁判确认四受检点及前序门均被接受，事件锁存至片尾。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察：视频 64.8 s、视频 65.0 s、视频 65.2 s，该门附近的机器人足部和雪板接触区域太小或已超出可辨范围；无状态接触日志，不能逐门确认实际支承及其他接触。" + } + ] + }, + "gate_25_passed": { + "method": "visual", + "intervals": [ + [ + 67.1, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 67.1, + "log_step": 224, + "text_zh": "观察：log step 224 的交付文件两次新进程回放报告均为 success=true、8341个控制步、仿真 83.41 s、相同终态哈希；其裁判事件包含 gate 25，事件时刻为仿真 67.095 s。最终评分器也报告 metric_gates=32、stages_reached=34、replay_success=1、deterministic=1。" + }, + { + "t": 67.0, + "log_step": null, + "text_zh": "观察：视频 67.0 s、视频 67.2 s，固定相机仍朝向起点和近段旗门，机器人及当前门的足部、板尖已不可辨；与成功参考录像相同的远景限制，画面不提供逐点过门证据。" + }, + { + "t": 67.1, + "log_step": 224, + "text_zh": "推断：这是交付文件回放的裁判阶段，而非搜索候选的结果。由8341控制步×0.01 s=仿真83.41 s与视频83.44 s时长相符，且已在视频4.24–4.60 s起点越红线及视频11.6–11.8 s首门附近滑行处核对，按同一回合时钟将该门事件定位为约视频 67.1 s（N6，精度约0.2 s）；成功裁判确认四受检点及前序门均被接受，事件锁存至片尾。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察：视频 67.0 s、视频 67.2 s，该门附近的机器人足部和雪板接触区域太小或已超出可辨范围；无状态接触日志，不能逐门确认实际支承及其他接触。" + } + ] + }, + "gate_26_passed": { + "method": "visual", + "intervals": [ + [ + 69.7, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 69.7, + "log_step": 224, + "text_zh": "观察：log step 224 的交付文件两次新进程回放报告均为 success=true、8341个控制步、仿真 83.41 s、相同终态哈希；其裁判事件包含 gate 26，事件时刻为仿真 69.673 s。最终评分器也报告 metric_gates=32、stages_reached=34、replay_success=1、deterministic=1。" + }, + { + "t": 69.6, + "log_step": null, + "text_zh": "观察：视频 69.6 s、视频 69.8 s，固定相机仍朝向起点和近段旗门，机器人及当前门的足部、板尖已不可辨；与成功参考录像相同的远景限制，画面不提供逐点过门证据。" + }, + { + "t": 69.7, + "log_step": 224, + "text_zh": "推断：这是交付文件回放的裁判阶段，而非搜索候选的结果。由8341控制步×0.01 s=仿真83.41 s与视频83.44 s时长相符，且已在视频4.24–4.60 s起点越红线及视频11.6–11.8 s首门附近滑行处核对，按同一回合时钟将该门事件定位为约视频 69.7 s（N6，精度约0.2 s）；成功裁判确认四受检点及前序门均被接受，事件锁存至片尾。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察：视频 69.6 s、视频 69.8 s，该门附近的机器人足部和雪板接触区域太小或已超出可辨范围；无状态接触日志，不能逐门确认实际支承及其他接触。" + } + ] + }, + "gate_27_passed": { + "method": "visual", + "intervals": [ + [ + 71.9, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 71.9, + "log_step": 224, + "text_zh": "观察：log step 224 的交付文件两次新进程回放报告均为 success=true、8341个控制步、仿真 83.41 s、相同终态哈希；其裁判事件包含 gate 27，事件时刻为仿真 71.890 s。最终评分器也报告 metric_gates=32、stages_reached=34、replay_success=1、deterministic=1。" + }, + { + "t": 71.8, + "log_step": null, + "text_zh": "观察：视频 71.8 s、视频 72.0 s，固定相机仍朝向起点和近段旗门，机器人及当前门的足部、板尖已不可辨；与成功参考录像相同的远景限制，画面不提供逐点过门证据。" + }, + { + "t": 71.9, + "log_step": 224, + "text_zh": "推断：这是交付文件回放的裁判阶段，而非搜索候选的结果。由8341控制步×0.01 s=仿真83.41 s与视频83.44 s时长相符，且已在视频4.24–4.60 s起点越红线及视频11.6–11.8 s首门附近滑行处核对，按同一回合时钟将该门事件定位为约视频 71.9 s（N6，精度约0.2 s）；成功裁判确认四受检点及前序门均被接受，事件锁存至片尾。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察：视频 71.8 s、视频 72.0 s，该门附近的机器人足部和雪板接触区域太小或已超出可辨范围；无状态接触日志，不能逐门确认实际支承及其他接触。" + } + ] + }, + "gate_28_passed": { + "method": "visual", + "intervals": [ + [ + 73.8, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 73.8, + "log_step": 224, + "text_zh": "观察：log step 224 的交付文件两次新进程回放报告均为 success=true、8341个控制步、仿真 83.41 s、相同终态哈希；其裁判事件包含 gate 28，事件时刻为仿真 73.840 s。最终评分器也报告 metric_gates=32、stages_reached=34、replay_success=1、deterministic=1。" + }, + { + "t": 73.6, + "log_step": null, + "text_zh": "观察：视频 73.6 s、视频 73.8 s、视频 74.0 s，固定相机仍朝向起点和近段旗门，机器人及当前门的足部、板尖已不可辨；与成功参考录像相同的远景限制，画面不提供逐点过门证据。" + }, + { + "t": 73.8, + "log_step": 224, + "text_zh": "推断：这是交付文件回放的裁判阶段，而非搜索候选的结果。由8341控制步×0.01 s=仿真83.41 s与视频83.44 s时长相符，且已在视频4.24–4.60 s起点越红线及视频11.6–11.8 s首门附近滑行处核对，按同一回合时钟将该门事件定位为约视频 73.8 s（N6，精度约0.2 s）；成功裁判确认四受检点及前序门均被接受，事件锁存至片尾。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察：视频 73.6 s、视频 73.8 s、视频 74.0 s，该门附近的机器人足部和雪板接触区域太小或已超出可辨范围；无状态接触日志，不能逐门确认实际支承及其他接触。" + } + ] + }, + "gate_29_passed": { + "method": "visual", + "intervals": [ + [ + 75.7, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 75.7, + "log_step": 224, + "text_zh": "观察：log step 224 的交付文件两次新进程回放报告均为 success=true、8341个控制步、仿真 83.41 s、相同终态哈希；其裁判事件包含 gate 29，事件时刻为仿真 75.666 s。最终评分器也报告 metric_gates=32、stages_reached=34、replay_success=1、deterministic=1。" + }, + { + "t": 75.6, + "log_step": null, + "text_zh": "观察：视频 75.6 s、视频 75.8 s，固定相机仍朝向起点和近段旗门，机器人及当前门的足部、板尖已不可辨；与成功参考录像相同的远景限制，画面不提供逐点过门证据。" + }, + { + "t": 75.7, + "log_step": 224, + "text_zh": "推断：这是交付文件回放的裁判阶段，而非搜索候选的结果。由8341控制步×0.01 s=仿真83.41 s与视频83.44 s时长相符，且已在视频4.24–4.60 s起点越红线及视频11.6–11.8 s首门附近滑行处核对，按同一回合时钟将该门事件定位为约视频 75.7 s（N6，精度约0.2 s）；成功裁判确认四受检点及前序门均被接受，事件锁存至片尾。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察：视频 75.6 s、视频 75.8 s，该门附近的机器人足部和雪板接触区域太小或已超出可辨范围；无状态接触日志，不能逐门确认实际支承及其他接触。" + } + ] + }, + "gate_30_passed": { + "method": "visual", + "intervals": [ + [ + 77.7, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 77.7, + "log_step": 224, + "text_zh": "观察：log step 224 的交付文件两次新进程回放报告均为 success=true、8341个控制步、仿真 83.41 s、相同终态哈希；其裁判事件包含 gate 30，事件时刻为仿真 77.680 s。最终评分器也报告 metric_gates=32、stages_reached=34、replay_success=1、deterministic=1。" + }, + { + "t": 77.6, + "log_step": null, + "text_zh": "观察：视频 77.6 s、视频 77.8 s，固定相机仍朝向起点和近段旗门，机器人及当前门的足部、板尖已不可辨；与成功参考录像相同的远景限制，画面不提供逐点过门证据。" + }, + { + "t": 77.7, + "log_step": 224, + "text_zh": "推断：这是交付文件回放的裁判阶段，而非搜索候选的结果。由8341控制步×0.01 s=仿真83.41 s与视频83.44 s时长相符，且已在视频4.24–4.60 s起点越红线及视频11.6–11.8 s首门附近滑行处核对，按同一回合时钟将该门事件定位为约视频 77.7 s（N6，精度约0.2 s）；成功裁判确认四受检点及前序门均被接受，事件锁存至片尾。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察：视频 77.6 s、视频 77.8 s，该门附近的机器人足部和雪板接触区域太小或已超出可辨范围；无状态接触日志，不能逐门确认实际支承及其他接触。" + } + ] + }, + "gate_31_passed": { + "method": "visual", + "intervals": [ + [ + 79.4, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 79.4, + "log_step": 224, + "text_zh": "观察：log step 224 的交付文件两次新进程回放报告均为 success=true、8341个控制步、仿真 83.41 s、相同终态哈希；其裁判事件包含 gate 31，事件时刻为仿真 79.366 s。最终评分器也报告 metric_gates=32、stages_reached=34、replay_success=1、deterministic=1。" + }, + { + "t": 79.2, + "log_step": null, + "text_zh": "观察：视频 79.2 s、视频 79.4 s、视频 79.6 s，固定相机仍朝向起点和近段旗门，机器人及当前门的足部、板尖已不可辨；与成功参考录像相同的远景限制，画面不提供逐点过门证据。" + }, + { + "t": 79.4, + "log_step": 224, + "text_zh": "推断：这是交付文件回放的裁判阶段，而非搜索候选的结果。由8341控制步×0.01 s=仿真83.41 s与视频83.44 s时长相符，且已在视频4.24–4.60 s起点越红线及视频11.6–11.8 s首门附近滑行处核对，按同一回合时钟将该门事件定位为约视频 79.4 s（N6，精度约0.2 s）；成功裁判确认四受检点及前序门均被接受，事件锁存至片尾。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察：视频 79.2 s、视频 79.4 s、视频 79.6 s，该门附近的机器人足部和雪板接触区域太小或已超出可辨范围；无状态接触日志，不能逐门确认实际支承及其他接触。" + } + ] + }, + "gate_32_passed": { + "method": "visual", + "intervals": [ + [ + 81.4, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 81.4, + "log_step": 224, + "text_zh": "观察：log step 224 的交付文件两次新进程回放报告均为 success=true、8341个控制步、仿真 83.41 s、相同终态哈希；其裁判事件包含 gate 32，事件时刻为仿真 81.434 s。最终评分器也报告 metric_gates=32、stages_reached=34、replay_success=1、deterministic=1。" + }, + { + "t": 81.2, + "log_step": null, + "text_zh": "观察：视频 81.2 s、视频 81.4 s、视频 81.6 s，固定相机仍朝向起点和近段旗门，机器人及当前门的足部、板尖已不可辨；与成功参考录像相同的远景限制，画面不提供逐点过门证据。" + }, + { + "t": 81.4, + "log_step": 224, + "text_zh": "推断：这是交付文件回放的裁判阶段，而非搜索候选的结果。由8341控制步×0.01 s=仿真83.41 s与视频83.44 s时长相符，且已在视频4.24–4.60 s起点越红线及视频11.6–11.8 s首门附近滑行处核对，按同一回合时钟将该门事件定位为约视频 81.4 s（N6，精度约0.2 s）；成功裁判确认四受检点及前序门均被接受，事件锁存至片尾。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察：视频 81.2 s、视频 81.4 s、视频 81.6 s，该门附近的机器人足部和雪板接触区域太小或已超出可辨范围；无状态接触日志，不能逐门确认实际支承及其他接触。" + } + ] + }, + "finish_crossed": { + "method": "visual", + "intervals": [ + [ + 83.4, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 83.4, + "log_step": 224, + "text_zh": "观察：log step 224 的交付文件两次新进程回放均报告全部32个 gate 和 finish 阶段、success=true、failed=null、8341控制步、仿真83.41 s且终态哈希相同；finish 裁判事件为仿真83.409 s。最终评分器记录success=1、replay_success=1、deterministic=1、metric_gates=32、metric_time_s=79.058。" + }, + { + "t": 83.4, + "log_step": null, + "text_zh": "观察：已检查视频82.4、82.6、82.8、83.0、83.2及83.40 s，画面仍为固定起点远景，终点与机器人双脚均无法辨认；近处黑色拱架是起点，不能把它当成终点证据。" + }, + { + "t": 83.4, + "log_step": 224, + "text_zh": "推断：交付回放的8341控制步×0.01 s=仿真83.41 s与视频83.44 s匹配，起滑及首门处也已核对同一时钟；故finish事件约为视频83.4 s。按本次裁判成功结果，全部前序门完成后双脚在终点规定范围内过线且被接受；成功为锁存事件，无需停稳。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察：视频82.4–83.40 s的固定远景不能显示终点附近机器人、足部、雪板或雪面接触；没有状态接触日志，实际接触方式未知。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "gates", + "unit": "count", + "threshold": 32, + "direction": "higher_better", + "value": 32.0 + }, + { + "name": "speed_m_s", + "unit": "m/s", + "threshold": 0.1, + "direction": "higher_better", + "value": 5.35 + }, + { + "name": "time_s", + "unit": "s", + "threshold": null, + "direction": "lower_better", + "value": 79.058 + } + ], + "failure": null, + "summary_zh": "交付轨迹自行起滑，依次通过32道旗门并越过终点，33/33节点最终达成；交付文件两次新进程回放成功且终态哈希一致，最终评分器确认32门、34个原生阶段及无摔倒，故无失败断点。逐门时间采用log step 224交付回放的裁判事件，利用8341控制步对应仿真83.41 s与视频83.44 s的时长及起滑、首门画面核对，终点约视频83.4 s，裁判净比赛用时79.058 s。已看完整段抽帧、各过门事件附近密集帧和视频最后一秒，并比较成功参考录像；没有状态日志，固定远景无法逐点辨认四点过门，后段接触不可见处保留未知，不把会话搜索中的摔倒当成本次失败。", + "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 32 门、34 阶段、未摔倒），全部 33 个节点达成、无失败断点，与评分器一致；控制步→视频时间映射（8341 步=83.41 s≈视频 83.44 s）有说明并已在起滑与首门核对；固定远景下逐门接触标未知合理。", + "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": "2026-10-29T10:04:56Z", + "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/embodied-olympics-fun-alpine-slalom-i00-privileged--ev2gd7x/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-alpine-slalom-i00-privileged--ev2gd7x/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "7b11adfb402406d898f9cf4311d4640bfe743a16443044c843d2405922d38150", + "video_sha256": "f4624b71b383accd420070c7fd58f1b23db22676944ec17fd4f5ba3a29187250", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/alpine_slalom/trajectories/embodied-olympics-fun-alpine-slalom-i00-standard--m7dzy5z.json b/phase2/embodied-olympics-fun/alpine_slalom/trajectories/embodied-olympics-fun-alpine-slalom-i00-standard--m7dzy5z.json new file mode 100644 index 0000000000000000000000000000000000000000..b00509a254420c44f7f50514eb6682b9bcc698e3 --- /dev/null +++ b/phase2/embodied-olympics-fun/alpine_slalom/trajectories/embodied-olympics-fun-alpine-slalom-i00-standard--m7dzy5z.json @@ -0,0 +1,614 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-alpine-slalom-i00-standard--m7dzy5z", + "taskKey": "embodied-olympics-fun/alpine_slalom", + "dag_sha256": "0382103c21732bd9a43317e1385d02d1f88b24ff2896d1d6062c6a895ff37a48", + "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": 2000, + "sim_steps": 4123, + "calls": 126, + "rows": 4120, + "resets": 5, + "reset_budget": 5, + "sim_startups": 1, + "elapsed_s": 1893.6, + "video_rendered": 1, + "first_success": 0, + "first_reason_code": 3, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.265, + "peak_processes": 11, + "peak_disk_gb": 1.106, + "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/embodied-olympics-fun-alpine-slalom-i00-standard--m7dzy5z/replay.mp4", + "duration_s": 20.04, + "fps": 25.0, + "frames": 501, + "sha256": "9717439cee76f977b72066149e9313c20dc1eac8768413869bf36df923707d05" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 33, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "gate_1_passed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 3.0, + "log_step": 100, + "text_zh": "观察：视频0.0 s机器人双脚及雪板在起点红线前；视频约2.4–4.8 s机器人躯干逐渐向下坡方向前倾，双脚仍在红线上坡侧，未滑向第1道旗门。log step 100的起滑代码命令踝关节前倾并摆动双臂。" + }, + { + "t": 19.96, + "log_step": 101, + "text_zh": "观察：视频19.0–19.96 s逐帧检查，机器人仍以前倾姿势停在起点红线上坡侧，双脚没有全部越线。log step 101所执行的最终episode 6从服务返回outcome=\"did not start\"、stages_reached=0、stages_total=34；该阶段计数包含起点、32门和终点。log step 103确认本回合2000个控制步且done=true。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：视频0.0–19.96 s机器人双脚经固定雪板支承在起点平台，腿和躯干前倾尝试起滑；雪板及双脚没有离开起点区域，双臂仅在空中摆动。" + } + ] + }, + "gate_2_passed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 19.96, + "log_step": 101, + "text_zh": "推断：受评视频0.0–19.96 s中机器人始终留在起点平台，末秒仍未以双脚跨过起点红线，故没有到达或通过第2道旗门；按DAG的顺序验收此前置也未成立。log step 101的最终episode 6服务裁判报告did not start及0/34原生阶段，与视频一致。" + } + ] + }, + "gate_3_passed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 19.96, + "log_step": 101, + "text_zh": "推断：受评视频0.0–19.96 s中机器人始终留在起点平台，末秒仍未以双脚跨过起点红线，故没有到达或通过第3道旗门；按DAG的顺序验收此前置也未成立。log step 101的最终episode 6服务裁判报告did not start及0/34原生阶段，与视频一致。" + } + ] + }, + "gate_4_passed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 19.96, + "log_step": 101, + "text_zh": "推断：受评视频0.0–19.96 s中机器人始终留在起点平台，末秒仍未以双脚跨过起点红线，故没有到达或通过第4道旗门；按DAG的顺序验收此前置也未成立。log step 101的最终episode 6服务裁判报告did not start及0/34原生阶段，与视频一致。" + } + ] + }, + "gate_5_passed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 19.96, + "log_step": 101, + "text_zh": "推断：受评视频0.0–19.96 s中机器人始终留在起点平台，末秒仍未以双脚跨过起点红线，故没有到达或通过第5道旗门；按DAG的顺序验收此前置也未成立。log step 101的最终episode 6服务裁判报告did not start及0/34原生阶段，与视频一致。" + } + ] + }, + "gate_6_passed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 19.96, + "log_step": 101, + "text_zh": "推断：受评视频0.0–19.96 s中机器人始终留在起点平台，末秒仍未以双脚跨过起点红线，故没有到达或通过第6道旗门；按DAG的顺序验收此前置也未成立。log step 101的最终episode 6服务裁判报告did not start及0/34原生阶段，与视频一致。" + } + ] + }, + "gate_7_passed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 19.96, + "log_step": 101, + "text_zh": "推断：受评视频0.0–19.96 s中机器人始终留在起点平台，末秒仍未以双脚跨过起点红线，故没有到达或通过第7道旗门；按DAG的顺序验收此前置也未成立。log step 101的最终episode 6服务裁判报告did not start及0/34原生阶段，与视频一致。" + } + ] + }, + "gate_8_passed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 19.96, + "log_step": 101, + "text_zh": "推断：受评视频0.0–19.96 s中机器人始终留在起点平台，末秒仍未以双脚跨过起点红线，故没有到达或通过第8道旗门；按DAG的顺序验收此前置也未成立。log step 101的最终episode 6服务裁判报告did not start及0/34原生阶段，与视频一致。" + } + ] + }, + "gate_9_passed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 19.96, + "log_step": 101, + "text_zh": "推断：受评视频0.0–19.96 s中机器人始终留在起点平台，末秒仍未以双脚跨过起点红线，故没有到达或通过第9道旗门；按DAG的顺序验收此前置也未成立。log step 101的最终episode 6服务裁判报告did not start及0/34原生阶段，与视频一致。" + } + ] + }, + "gate_10_passed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 19.96, + "log_step": 101, + "text_zh": "推断：受评视频0.0–19.96 s中机器人始终留在起点平台，末秒仍未以双脚跨过起点红线，故没有到达或通过第10道旗门；按DAG的顺序验收此前置也未成立。log step 101的最终episode 6服务裁判报告did not start及0/34原生阶段，与视频一致。" + } + ] + }, + "gate_11_passed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 19.96, + "log_step": 101, + "text_zh": "推断：受评视频0.0–19.96 s中机器人始终留在起点平台，末秒仍未以双脚跨过起点红线，故没有到达或通过第11道旗门；按DAG的顺序验收此前置也未成立。log step 101的最终episode 6服务裁判报告did not start及0/34原生阶段，与视频一致。" + } + ] + }, + "gate_12_passed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 19.96, + "log_step": 101, + "text_zh": "推断：受评视频0.0–19.96 s中机器人始终留在起点平台，末秒仍未以双脚跨过起点红线，故没有到达或通过第12道旗门；按DAG的顺序验收此前置也未成立。log step 101的最终episode 6服务裁判报告did not start及0/34原生阶段，与视频一致。" + } + ] + }, + "gate_13_passed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 19.96, + "log_step": 101, + "text_zh": "推断：受评视频0.0–19.96 s中机器人始终留在起点平台，末秒仍未以双脚跨过起点红线，故没有到达或通过第13道旗门；按DAG的顺序验收此前置也未成立。log step 101的最终episode 6服务裁判报告did not start及0/34原生阶段，与视频一致。" + } + ] + }, + "gate_14_passed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 19.96, + "log_step": 101, + "text_zh": "推断：受评视频0.0–19.96 s中机器人始终留在起点平台，末秒仍未以双脚跨过起点红线，故没有到达或通过第14道旗门；按DAG的顺序验收此前置也未成立。log step 101的最终episode 6服务裁判报告did not start及0/34原生阶段，与视频一致。" + } + ] + }, + "gate_15_passed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 19.96, + "log_step": 101, + "text_zh": "推断：受评视频0.0–19.96 s中机器人始终留在起点平台，末秒仍未以双脚跨过起点红线，故没有到达或通过第15道旗门；按DAG的顺序验收此前置也未成立。log step 101的最终episode 6服务裁判报告did not start及0/34原生阶段，与视频一致。" + } + ] + }, + "gate_16_passed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 19.96, + "log_step": 101, + "text_zh": "推断：受评视频0.0–19.96 s中机器人始终留在起点平台，末秒仍未以双脚跨过起点红线，故没有到达或通过第16道旗门；按DAG的顺序验收此前置也未成立。log step 101的最终episode 6服务裁判报告did not start及0/34原生阶段，与视频一致。" + } + ] + }, + "gate_17_passed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 19.96, + "log_step": 101, + "text_zh": "推断：受评视频0.0–19.96 s中机器人始终留在起点平台，末秒仍未以双脚跨过起点红线，故没有到达或通过第17道旗门；按DAG的顺序验收此前置也未成立。log step 101的最终episode 6服务裁判报告did not start及0/34原生阶段，与视频一致。" + } + ] + }, + "gate_18_passed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 19.96, + "log_step": 101, + "text_zh": "推断：受评视频0.0–19.96 s中机器人始终留在起点平台，末秒仍未以双脚跨过起点红线，故没有到达或通过第18道旗门；按DAG的顺序验收此前置也未成立。log step 101的最终episode 6服务裁判报告did not start及0/34原生阶段，与视频一致。" + } + ] + }, + "gate_19_passed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 19.96, + "log_step": 101, + "text_zh": "推断：受评视频0.0–19.96 s中机器人始终留在起点平台，末秒仍未以双脚跨过起点红线，故没有到达或通过第19道旗门；按DAG的顺序验收此前置也未成立。log step 101的最终episode 6服务裁判报告did not start及0/34原生阶段，与视频一致。" + } + ] + }, + "gate_20_passed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 19.96, + "log_step": 101, + "text_zh": "推断：受评视频0.0–19.96 s中机器人始终留在起点平台，末秒仍未以双脚跨过起点红线，故没有到达或通过第20道旗门；按DAG的顺序验收此前置也未成立。log step 101的最终episode 6服务裁判报告did not start及0/34原生阶段，与视频一致。" + } + ] + }, + "gate_21_passed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 19.96, + "log_step": 101, + "text_zh": "推断：受评视频0.0–19.96 s中机器人始终留在起点平台，末秒仍未以双脚跨过起点红线，故没有到达或通过第21道旗门；按DAG的顺序验收此前置也未成立。log step 101的最终episode 6服务裁判报告did not start及0/34原生阶段，与视频一致。" + } + ] + }, + "gate_22_passed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 19.96, + "log_step": 101, + "text_zh": "推断：受评视频0.0–19.96 s中机器人始终留在起点平台，末秒仍未以双脚跨过起点红线，故没有到达或通过第22道旗门；按DAG的顺序验收此前置也未成立。log step 101的最终episode 6服务裁判报告did not start及0/34原生阶段，与视频一致。" + } + ] + }, + "gate_23_passed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 19.96, + "log_step": 101, + "text_zh": "推断：受评视频0.0–19.96 s中机器人始终留在起点平台，末秒仍未以双脚跨过起点红线，故没有到达或通过第23道旗门；按DAG的顺序验收此前置也未成立。log step 101的最终episode 6服务裁判报告did not start及0/34原生阶段，与视频一致。" + } + ] + }, + "gate_24_passed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 19.96, + "log_step": 101, + "text_zh": "推断：受评视频0.0–19.96 s中机器人始终留在起点平台，末秒仍未以双脚跨过起点红线，故没有到达或通过第24道旗门；按DAG的顺序验收此前置也未成立。log step 101的最终episode 6服务裁判报告did not start及0/34原生阶段，与视频一致。" + } + ] + }, + "gate_25_passed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 19.96, + "log_step": 101, + "text_zh": "推断：受评视频0.0–19.96 s中机器人始终留在起点平台，末秒仍未以双脚跨过起点红线，故没有到达或通过第25道旗门；按DAG的顺序验收此前置也未成立。log step 101的最终episode 6服务裁判报告did not start及0/34原生阶段，与视频一致。" + } + ] + }, + "gate_26_passed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 19.96, + "log_step": 101, + "text_zh": "推断：受评视频0.0–19.96 s中机器人始终留在起点平台，末秒仍未以双脚跨过起点红线，故没有到达或通过第26道旗门；按DAG的顺序验收此前置也未成立。log step 101的最终episode 6服务裁判报告did not start及0/34原生阶段，与视频一致。" + } + ] + }, + "gate_27_passed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 19.96, + "log_step": 101, + "text_zh": "推断：受评视频0.0–19.96 s中机器人始终留在起点平台，末秒仍未以双脚跨过起点红线，故没有到达或通过第27道旗门；按DAG的顺序验收此前置也未成立。log step 101的最终episode 6服务裁判报告did not start及0/34原生阶段，与视频一致。" + } + ] + }, + "gate_28_passed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 19.96, + "log_step": 101, + "text_zh": "推断：受评视频0.0–19.96 s中机器人始终留在起点平台，末秒仍未以双脚跨过起点红线，故没有到达或通过第28道旗门；按DAG的顺序验收此前置也未成立。log step 101的最终episode 6服务裁判报告did not start及0/34原生阶段，与视频一致。" + } + ] + }, + "gate_29_passed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 19.96, + "log_step": 101, + "text_zh": "推断：受评视频0.0–19.96 s中机器人始终留在起点平台，末秒仍未以双脚跨过起点红线，故没有到达或通过第29道旗门；按DAG的顺序验收此前置也未成立。log step 101的最终episode 6服务裁判报告did not start及0/34原生阶段，与视频一致。" + } + ] + }, + "gate_30_passed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 19.96, + "log_step": 101, + "text_zh": "推断：受评视频0.0–19.96 s中机器人始终留在起点平台，末秒仍未以双脚跨过起点红线，故没有到达或通过第30道旗门；按DAG的顺序验收此前置也未成立。log step 101的最终episode 6服务裁判报告did not start及0/34原生阶段，与视频一致。" + } + ] + }, + "gate_31_passed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 19.96, + "log_step": 101, + "text_zh": "推断：受评视频0.0–19.96 s中机器人始终留在起点平台，末秒仍未以双脚跨过起点红线，故没有到达或通过第31道旗门；按DAG的顺序验收此前置也未成立。log step 101的最终episode 6服务裁判报告did not start及0/34原生阶段，与视频一致。" + } + ] + }, + "gate_32_passed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 19.96, + "log_step": 101, + "text_zh": "推断：受评视频0.0–19.96 s中机器人始终留在起点平台，末秒仍未以双脚跨过起点红线，故没有到达或通过第32道旗门；按DAG的顺序验收此前置也未成立。log step 101的最终episode 6服务裁判报告did not start及0/34原生阶段，与视频一致。" + } + ] + }, + "finish_crossed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 19.96, + "log_step": 101, + "text_zh": "推断：受评视频0.0–19.96 s中机器人始终留在起点平台，末秒仍未以双脚跨过起点红线，故没有到达或通过终点；按DAG的顺序验收此前置也未成立。log step 101的最终episode 6服务裁判报告did not start及0/34原生阶段，与视频一致。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "gates", + "unit": "count", + "threshold": 32, + "direction": "higher_better", + "value": null + }, + { + "name": "speed_m_s", + "unit": "m/s", + "threshold": 0.1, + "direction": "higher_better", + "value": null + }, + { + "name": "time_s", + "unit": "s", + "threshold": null, + "direction": "lower_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "gate_1_passed", + "t": 3.0, + "log_step": 101, + "where": "操作中", + "evidence_zh": "观察：视频约2.4–4.8 s机器人执行起滑动作，躯干明显前倾但双脚未越起点红线；视频19.0–19.96 s仍停在原处，首门的起滑验收始终未完成。log step 101最终episode 6服务裁判报告did not start、0/34阶段，log step 103确认回合终结；首门节点包含自行起滑，因此断点为gate_1_passed。" + }, + "phenomena": [ + { + "cat": "全身", + "item": "步态", + "t": 3.0, + "log_step": 101, + "evidence_zh": "观察：视频约2.4–4.8 s机器人由直立转为明显前倾，雪板只在起点附近小幅变化，未形成离开平台的前进；视频19.0–19.96 s双脚仍在红线上坡侧，机器人保持该姿势，没有出现倒地过程。log step 101服务裁判以did not start终结。" + } + ], + "root_causes": [ + { + "cat": "认知", + "item": "推理/规划", + "source_zh": "日志 / agent 自述 / 视频", + "evidence_zh": "推断：agent为水平平台选择了“前倾使雪板滑动，再进入下坡滑行”的起滑策略，但该策略未产生所需位移。log step 98脚本说明明确依赖lean forward to slip off the platform；log step 100的实际修改仍是加大踝前倾及空中摆臂，并允许起滑阶段经过7秒后直接切到GLIDE，而非确认双脚过起点。视频3.0–19.96 s显示前倾后双脚始终未出发，log step 101服务返回did not start；log step 56的先前测试曾用walk获得1/34阶段，说明此前已有可越起点的会话证据，但受评起滑方案未采用它。agent在log step 118自述错误假定平台会让雪板滑动；此处归因为起滑方案及阶段切换设计错误，不将其“高摩擦”解释当作已测得的环境参数。" + } + ], + "gap_zh": "没有逐帧状态日志；固定远景可判断双脚未越起点，但不能独立测定雪板接触力、平台摩擦系数或踝关节饱和机理。log step 113为agent对保存本体数据的后处理，不作为独立状态日志；根因细节因此保留为结合代码与画面的推断。DAG的gates、speed_m_s和time_s均未作为该次reward字段报告，native_metrics的value置null。", + "unmet": { + "gate_1_passed": false, + "gate_2_passed": false, + "gate_3_passed": false, + "gate_4_passed": false, + "gate_5_passed": false, + "gate_6_passed": false, + "gate_7_passed": false, + "gate_8_passed": false, + "gate_9_passed": false, + "gate_10_passed": false, + "gate_11_passed": false, + "gate_12_passed": false, + "gate_13_passed": false, + "gate_14_passed": false, + "gate_15_passed": false, + "gate_16_passed": false, + "gate_17_passed": false, + "gate_18_passed": false, + "gate_19_passed": false, + "gate_20_passed": false, + "gate_21_passed": false, + "gate_22_passed": false, + "gate_23_passed": false, + "gate_24_passed": false, + "gate_25_passed": false, + "gate_26_passed": false, + "gate_27_passed": false, + "gate_28_passed": false, + "gate_29_passed": false, + "gate_30_passed": false, + "gate_31_passed": false, + "gate_32_passed": false, + "finish_crossed": false + } + }, + "summary_zh": "受评的是log step 101运行的最终episode 6：机器人尝试前倾起滑，却始终未以双脚越过起点红线，32道旗门及终点均未达到，节点进度为0/33。视频约3.0 s起可见前倾而不前进，末秒仍未恢复，服务裁判报告did not start、0/34原生阶段；脚本依赖前倾滑出平台且按经过时间切换阶段，支持起滑策略与阶段设计错误。缺少逐帧状态及接触力证据，平台高摩擦的具体物理解释仅为agent自述，未独立验证。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "受评episode 6 的起滑失败：视频全程机器人停在起点红线上坡侧，服务裁判 did not start、0/34，33个节点全为 false，断点 gate_1_passed、阶段/根因与 log step 98/100/101/118 一致，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": "2026-10-29T10:04:00Z", + "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/embodied-olympics-fun-alpine-slalom-i00-standard--m7dzy5z/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-alpine-slalom-i00-standard--m7dzy5z/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "e645351e336c06d9d28160118b2f9fc03988ffd9f72594e19fbb981e4e0aa64a", + "video_sha256": "9717439cee76f977b72066149e9313c20dc1eac8768413869bf36df923707d05", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/badminton_serve/task_dag.json b/phase2/embodied-olympics-fun/badminton_serve/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..3c29692ce23603c43a649e94b8679510f699744f --- /dev/null +++ b/phase2/embodied-olympics-fun/badminton_serve/task_dag.json @@ -0,0 +1,302 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "embodied-olympics-fun/badminton_serve", + "suite": "embodied-olympics-fun", + "base": "embodied-olympics-fun-badminton-serve-i00", + "title_zh": "羽毛球合法发球", + "summary_zh": "机器人从己方右发球区合法放球并用右手球拍击球，使羽毛球越网后首次落在对角接发球区且无裁判违例。", + "embodiment_zh": "两模式均为 Unitree G1 双足人形机器人（29 个本体关节，约 1.32 m、35 kg），双臂各装 Unitree Dex3-1 三指灵巧手；初始左手握羽毛球羽毛、右手靠摩擦握自由球拍拍柄，非腕部固定工具，双腿在球场支撑，无轮式底盘。", + "task_class": "atomic", + "analysis_level": "step", + "spec": { + "version": "1.2", + "task": "embodied-olympics-fun-badminton-serve-i00", + "bench": "embodied-olympics-fun", + "progress": "live", + "entities": {}, + "milestones": [ + { + "id": "serve_in_target", + "zh": "合法发球入区", + "mode": "event", + "when": null, + "after": [], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "场地远景中，机器人在近侧右发球区左手放开白色羽毛球、右手挥拍将其向上击出，球从黑色球网上方两柱之间进入远侧对角、淡黄色标示的接发球区，首次触地接触点须满足 x∈[4.43,9.15] m、y∈[0.43,3.04] m（线内含；原点为初始骨盆正下方、+x 朝网、+y 向左），至此才完成；只放球、击球、过网而尚未合法落地，或落点错误均未完成。必须按 release→hit→over the net→in 依次验收：放球前球在左手且随后脱离机器人接触满 0.03 s，首次击球为球托先触拍、全球最高处 <1.15 m、入拍前速 ≤4.0 m/s、拍头有效向网前挥（>0.5 m/s 开始、<0.1 m/s 停止，放球后才开始的前挥不得在击球前停止），击球结束后 0.03 s 球速大于入拍前速且仍上升；接触相隔 <0.03 s 算一击且接触持续 ≤0.02 s，过网参考点在 x=2.45 m 处满足 |y−0.45|≤3.05 m、z>1.524+0.026((y−0.45)/3.05)^2 m，同时从发球开始到击球双脚不踩己方发球区界线、离地间隙 ≤0.02 s、水平位移 ≤0.03 m，右手始终握拍（间隙 ≤0.05 s）且全程无所列违例；远景不能分辨的球托初触、瞬时速度和接触时长须以裁判事件及指标核实。", + "step": { + "cat": "动态", + "item": "击打" + }, + "contact": [ + { + "who": "手持工具", + "how": "其他", + "note_zh": "候选：击打；右手握持的球拍拍面瞬时击打羽毛球球托，现有接触方式没有击打，不能用推/扫或压/擦代替；握拍链路不重复标注。", + "plan": 1 + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "机器人双脚同时支撑球场地面，在发球开始至击球期间保持原地站立。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:3-5", + "inputs/task/standard/instruction.md:3-8", + "inputs/task/privileged/tests/rfo_task.py:404-412", + "inputs/task/privileged/tests/rfo_task.py:483-491", + "inputs/task/privileged/tests/rfo_task.py:542-564", + "inputs/task/privileged/tests/rfo_task.py:617-744", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178", + "inputs/oracle/replay.mp4:0.00-2.96 s", + "inputs/oracle/run.json:reward; steps[191].out.text" + ], + "conf": 6, + "note_zh": "按 S10，一次发球的放球、击球、过网、首次落地无法在中间暂停后继续，四个原生阶段合成一个事件目标；击球过程与落点判定成功后锁存，无终态静置窗口。左手此前握住羽毛球羽毛并张开放球，右手随后经球拍击打，二者先后发生而不增加同时接触条目；左右手初始已分别持球、持拍，无工具拾取前置步骤。" + } + ], + "constraints": [ + { + "id": "robot_fell", + "zh": "机器人跌倒", + "never": null, + "criterion_zh": "尝试尚未成功时，骨盆高度低于 0.35 m，或双脚以外的机器人部位接触地面；视频可见身体下坠、手膝或躯干撑地。通用裁判在控制步末检查。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:272-282", + "inputs/task/privileged/tests/rfo_task.py:478-480" + ] + }, + { + "id": "racket_lost", + "zh": "球拍脱手触地", + "never": null, + "criterion_zh": "右手未以手指与拇指或手掌共同接触球拍，且离开有效握持超过 0.05 s，或球拍任何时候触地，均判失败；视频可见球拍脱离右手或落到地面。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:527-531", + "inputs/task/privileged/tests/rfo_task.py:623-628" + ] + }, + { + "id": "net_or_post_contact", + "zh": "人体球拍触网", + "never": null, + "criterion_zh": "机器人任一部位或球拍接触球网或网柱即失败；羽毛球在合法击球后擦网本身允许，但触网柱不允许。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:617-622", + "inputs/task/privileged/tests/rfo_task.py:693-698" + ] + }, + { + "id": "service_feet_fault", + "zh": "发球脚步违例", + "never": null, + "criterion_zh": "从放球与导致击球的前挥开始两者较早时刻到击球，双脚接地点必须严格在己方发球区内部：-4.21 < x < 0.43 m、-2.10 < y < 0.43 m，不踩线；每只脚离地间隙不得超过 0.02 s，脚部位置相对发球开始的最大水平位移不得超过 0.03 m。脚踩线、出区、持续离地或明显移步为失败，裁判在击球时追溯该段历史。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:483-491", + "inputs/task/privileged/tests/rfo_task.py:601-615", + "inputs/task/privileged/tests/rfo_task.py:632-648", + "inputs/task/privileged/tests/rfo_task.py:735-741", + "inputs/task/privileged/tests/rfo_task.py:405-412" + ] + }, + { + "id": "shuttle_contact_fault", + "zh": "羽球接触违例", + "never": null, + "criterion_zh": "羽毛球尚未脱离机器人接触满 0.03 s 前，球拍不得触球；确认放球后，羽毛球不得再碰机器人。合法击球后不得二次击球或持球：相隔至少 0.03 s 的球拍接触另计一击，同一次接触持续超过 0.02 s 判持球，羽毛球触网柱也失败。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:649-668", + "inputs/task/privileged/tests/rfo_task.py:679-696", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:304-347" + ] + }, + { + "id": "service_stroke_fault", + "zh": "发球击打违例", + "never": null, + "criterion_zh": "放球后的拍头前挥以向网速度超过 0.5 m/s 开始、低于 0.1 m/s 停止，开始后若在击球前停止即失败；首次击球必须球托接触球拍（羽毛可同时接触）、全球最高处严格低于 1.15 m、入拍前速度不超过 4.0 m/s，且拍头仍处在有效前挥中。击球最后一次接触后 0.03 s 验收时，羽毛球速度须大于入拍前速度且竖直速度大于 0；任一条件不满足即失败。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:632-643", + "inputs/task/privileged/tests/rfo_task.py:684-692", + "inputs/task/privileged/tests/rfo_task.py:720-744", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:325-347" + ] + }, + { + "id": "serve_path_fault", + "zh": "羽球飞行落点错", + "never": null, + "criterion_zh": "未击球前羽毛球触地、触网或穿越网面即失败；击球后越过网面时须在两柱之间且参考点高于该横向位置的网顶，首次触地接触点须在接发球区内。从网下、柱外通过，或落在己方、对方其他区域、场外均失败；落点判界线内含，合法击球后擦网允许。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:672-677", + "inputs/task/privileged/tests/rfo_task.py:695-718" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "badminton_serve_in", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178; inputs/task/privileged/tests/rfo_task.py:699-744; 需逐帧记录原生 ref.stages、ref.success、ref.failed 及球拍/球接触事件；当前无状态日志，未绑定任何路径。" + }, + { + "atom": "robot_fell", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:272-282; inputs/task/privileged/tests/rfo_task.py:478-480; 需记录原生裁判失败原因及对应接触、脚步和历史跟踪量，当前无状态日志。" + }, + { + "atom": "racket_lost", + "level": "new_code", + "source": "inputs/task/privileged/tests/rfo_task.py:527-531; inputs/task/privileged/tests/rfo_task.py:623-628; 需记录原生裁判失败原因及对应接触、脚步和历史跟踪量，当前无状态日志。" + }, + { + "atom": "net_or_post_contact", + "level": "new_code", + "source": "inputs/task/privileged/tests/rfo_task.py:617-622; inputs/task/privileged/tests/rfo_task.py:693-698; 需记录原生裁判失败原因及对应接触、脚步和历史跟踪量，当前无状态日志。" + }, + { + "atom": "service_feet_fault", + "level": "new_code", + "source": "inputs/task/privileged/tests/rfo_task.py:483-491; inputs/task/privileged/tests/rfo_task.py:601-615; inputs/task/privileged/tests/rfo_task.py:632-648; inputs/task/privileged/tests/rfo_task.py:735-741; inputs/task/privileged/tests/rfo_task.py:405-412; 需记录原生裁判失败原因及对应接触、脚步和历史跟踪量，当前无状态日志。" + }, + { + "atom": "shuttle_contact_fault", + "level": "new_code", + "source": "inputs/task/privileged/tests/rfo_task.py:649-668; inputs/task/privileged/tests/rfo_task.py:679-696; inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:304-347; 需记录原生裁判失败原因及对应接触、脚步和历史跟踪量，当前无状态日志。" + }, + { + "atom": "service_stroke_fault", + "level": "new_code", + "source": "inputs/task/privileged/tests/rfo_task.py:632-643; inputs/task/privileged/tests/rfo_task.py:684-692; inputs/task/privileged/tests/rfo_task.py:720-744; inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:325-347; 需记录原生裁判失败原因及对应接触、脚步和历史跟踪量，当前无状态日志。" + }, + { + "atom": "serve_path_fault", + "level": "new_code", + "source": "inputs/task/privileged/tests/rfo_task.py:672-677; inputs/task/privileged/tests/rfo_task.py:695-718; 需记录原生裁判失败原因及对应接触、脚步和历史跟踪量，当前无状态日志。" + } + ] + }, + "native_metrics": [ + { + "name": "hit_height_m", + "unit": "m", + "threshold": 1.15, + "direction": "lower_better", + "note_zh": "严格小于；全球最高点，来源：inputs/task/privileged/tests/rfo_task.py:723-743" + }, + { + "name": "shuttle_speed_m_s", + "unit": "m/s", + "threshold": null, + "direction": "higher_better", + "note_zh": "击球结束时速度须严格大于本次入拍前速度，无固定下限；入拍前另限 ≤4.0 m/s。来源：inputs/task/privileged/tests/rfo_task.py:665,684-692,731-732" + }, + { + "name": "shuttle_angle_deg", + "unit": "deg", + "threshold": 0, + "direction": "higher_better", + "note_zh": "击球结束时由速度计算的仰角，裁判实际检查竖直速度 >0。来源：inputs/task/privileged/tests/rfo_task.py:684-692" + }, + { + "name": "landing_x_m", + "unit": "m", + "threshold": [ + 4.43, + 9.15 + ], + "direction": "target", + "note_zh": "首次触地接触点，闭区间，按实例参数计算；来源：inputs/task/privileged/tests/rfo_task.py:483-491,708-718" + }, + { + "name": "landing_y_m", + "unit": "m", + "threshold": [ + 0.43, + 3.04 + ], + "direction": "target", + "note_zh": "首次触地接触点，闭区间；来源：inputs/task/privileged/tests/rfo_task.py:483-491,708-718" + } + ], + "open_questions_zh": [ + "状态证据缺口：inputs/states/index.json 的 logs 为空，成功参考也无状态日志；按 P8，节点 when、约束 never 与 success 均为 null，atoms 仅列需补录的原生裁判量，不能据源码字段假装已绑定。", + "视频证据限制：参考远景可见放球、挥拍和白色羽毛球向远侧飞行，但球很小，无法核实毫秒级球托初触、接触持续、脚部位移与瞬时速度。渲染每 4 个控制步采样一次（inputs/grader/efe_embodied-olympics-fun-files/rcb_rfo.py:216-241）；末帧对应约 2.96 s，早于参考裁判首次落地 2.973 s，因此落地成功同时参考 run.json 的 reward 和最终交付回放输出，不声称末帧已直接拍到触地。", + "模式验收差异：任务规则代码在 privileged tests、privileged environment 和 standard sim 中逐字相同；privileged 另要求同一提交在两个新进程中都成功且末态哈希相同（inputs/grader/efe_embodied-olympics-fun-files/grade.py:94-110），standard 评分最后一次实时尝试（inputs/grader/efe_standard-files/v2_1/grade.py:93-99）。交付、复现和基础设施有效性不是世界状态，不增加节点；无日志时 success 不能自动覆盖这些交付检查。", + "工具前置核对（D2）：文本与评分器都要求右手用球拍发球，且右手握拍须贯穿尝试（inputs/task/privileged/tests/rfo_task.py:527-531,623-628）；reset 已将球拍置于右手并闭合灵巧手（同文件:542-564），所以不增设拿起球拍节点。", + "粒度采用 S10 与 D11 的连续投击阅读：四个原生阶段仅验收一次完整合法发球，不形成四个语义步骤；15 s / 1500 控制步是时限而非约束（P21）。" + ], + "conf": 6, + "conf_reason_zh": "一个连续发球步骤直接来自文本，阈值、初始持物、两模式共同裁判与成功回放均已核对；没有状态日志，视频也不能核验精细接触和最终落地瞬间，故保留空条件并降低置信度。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "单个连续发球步骤（放球→击球→过网→入区）符合 S10：球放开后不能停下再继续，过网和落地是飞行结果而非机器人可暂停的步骤；task_class=atomic 与之一致。无状态日志，按 P8 when/never/success 为 null，已在 open_questions_zh 说明；约束均为立即终止，来源行号与指令文本一致，未发现多余或缺失的节点、边、阈值。", + "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：['serve_in_target']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:48:37Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__badminton_serve/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__badminton_serve/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "248bafc839e43a1099e3fe8b69440a1b297b88b48d3dd2395d36809406933a28", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/embodied-olympics-fun/badminton_serve/trajectories/embodied-olympics-fun-badminton-serve-i00-privileged--xh7w7ov.json b/phase2/embodied-olympics-fun/badminton_serve/trajectories/embodied-olympics-fun-badminton-serve-i00-privileged--xh7w7ov.json new file mode 100644 index 0000000000000000000000000000000000000000..45439287789f6a39d544f850e06ff561d3087211 --- /dev/null +++ b/phase2/embodied-olympics-fun/badminton_serve/trajectories/embodied-olympics-fun-badminton-serve-i00-privileged--xh7w7ov.json @@ -0,0 +1,206 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-badminton-serve-i00-privileged--xh7w7ov", + "taskKey": "embodied-olympics-fun/badminton_serve", + "dag_sha256": "e5772140aff5f054a271f90fd295419869c755e083b6526601e08a99bee573e1", + "outcome": { + "success": 1, + "reward": 1.0, + "source": "catalog", + "raw": { + "success": 1, + "replay_success": 1, + "deterministic": 1, + "n_actions": 225, + "sim_steps": 225, + "horizon": 1500, + "episode_over": 1, + "progress": 1.0, + "stages_reached": 4, + "stages_total": 4, + "fell": 0, + "replay_s": 8.1, + "bad_trajectory": 0, + "metric_hit_height_m": 0.6018, + "metric_landing_x_m": 5.731, + "metric_landing_y_m": 1.47, + "metric_shuttle_angle_deg": 44.55, + "metric_shuttle_speed_m_s": 11.7617, + "grader_error": 0, + "video_rendered": 1, + "agent_sessions": 44, + "agent_resets": 154, + "agent_replays": 4, + "agent_cost_self_reported": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.745, + "peak_processes": 9, + "peak_disk_gb": 0.252, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0, + "session_log_tampered": 0, + "session_log_from_monitor": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-badminton-serve-i00-privileged--xh7w7ov/replay.mp4", + "duration_s": 2.28, + "fps": 25.0, + "frames": 57, + "sha256": "54506ffab73c0dee841aef7a658e2d1b2b08c4e146d7fa750edc06f5b8838d3d" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "serve_in_target": { + "method": "visual", + "intervals": [ + [ + 2.25, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.92, + "log_step": 95, + "text_zh": "观察：逐帧检查视频 0.00–1.24 s，机器人在近侧站立准备，左臂随后抬开，右手球拍于视频约 0.84–0.92 s 向前上方挥出；视频 0.92–1.24 s 可见白色羽毛球离开机器人向球网上方飞行。log step 95 的最终交付回放由基准裁判报告 release、hit、over the net、in 四阶段全部完成，failed=null；该裁判是仿真任务的规则检查器。" + }, + { + "t": 2.24, + "log_step": 95, + "text_zh": "观察：逐帧检查视频 1.24–2.24 s（含末秒全部帧），机器人双脚保持支撑、右手球拍仍在，羽毛球飞向远侧区域后变小并受球网遮挡，末帧不能辨清首次触地。inputs/run.json 的 reward 报告 success=1、replay_success=1、deterministic=1、stages_reached=4；击球高度 0.6018 m、出球速度 11.7617 m/s、仰角 44.55°，首次落点 (5.731, 1.470) m 位于任务规定目标区内。" + }, + { + "t": 2.25, + "log_step": 95, + "text_zh": "推断：log step 94 执行最终生成脚本并回放交付文件，log step 95 报告该文件执行 225 个控制步、仿真时长 2.25 s，生成脚本（log step 92）注明 60 步为仿真 0.60 s，故控制率为 100 Hz；225/100=2.25 s 与视频总长 2.28 s 相合，且裁判仿真 0.862 s 击球事件已在视频 0.84–0.92 s 的挥拍与出球处核对。按此近似映射，裁判仿真 2.247 s 的 in/success 对应视频约 2.25 s，作为节点达成锚点；该时刻略晚于视频 2.24 s 末帧，成功及合法首次落点由交付回放裁判与最终评分字段共同确认，并非声称末帧直接拍到触地。" + } + ], + "contact_actual": [ + { + "who": "手持工具", + "how": "其他", + "note_zh": "候选：击打；视频约 0.84–0.92 s 机器人右手持拍前上挥，随后白色羽毛球向上飞出，结合 log step 95 交付回放裁判的 hit 事件，判定实际发生球拍对羽毛球的击打；现有方式无击打，推/扫或压/擦均不适用。此前左手持球后放开并撤离，右手握拍链路不重复标注。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "视频 0.00–2.24 s 可见机器人双脚在近侧球场承重支撑，准备及挥拍过程中保持站立，未见跌倒。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "hit_height_m", + "value": 0.6018, + "unit": "m", + "threshold": 1.15, + "direction": "lower_better" + }, + { + "name": "shuttle_speed_m_s", + "value": 11.7617, + "unit": "m/s", + "threshold": null, + "direction": "higher_better" + }, + { + "name": "shuttle_angle_deg", + "value": 44.55, + "unit": "deg", + "threshold": 0, + "direction": "higher_better" + }, + { + "name": "landing_x_m", + "value": 5.731, + "unit": "m", + "threshold": [ + 4.43, + 9.15 + ], + "direction": "target" + }, + { + "name": "landing_y_m", + "value": 1.47, + "unit": "m", + "threshold": [ + 0.43, + 3.04 + ], + "direction": "target" + } + ], + "failure": null, + "summary_zh": "机器人完成放球、右手持拍向前上方击球及越网，最终交付回放裁判和评分器均确认合法发球入区，节点完成 1/1，首次落点为 (5.731, 1.470) m。节点达成标在视频约 2.25 s，由交付文件控制步与裁判事件映射，已用视频约 0.84–0.92 s 的挥拍出球核对；成功参考视频也呈现同类站立发球与越网过程。无状态日志，远景不能核实毫秒级接触和脚步阈值，且视频 2.24 s 末帧略早于裁判落地事件，相关合法性与最终触地依靠本次交付回放及评分字段确认。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "成功运行，单节点 serve_in_target 达成，时间锚点、接触标注（击打=其他、腿/躯干支撑）和裁判字段一致，无失败层，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": "2026-10-29T10:04:33Z", + "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/embodied-olympics-fun-badminton-serve-i00-privileged--xh7w7ov/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-badminton-serve-i00-privileged--xh7w7ov/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "8be181675f60a981106b1f79f3ea6321f2ddaf304100c9a6f52e1b2947cf7399", + "video_sha256": "54506ffab73c0dee841aef7a658e2d1b2b08c4e146d7fa750edc06f5b8838d3d", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/badminton_serve/trajectories/embodied-olympics-fun-badminton-serve-i00-standard--dearvde.json b/phase2/embodied-olympics-fun/badminton_serve/trajectories/embodied-olympics-fun-badminton-serve-i00-standard--dearvde.json new file mode 100644 index 0000000000000000000000000000000000000000..501ff79caf6ad85f94011fdfa327e8c95c6f5a65 --- /dev/null +++ b/phase2/embodied-olympics-fun/badminton_serve/trajectories/embodied-olympics-fun-badminton-serve-i00-standard--dearvde.json @@ -0,0 +1,228 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-badminton-serve-i00-standard--dearvde", + "taskKey": "embodied-olympics-fun/badminton_serve", + "dag_sha256": "e5772140aff5f054a271f90fd295419869c755e083b6526601e08a99bee573e1", + "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": 271, + "sim_steps": 532, + "calls": 5, + "rows": 532, + "resets": 1, + "reset_budget": 5, + "sim_startups": 1, + "elapsed_s": 6444.2, + "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": 0.0, + "peak_memory_gb": 1.219, + "peak_processes": 11, + "peak_disk_gb": 2.597, + "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/embodied-olympics-fun-badminton-serve-i00-standard--dearvde/replay.mp4", + "duration_s": 2.72, + "fps": 25.0, + "frames": 68, + "sha256": "6b7231cc3ead0806d6fb281686b711be780321c4c17319709116a5d13abc8f44" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "serve_in_target": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 140, + "text_zh": "观察：视频 0.00–2.20 s 中机器人站在近侧发球区，左手带着白色羽球抬到身前，右手握拍后摆准备；log step 140 显示 agent 重置后提交 a2，episode=2，是最后一次实时尝试。此前 a1 的结果不用于本节点判定。" + }, + { + "t": 2.48, + "log_step": 141, + "text_zh": "观察：视频 2.36–2.68 s 中白色羽球离开左手后从身前持续下落，右臂挥拍，但没有出现羽球向上击出并飞向远侧接发球区的轨迹。log step 141 读取 a2 的服务裁判结果：stages_reached=1/4，outcome 为“the racket's forward movement stopped before the hit”，done=True；服务只确认放球，未确认击球。" + }, + { + "t": 2.68, + "log_step": 141, + "text_zh": "推断：本次视频末段的下落羽球与 log step 141 的未击球判定一致，合法击球、过网、入区这条验收链没有完成，因此该事件节点全程未达到；不是仅由 success=0 推定。成功参考视频 1.76–1.96 s 可见羽球上升飞离机器人，本次视频没有该表现；参考只用于识别可见线索。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "观察：视频 0.00–2.20 s 的准备阶段，机器人左手持羽球、右手握球拍拍柄，两臂配合移动到放球和挥拍位置；随后左手释放，右手继续持拍挥动。实际持物接触限于手与各自物体，log step 141 未确认球拍击中羽球，因此不标手持工具的击打接触。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：视频 0.00–2.68 s 中机器人两脚支撑球场地面，身体保持站立完成准备、放球和挥拍尝试。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "hit_height_m", + "unit": "m", + "threshold": 1.15, + "direction": "lower_better", + "value": null + }, + { + "name": "shuttle_speed_m_s", + "unit": "m/s", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "shuttle_angle_deg", + "unit": "deg", + "threshold": 0, + "direction": "higher_better", + "value": null + }, + { + "name": "landing_x_m", + "unit": "m", + "threshold": [ + 4.43, + 9.15 + ], + "direction": "target", + "value": null + }, + { + "name": "landing_y_m", + "unit": "m", + "threshold": [ + 0.43, + 3.04 + ], + "direction": "target", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "serve_in_target", + "t": 2.56, + "log_step": 141, + "where": "操作中", + "evidence_zh": "观察：视频 2.36–2.68 s 中机器人左手放球后右手挥拍，羽球持续下落而未被击出，视频 2.52–2.60 s 已降至右手附近及其下方。log step 141 的 a2 服务裁判报告“球拍前挥在击球前停止”、只达到放球阶段；该次发球未修复，断点为正在尝试的合法发球入区步骤，视频 2.56 s 为挥空过程的近似导航锚点。" + }, + "unmet": { + "serve_in_target": false + }, + "phenomena": [ + { + "cat": "运动", + "item": "动态时机", + "t": 2.56, + "log_step": 141, + "evidence_zh": "观察：视频 2.36–2.68 s 中下落的羽球经过机器人身前，右手球拍挥动后羽球仍向下，没有转成上升向网的飞行；log step 141 的裁判结果确认击球前前挥已停止。推断：机器人放球后的挥拍没有与运动中的羽球形成有效拦截，属于动态目标的击球配合失败。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "推断：agent 的脚本在 log step 123 将放球、挥拍及随挥编成一次提交的开环 stand 指令序列，log step 129 写入关节目标，log step 139 修改接触姿态，log step 140 用 H2 姿态及固定时序提交最终 a2；视频 2.36–2.68 s 的实际羽球下落和挥拍未形成击球，log step 141 确认前挥先停止。证据支持放球与球拍运动的控制配合失效，但不能进一步区分接触几何、关节跟踪或时序参数的具体责任，保留在控制大类。", + "source_zh": "日志" + } + ], + "gap_zh": "无状态日志；本次视频仅有远景，25 fps 的画面不能分辨球托初触、毫秒级接触持续、拍头瞬时前向速度及精确三维错失距离。a2 的服务裁判结果支持未击球和前挥停止，但不足以把控制根因细分；DAG 的五项原生指标未在评分字段中报告，均填 null。视频末帧为视频 2.68 s，log step 141 报告终止于仿真 2.71 s，不能声称末帧直接显示裁判终止瞬间。" + }, + "summary_zh": "评分采用 log step 140 重置后的第二次实时尝试 a2；机器人保持站立、完成放球并挥拍，但合法发球入区节点未达到，最终进度为 0/1。视频 2.36–2.68 s 可见羽球下落而未被击出，log step 141 裁判只确认放球并判定击球前前挥停止，断点为发球操作中的动态时机失败，根因保留为控制。缺少状态日志和足够近的接触画面，无法细分具体控制误差，原生击球与落点指标也未报告。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "节点 serve_in_target 未达到、最终 0/1、断点在发球操作中、现象为动态时机（接球目标为运动中的羽球，对应 CASES 431 的同类判例）、根因为控制，均与视频帧和 log step 140/141 的服务裁判结果（stages 1/4，球拍前挥在击球前停止）一致；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": "2026-10-29T10:05:22Z", + "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/embodied-olympics-fun-badminton-serve-i00-standard--dearvde/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-badminton-serve-i00-standard--dearvde/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "ad66578dc2b91db748a6fa2fed1cf26d7824d5c55b84f37c8b3330803ef9a99c", + "video_sha256": "6b7231cc3ead0806d6fb281686b711be780321c4c17319709116a5d13abc8f44", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/bowling/task_dag.json b/phase2/embodied-olympics-fun/bowling/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..8e8ab7289b5a37e03406afab50716f4e6ed2aaff --- /dev/null +++ b/phase2/embodied-olympics-fun/bowling/task_dag.json @@ -0,0 +1,271 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "embodied-olympics-fun/bowling", + "suite": "embodied-olympics-fun", + "base": "embodied-olympics-fun-bowling-i00", + "title_zh": "保龄球全中", + "summary_zh": "在犯规线后合法送出一球，首次有效碰瓶满五秒的点数时十只球瓶全部倒下，且未触发犯规、入沟、非法触球或跌倒失败。", + "embodiment_zh": "两模式均为 Unitree H1-2 双足人形机器人，27 个关节，双臂末端各为无工作手指的平橡胶掌垫，无有效抓握机构、夹爪或吸盘；开局球靠双掌夹压和摩擦保持，脚在助走区承重，无轮式底盘。", + "task_class": "atomic", + "analysis_level": "step", + "spec": { + "version": "1.2", + "task": "embodied-olympics-fun-bowling-i00", + "bench": "embodied-olympics-fun", + "progress": "live", + "entities": {}, + "milestones": [ + { + "id": "strike_completed", + "zh": "投球全中", + "mode": "state", + "when": null, + "after": [], + "hold_s": 0, + "goal": true, + "freeze_held": false, + "criterion_zh": "机器人从犯规线后送出紫色球，球连续 0.05 秒不接触机器人后确认离手，球心越过犯规线（x > 0）并首次有效碰瓶，在该次碰瓶满 5.0 秒后的首次控制步点数时十瓶均倒下：每瓶的竖轴偏离世界竖直超过 20°，或瓶体原点已离开球道范围（|y| > 0.52705 米、x < 0 或 x > 19.157 米、或 z <= 0.09 米）。球瓶视角应见十只白色红颈球瓶全部倾倒或离开木质瓶台；只有球离手、途中碰瓶、部分瓶倒下或尚未到点数时刻均未完成，点数前球瓶恢复直立且回到球道内也不算倒瓶，并须至点数时未触发犯规、入沟、非法触球或跌倒失败。", + "step": { + "cat": "动态", + "item": "击打" + }, + "contact": [ + { + "who": "双臂协作", + "how": "其他", + "note_zh": "候选：夹持；左右平掌垫从两侧夹压保龄球并送出，两手没有能闭合的手指，不能标抓握（D12）；随后球离手，由球与球瓶及球瓶之间的碰撞完成倒瓶。", + "plan": 1 + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚接触助走区支撑身体，配合双掌夹压送出保龄球。", + "plan": 1 + }, + { + "who": "双臂协作", + "how": "托/舀", + "note_zh": "左右平掌垫从球的下后方共同承托并送出保龄球；开局先由两掌侧向夹压，随后转成承托，二者依次发生；成功参考采用此方案。", + "plan": 2 + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚接触助走区支撑身体，配合双掌承托送出保龄球。", + "plan": 2 + }, + { + "who": "手指/夹爪", + "how": "托/舀", + "note_zh": "一只平掌垫在开局双掌夹球之后承托并送出保龄球；规则允许离手前球仅接触一只手，不要求始终双手共同送球。", + "plan": 3 + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚接触助走区支撑身体，配合单掌承托送出保龄球。", + "plan": 3 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "task/privileged/tests/rfo_task.py:149-179", + "task/privileged/tests/rfo_task.py:269-294", + "task/privileged/tests/rfo_task.py:310-368", + "grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178", + "task/standard/environment/sim/rfo_task.py:310-368", + "oracle/replay.mp4 @ 0.0, 0.5, 1.0, 4.5, 5.0, 9.64 s", + "oracle/run.json: reward; steps[id=83,84,87]" + ], + "conf": 6, + "note_zh": "一次合法投球造成全中，delivery、pins、strike 只作为本步的有序验收，不拆蓄力、离手、飞行、碰瓶、等待步骤，也不强制参考轨迹的静止站位、下手式摆臂或双手出球。逐瓶判定的十项是同一次动作的连锁结果；此粒度取舍见疑问。5 秒是首次碰瓶后的点数延迟，不是要求十瓶连续倒下 5 秒，因此 hold_s 为 0；点数以前倒瓶不提前记完成。远景录像中球瓶很小，无法精确量倾角或数清遮挡球瓶时应参考球瓶相机或原生逐瓶状态，不能凭球飞走推断全中。" + } + ], + "constraints": [ + { + "id": "foul_contact", + "zh": "越线触碰", + "never": null, + "criterion_zh": "整次尝试中，机器人任一部位与球及机器人自身以外的物体产生接触，且任一接触点位于犯规线或其前方（场景坐标 x >= 0），立即犯规；人在空中越线但没有接触不等同犯规。录像看机器人是否接触线前球道、沟槽或其他场物，细微接触需接触记录确认。", + "consequence": "立即终止", + "sources": [ + "task/privileged/tests/rfo_task.py:322-327" + ] + }, + { + "id": "pre_release_body_touch", + "zh": "出手前碰身", + "never": null, + "criterion_zh": "从开局至确认离手前，球不得接触左右手以外的机器人部位；手包括掌垫和归为手的腕部。球碰到前臂、髋部或躯干即违规，确认离手要求连续 0.05 秒不接触机器人，离手时刻回记为最后手部接触时刻。", + "consequence": "立即终止", + "sources": [ + "task/privileged/tests/rfo_task.py:156", + "task/privileged/tests/rfo_task.py:330-339", + "grader/efe_embodied-olympics-fun-files/rfo/referee.py:25-43" + ] + }, + { + "id": "post_release_ball_touch", + "zh": "出手后触球", + "never": null, + "criterion_zh": "确认离手以后、首次有效碰瓶以前，球再次接触机器人任一部位即违规；录像可见已分离的球又碰到掌垫、腕部或其他身体部位。任务文字禁止出手后全程再碰球，但代码在首次碰瓶后跳过此检查，差异见疑问。", + "consequence": "立即终止", + "sources": [ + "task/privileged/tests/rfo_task.py:328-342" + ] + }, + { + "id": "gutter_before_pins", + "zh": "碰瓶前入沟", + "never": null, + "criterion_zh": "确认离手后、首次有效碰瓶以前，球一旦与左右沟槽发生接触即失败；仅靠近沟边不算，首次有效碰瓶后球进入沟槽不受这一检查否决。录像表现为球沿木质球道外的灰色沟槽运动，边界擦碰需接触记录确认。", + "consequence": "立即终止", + "sources": [ + "task/privileged/tests/rfo_task.py:303", + "task/privileged/tests/rfo_task.py:328-350" + ] + }, + { + "id": "robot_fallen", + "zh": "机器人跌倒", + "never": null, + "criterion_zh": "点数结束前，骨盆世界高度低于 0.60 米（助走区表面高 0.10 米，阈值为其上 0.50 米），或左右脚以外的部位接触助走区或建筑地板，即跌倒失败；录像看髋部明显下坠或手、膝、躯干触地，单纯身体摇晃不算。", + "consequence": "立即终止", + "sources": [ + "task/privileged/tests/rfo_task.py:33", + "task/privileged/tests/rfo_task.py:152-156", + "task/privileged/tests/rfo_task.py:211-223", + "grader/efe_embodied-olympics-fun-files/rfo/task.py:20-22", + "grader/efe_embodied-olympics-fun-files/rfo/referee.py:272-289" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "pins_down", + "level": "new_code", + "source": "task/privileged/tests/rfo_task.py:310-320,370-375; log Task.privileged(w)[\"pins_down\"] as ten per-pin booleans" + }, + { + "atom": "legal_delivery", + "level": "new_code", + "source": "task/privileged/tests/rfo_task.py:330-356; log trackers.released and referee stages delivery,pins" + }, + { + "atom": "pin_count_due", + "level": "new_code", + "source": "task/privileged/tests/rfo_task.py:305,358-368; log trackers.t_pins and referee time; compare elapsed >= params.settle_s" + }, + { + "atom": "foul_contact", + "level": "new_code", + "source": "task/privileged/tests/rfo_task.py:324-327; log robot-external contact point x, excluding ball" + }, + { + "atom": "pre_release_body_touch", + "level": "new_code", + "source": "task/privileged/tests/rfo_task.py:330-339; log ball contact robot parts and trackers.released,last_touch" + }, + { + "atom": "post_release_ball_touch", + "level": "new_code", + "source": "task/privileged/tests/rfo_task.py:328-342; log ball-robot contact and trackers.released,t_pins" + }, + { + "atom": "gutter_before_pins", + "level": "new_code", + "source": "task/privileged/tests/rfo_task.py:328-350; log ball-gutter contacts and trackers.released,t_pins" + }, + { + "atom": "robot_fallen", + "level": "new_code", + "source": "grader/efe_embodied-olympics-fun-files/rfo/referee.py:276-282; log pelvis world z and parts_touching(ground)" + } + ] + }, + "native_metrics": [ + { + "name": "pins_down", + "unit": "个", + "threshold": 10, + "direction": "higher_better", + "sources": [ + "task/privileged/tests/rfo_task.py:310-320", + "task/privileged/tests/rfo_task.py:358-368" + ] + } + ], + "open_questions_zh": [ + "inputs/states/index.json 的 logs 为空；按 P8，所有节点的 when、所有约束的 never 和 success 均为 null，不能安全绑定未提供的逐帧字段；atoms 列出后续适配器须记录的世界状态。无状态回放可用于自动 reach-all 或 success==grader 检查。", + "粒度边界（S5、S10、D6）：文本要求一球击倒十瓶，评分逐瓶计算真假再求和，但十瓶倒下是一次不可暂停后继续的送球动作产生的连锁目标，按 Step 3.7 的一次动作产生多个目标与 S10 合为一个投球全中节点，task_class 在文字步骤复核中定为 atomic；不把它解释为十次独立击倒操作。若将 S5 对所有计数目标无条件优先，则会得到十个计数节点；这一规则边界仍有疑义。", + "任务设计问题：文本说离手后任何部位不得再触球；但 task/privileged/tests/rfo_task.py:328-329 在首次有效碰瓶后直接返回，跳过再次触球和入沟检查（341-347）。DAG 的约束按代码注明检查区间，success 对应实际裁判，不能额外否决代码允许的碰瓶后触球。", + "文字与代码边界：文本称球碰瓶前入沟无效；代码仅在确认离手后检测沟槽（task/privileged/tests/rfo_task.py:330-347），持续手触球时提前返回。犯规线检查始终运行，但过线阶段仅检验确认离手后球心 x > 0，并不检查一次从线后的跨越事件（344-345）。不另造过线步骤。", + "两模式的任务模块逐字相同、参数和冻结初始骨盆位置相同，使用同一物理裁判；交付评分不同：privileged 额外要求两次独立回放成功且终态哈希相同（grader/efe_embodied-olympics-fun-files/grade.py:101-115），standard 判最后一次在线尝试的裁判结果（grader/efe_embodied-olympics-fun-files/standard/server.py:157-178）。回放确定性及文件有效性不作为世界状态节点，当前 success 无日志可表达而为 null；INDEX 所列 standard 通用 grade.py 未实际包含在本目录中，外层评分仅能由服务端代码及说明核对。", + "接触词汇缺口 D12：无手指双掌夹压球使用双臂协作×其他，候选夹持；参考动作先夹压再转为双掌承托，后者为托/舀；文本允许一手或双手触球，故同时保留单掌承托的替代方案，不把双手共同出球规定为必经。", + "成功参考仅提供 640×360 场地远景视频，0 秒见双掌持球，约 1 秒球已离手，约 4.7 秒抵达瓶区，结束仍双脚站立；远端球瓶过小，不能逐瓶可靠量角或辨认全部十瓶。全中及三阶段顺序由该参考的原生回放记录佐证（oracle/run.json 的 reward 与 steps[id=83]），不由远景强推十个逐瓶完成时刻。", + "release_speed_m_s、entry_speed_m_s、entry_y_m、entry_angle_deg 为原生记录指标（task/privileged/tests/rfo_task.py:340,353-355），没有成功阈值，不作为步骤或约束；参考的 5.7 米每秒离手速度、4.25 米每秒入瓶速度和偏移仅属该策略。", + "40 秒或 4000 控制步是尝试时限，按 P21 不列为约束；5 秒点数延迟包含在全中验收中，不能误设为倒瓶持续保持五秒。" + ], + "conf": 6, + "conf_reason_zh": "目标、初态、逐瓶阈值和终止规则有源码依据，标准与特权物理裁判一致，参考回放记录确认全中；无状态日志且远景不能逐瓶复核，单球多瓶的 S5/S10 粒度边界以及文本与代码的触球检查区间差异降低置信度。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "单节点“投球全中”（动态·击打）与指令“一球击倒十瓶”及 bowling_realgrasp 先例（STEP-8）一致，十瓶是一次不可暂停的送球产生的连锁目标，按 S10 合并可接受；约束覆盖犯规、碰身、出手后触球、入沟、跌倒，阈值（20°、0.60 m）与文字及代码一致；无状态日志故 when/never/success 为 null 合规，且 open_questions_zh 已列出文字与代码差异。无需修改内容。", + "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：['strike_completed']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:46:11Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__bowling/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__bowling/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "9ab0ee6133c1dc9c14f4f92ddb61e9415fea366cb83c83bb8ffed700012b8d04", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/embodied-olympics-fun/bowling/trajectories/embodied-olympics-fun-bowling-i00-privileged--nd6fgot.json b/phase2/embodied-olympics-fun/bowling/trajectories/embodied-olympics-fun-bowling-i00-privileged--nd6fgot.json new file mode 100644 index 0000000000000000000000000000000000000000..fd8eae4d1eec32b9221ee9408abd12e58ffd5119 --- /dev/null +++ b/phase2/embodied-olympics-fun/bowling/trajectories/embodied-olympics-fun-bowling-i00-privileged--nd6fgot.json @@ -0,0 +1,177 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-bowling-i00-privileged--nd6fgot", + "taskKey": "embodied-olympics-fun/bowling", + "dag_sha256": "958dbc6f1b15aea49b7fbb6f67f6fe11ebb2ebd0b8c66ca9ec74face659152c2", + "outcome": { + "success": 1, + "reward": 1.0, + "source": "catalog", + "raw": { + "success": 1, + "replay_success": 1, + "deterministic": 1, + "n_actions": 1022, + "sim_steps": 1022, + "horizon": 4000, + "episode_over": 1, + "progress": 1.0, + "stages_reached": 3, + "stages_total": 3, + "fell": 0, + "replay_s": 15.85, + "bad_trajectory": 0, + "metric_entry_angle_deg": -0.47, + "metric_entry_speed_m_s": 3.83, + "metric_entry_y_m": -0.12, + "metric_pins_down": 10.0, + "metric_release_speed_m_s": 5.06, + "grader_error": 0, + "video_rendered": 1, + "agent_sessions": 1026, + "agent_resets": 10967, + "agent_replays": 5, + "agent_cost_self_reported": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 10.398, + "peak_processes": 24, + "peak_disk_gb": 1.384, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0, + "session_log_tampered": 0, + "session_log_from_monitor": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-bowling-i00-privileged--nd6fgot/replay.mp4", + "duration_s": 10.24, + "fps": 25.0, + "frames": 256, + "sha256": "e82225205977aa60b0585d7e8557180a4c97994e0d0709718c9a9b653f913dc3" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "strike_completed": { + "method": "visual", + "intervals": [ + [ + 10.22, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.8, + "log_step": null, + "text_zh": "观察：视频 0.0–0.8 s，机器人双脚站在犯规线后的助走区，两只平掌垫在髋前两侧夹压紫色保龄球；视频 0.88–0.92 s 双臂向前送球，视频 0.96–1.12 s 球已与掌垫分离并向球道前方运动。" + }, + { + "t": 5.28, + "log_step": null, + "text_zh": "观察：视频 4.8–5.2 s 紫色球沿木质球道抵近球瓶，视频 5.28–6.0 s 瓶阵开始散开和倾倒；视频 8.0 s 及 9.2–10.2 s 瓶台上的白色瓶体呈倒伏状，机器人始终双脚支撑站立。远景无法逐瓶精确测量倾角或排除重叠遮挡。" + }, + { + "t": 10.2, + "log_step": 148, + "text_zh": "观察：inputs/run.json 的 reward 报告 success=1、replay_success=1、deterministic=1、stages_reached=3、metric_pins_down=10、fell=0。log step 133 的脚本将已通过两次独立回放的 1022 行动作保存为交付文件；log step 148 输出交付动作数组与该候选完全一致，两次回放均由仿真裁判报告 delivery、pins、strike 三阶段及十瓶全倒，终态哈希相同，故这些回放对应本次交付。" + }, + { + "t": 10.22, + "log_step": 148, + "text_zh": "推断：节点在视频约 10.22 s 点数完成时才成立。log step 6 的工具说明记载每控制步为 0.01 s；本次交付回放在 log step 148 执行 1022 步，即仿真 10.22 s，与视频 10.24 s 总时长及最后一帧视频 10.20 s 一致。该回放事件的仿真 0.97 s 过线、5.212 s 首次碰瓶，已分别用视频 0.96–1.12 s 球脱手前行和 5.20–5.36 s 瓶阵开始变化核对，采用视频时间约等于控制步数乘 0.01 s 的映射；首次碰瓶后五秒的 strike 事件为仿真 10.22 s，评分器十瓶全倒及成功判定支持节点终值为真。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "其他", + "note_zh": "候选：夹持。观察：视频 0.0–0.88 s，机器人两只平掌垫从两侧夹压紫色保龄球，随后两臂前送，视频 0.96 s 球已离手；没有手指闭合或夹爪包围，故不标抓握。背后远景不能可靠细分离手瞬间各掌的接触点，不据成功参考另加承托接触。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：视频 0.0–1.12 s 投球过程中，机器人双脚接触犯规线后木质助走区承重；视频 1.5–10.2 s 仍站立支撑，等待球瓶倒伏及点数。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "pins_down", + "value": 10.0, + "unit": "个", + "threshold": 10, + "direction": "higher_better" + } + ], + "failure": null, + "summary_zh": "机器人在犯规线后用双掌夹压持球并向前送出，球于视频约 5.2 s 碰瓶；首次碰瓶后的五秒点数结束时，投球全中节点在视频约 10.22 s 达成，终态为 1/1，评分器记录十瓶全倒及确定性回放成功。视频检查至最后一帧视频 10.20 s，机器人仍站立；没有失败断点。无状态日志，远景与成功参考一样不足以逐瓶精确量角，十瓶验收及合法投球由本次交付的裁判结果佐证，离手瞬间各掌接触点仍受遮挡。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "标签与视频和评分器结果一致：球在约 1 s 离手，约 5.3 s 碰瓶，success=1、十瓶全倒，节点终值为真；时钟映射有核对，接触标注合理，无失败断点。", + "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": "2026-10-29T10:03:34Z", + "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/embodied-olympics-fun-bowling-i00-privileged--nd6fgot/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-bowling-i00-privileged--nd6fgot/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "956bd427ba7710d554847f224ff07596f196a21a23391716dae3d2b7e17cd6f7", + "video_sha256": "e82225205977aa60b0585d7e8557180a4c97994e0d0709718c9a9b653f913dc3", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/bowling/trajectories/embodied-olympics-fun-bowling-i00-standard--gtnmg2p.json b/phase2/embodied-olympics-fun/bowling/trajectories/embodied-olympics-fun-bowling-i00-standard--gtnmg2p.json new file mode 100644 index 0000000000000000000000000000000000000000..48778c0c57a9c0e0d13569a011ded5d25960f1f1 --- /dev/null +++ b/phase2/embodied-olympics-fun/bowling/trajectories/embodied-olympics-fun-bowling-i00-standard--gtnmg2p.json @@ -0,0 +1,197 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-bowling-i00-standard--gtnmg2p", + "taskKey": "embodied-olympics-fun/bowling", + "dag_sha256": "958dbc6f1b15aea49b7fbb6f67f6fe11ebb2ebd0b8c66ca9ec74face659152c2", + "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": 1635, + "sim_steps": 25065, + "calls": 190, + "rows": 9411, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1462.0, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.512, + "peak_processes": 13, + "peak_disk_gb": 1.167, + "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/embodied-olympics-fun-bowling-i00-standard--gtnmg2p/replay.mp4", + "duration_s": 16.36, + "fps": 25.0, + "frames": 409, + "sha256": "c392fbe49c9c92b56d321e02d66fa7b4b205cfb0fd7a862bc6bd5b3c805f39ab" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "strike_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察：视频 0.00–0.52 s，机器人双脚站在犯规线后的助走区，两只平掌垫夹压紫色球；视频 0.64–0.80 s 双臂向前上方摆动，掌垫转到球的下侧，视频 0.88–1.00 s 球与双掌分离。完成送球动作本身尚不满足投球全中。" + }, + { + "t": 16.32, + "log_step": 146, + "text_zh": "观察：视频 8.00–16.32 s，紫色球沿球道向远端运动并逐渐靠近画面左侧沟边；视频 15.52、15.84、16.08、16.32 s 球仍在球瓶前方的边缘位置，远端球瓶保持直立排列，没有碰瓶后倒瓶过程。log step 146 的机器人服务端最终观测返回 done=True、reason=failed、outcome=a gutter ball；log step 144 确认这是最后的 episode 51，重置次数剩余为零。" + }, + { + "t": 16.32, + "log_step": 138, + "text_zh": "推断：log step 136 发起的最后一组测试以 p0.11 为末次投球，log step 138 返回该次 a gutter ball、stages=1/3，随后 log step 142 重置被拒绝，故这次评分仅有送球阶段，未进入有效碰瓶和全中验收。成功参考在参考视频 4.80 s 出现球瓶受撞运动、参考视频 9.68 s 瓶台上球瓶倒伏，本次视频没有相应变化；结合本次服务端入沟判定，投球全中始终未达成。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "其他", + "note_zh": "观察：视频 0.00–0.52 s，左右无手指平掌垫从两侧夹压紫色保龄球。候选：夹持；两掌之间夹压，单手没有合拢包住球，不符合抓握，现有方式无准确对应项。" + }, + { + "who": "双臂协作", + "how": "托/舀", + "note_zh": "观察：在初始侧向夹压之后，视频 0.64–0.80 s，机器人双掌位于球的下侧，配合向前上方摆臂承托送球；视频约 0.9 s 球离手。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：视频 0.00–16.32 s，机器人双脚接触助走区承重，配合双臂持球和送球，之后仍站立。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "pins_down", + "value": null, + "unit": "个", + "threshold": 10, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "strike_completed", + "t": 16.0, + "log_step": 146, + "where": "操作中", + "evidence_zh": "观察：机器人在本步中已送出球，但视频 15.52–16.32 s 球沿画面左侧球道边缘接近瓶区，未碰到仍直立的球瓶；log step 146 的服务端最终观测明确给出 a gutter ball，log step 144 确认末次 episode 51 已失败。断点锚定在视频约 16 s 的贴沟边运动，确切沟槽接触点在远景中无法辨认。" + }, + "unmet": { + "strike_completed": false + }, + "phenomena": [ + { + "cat": "运动", + "item": "位姿不准", + "t": 16.0, + "log_step": 146, + "evidence_zh": "观察：视频 2.00–16.32 s，机器人送出的紫色球没有沿能撞到球瓶的位置保持球路，而是逐步逼近画面左侧沟边；末段球仍位于瓶阵前侧边缘，球瓶保持直立，log step 146 的服务端结果为 a gutter ball。目标球瓶静止，偏离的是送球方向及后续球路。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志 / 视频", + "evidence_zh": "推断：log step 120 的投球脚本固定左右肩俯仰摆动，并通过 torso 目标改变瞄准方向；log step 136 以相同摆臂方案依次试 torso=0.05、0.08、0.11，log step 138 三次均返回 a gutter ball，末次即最终评分回合。视频 0.64–1.00 s 机器人实际摆臂送出球，视频 8.00–16.32 s 球路持续偏向沟边，支持送球控制未能产生通向瓶阵的球路。没有足够数据区分掌垫接触、释放状态、关节跟踪或瞄准参数中的具体机制，保留控制大类。" + } + ], + "gap_zh": "没有逐帧状态日志或球—沟槽接触记录，远景无法精确定位首次入沟接触，断点时间仅为视频约 16 s 的导航锚点。未提供末次投球的力、释放速度及完整关节跟踪数据，无法进一步确定控制子类；run.json 的 reward 未报告 pins_down，原生指标保留 null。agent 在 log step 188 自述球向机器人右侧偏，但本标注仅按本次画面描述为画面左侧沟边，不将其自述方向或摩擦推测作为事实。" + }, + "summary_zh": "本次评分对应最后的在线回合 episode 51：机器人双掌夹球后承托向前送出，约视频 0.9 s 离手并持续站立，但投球全中节点始终未达到，最终为 0/1。球路逐渐偏向画面左侧沟边，视频末段尚未碰瓶，服务端明确判为入沟；现象为位姿不准，根因保留控制大类。缺少接触及释放状态记录，无法精确定位入沟瞬间或细分控制机制，评分字段也未报告倒瓶数。", + "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 strike_completed 未达成（服务端 a gutter ball、末段球贴左沟边、瓶全部直立）与视频一致；断点、位姿不准现象、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": "2026-10-29T10:04:53Z", + "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/embodied-olympics-fun-bowling-i00-standard--gtnmg2p/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-bowling-i00-standard--gtnmg2p/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "bc5ffd991a4beb980c9cb85f3d1c5863f40542d44fd4256cb1e8db65f410cc54", + "video_sha256": "c392fbe49c9c92b56d321e02d66fa7b4b205cfb0fd7a862bc6bd5b3c805f39ab", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/box_jump/task_dag.json b/phase2/embodied-olympics-fun/box_jump/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..bc1f18927277e18f51ad5ecc60a074085bb76c75 --- /dev/null +++ b/phase2/embodied-olympics-fun/box_jump/task_dag.json @@ -0,0 +1,260 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "embodied-olympics-fun/box_jump", + "suite": "embodied-olympics-fun", + "base": "embodied-olympics-fun-box-jump-i00", + "title_zh": "双脚跳箱并站直", + "summary_zh": "机器人按合规双脚起跳和箱顶落地规则完成一次跳箱，并在箱顶双脚支撑、伸直站稳达到连续站立窗口。", + "embodiment_zh": "两种模式均为 Unitree H1-2 双足人形机器人，27 个关节，身高约 1.78 m、质量约 67 kg；双臂腕端为不可抓取的固定拳垫，无可动手指，本任务由双腿和躯干完成，双脚支撑，无轮式底盘。", + "task_class": "atomic", + "analysis_level": "step", + "spec": { + "version": "1.2", + "task": "embodied-olympics-fun-box-jump-i00", + "bench": "embodied-olympics-fun", + "progress": "live", + "entities": {}, + "milestones": [ + { + "id": "box_jump_stood", + "zh": "跳上箱并站直", + "mode": "state", + "when": null, + "after": [], + "hold_s": 1.99, + "freeze_held": false, + "goal": true, + "criterion_zh": "场地侧视中，机器人从棕色箱前的地面双脚蹬起，未经触箱先有至少 0.10 s 的完全腾空（各脚在起跳前 0.10 s 内触地，骨盆较离地时上升至少 0.10 m），首次落地接触后的 0.50 s 内只许脚接触箱顶且窗口结束时双脚均在顶面，每处脚箱接触的箱体局部高度不低于顶面减 0.02 m、法线偏离顶面外法线不超过 20°，之后双脚留在顶面、无其他身体部位触地或触箱，伸直双腿使骨盆高于箱顶至少 1.00 m、躯干偏离竖直不超过 20°、骨盆速度不超过 0.35 m/s，连续保持 200 个 0.01 s 控制步才完成。箱前深蹲、空中收腿或箱顶蹲伏都尚未完成；站直后若失去上述支撑、姿态或速度条件即不再满足，站立计数重置，触地或非脚触箱则立即判负。", + "step": { + "cat": "本体", + "item": "跳跃" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "plan": 1, + "note_zh": "机器人左右脚在地面蹬起，并以左右脚落到箱顶承重、站直；中间腾空无外部接触，手臂可摆动但不接触箱子。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task.json:10", + "inputs/task/privileged/instruction.md:4", + "inputs/task/standard/instruction.md:5", + "inputs/task/privileged/environment/rfo_task.py:35-42", + "inputs/task/privileged/environment/rfo_task.py:72-84", + "inputs/task/privileged/environment/rfo_task.py:103-175", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-184", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:291-301", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:353-389", + "inputs/oracle/replay.mp4:0.00-4.12s" + ], + "conf": 6, + "note_zh": "按 S10 合并一次连续跳跃的起跳、落地和站直，裁判 take-off → landing → stand 仅作验收顺序；不因原生三阶段而拆节点。hold_s=1.99 来自 D31：(200-1)×0.01 s，文本称保持 2.0 s；when 为 null，实际由视频及裁判记录核实整个历史和保持窗口，不额外等待 1.99 s。箱子可移动，顶面条件随箱体当前位姿计算，不固定为世界 z=0.61 m。" + } + ], + "constraints": [ + { + "id": "prejump_fall", + "zh": "起跳前跌倒", + "never": null, + "criterion_zh": "有效起跳阶段尚未成立时，脚以外的任一身体部位接触地面即失败；深蹲和骨盆低不单独算跌倒。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:126-128", + "inputs/task/privileged/environment/rfo_task.py:37" + ] + }, + { + "id": "box_touched_before_jump", + "zh": "跳前触箱", + "never": null, + "criterion_zh": "首个至少腾空 0.10 s 的飞行结束前，机器人任何部位接触箱子即失败，不能踏上或攀爬箱子。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:132-144" + ] + }, + { + "id": "invalid_takeoff", + "zh": "起跳不合规", + "never": null, + "criterion_zh": "首个有效飞行结束时，若任一脚在离地前 0.10 s 内没有接触地面，或飞行顶点骨盆相对起跳时上升不足 0.10 m，立即失败。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:134-141", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:368-389" + ] + }, + { + "id": "invalid_landing", + "zh": "落地不合规", + "never": null, + "criterion_zh": "起跳阶段确认后至落地确认前，任何部位触地、脚以外部位触箱或脚接触箱侧面均立即失败；首次接触后满 0.50 s 时若没有双脚同时接触合规箱顶，也立即失败，脚接触的局部高度须不低于顶面减 0.02 m，接触法线与箱顶向外法线夹角须不超过 20°。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:103-120", + "inputs/task/privileged/environment/rfo_task.py:147-161" + ] + }, + { + "id": "forbidden_contact_after_landing", + "zh": "落地后非法接触", + "never": null, + "criterion_zh": "落地阶段确认后，机器人任一部位接触地面，或脚以外任一部位接触箱子，即立即失败；站立姿态或速度短暂不达标只重置站立计数，不单独触发此约束。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:147-152", + "inputs/task/privileged/environment/rfo_task.py:163-175", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:180-184" + ] + }, + { + "id": "simulation_unstable", + "zh": "仿真发散", + "never": null, + "criterion_zh": "物理步出现 MuJoCo 加速度异常警告、非有限位置，或机器人或道具任一被检查的线速度分量绝对值超过 150 m/s，判为仿真发散并结束失败。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:108-109", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:866-885" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "dyn.box_jump_legal_stand", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:103-175; inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:180-184,291-301; 待逐控制步记录合规起跳与落地历史、当前双脚顶面接触、骨盆高度与速度、躯干竖直条件，并记录原生站立计数；当前没有状态日志，不能绑定。" + }, + { + "atom": "dyn.prejump_fall", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:126-128; inputs/task/privileged/environment/rfo_task.py:37; 待记录该违规的逐步布尔量，当前没有状态日志。" + }, + { + "atom": "dyn.box_touched_before_jump", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:132-144; 待记录该违规的逐步布尔量，当前没有状态日志。" + }, + { + "atom": "dyn.invalid_takeoff", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:134-141; inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:368-389; 待记录该违规的逐步布尔量，当前没有状态日志。" + }, + { + "atom": "dyn.invalid_landing", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:103-120; inputs/task/privileged/environment/rfo_task.py:147-161; 待记录该违规的逐步布尔量，当前没有状态日志。" + }, + { + "atom": "dyn.forbidden_contact_after_landing", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:147-152; inputs/task/privileged/environment/rfo_task.py:163-175; inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:180-184; 待记录该违规的逐步布尔量，当前没有状态日志。" + }, + { + "atom": "dyn.simulation_unstable", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:108-109; inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:866-885; 待记录该违规的逐步布尔量，当前没有状态日志。" + } + ] + }, + "native_metrics": [ + { + "name": "air_s", + "unit": "s", + "threshold": 0.1, + "direction": "higher_better", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:383-389", + "inputs/task/privileged/environment/rfo_task.py:40", + "inputs/task/privileged/environment/rfo_task.py:135" + ] + }, + { + "name": "pelvis_rise_m", + "unit": "m", + "threshold": 0.1, + "direction": "higher_better", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:134-140" + ] + }, + { + "name": "pelvis_above_box_m", + "unit": "m", + "threshold": 1.0, + "direction": "higher_better", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:168-170" + ] + } + ], + "open_questions_zh": [ + "inputs/states/index.json 的 logs 为空；依 P8，唯一节点 when、所有约束 never 和 success 均为 null。成功视频可核对跳上箱、蹲伏到站直和站立保持，但毫米级接触高度、法线角度及精确速度须依裁判或新增逐帧日志，不能从画面安全绑定。", + "模式差异：运动判据和实例参数一致，但 privileged 用两次新进程重放均成功且最终状态哈希相同作为交付验收（inputs/grader/efe_embodied-olympics-fun-files/grade.py:105-115），standard 采用最后一次实时回合的裁判结果，重放仅生成视频（inputs/grader/efe_embodied-olympics-fun-files/standard/server.py:6-12,157-178）。这些交付与有效性检查不构造世界状态节点；本文件 success 因无状态日志留空。" + ], + "conf": 6, + "conf_reason_zh": "文本、两模式同源任务代码、初态与成功视频一致，按 S10 保留一个连续跳跃步骤，所有数值来自评分器；缺少状态日志，精确接触、速度、历史阶段与保持窗口尚不能逐帧验证。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "单节点“跳上箱并站直”（本体·跳跃，atomic）与描述、裁判阶段（起跳→落地→站立）及 backflip/pole_vault 的 S10 先例一致；阈值（0.10 s、0.10 m、0.02 m、20°、1.00 m、0.35 m/s、2.0 s）与 rfo_task.py 参数一致；无状态日志，when/never/success 为 null 合规，约束均为立即终止，open_questions_zh 已写明模式差异与无日志。无需修改。", + "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：['box_jump_stood']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:49:46Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__box_jump/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__box_jump/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "c44a30d05d724b3bf97601b77cb608da868b77bb8075bfbd85cb1bbec2c31164", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/embodied-olympics-fun/box_jump/trajectories/embodied-olympics-fun-box-jump-i00-standard--gdqaf9u.json b/phase2/embodied-olympics-fun/box_jump/trajectories/embodied-olympics-fun-box-jump-i00-standard--gdqaf9u.json new file mode 100644 index 0000000000000000000000000000000000000000..240d60603482fa82e769fdf505b270b9580289d0 --- /dev/null +++ b/phase2/embodied-olympics-fun/box_jump/trajectories/embodied-olympics-fun-box-jump-i00-standard--gdqaf9u.json @@ -0,0 +1,201 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-box-jump-i00-standard--gdqaf9u", + "taskKey": "embodied-olympics-fun/box_jump", + "dag_sha256": "94c6447a4b03c43587761e86e851339c4316a0cf5855b294978de845b6362c39", + "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": 354, + "sim_steps": 17514, + "calls": 208, + "rows": 17514, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3696.4, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 1.068, + "peak_processes": 12, + "peak_disk_gb": 2.935, + "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/embodied-olympics-fun-box-jump-i00-standard--gdqaf9u/replay.mp4", + "duration_s": 3.56, + "fps": 25.0, + "frames": 89, + "sha256": "6330fe4f801d088934002d5f0bf4be21696b2710cdf732e3dbb6508645a29d6e" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "box_jump_stood": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察：视频 0.00–2.68 s，机器人在棕色箱前地面走近并保持站立，尚未到箱顶；视频 2.96–3.24 s 下蹲，视频 3.28–3.52 s 蹬伸、离地并开始收腿，末帧脚仍低于箱顶且位于箱前缘旁，未出现箱顶双脚支撑和站直保持。" + }, + { + "t": 3.52, + "log_step": 574, + "text_zh": "观察：log step 574 明确执行最后的第 51 回合，机器人服务返回 failed、outcome=\"landed on the side of the box\"、stages_reached=1/3、仿真 3.54 s；log step 580 确认该回合已结束、control_steps=354。该回合先走近、站定再起跳的顺序与视频一致，合规起跳阶段通过，但合规落箱与站立未完成。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：视频 0.00–3.40 s，机器人以左右脚接触地面完成走近、站定、下蹲及蹬地起跳；未建立箱顶双脚支撑。" + }, + { + "who": "腿/躯干", + "how": "其他", + "note_zh": "推断：候选“非预期侧碰”。视频 3.48–3.52 s，机器人收腿时脚位于箱前侧且低于箱顶，结合 log step 574 裁判报告 landed on the side of the box，可判定脚与箱侧面发生接触。该碰触不是承重支撑、推扫或有意击打，现有方式无准确对应；不能确定先触箱的是哪只脚。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "air_s", + "unit": "s", + "threshold": 0.1, + "direction": "higher_better", + "value": null + }, + { + "name": "pelvis_rise_m", + "unit": "m", + "threshold": 0.1, + "direction": "higher_better", + "value": null + }, + { + "name": "pelvis_above_box_m", + "unit": "m", + "threshold": 1.0, + "direction": "higher_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "box_jump_stood", + "t": 3.52, + "log_step": 574, + "where": "操作中", + "evidence_zh": "观察：视频 3.48–3.52 s，机器人正向箱子起跳收腿，脚仍在箱顶以下、贴近箱前缘；log step 574 的最后第 51 回合返回“landed on the side of the box”，只通过起跳阶段即终止。机器人未完成跳上箱并站直，此失败在该评分回合内没有修复。" + }, + "unmet": { + "box_jump_stood": false + }, + "phenomena": [ + { + "cat": "运动", + "item": "被挡/碰撞", + "t": 3.52, + "log_step": 574, + "evidence_zh": "观察：视频 3.48–3.52 s，机器人脚在接近箱体时仍低于顶面，没有越过箱前缘形成箱顶落脚；log step 574 中该回合裁判明确报告脚落在箱侧面并结束失败。脚与箱侧面的错误接触阻断了合规落箱。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "推断：log step 573 的最终脚本预先按时间生成髋、膝、踝目标，并将整段 A_jump 一次提交给 joints 模式；其目标是蹬地、收腿、箱顶落脚和站直。视频 3.48–3.52 s 的实际抬脚未能在接近箱前缘时获得箱顶净空，log step 574 也确认侧面落脚，说明这段腿部运动控制未实现所需清障与落脚。现有证据不足以把问题细分为目标轨迹不当、关节跟踪误差或具体参数错误。", + "source_zh": "日志 / 视频" + } + ], + "gap_zh": "无状态日志；视频只有场地侧视，末帧为视频 3.52 s，裁判在仿真 3.54 s 终止，无法逐控制步核实先触箱的脚、接触高度和法线，也无法区分目标轨迹与实际跟踪误差。评分器及最后回合输出未报告 air_s、pelvis_rise_m、pelvis_above_box_m 的数值，原生指标留 null。" + }, + "summary_zh": "机器人在最后第 51 回合完成走近、下蹲和起跳，但未完成“跳上箱并站直”，节点最终 0/1。视频 3.48–3.52 s 显示脚尚低于箱顶，log step 574 的该回合裁判报告落在箱侧面，仅通过起跳阶段；失败归于跳跃控制未实现清障与落脚。缺少状态日志，无法确认精确接触几何及控制误差的具体来源。", + "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 box_jump_stood、被挡/碰撞与控制根因均与视频末段（脚低于箱顶、收腿贴箱前缘）及 log step 574 裁判“landed on the side of the box”一致；时钟和证据标注无误。", + "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": "2026-10-29T10:07:28Z", + "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/embodied-olympics-fun-box-jump-i00-standard--gdqaf9u/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-box-jump-i00-standard--gdqaf9u/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "18c9a5fad4935832649a9c33ab806d60f3b285424ea9a43d8dda9f0a7809cf52", + "video_sha256": "6330fe4f801d088934002d5f0bf4be21696b2710cdf732e3dbb6508645a29d6e", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/box_jump/trajectories/embodied-olympics-fun-box-jump-i00-standard--o6qchzm.json b/phase2/embodied-olympics-fun/box_jump/trajectories/embodied-olympics-fun-box-jump-i00-standard--o6qchzm.json new file mode 100644 index 0000000000000000000000000000000000000000..7062b4b4d8ed95ba35e7c51596733d60c7c97c22 --- /dev/null +++ b/phase2/embodied-olympics-fun/box_jump/trajectories/embodied-olympics-fun-box-jump-i00-standard--o6qchzm.json @@ -0,0 +1,189 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-box-jump-i00-standard--o6qchzm", + "taskKey": "embodied-olympics-fun/box_jump", + "dag_sha256": "94c6447a4b03c43587761e86e851339c4316a0cf5855b294978de845b6362c39", + "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": 509, + "sim_steps": 1374, + "calls": 1337, + "rows": 1374, + "resets": 5, + "reset_budget": 5, + "sim_startups": 1, + "elapsed_s": 3925.9, + "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": 0.0, + "peak_memory_gb": 7.697, + "peak_processes": 19, + "peak_disk_gb": 2.012, + "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/embodied-olympics-fun-box-jump-i00-standard--o6qchzm/replay.mp4", + "duration_s": 5.12, + "fps": 25.0, + "frames": 128, + "sha256": "5d4cb8df8131d32379f0dd6008ee26c1afae1901c6dfc89c7b09e95d5279fe4d" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "box_jump_stood": { + "method": "visual", + "intervals": [ + [ + 5.09, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察：视频 0.00 s 机器人双脚站在棕色箱前地面，尚未跳上箱；视频 0.32–0.68 s 机器人屈髋屈膝下蹲后伸腿蹬地。" + }, + { + "t": 0.84, + "log_step": null, + "text_zh": "观察：逐帧检查视频 0.68–1.28 s；视频 0.84–1.00 s 机器人双脚均离开地面，向前上方腾空并收腿，视频约 1.04–1.12 s 脚落到箱顶，随后双脚在顶面承重。画面未见手臂或躯干触箱。" + }, + { + "t": 4.0, + "log_step": null, + "text_zh": "观察：视频 1.28–4.32 s 机器人双脚留在箱顶，由屈膝姿态逐渐伸腿抬高骨盆、挺直躯干；视频 4.40–5.08 s 的逐帧画面持续显示箱顶双脚支撑和近直立姿态，未见跌落。" + }, + { + "t": 5.08, + "log_step": 131, + "text_zh": "观察：log step 131 的机器人实时服务返回 episode=6、control_steps=509、sim_time_s=5.09、done=true、reason=success、stages_reached=3、stages_total=3；inputs/run.json 的评分器字段 live_success=1、graded_by_live=1、finalized=1，表明该最后回合是评分回合。视频 5.08 s 机器人仍直立在箱顶。" + }, + { + "t": 5.09, + "log_step": 131, + "text_zh": "推断：log step 3 的服务协议给出 control_dt=0.01 s，log step 131 最后评分回合的 509 个控制步对应视频约 509×0.01=5.09 s，与视频总长 5.12 s 相符；同一脚本输出的落地时刻仿真 1.15 s 已与视频 1.04–1.20 s 的箱顶落地画面交叉核对。故以视频约 5.09 s 为整个合规跳跃及连续站立窗口通过的时刻（帧间精度约 0.04 s），不是刚触箱时即算完成；精确接触高度、法线、骨盆速度和连续计数依据裁判成功判定。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：机器人左右脚先在地面承重并蹬地起跳，腾空后以双脚落在棕色箱顶，持续支撑身体并逐渐伸腿站直；见视频 0.00–1.28 s 与视频 2.56–5.08 s。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "air_s", + "unit": "s", + "threshold": 0.1, + "direction": "higher_better", + "value": null + }, + { + "name": "pelvis_rise_m", + "unit": "m", + "threshold": 0.1, + "direction": "higher_better", + "value": null + }, + { + "name": "pelvis_above_box_m", + "unit": "m", + "threshold": 1.0, + "direction": "higher_better", + "value": null + } + ], + "failure": null, + "summary_zh": "评分采用最后第 6 次实时回合：机器人双脚从地面蹬起、腾空后落在箱顶，再逐渐站直，唯一节点“跳上箱并站直”于视频约 5.09 s 完成，终态为 1/1。log step 131 的服务裁判返回三阶段全部通过且成功，与视频末尾的箱顶站立一致；会话前五次测试不属于这段评分视频。没有状态日志，毫米级脚箱接触位置、法线、精确速度与连续站立计数无法从侧视画面逐项测量，依本次裁判成功结果确认，三个原生指标未报告数值。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "单节点“跳上箱并站直”于视频约 5.09 s 达成、终态 1/1，与帧和 log 的服务读数（control_steps=509、stages_reached=3）一致；contact_actual（腿/躯干·支撑）与观察一致；成功运行 failure 为 null 合规，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": "2026-10-29T10:07:47Z", + "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/embodied-olympics-fun-box-jump-i00-standard--o6qchzm/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-box-jump-i00-standard--o6qchzm/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "7b1324dc1477ae3ee1d56dd6f22d3a7eaf96a98ad044a67b53a7ae7b53ddb748", + "video_sha256": "5d4cb8df8131d32379f0dd6008ee26c1afae1901c6dfc89c7b09e95d5279fe4d", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/boxing/task_dag.json b/phase2/embodied-olympics-fun/boxing/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..aea4f764624c1acbf1afa9cbdea9223cd93caa93 --- /dev/null +++ b/phase2/embodied-olympics-fun/boxing/task_dag.json @@ -0,0 +1,295 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "embodied-olympics-fun/boxing", + "suite": "embodied-olympics-fun", + "base": "embodied-olympics-fun-boxing-i00", + "title_zh": "击袋后护架站稳", + "summary_zh": "右拳拳峰在白色标记带完成一次冲量至少 12.0 N·s、接触不超过 0.10 s 的击打，随后收拳至右肩附近并站稳连续 2.0 s，期间无违规触袋或跌倒。", + "embodiment_zh": "两种模式均为双足 Unitree H1-2 人形机器人，27 个关节，双臂各七关节；末端是无可动手指的固定拳垫，非可抓握夹爪或灵巧手，右臂出拳，双腿和躯干支撑平衡，无轮式底盘。", + "task_class": "compositional", + "analysis_level": "step", + "spec": { + "version": "1.2", + "task": "embodied-olympics-fun-boxing-i00", + "bench": "embodied-olympics-fun", + "progress": "live", + "entities": {}, + "milestones": [ + { + "id": "bag_struck", + "zh": "击中沙袋", + "mode": "event", + "when": null, + "after": [], + "hold_s": 0, + "goal": true, + "criterion_zh": "场馆侧视可见右拳迅速打中红色沙袋的白色环带，沙袋随即摆离、拳袋分离至少 0.03 s 后才完成此步：首次接触须在带内，接触历时不超过 0.10 s，按挂钩角动量变化计算的有效冲量至少 12.0 N·s。仅伸手靠近、贴袋推挤或轻碰均未完成；摆动只是视觉线索，精确冲量须看裁判的 strike 事件及 impulse_n_s，指定拳峰和禁用其他部位见约束。", + "step": { + "cat": "动态", + "item": "击打" + }, + "contact": [ + { + "who": "手指/夹爪", + "how": "其他", + "plan": 1, + "note_zh": "右手末端固定拳垫的拳峰击打沙袋白带；候选接触名为“击打”，现有怎么碰条目没有短促冲击，不能用抓握或推/扫替代。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "plan": 1, + "note_zh": "左右脚与地面接触支承机器人出拳，腿和躯干维持站立。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "task/privileged/instruction.md:4", + "task/privileged/environment/rfo_task.py:94-98", + "task/privileged/environment/rfo_task.py:147-190", + "grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159" + ], + "conf": 6, + "note_zh": "一个有效击打结果；击中、接触结束和冲量结算是一击的验收，不拆节点。event 保留已完成的一击，后续再次触袋由约束判失败。右拳拳峰是文本和评分器共同限定的手段，不另设准备或握拳步骤。" + }, + { + "id": "guard_held", + "zh": "护架站稳", + "mode": "final", + "when": null, + "after": [ + "bag_struck" + ], + "hold_s": 1.99, + "drop_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "有效击打后，场馆侧视可见右肘弯曲、右拳从袋前收回右肩附近，身体直立、仅脚接地且基本不平移；右掌至右肩关节距离须不超过 0.35 m，躯干竖直方向余弦至少 0.94，骨盆三维速度不超过 0.35 m/s，连续满足 200 个 0.01 s 控制步才完成。仍向袋伸拳、明显倾斜或移动、护架保持不足，或中途失去任一条件，都不能完成，连续计时须重新开始；全程还须满足禁再触袋及不跌倒约束。", + "step": { + "cat": "本体", + "item": "保持平衡" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "plan": 1, + "note_zh": "左右脚支撑地面，腿和躯干使机器人收回右拳后直立站稳；右手收至右肩附近且不接触沙袋。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "task/privileged/instruction.md:27-31", + "task/privileged/environment/rfo_task.py:192-199", + "grader/efe_embodied-olympics-fun-files/rfo/referee.py:180-184", + "grader/efe_embodied-olympics-fun-files/rfo/referee.py:292-301", + "grader/efe_embodied-olympics-fun-files/rfo/world.py:1103-1105" + ], + "conf": 7, + "note_zh": "文本独立要求击打后的护架站立状态，击打完成与收拳保持之间可以停顿，故为第二步；并非机器人回到开局关节角。final 从有效击打后才验收，不在开局计分；200 步按 D31 对应首末采样跨度 1.99 s，criterion 仍以原生 2.0 s 计数为准。drop_s=0 对应 ref.hold 任一控制步不满足即清零。" + } + ], + "constraints": [ + { + "id": "wrong_bag_contact", + "zh": "违规部位触袋", + "never": null, + "criterion_zh": "机器人任何非右拳拳垫部位触及沙袋，或右拳接触点沿前臂距腕安装点不足 0.07 m（不在拳垫末端 0.03 m 的拳峰区），即违规；录像中左手、前臂或躯干碰袋也属此项。", + "consequence": "立即终止", + "sources": [ + "task/privileged/environment/rfo_task.py:138-164" + ] + }, + { + "id": "hit_outside_band", + "zh": "击中带外", + "never": null, + "criterion_zh": "首次击袋接触点的平均位置不在白色环带内即违规；代码用接触点在沙袋自身轴向距挂钩的深度判断，合法深度为 0.65 至 1.05 m（静止沙袋对应离地 1.20 至 1.60 m）。", + "consequence": "立即终止", + "sources": [ + "task/privileged/environment/rfo_task.py:96-98", + "task/privileged/environment/rfo_task.py:165-171", + "task/privileged/environment/rfo_task.py:50-53" + ] + }, + { + "id": "bag_pushed", + "zh": "推袋超时", + "never": null, + "criterion_zh": "首次击袋尚未结束时，从第一次接触起经过超过 0.10 s 仍与袋接触，即判为推袋；持续贴住白带向前推而不迅速分离即违规。", + "consequence": "立即终止", + "sources": [ + "task/privileged/environment/rfo_task.py:177-179" + ] + }, + { + "id": "punch_too_weak", + "zh": "击打过弱", + "never": null, + "criterion_zh": "拳袋最后接触后连续分离至少 0.03 s 时结算：沙袋绕挂钩角动量变化的水平分量模长除以此次接触的最大轴向深度，所得冲量小于 12.0 N·s 即失败；沙袋轻摆不能证明冲量达标。", + "consequence": "立即终止", + "sources": [ + "task/privileged/environment/rfo_task.py:173-188" + ] + }, + { + "id": "bag_touched_again", + "zh": "再次触袋", + "never": null, + "criterion_zh": "首次击打结算后，机器人任何部位再次触及沙袋即失败，包含沙袋摆回后撞到仍伸出的手；结算前不足 0.03 s 的接触间断仍属于同一次击打。", + "consequence": "立即终止", + "sources": [ + "task/privileged/environment/rfo_task.py:149-153", + "task/privileged/environment/rfo_task.py:180-190" + ] + }, + { + "id": "robot_fallen", + "zh": "机器人跌倒", + "never": null, + "criterion_zh": "骨盆离地低于 0.50 m，或除左右脚以外的任一机器人部位接触地面，即跌倒并终止；录像中膝、手、躯干落地均属违规。", + "consequence": "立即终止", + "sources": [ + "task/privileged/environment/rfo_task.py:95-98", + "grader/efe_embodied-olympics-fun-files/rfo/referee.py:272-289", + "grader/efe_embodied-olympics-fun-files/rfo/task.py:21-23" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "bag_struck", + "level": "new_code", + "source": "task/privileged/environment/rfo_task.py:180-190; 需记录 ref.stages 中 strike、trackers 的接触起止与深度以及 ref.metrics.impulse_n_s；现无状态日志。" + }, + { + "atom": "guard_held", + "level": "new_code", + "source": "task/privileged/environment/rfo_task.py:196-199; grader/efe_embodied-olympics-fun-files/rfo/referee.py:180-184,292-301; 需记录 ref._hold.guard、右掌至右肩距离、standing 及组成条件；现无状态日志。" + }, + { + "atom": "wrong_bag_contact", + "level": "new_code", + "source": "task/privileged/environment/rfo_task.py:138-164; 需记录对应接触、冲量或跌倒判据及 ref.failed；现无状态日志。" + }, + { + "atom": "hit_outside_band", + "level": "new_code", + "source": "task/privileged/environment/rfo_task.py:96-98; task/privileged/environment/rfo_task.py:165-171; task/privileged/environment/rfo_task.py:50-53; 需记录对应接触、冲量或跌倒判据及 ref.failed；现无状态日志。" + }, + { + "atom": "bag_pushed", + "level": "new_code", + "source": "task/privileged/environment/rfo_task.py:177-179; 需记录对应接触、冲量或跌倒判据及 ref.failed；现无状态日志。" + }, + { + "atom": "punch_too_weak", + "level": "new_code", + "source": "task/privileged/environment/rfo_task.py:173-188; 需记录对应接触、冲量或跌倒判据及 ref.failed；现无状态日志。" + }, + { + "atom": "bag_touched_again", + "level": "new_code", + "source": "task/privileged/environment/rfo_task.py:149-153; task/privileged/environment/rfo_task.py:180-190; 需记录对应接触、冲量或跌倒判据及 ref.failed；现无状态日志。" + }, + { + "atom": "robot_fallen", + "level": "new_code", + "source": "task/privileged/environment/rfo_task.py:95-98; grader/efe_embodied-olympics-fun-files/rfo/referee.py:272-289; grader/efe_embodied-olympics-fun-files/rfo/task.py:21-23; 需记录对应接触、冲量或跌倒判据及 ref.failed；现无状态日志。" + } + ] + }, + "native_metrics": [ + { + "name": "impulse_n_s", + "unit": "N*s", + "threshold": 12.0, + "direction": "higher_better" + }, + { + "name": "contact_s", + "unit": "s", + "threshold": 0.1, + "direction": "lower_better" + }, + { + "name": "hit_height_m", + "unit": "m", + "threshold": [ + 1.2, + 1.6 + ], + "direction": "target" + } + ], + "open_questions_zh": [ + "inputs/states/index.json 的 logs 为空，成功录像亦无状态日志；按 P8 所有 when、never 和 success 均为 null，不能执行状态条件与评分器的一致性验证，所列 new_code 仅说明需要补记的状态，未虚构已有字段。录像可判击中、摆动和护架，不能单凭摆幅精确验证 12.0 N·s 或拳垫末端 0.03 m。", + "按 S10 将有效击袋和护架站稳分为两步：击打结算后可暂停再收拳；护架不是击打在途的一帧，也不是开局站立的评估日程。护架用 final 防止开局提前计分。", + "任务设计问题：评分器强制只许右拳拳峰触袋，其他部位即使产生同样冲量也立即失败（rfo_task.py:155-164）；文本亦明确限定该手段。此限制放入违规触袋约束，不增加手段节点，success 因无状态日志为 null。", + "contact_impulse_n_s 是法向接触力积分的记录值，不是通过阈值；实际判断使用沙袋挂钩角动量变化计算的 impulse_n_s（rfo_task.py:173-190），故不把接触力积分设为约束或目标。hit_height_m 只报告首次接触点平均世界高度；带内判定实际使用沙袋局部轴向深度，静止时对应 1.20 至 1.60 m。", + "两种模式的实例与任务模块逐字一致，语义判据共用同一裁判；privileged 外层还要求两次新进程重放成功且终态哈希相同（grader/efe_embodied-olympics-fun-files/grade.py:106-115），standard 接受最后一次现场裁判结果（grader/efe_standard-files/v2_1/grade.py:93-100）。重放确定性、文件有效性和仿真数值发散不作为世界状态步骤；15 s 为回合时限而非约束。" + ], + "conf": 6, + "conf_reason_zh": "两步均直接来自文本，顺序、阈值、连续保持和终止条件都有源码依据，两模式同规则；成功录像与奖励记录相符，但没有状态日志，精确冲量及接触点只能借助裁判记录而不能从录像独立测定。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "文本步骤为“右拳拳峰击打白带≥12 N·s、接触≤0.1 s”→“收拳护架站稳2 s”，两步与DAG一致；护架判据（0.35 m、20°、0.35 m/s、2.0 s）与文本一致，final+after击打可避免开局计分；约束对应评分器全部失败条件且均为立即终止；无状态日志故when/success为null并已在open_questions_zh说明。局限：未逐行核对评分器与录像帧，仅依据说明文本与DAG自洽性审查。", + "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：['bag_struck', 'guard_held']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:47:56Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__boxing/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__boxing/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "9cbefe669fd5177b532a04291683bdd79e35e6072067cbcd5ee27c64a52e1537", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/embodied-olympics-fun/boxing/trajectories/embodied-olympics-fun-boxing-i00-privileged--os2vhh2.json b/phase2/embodied-olympics-fun/boxing/trajectories/embodied-olympics-fun-boxing-i00-privileged--os2vhh2.json new file mode 100644 index 0000000000000000000000000000000000000000..c272ff85f6f2ce5c0d1c5807a9fd9bdb0fc7df0f --- /dev/null +++ b/phase2/embodied-olympics-fun/boxing/trajectories/embodied-olympics-fun-boxing-i00-privileged--os2vhh2.json @@ -0,0 +1,222 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-boxing-i00-privileged--os2vhh2", + "taskKey": "embodied-olympics-fun/boxing", + "dag_sha256": "5fe95f4fd88e6dbc49fa1e5ba7230e5fd61571eeb504fd1c0835f0ca7840e98d", + "outcome": { + "success": 1, + "reward": 1.0, + "source": "catalog", + "raw": { + "success": 1, + "replay_success": 1, + "deterministic": 1, + "n_actions": 332, + "sim_steps": 332, + "horizon": 1500, + "episode_over": 1, + "progress": 1.0, + "stages_reached": 2, + "stages_total": 2, + "fell": 0, + "replay_s": 3.68, + "bad_trajectory": 0, + "metric_contact_impulse_n_s": 14.26, + "metric_contact_s": 0.091, + "metric_hit_height_m": 1.273, + "metric_impulse_n_s": 14.78, + "grader_error": 0, + "video_rendered": 1, + "agent_sessions": 97, + "agent_resets": 15, + "agent_replays": 12, + "agent_cost_self_reported": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 1.103, + "peak_processes": 12, + "peak_disk_gb": 0.24, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0, + "session_log_tampered": 0, + "session_log_from_monitor": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-boxing-i00-privileged--os2vhh2/replay.mp4", + "duration_s": 3.36, + "fps": 25.0, + "frames": 84, + "sha256": "cf2dfbe8fccbc51f067f9b0011e0732294430d769176085267c827269b00e8a1" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 2, + "final": 2, + "nodes": { + "bag_struck": { + "method": "visual", + "intervals": [ + [ + 1.0, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.88, + "log_step": null, + "text_zh": "观察：视频 0.80–0.92 s，机器人右臂向前出拳，右手末端在红色沙袋白色环带下部接触袋面；视频 1.00–1.24 s 沙袋逐渐摆离，拳袋间距增大，机器人双脚仍支撑地面。" + }, + { + "t": 1.0, + "log_step": 154, + "text_zh": "观察：inputs/run.json 的评分器字段为 metric_impulse_n_s=14.78、metric_contact_s=0.091、metric_hit_height_m=1.273，stages_reached=2、replay_success=1；log step 154 对交付文件的两次重放均由裁判返回 strike、guard 两阶段及相同指标。" + }, + { + "t": 1.0, + "log_step": null, + "text_zh": "推断：结合视频中短促接触、沙袋摆离及评分器通过结果，将有效击打结算锚定在视频约 1.00 s，时间约有 0.1 s 误差；录像不足以独立测出冲量、精确接触历时及拳峰末端区域，精确验收依赖评分器字段。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "其他", + "note_zh": "观察：视频 0.80–0.92 s，机器人右手末端固定拳垫击打沙袋白带；候选接触名为“击打”。现有怎么碰条目没有短促冲击，故用其他。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：视频 0.00–1.24 s，机器人左右脚接地支撑出拳，双腿与躯干维持站立。" + } + ] + }, + "guard_held": { + "method": "visual", + "intervals": [ + [ + 3.32, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 1.44, + "log_step": null, + "text_zh": "观察：视频 1.16–1.44 s，机器人右肘弯曲、右拳从袋前收至肩侧；视频 1.44–3.32 s，机器人保持右拳靠近肩部的姿态，双脚接地，身体基本留在原位。已逐帧检查视频 2.36–3.32 s，未见跌倒或再次触袋。" + }, + { + "t": 3.32, + "log_step": 154, + "text_zh": "观察：log step 154 是交付轨迹文件的裁判重放输出，两次均记录 steps=332、time_s=3.32、stages=[strike, guard]、success=true，终态哈希相同；最终评分器同时报告 deterministic=1、stages_reached=2、fell=0。" + }, + { + "t": 3.32, + "log_step": 154, + "text_zh": "推断：护架节点要求连续保持达标，不能从开始收拳时计为完成；以已检查的末帧视频 3.32 s 为完成锚点，结合 log step 154 交付文件在第 332 个控制步完成两阶段的结果，判为最终完成。此处控制步为仿真控制计数而非 agent 时间，332×0.01 s=3.32 s，与视频 3.32 s 的站立护架末帧核对；连续计数及速度、躯干方向余弦阈值依赖裁判通过结果，完成时刻约有一帧误差。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：视频 1.16–3.32 s，机器人左右脚在地面支撑身体完成收拳及站稳，右拳收至肩侧后未再触袋。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "impulse_n_s", + "unit": "N*s", + "threshold": 12.0, + "direction": "higher_better", + "value": 14.78 + }, + { + "name": "contact_s", + "unit": "s", + "threshold": 0.1, + "direction": "lower_better", + "value": 0.091 + }, + { + "name": "hit_height_m", + "unit": "m", + "threshold": [ + 1.2, + 1.6 + ], + "direction": "target", + "value": 1.273 + } + ], + "failure": null, + "summary_zh": "观察：本次交付重放完成击中沙袋和护架站稳两节点，最终为 2/2；视频约 0.80–0.92 s 右拳触及白带，随后沙袋摆离、机器人收拳并直立至视频 3.32 s。观察：评分器记录有效冲量 14.78 N·s、接触时长 0.091 s、击中高度 1.273 m，两阶段均通过，交付文件两次重放成功且终态哈希一致（log step 154）。推断：击打完成约在视频 1.00 s、护架连续保持完成约在视频 3.32 s；无状态日志，精确拳峰接触区域、连续保持计数及数值阈值由裁判结果补足，时间为帧级近似。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "成功运行，两节点均达成（击打约视频 1.0 s，护架约视频 3.32 s），与帧及评分器字段一致；contact_actual 合理（击打用其他，脚支撑），无失败标签，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": "2026-10-29T10:04:33Z", + "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/embodied-olympics-fun-boxing-i00-privileged--os2vhh2/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-boxing-i00-privileged--os2vhh2/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "2d13ecb4ad05033f1bf9319205a091d770f3fc7117ef0f1a530e4a4eb7cff1dd", + "video_sha256": "cf2dfbe8fccbc51f067f9b0011e0732294430d769176085267c827269b00e8a1", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/boxing/trajectories/embodied-olympics-fun-boxing-i00-standard--cyq2qzc.json b/phase2/embodied-olympics-fun/boxing/trajectories/embodied-olympics-fun-boxing-i00-standard--cyq2qzc.json new file mode 100644 index 0000000000000000000000000000000000000000..abb72f4e9852b0f27974af4426ac8d9093c1b54f --- /dev/null +++ b/phase2/embodied-olympics-fun/boxing/trajectories/embodied-olympics-fun-boxing-i00-standard--cyq2qzc.json @@ -0,0 +1,231 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-boxing-i00-standard--cyq2qzc", + "taskKey": "embodied-olympics-fun/boxing", + "dag_sha256": "5fe95f4fd88e6dbc49fa1e5ba7230e5fd61571eeb504fd1c0835f0ca7840e98d", + "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": 158, + "sim_steps": 1127, + "calls": 59, + "rows": 1127, + "resets": 5, + "reset_budget": 5, + "sim_startups": 1, + "elapsed_s": 824.6, + "video_rendered": 1, + "first_success": 0, + "first_reason_code": 3, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.326, + "peak_processes": 9, + "peak_disk_gb": 0.837, + "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/embodied-olympics-fun-boxing-i00-standard--cyq2qzc/replay.mp4", + "duration_s": 1.6, + "fps": 25.0, + "frames": 40, + "sha256": "940724b6216597751b83774bc50d4ec75c853ae563acc1509ee5ea00be21a0a4" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "bag_struck": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.56, + "log_step": 97, + "text_zh": "观察：视频 0.00–0.52 s 机器人双脚支撑、右前臂朝向袋前；视频 0.56–1.04 s 右手向下收回。log step 95 重置后，log step 96–97 的脚本先保持零肩肘目标，再令右肘目标为 0.9 rad，此动作顺序与录像一致，所标为最后一回合，不包含早先重置前的试验。" + }, + { + "t": 1.12, + "log_step": 100, + "text_zh": "观察：视频 1.08–1.20 s 机器人抬起右臂，右手末端在侧视投影中抵近白带下方的红色袋身；视频 1.24–1.56 s 右手仍处袋前低位，没有出现白带处短促击中后沙袋明显摆离、右拳撤回的完整过程。" + }, + { + "t": 1.56, + "log_step": 106, + "text_zh": "推断：本回合未完成有效击袋。视频 0.00–1.56 s 未见白带上的有效击打过程；agent 在 log step 104、106 自述裁判判为非右拳拳峰触袋且阶段为 0/2，与所见低位抬手过程一致；run.json 的 live_success=0、graded_by_live=1 支持最终失败，但不据零分单独推断节点未达。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "其他", + "note_zh": "推断：视频 1.08–1.56 s 右手末端抬向沙袋并在红色袋身边缘发生误触，结合 agent 在 log step 104、106 所述非拳峰触袋判罚；精确接触点无法从侧视分辨。候选接触名“误触”，未见抓握、持续推扫或压擦，也未形成有效的拳峰击打。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：视频 0.00–1.56 s 机器人以双脚接地支撑抬臂和收肘，身体保持站立，未见跌倒。" + } + ] + }, + "guard_held": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.56, + "log_step": 106, + "text_zh": "观察：视频 0.00–1.56 s 只有击袋准备中的收肘、抬臂，末帧右手仍在袋前，未见有效击打后将右拳收至右肩并连续保持 2.0 s 的阶段；整个录像时长也只有 1.60 s。" + }, + { + "t": 1.56, + "log_step": 103, + "text_zh": "推断：护架站稳未达，且尚未开始该后续步骤；前置有效击打未达，log step 95–103 的最后一回合动作只到抬肩准备，随后 agent 在 log step 104 报告该回合已判负。开局站立不计为击打后的护架。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "impulse_n_s", + "unit": "N*s", + "threshold": 12.0, + "direction": "higher_better", + "value": null + }, + { + "name": "contact_s", + "unit": "s", + "threshold": 0.1, + "direction": "lower_better", + "value": null + }, + { + "name": "hit_height_m", + "unit": "m", + "threshold": [ + 1.2, + 1.6 + ], + "direction": "target", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "bag_struck", + "t": 1.5, + "log_step": 103, + "where": "操作中", + "evidence_zh": "观察：视频 1.08–1.56 s 机器人在击袋准备中抬起右臂，右手位于白带下方的袋身前，录像止于袋前手位，未完成有效击打。log step 103 的脚本在右肘目标 0.9 rad 下把右肩目标继续调到 -0.8 rad；agent 在 log step 104 自述此次抬肩触发非拳峰触袋判罚，后续没有重置或修复。视频约 1.5 s 为末次抬肩失败的粗略导航锚点，非精确碰撞时刻。" + }, + "unmet": { + "bag_struck": false, + "guard_held": false + }, + "phenomena": [ + { + "cat": "运动", + "item": "位姿不准", + "t": 1.5, + "log_step": 103, + "evidence_zh": "观察：视频 1.08–1.56 s 机器人右手抬起后落在白带下方红色袋身前，末端未对准白带完成规定的拳峰击打；没有随后收拳护架的过程。agent 在 log step 104、106 自述裁判判为非右拳拳峰触袋，该自述作为补充，不能单独确认具体触碰部位。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "t": 1.5, + "log_step": 103, + "source_zh": "日志", + "evidence_zh": "推断：击袋准备时的右臂运动控制未把手端引至合法击打位姿。log step 97、100、103 的脚本先设右肘目标 0.9 rad，再将右肩目标依次设为 -0.45、-0.8 rad；视频 0.56–1.56 s 与收肘后向袋身低位抬手的顺序相符，未执行出对准白带的有效拳峰击打。log step 104 的 agent 自述称回合在继续抬肩时判负。缺少末回合原始关节反馈和接触记录，不细分为参数、跟踪或接触设计问题。" + } + ], + "gap_zh": "无状态日志；侧视可以判右臂运动及未完成击打后护架，但不能精确区分右手拳峰与其他部位的接触点，也不能测定冲量与接触历时。log step 95–104 的服务调用原始输出为空，末回合“非右拳拳峰触袋、0/2”的具体判词仅见 agent 自述及其写入笔记的文本，不能当作直接服务读数；log step 89、93 的原始带外判罚属于更早回合，未用于本回合判罚。评分字段未报告 impulse_n_s、contact_s、hit_height_m，故原生指标均置 null。" + }, + "summary_zh": "本次评分录像对应最后一次重置后的收肘、抬肩准备过程，未完成有效击袋，护架站稳也未开始，终态为 0/2。机器人在视频约 1.5 s 仍以低位右手接近袋身，断在“击中沙袋”；动作控制未达合法击打位姿，agent 自述最终因非拳峰触袋被判负。缺少末回合原始接触读数和冲量指标，不能独立核实精确违规部位或进一步细分控制根因。", + "conf": 6, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "视频仅 1.6 s，为最后一回合；两节点均未达、断点 bag_struck/操作中、位姿不准 + 控制根因都与可见画面和 log 一致，未发现需改内容的错误。末回合判词仅为 agent 自述，标注已如实说明。", + "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": "2026-10-29T10:05:18Z", + "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/embodied-olympics-fun-boxing-i00-standard--cyq2qzc/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-boxing-i00-standard--cyq2qzc/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "7a100bd2647b2ea8176c561a8304a546e0d183598c7eed116899439b1fbeb7a3", + "video_sha256": "940724b6216597751b83774bc50d4ec75c853ae563acc1509ee5ea00be21a0a4", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/catch/task_dag.json b/phase2/embodied-olympics-fun/catch/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..ec5360f97cd64d7dd6247f3e5dd01c794960c966 --- /dev/null +++ b/phase2/embodied-olympics-fun/catch/task_dag.json @@ -0,0 +1,278 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "embodied-olympics-fun/catch", + "suite": "embodied-olympics-fun", + "base": "embodied-olympics-fun-catch-i00", + "title_zh": "单手接稳飞球", + "summary_zh": "在飞球过顶且未落地或碰到禁触身体部位前用任一只裸手包住棒球，并由该手单独控制满规定的一秒窗口，全程不摔倒。", + "embodiment_zh": "三个模式均为Unitree H1-2人形机器人，27个本体关节、双臂，各臂末端为Inspire五指灵巧手（每手六个驱动、单一闭合指令），双足移动，无轮式底盘；任一手可接球，无手套或承接工具。", + "task_class": "atomic", + "analysis_level": "step", + "spec": { + "version": "1.2", + "task": "embodied-olympics-fun-catch-i00", + "bench": "embodied-olympics-fun", + "progress": "live", + "entities": {}, + "milestones": [ + { + "id": "fly_ball_caught", + "zh": "接稳飞球", + "mode": "state", + "when": null, + "after": [], + "hold_s": 0, + "drop_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "场地远景中白色小球从空中下降到伸出的裸手，并留在该手随手移动；须已过最高点，在首次触身后0.5秒内由任一只手的四指组与拇指或掌部组同时从相对两侧接触球（两组平均接触方向夹角至少90°），随后同手包住球满100个10毫秒控制步，包围接触每次中断不超过0.02秒，期间球不接触该手以外的物体且机器人不摔倒。仅碰到球、球躺在张开的掌上或持球未满窗口均未完成，球脱手超过0.02秒、转到另一只手或碰到身体和地面即失去；远景不能精确分辨接触夹角和毫秒中断，需原生裁判记录佐证。", + "step": { + "cat": "动态", + "item": "接住" + }, + "contact": [ + { + "who": "手指/夹爪", + "how": "抓握", + "note_zh": "左手或右手的四指与拇指或掌部从相对两侧合拢包住棒球；接球手任选，另一手不能在接稳阶段触球。", + "plan": 1 + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "机器人双脚与腿在场地地面支撑身体，移动至棒球落点附近并在同手持球期间保持不摔倒。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:3-27", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch.py:749-770", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch.py:799-807", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch.py:852-882", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch.py:884-947", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-184", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:272-289", + "inputs/task/privileged/environment/instance.json:4-27", + "inputs/oracle/replay.mp4:3.6-4.92s", + "inputs/oracle/run.json:steps[id=57,60,62]" + ], + "conf": 6, + "note_zh": "一个接球步骤，catch和hold是其先后验收阶段；伸手、走到落点、闭手及持球计时不拆为节点。初态站立、双手张开且尚未接球；成功录像使用右手，但左手同样合法。hold_s为0，因为完成判据已包含原生100步保持计数，不重复施加等待；裁判在接住当个控制步开始计数，首末采样相差0.99秒。drop_s为0，对应脱手或禁触违规即终止的规则；freeze_held为false，因为验收要求球被手握住。" + } + ], + "constraints": [ + { + "id": "ball_ground_before_catch", + "zh": "接住前球落地", + "never": null, + "criterion_zh": "球离开发球机后、尚未被任一只手包住时，球接触场地地面即失败。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch.py:894-908" + ] + }, + { + "id": "ball_body_before_catch", + "zh": "接住前球碰身体", + "never": null, + "criterion_zh": "接住前棒球接触双手和双前臂以外的任何机器人部位即失败；不能靠胸口等部位夹停球。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch.py:745", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch.py:869-879", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch.py:909-911" + ] + }, + { + "id": "ball_touched_before_apex", + "zh": "过顶前触球", + "never": null, + "criterion_zh": "发射后、球尚未过最高点时，任意机器人部位接触球即失败；过顶由球竖直速度曾小于零判定。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch.py:901-902", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch.py:912-918" + ] + }, + { + "id": "catch_window_expired", + "zh": "触球后未及时握住", + "never": null, + "criterion_zh": "球首次接触机器人后超过0.5秒仍未被任一只手从相对两侧包住即失败；球随后离开机器人也继续计时。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch.py:805-806", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch.py:912-930" + ] + }, + { + "id": "ball_left_catching_hand", + "zh": "接住后球脱手", + "never": null, + "criterion_zh": "接住后，同一只接球手不再满足包围接触，且距上一次满足已超过0.02秒即失败；再接住不能补救。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch.py:884-889", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch.py:933-936" + ] + }, + { + "id": "ball_touched_elsewhere_after_catch", + "zh": "接住后球碰他物", + "never": null, + "criterion_zh": "接住后，棒球碰到接球手以外的任意物体即失败，包括另一只手、任一前臂、躯干或地面。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch.py:875-879", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch.py:937-940" + ] + }, + { + "id": "robot_fell", + "zh": "机器人摔倒", + "never": null, + "criterion_zh": "成功前骨盆高度低于0.50米，或双脚以外的机器人部位触地，即判摔倒并终止。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch.py:802-806", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch.py:852-855", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/task.py:21-23", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:272-282" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "dyn.catch_hold_complete", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch.py:884-947 与 inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-184；需逐帧导出原生接住标记、接球手、包围接触和容许中断后的hold计数（100个控制步），作为接稳完成条件；目前没有状态日志。" + }, + { + "atom": "dyn.ball_ground_before_catch", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch.py:894-908；需记录球已发射、接住手标记及球与地面的逐物理步接触。" + }, + { + "atom": "dyn.ball_body_before_catch", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch.py:909-911；需记录接住前球与禁触身体部位的接触。" + }, + { + "atom": "dyn.ball_touched_before_apex", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch.py:901-918；需记录下降标记及机器人与球的接触。" + }, + { + "atom": "dyn.catch_window_expired", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch.py:912-930；需记录首次触球时刻、接住手标记和当前物理时间。" + }, + { + "atom": "dyn.ball_left_catching_hand", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch.py:933-936；需记录接球手、上次包围接触时刻及当前物理时间。" + }, + { + "atom": "dyn.ball_touched_elsewhere_after_catch", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch.py:937-940；需记录球与接球手以外几何体的逐物理步接触。" + }, + { + "atom": "dyn.robot_fell", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:276-282；需记录骨盆高度和非足部触地，或原生摔倒标记。" + } + ] + }, + "native_metrics": [ + { + "name": "catch_height_m", + "unit": "m", + "threshold": null, + "direction": null + }, + { + "name": "catch_speed_m_s", + "unit": "m/s", + "threshold": null, + "direction": null + }, + { + "name": "landed_from_start_m", + "unit": "m", + "threshold": null, + "direction": null + } + ], + "open_questions_zh": [ + "inputs/states/index.json中没有状态日志，按DECISIONS.md P8所有节点when、约束never及success均为null；atoms列出需要新增记录的原生世界状态，不能用交付或复现有效性替代。无法进行状态日志的到达率和评分一致性实证核对。", + "录像为640×360、25帧每秒的场地远景，约3.9秒可见小白球到达机器人伸出的右手，此后留在手边至片尾；无法从这些帧精确判断90°接触夹角或0.02秒中断，后续应结合原生接触及裁判计数，不能仅凭手靠近球判完成。", + "模式差异：接球模块、阈值和冻结初态相同；privileged评分要求两个新进程重放均成功且末态哈希一致（inputs/grader/efe_embodied-olympics-fun-files/grade.py:106-115），standard及standard-realtime使用最后一次现场尝试的裁判结论（inputs/grader/efe_embodied-olympics-fun-files/standard/server.py:3-12；inputs/grader/efe_standard-files/v2_1/grade.py:93-99）。这些交付和运行有效性检查不建世界状态节点。", + "文本与评分器的时间离散化：文本说接住后保持1.0秒，评分器每个10毫秒控制步将hold计数加一，100次即通过，首个和最后一个采样间隔为0.99秒（catch.py:942-947；referee.py:180-184）。采用原生计数验收，不额外延后或按录像调阈值。", + "原生catch_height_m、catch_speed_m_s及失败时的landed_from_start_m仅记录数值，没有成功阈值或优化方向（catch.py:905-908、927-929），故native_metrics的threshold和direction均为null；场地所述落点距初位3米内不是另一项得分要求。", + "成功运行记录的早期任务文本写手指10牛、拇指15牛，当前三个模式文本写20牛、30牛；这提示手部模型版本可能不同，现有材料不足以验证全部动力学一致性。该录像仅用于接球状态及视角参考，不据此改变当前规则阈值或限定接球手（inputs/oracle/run.json:steps[id=2]；inputs/task/privileged/instruction.md:18）。" + ], + "conf": 6, + "conf_reason_zh": "单一接球步骤、接触类型、初态及全部失败规则均有文本和代码支持，三个模式共享规则，成功录像与原生阶段报告相符；但无状态日志，远景不能验证微观接触及保持计数，且成功运行的手部力参数存在版本差异。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "文本只描述一个步骤（接住飞球并保持），catch→hold是同一连续动作的裁判阶段，单节点atomic合理；七个失败条件与catch.py/referee.py逐条对应（0.5秒窗口、0.02秒中断、90°、骨盆0.50米、过顶判定），均为立即终止。无状态日志，按P8 when/never/success为null并以atoms列出，符合规则；open_questions已列出文本/评分器的1.0秒与100步离散差异及无日志限制。", + "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：['fly_ball_caught']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:48:34Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__catch/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__catch/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "7d5a1601296682cc28ab86385b1caa097778e4661db249ad6e1f96ecbf0fa2cd", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/embodied-olympics-fun/catch/trajectories/embodied-olympics-fun-catch-i00-privileged--bnvu2kt.json b/phase2/embodied-olympics-fun/catch/trajectories/embodied-olympics-fun-catch-i00-privileged--bnvu2kt.json new file mode 100644 index 0000000000000000000000000000000000000000..284c39d0198b487cbe2bffe7bb0b92c4bc49061c --- /dev/null +++ b/phase2/embodied-olympics-fun/catch/trajectories/embodied-olympics-fun-catch-i00-privileged--bnvu2kt.json @@ -0,0 +1,186 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-catch-i00-privileged--bnvu2kt", + "taskKey": "embodied-olympics-fun/catch", + "dag_sha256": "61132849363e640170264bbd47c4c4cebc23f7ecfe683cae11a4abb07d10d44b", + "outcome": { + "success": 1, + "reward": 1.0, + "source": "catalog", + "raw": { + "success": 1, + "replay_success": 1, + "deterministic": 1, + "n_actions": 497, + "sim_steps": 497, + "horizon": 900, + "episode_over": 1, + "progress": 1.0, + "stages_reached": 2, + "stages_total": 2, + "fell": 0, + "replay_s": 6.34, + "bad_trajectory": 0, + "metric_catch_height_m": 0.955, + "metric_catch_speed_m_s": 9.612, + "grader_error": 0, + "video_rendered": 1, + "agent_sessions": 93, + "agent_resets": 3045, + "agent_replays": 17, + "agent_cost_self_reported": 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/embodied-olympics-fun-catch-i00-privileged--bnvu2kt/replay.mp4", + "duration_s": 5.0, + "fps": 25.0, + "frames": 125, + "sha256": "45ed1e5fa236164ac3aab0f467bd6b855e9ab14b49d3fe7b98d94562cafa07a4" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "fly_ball_caught": { + "method": "visual", + "intervals": [ + [ + 4.96, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察：视频 0.00 s 机器人站在场地上，手中尚无白球；视频 0.40–3.60 s 机器人迈步调整位置并伸出右手准备接球。" + }, + { + "t": 3.96, + "log_step": null, + "text_zh": "观察：视频 3.88–3.96 s 白色小球向机器人伸出的右手下降；视频 4.00 s 小球到达右手位置，逐帧检查视频 4.00–4.96 s 可见白点留在该手边随手移动，没有继续落到地面，机器人保持站立。" + }, + { + "t": 4.96, + "log_step": 164, + "text_zh": "观察：log step 164 读取交付文件的新进程重放结果，原生裁判报告 success=true、executed_rows=497、failed=null，仿真 3.974 s 进入 catch 阶段且 hand=right，仿真 4.97 s 进入 hold 阶段并成功；log step 171 的交付一致性检查显示动作校验和匹配、两次新进程重放均成功且末态相同。inputs/run.json 的最终评分字段亦为 success=1、replay_success=1、deterministic=1、sim_steps=497、stages_reached=2、fell=0。" + }, + { + "t": 4.96, + "log_step": 164, + "text_zh": "推断：交付重放的 497 个控制步乘以每步 0.01 秒为仿真 4.97 s，与视频总长 5.00 s 相差不足一帧；仿真 3.974 s 的接球事件已在视频 3.96–4.00 s 球到手处核对，因此把原生 hold 成功定位到视频约 4.96 s 的末帧，时间精度约 0.04 秒。该节点包含规定保持窗口，首次球到手尚不算完成；成功裁判结果佐证同一右手包围接触满足原生窗口，节点完成后持续至片尾。无逐帧状态日志，远景本身不能分辨接触夹角和毫秒级中断。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "抓握", + "note_zh": "机器人右手接住棒球并持有；视频 4.00–4.96 s 白球留在伸出的右手边，结合 log step 164 原生裁判的 right-hand catch 和 hold 成功，推断右手手指与拇指或掌部形成包围抓握。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "机器人双脚与腿在场地地面支撑身体，视频 0.40–3.60 s 迈步调整接球位置，视频 4.00–4.96 s 持球期间继续保持站立。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "catch_height_m", + "unit": "m", + "threshold": null, + "direction": null, + "value": 0.955 + }, + { + "name": "catch_speed_m_s", + "unit": "m/s", + "threshold": null, + "direction": null, + "value": 9.612 + }, + { + "name": "landed_from_start_m", + "unit": "m", + "threshold": null, + "direction": null, + "value": null + } + ], + "failure": null, + "summary_zh": "机器人迈步调整位置并以右手接住下降的白球，视频 4.00–4.96 s 球留在右手边且机器人保持站立，包含保持窗口的“接稳飞球”节点约在视频 4.96 s 完成，最终为 1/1。交付文件的原生重放报告接球与保持均通过，最终评分 success=1 且确定性重放成功。无状态日志，远景无法独立核验接触夹角和毫秒级中断，微观接触及保持窗口依据本次交付的原生裁判结果，完成时刻取末帧附近的近似值。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "单节点成功任务，节点时间 4.96 s（497 步×0.01 s≈片尾）与 log step 164 的原生裁判结果一致，接触和推断标注合理，failure 为空正确。", + "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": "2026-10-29T10:06:13Z", + "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/embodied-olympics-fun-catch-i00-privileged--bnvu2kt/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-catch-i00-privileged--bnvu2kt/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "651873f3d14c35a227d4e4f99f8e3638e628dc6fc0c51dd182e1f600d7841c73", + "video_sha256": "45ed1e5fa236164ac3aab0f467bd6b855e9ab14b49d3fe7b98d94562cafa07a4", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/catch/trajectories/embodied-olympics-fun-catch-i00-standard-realtime--y3yz63f.json b/phase2/embodied-olympics-fun/catch/trajectories/embodied-olympics-fun-catch-i00-standard-realtime--y3yz63f.json new file mode 100644 index 0000000000000000000000000000000000000000..4e6424b009c53f1de3d07c5451325b1be1d1b31f --- /dev/null +++ b/phase2/embodied-olympics-fun/catch/trajectories/embodied-olympics-fun-catch-i00-standard-realtime--y3yz63f.json @@ -0,0 +1,221 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-catch-i00-standard-realtime--y3yz63f", + "taskKey": "embodied-olympics-fun/catch", + "dag_sha256": "61132849363e640170264bbd47c4c4cebc23f7ecfe683cae11a4abb07d10d44b", + "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": 396, + "sim_steps": 2390, + "calls": 17, + "rows": 2410, + "resets": 5, + "reset_budget": 5, + "sim_startups": 1, + "elapsed_s": 2233.7, + "video_rendered": 1, + "first_success": 0, + "first_reason_code": 2, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 23.895, + "max_lag_s": 0.2227, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 374, + "late_steps": 0, + "requests_refused": 1, + "request_busy_s": 0.495, + "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/embodied-olympics-fun-catch-i00-standard-realtime--y3yz63f/replay.mp4", + "duration_s": 4.0, + "fps": 25.0, + "frames": 100, + "sha256": "620bb5fe1668a65508d74e0b8069e8d18aefe41236fae4ad595f641ac536e89d" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "fly_ball_caught": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察：视频 0.00 s 机器人站在场地内，双手未持球；视频 0.00–2.88 s 机器人随后向画面右侧移动，尚未出现接球。" + }, + { + "t": 3.8, + "log_step": 56, + "text_zh": "观察：逐帧检查视频 3.00–3.96 s，机器人移动并抬起右臂；视频 3.80–3.96 s 小白球下降到机器人上身附近，未见球在右手中停住并随手保持。log step 56 的现场机器人服务返回 episode 6 的裁判结果“the ball touched the body before it was caught”，stages 0/2、control steps 396；log step 57 确认该次尝试已结束。" + }, + { + "t": 3.96, + "log_step": 57, + "text_zh": "推断：视频末段与 log step 56–57 的最后现场尝试相符；原生裁判在接住前碰身即终止，故本次未建立合法接住阶段，更未完成同手保持窗口，接稳飞球全程未达成。log step 6 给出控制周期 0.01 s，episode 6 的396步对应仿真 3.96 s，与视频 3.96 s 的最后画面核对相符；本标签的时间锚点取自视频帧，未用 agent 工作时间换算。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：视频 0.00–3.96 s 机器人双脚交替接触场地地面，支撑身体并向接球位置移动，片尾仍站立。" + }, + { + "who": "未知", + "how": "未知", + "note_zh": "观察：log step 56 的原生裁判记录球在接住前接触机器人身体；视频 3.92–3.96 s 球接近上身，但远景不能分清具体触球部位，也不能确认右手是否先有短暂接触；未观察到手指从相对两侧包住球，故不标抓握。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "catch_height_m", + "unit": "m", + "threshold": null, + "direction": null, + "value": null + }, + { + "name": "catch_speed_m_s", + "unit": "m/s", + "threshold": null, + "direction": null, + "value": null + }, + { + "name": "landed_from_start_m", + "unit": "m", + "threshold": null, + "direction": null, + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "fly_ball_caught", + "t": 3.92, + "log_step": 56, + "where": "操作中", + "evidence_zh": "观察：视频 3.80–3.96 s 机器人正抬右臂迎接下降球，球到达身体附近而没有留在手中；log step 56 的机器人服务裁判报告接住前球碰身体、stages 0/2，并终止最后一次现场尝试。失败发生在接稳飞球操作中。" + }, + "unmet": { + "fly_ball_caught": false + }, + "phenomena": [ + { + "cat": "运动", + "item": "动态时机", + "t": 3.92, + "log_step": 56, + "evidence_zh": "观察：视频 3.60–3.96 s 机器人右臂保持抬起，下降小球接近上身时没有被该手截住并停留；log step 56 的服务裁判报告球在接住前碰到身体。推断：机器人接球手的迎球位置与下降球到达时刻未匹配；这是运动目标的漏接，未观察到曾持球后再脱手。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "t": 3.92, + "log_step": 56, + "source_zh": "日志 / 视频", + "evidence_zh": "观察：log step 56 中 agent 的脚本预先设定斜向步行、右臂关节姿态及闭手时序，并一次提交最后尝试；视频 3.00–3.96 s 机器人移动并抬臂，却没有让右手截住下降球，服务裁判记录接住前碰身。推断：这套步行、摆臂和闭手控制未将接球手与球轨迹协调到位；现有材料不能区分行走位移误差、臂姿误差或闭手时序误差，故保留上位项“控制”。" + } + ], + "gap_zh": "无状态日志或原生逐物理步接触记录，640×360远景不能核实具体禁触身体部位、手球短暂接触、接触夹角或毫秒级中断。log step 62–63 中 agent 将失败归于步行超调并自述位移约1.8米，这些尺度及胸肩碰撞位置未获独立三维证据核实，不据此细分控制根因；三项原生指标在评分器字段中均未报告。" + }, + "summary_zh": "最后一次现场尝试（episode 6）中，机器人移动并抬右臂迎球，但未完成接稳飞球，节点终态为0/1。视频约3.92–3.96 s 球到达机器人上身附近，现场服务裁判记录接住前碰身、阶段0/2，表明迎球动作未匹配下降球，归因到控制。远景和缺少状态日志使具体触球部位及控制误差来源无法进一步核实。", + "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、log step 56/57 一致：最后一次现场尝试（episode 6，396 步=3.96 s）接住前球碰身，节点未达成，断点在 fly_ball_caught，时间为视频时钟，未换算 agent 时间；动态时机 + 控制证据有限但标注诚实。", + "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": "2026-10-29T10:06:04Z", + "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/embodied-olympics-fun-catch-i00-standard-realtime--y3yz63f/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-catch-i00-standard-realtime--y3yz63f/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "96b0ffdba9865e4fdbf0d0573f185f749cab0df5272b4c0822f12146a96607fd", + "video_sha256": "620bb5fe1668a65508d74e0b8069e8d18aefe41236fae4ad595f641ac536e89d", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/catch_and_shoot/task_dag.json b/phase2/embodied-olympics-fun/catch_and_shoot/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..d4508cf0f26522a300b7f5082e1de70916454b76 --- /dev/null +++ b/phase2/embodied-olympics-fun/catch_and_shoot/task_dag.json @@ -0,0 +1,335 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "embodied-olympics-fun/catch_and_shoot", + "suite": "embodied-olympics-fun", + "base": "embodied-olympics-fun-catch-and-shoot-i00", + "title_zh": "双手接球后投篮", + "summary_zh": "双手在空中接住左翼来球后合法出手，使球从上方穿过篮圈并下降一球直径，期间无犯规或摔倒。", + "embodiment_zh": "三种模式均为 Unitree H1-2 人形机器人，27 个本体关节、双臂，左、右末端各为 Inspire RH56DFX 五指灵巧手，每手六个驱动由一个闭合量控制；双腿足式支撑，无轮式底盘，篮球仅靠真实接触与摩擦持握。", + "task_class": "compositional", + "analysis_level": "step", + "spec": { + "version": "1.2", + "task": "embodied-olympics-fun-catch-and-shoot-i00", + "bench": "embodied-olympics-fun", + "progress": "live", + "entities": {}, + "milestones": [ + { + "id": "ball_caught", + "zh": "双手接住球", + "mode": "event", + "when": null, + "after": [], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "场地外部视角可见橙色篮球从机器人左侧传球机飞来，未先触地便被胸腹前两只手共同承接并控制；仅单手碰到、短暂拍过或球弹走均未完成。裁判要求两只手各自的有效接触同时连续保持至少 0.10 秒（每手自最后一次实际接触起不超过 0.03 秒仍算接触），完成后即记住该事件，投篮离手不取消接球进度。", + "step": { + "cat": "动态", + "item": "接住" + }, + "contact": [ + { + "who": "双臂协作", + "how": "其他", + "plan": 1, + "note_zh": "候选：夹持；左掌与右掌从篮球两侧相向夹压，靠接触和摩擦承接、持住或加速篮球；单只手不合拢包住球，故不标抓握，现有接触条目没有双掌相向夹持（暂行：DECISIONS.md D12）。" + }, + { + "who": "双臂协作", + "how": "抓握", + "plan": 2, + "note_zh": "左、右灵巧手的手指分别与拇指或掌面合拢夹住篮球两侧的局部球面，共同持球；对应任务明确允许的手指闭合包住球面的方式，区别于仅用两张开放手掌相向夹压。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:4", + "inputs/task/privileged/instruction.md:10", + "inputs/task/privileged/instruction.md:16-26", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1047-1053", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1116-1139", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1141-1144", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1152-1181", + "inputs/oracle/replay.mp4:2.4-3.0 s" + ], + "conf": 6, + "note_zh": "接住后能持球暂停，再继续投篮，故与投篮分开；接触的 0.10 秒窗口由裁判在物理子步内累计，拟记录的是已完成的 catch 事件，hold_s=0 避免在事件之后重复等待。初态双手空、球在传球机中；无需额外准备节点。" + }, + { + "id": "shot_through_ring", + "zh": "投篮命中", + "mode": "event", + "when": null, + "after": [ + "ball_caught" + ], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "场地外部视角可见机器人将接住的橙色球投出，球与双手分离后飞向画面右侧篮板下的篮圈，从上方落入并继续下降到圈下；仅球已出手、接近篮筐、碰圈后弹开或从下方上穿都不算。按裁判顺序须先接球再释放（两手有效接触均消失），球心向下穿越高 3.05 米的圈面时距圈心小于 0.225 米，未向上回穿圈面，随后向下穿越高 2.811 米的平面时距圈心小于 0.3445 米，且触发成功前不得出现约束中的犯规或摔倒。", + "step": { + "cat": "动态", + "item": "抛/投" + }, + "contact": [ + { + "who": "双臂协作", + "how": "其他", + "plan": 1, + "note_zh": "候选：夹持；左掌与右掌从篮球两侧相向夹压，靠接触和摩擦承接、持住或加速篮球；单只手不合拢包住球，故不标抓握，现有接触条目没有双掌相向夹持（暂行：DECISIONS.md D12）。" + }, + { + "who": "双臂协作", + "how": "抓握", + "plan": 2, + "note_zh": "左、右灵巧手的手指分别与拇指或掌面合拢夹住篮球两侧的局部球面，共同持球；对应任务明确允许的手指闭合包住球面的方式，区别于仅用两张开放手掌相向夹压。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:4", + "inputs/task/privileged/instruction.md:26-30", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:744", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1047-1053", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1135-1139", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1182-1206", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:429-458", + "inputs/oracle/replay.mp4:4.4-6.6 s" + ], + "conf": 6, + "note_zh": "出手和飞行是投篮命中的连续过程，不单设球已离手节点；释放后能短暂再触球的裁判宽限另见约束与疑问。成功瞬间结束，穿圈作为事件锁存，不要求篮球永远停在篮筐里；接触标送球动作，离手后的无接触飞行记在本说明中。成功视频使用双手下手投篮，任务也允许其他投篮样式，不限制挥臂方向或轨迹。" + } + ], + "constraints": [ + { + "id": "ball_touched_before_pass", + "zh": "提前触球", + "never": null, + "criterion_zh": "篮球尚未离开传球机滚轮前，机器人任意部位碰到篮球即违规；场地视角可见机器人靠近传球机并碰到仍在机内的球。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1152-1158" + ] + }, + { + "id": "nonhand_ball_contact", + "zh": "非手部触球", + "never": null, + "criterion_zh": "传球已离机且尚未被裁判记为释放时，篮球接触机器人除双手以外的部位即违规；手包括手指、拇指、掌和腕，球靠在胸腹、上臂、腿或脚上均不允许。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/instruction.md:18", + "inputs/task/privileged/instruction.md:26", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1131-1134", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1159-1165" + ] + }, + { + "id": "ball_ground_before_goal", + "zh": "球提前触地", + "never": null, + "criterion_zh": "从传球离机到投篮成功前，篮球任何一次接触地板均判失败；包括来球先落地、接住后持球触地及出手后未进筐便落地，允许球碰篮圈或篮板。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1179-1180", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1190-1192", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1200-1206" + ] + }, + { + "id": "travelling", + "zh": "走步", + "never": null, + "criterion_zh": "从接球成立到裁判记为释放检查脚步：双脚着地接球时，首次抬脚确定另一脚为轴心脚，轴心脚离地后出手前任何脚再落地即违规；双脚均未抬起前任一脚、之后轴心脚沿地滑移超过 0.20 米也违规。若接球时有脚离地，则最多两步且不得同脚连续两次落地，仍着地的脚不得从其锚点滑移超过 0.20 米；脚离地至少 0.03 秒后再接触地面算落地，异脚落地间隔不超过 0.03 秒合为一步，同脚重复落地在超过合并窗口或释放时确认。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:847-948", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1052", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1172-1175", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1187-1198" + ] + }, + { + "id": "post_release_retouch", + "zh": "出手后再触球", + "never": null, + "criterion_zh": "裁判记为释放后超过 0.05 秒，机器人任意部位再次接触篮球即判失败；画面中球离手飞行后被手、躯干或腿重新碰到均属此项，释放时间按每手 0.03 秒接触宽限后的有效接触消失计算。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1052", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1141-1144", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1194-1201" + ] + }, + { + "id": "upward_ring_entry", + "zh": "从下方穿篮", + "never": null, + "criterion_zh": "接球后篮球尚未登记向下入圈时，球心从下向上穿越高 3.05 米的篮圈平面，穿越点到圈心的水平距离小于 0.225 米，立即判违规。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1047", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1182-1186", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:400-423" + ] + }, + { + "id": "robot_fell", + "zh": "机器人摔倒", + "never": null, + "criterion_zh": "投篮成功前，机器人骨盆高度低于 0.35 米，或双脚以外的身体部位接触地面，裁判判为摔倒并终止；外部视角可见躯干或手臂落地，单纯身体晃动但未满足这些条件不算。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1038-1040", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1053", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/task.py:21-23", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:272-289" + ] + }, + { + "id": "simulation_unstable", + "zh": "仿真失稳", + "never": null, + "criterion_zh": "物理积分出现坏加速度警告、非有限位姿，或机器人或自由物体的平移速度分量绝对值超过 150 米每秒时，模拟器以仿真失稳终止；视频可能突然跳变或停止，单凭视频停止不能确认，需查看该失败原因。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:108-109", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:866-885" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "catch_completed", + "level": "new_code", + "source": "dyn.catch_completed <- ref.has(\"catch\"); inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1168-1175" + }, + { + "atom": "shot_through_ring", + "level": "new_code", + "source": "dyn.shot_through_ring <- ref.has(\"through the ring\"); inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1182-1206" + }, + { + "atom": "task_success", + "level": "new_code", + "source": "dyn.task_success <- world.ref.success (per-frame native checker flag); inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:174-178" + }, + { + "atom": "ball_touched_before_pass", + "level": "new_code", + "source": "dyn.ball_touched_before_pass <- per-frame native rule predicate / latched ref.failed; inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1152-1158" + }, + { + "atom": "nonhand_ball_contact", + "level": "new_code", + "source": "dyn.nonhand_ball_contact <- per-frame native rule predicate / latched ref.failed; inputs/task/privileged/instruction.md:18; inputs/task/privileged/instruction.md:26; inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1131-1134; inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1159-1165" + }, + { + "atom": "ball_ground_before_goal", + "level": "new_code", + "source": "dyn.ball_ground_before_goal <- per-frame native rule predicate / latched ref.failed; inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1179-1180; inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1190-1192; inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1200-1206" + }, + { + "atom": "travelling", + "level": "new_code", + "source": "dyn.travelling <- per-frame native rule predicate / latched ref.failed; inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:847-948; inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1052; inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1172-1175; inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1187-1198" + }, + { + "atom": "post_release_retouch", + "level": "new_code", + "source": "dyn.post_release_retouch <- per-frame native rule predicate / latched ref.failed; inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1052; inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1141-1144; inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1194-1201" + }, + { + "atom": "upward_ring_entry", + "level": "new_code", + "source": "dyn.upward_ring_entry <- per-frame native rule predicate / latched ref.failed; inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1047; inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1182-1186; inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:400-423" + }, + { + "atom": "robot_fell", + "level": "new_code", + "source": "dyn.robot_fell <- per-frame native rule predicate / latched ref.failed; inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1038-1040; inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/catch_and_shoot.py:1053; inputs/grader/efe_embodied-olympics-fun-files/rfo/task.py:21-23; inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:272-289" + }, + { + "atom": "simulation_unstable", + "level": "new_code", + "source": "dyn.simulation_unstable <- per-frame native rule predicate / latched ref.failed; inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:108-109; inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:866-885" + } + ] + }, + "native_metrics": [], + "open_questions_zh": [ + "inputs/states/index.json 的 logs 为空，成功示例也没有状态日志；依 DECISIONS.md P8，两节点 when、各约束 never 和 success 均为 null，不能据未记录的字段伪造绑定。atoms 列出应新增记录的原生事件与规则量；精确的接触时长、穿圈几何及全部约束不能仅靠远景逐帧可靠测量，当前只能结合视频和裁判代码判读。", + "接触词表缺口（DECISIONS.md D12、P14）：开放双掌相向夹球标双臂协作×其他，候选夹持；任务文字也允许手指合拢包住球面，单列抓握为替代方案，不把两种方案当作同时发生。", + "文本与代码差异：instruction.md 的规则偏离段说从传球直到进筐均禁止非手部触球，但详细失败列表及 catch_and_shoot.py:1159-1165 只在裁判释放前检查非手部触球，释放超过 0.05 秒才检查任意部位再触球（1200-1201）；释放又使用 0.03 秒的每手接触宽限。约束按实际裁判分段写出，不能将宽限内接触擅自当作代码已判失败。", + "原地投篮的文字未给位置容差；评分器仅以接球至释放期间的走步规则约束移动，不检查投篮位置距初始位置的独立半径。具体规则还允许接球时有脚离地的两步路线，故不添加始终双脚不动的条件或站立步骤（catch_and_shoot.py:872-948、1187-1198）。", + "模式差异：物理规则、参数和阶段顺序三模式相同；privileged 的 grade.py:87-115 另要求有效有限动作文件、两次成功重放且末态哈希相同，standard 与 standard-realtime 则由服务按最后一次现场尝试的裁判结果评分（standard/server.py:157-178；inputs/grader/efe_standard-files/v2_1/grade.py:93-99）。这些交付、复现和基础设施条件不是语义世界状态，不建节点，也无可用日志表达，故 success 保持 null。", + "仅记录不评分：catch_and_shoot.py:1208-1209 的 ball_height_m 只是当前球高，没有单独通过阈值或单调优化方向，故 native_metrics 为空，不增加约束；标准实时模式的时钟滞后与第一尝试统计也不改变成绩（inputs/grader/efe_standard-files/v2_1/grade.py:14-19）。15 秒是任务时限，不建约束（DECISIONS.md P21）。", + "依 DECISIONS.md D10 当前读法保留双手接住球与投篮命中两步；接球后能够持球暂停，符合 S10，裁判的 release 阶段仅作为投篮验收的一部分。成功视频是双手下手投篮，仅用于识别画面线索，不固定投篮样式。" + ], + "conf": 6, + "conf_reason_zh": "两个步骤直接来自三模式一致的任务文字，顺序、时长、穿圈阈值与失败条件均有代码来源，初态和成功视频支持该拆分；但没有任何状态日志可绑定，远景不足以独立核实毫秒接触及精确穿圈位置，并存在文字触球范围与代码宽限的差异。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "接球→投篮命中两步与文本一致（catch→release→through the ring，release 并入投篮），约束与裁判代码分段（释放前非手触球、释放后0.05秒再触球、走步、触地、下方穿圈、摔倒、失稳）对应，无状态日志故 when/success 为 null 并已在 open_questions_zh 说明，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：['ball_caught', 'shot_through_ring']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:49:09Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__catch_and_shoot/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__catch_and_shoot/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "968d604e21b2b8f1838c31f970e752fbcb7d46b357dd01f46af5dc12b84f56ba", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/embodied-olympics-fun/catch_and_shoot/trajectories/embodied-olympics-fun-catch-and-shoot-i00-privileged--qvqbash.json b/phase2/embodied-olympics-fun/catch_and_shoot/trajectories/embodied-olympics-fun-catch-and-shoot-i00-privileged--qvqbash.json new file mode 100644 index 0000000000000000000000000000000000000000..4b6ba59b1455beb37e8e930b16087c569140cd7e --- /dev/null +++ b/phase2/embodied-olympics-fun/catch_and_shoot/trajectories/embodied-olympics-fun-catch-and-shoot-i00-privileged--qvqbash.json @@ -0,0 +1,192 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-catch-and-shoot-i00-privileged--qvqbash", + "taskKey": "embodied-olympics-fun/catch_and_shoot", + "dag_sha256": "f6a6d84d578d7c1084b9432c694b8d43d8fa3421ae8d0142dbb9a316794fb326", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "replay_success": 0, + "deterministic": 1, + "n_actions": 291, + "sim_steps": 278, + "horizon": 1500, + "episode_over": 1, + "progress": 0.0, + "stages_reached": 0, + "stages_total": 3, + "fell": 0, + "replay_s": 5.19, + "bad_trajectory": 0, + "metric_ball_height_m": 0.106, + "grader_error": 0, + "video_rendered": 1, + "agent_sessions": 81, + "agent_resets": 30, + "agent_replays": 4, + "agent_cost_self_reported": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.703, + "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, + "session_log_tampered": 0, + "session_log_from_monitor": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-catch-and-shoot-i00-privileged--qvqbash/replay.mp4", + "duration_s": 2.8, + "fps": 25.0, + "frames": 70, + "sha256": "83025b898e21d2d74141fa5516b1fa6e8f7d7af300cb97e6dc502e1553da168b" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "ball_caught": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 2.48, + "log_step": null, + "text_zh": "观察：本次视频 2.24–2.44 s 橙色篮球从画面左侧传球机飞到机器人胸前；视频 2.44–2.52 s 球短暂贴近右手并改变运动方向，左手未与右手共同把球留在胸腹前，随后球向右下方离开双手。" + }, + { + "t": 2.76, + "log_step": 120, + "text_zh": "观察：本次视频 2.56–2.76 s 篮球持续降到地板附近，机器人没有重新接住；log step 120 对最终保存文件的新进程重放返回 stages=[]，失败原因为“the pass was not caught: the ball touched the floor”；log step 126 确认最终文件仍与已核验的 v41 动作逐元素相同，log step 127 再次重放结果相同。" + }, + { + "t": 2.6, + "log_step": null, + "text_zh": "推断：本次视频没有成功示例在其视频 2.44–2.96 s 所显示的双手持续承接并控制球的画面线索；结合最终文件重放明确未接住的裁判结果，双手接住球事件未成立。成功示例仅用于辨认持球外观。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "其他", + "note_zh": "观察：视频 2.44–2.52 s 机器人右手短暂挡碰篮球，球随后向右下方离开；未见右手包住球或左右手形成夹持。候选：击打（单手挡拍）；现有接触方式没有短促挡拍项，不按抓握记录。左手虽向球靠拢，但未见实际共同持球。" + } + ] + }, + "shot_through_ring": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 2.76, + "log_step": 120, + "text_zh": "观察：本次视频全段 0–2.76 s 机器人没有接住球后朝篮圈出手；视频末段球从胸前下降到机器人前方地板附近，篮圈处没有篮球穿越。log step 120 的最终文件重放阶段列表为空，接球前即因球触地结束，因此投篮命中未成立。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [], + "failure": { + "breakpoint": { + "ref": "ball_caught", + "t": 2.52, + "log_step": 120, + "where": "操作中", + "evidence_zh": "观察：视频 2.44–2.52 s 机器人正在用双手迎接来球，右手短暂挡碰后，另一只手未及时合围；视频 2.56–2.76 s 球脱离双手向地板下降，接球再未完成。log step 120 对交付文件重放明确报“the pass was not caught: the ball touched the floor”，故断点属于双手接住球。" + }, + "unmet": { + "ball_caught": false, + "shot_through_ring": false + }, + "phenomena": [ + { + "cat": "运动", + "item": "动态时机", + "t": 2.52, + "log_step": null, + "evidence_zh": "观察：视频 2.44–2.68 s，机器人右手迎到运动中的篮球后向下追随，左手靠拢时球已从两手工作区域向右下方移走；机器人双手未及时共同截停来球。视频 2.76 s 球已接近地板，未出现双手持球阶段。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "推断：log step 118 的交付说明记载 agent 的控制器先以右拳柔顺退让挡球，再让两手到预测落点合围；log step 126 确认最终交付仍为该 v41 动作。视频 2.44–2.68 s 显示机器人确实迎球并追球，但单手接触未接续成双手控制，表明接球动作的控制没有实现预期衔接；现有证据不能进一步区分接触布局、轨迹跟踪或控制增益问题。", + "source_zh": "视频 / 日志" + } + ], + "gap_zh": "没有状态日志；视频为单一远景，能判断来球未受双手持续控制及最后落地，但不能精确确认各指接触对、接触力或毫秒级接触窗口。log step 112 的生成期诊断属于会话实验，其球运动时间不能直接替代交付重放的视频时间；此处时间全部取自实际视频，根因只标到控制。" + }, + "summary_zh": "机器人尝试接住左侧来球，右手短暂挡碰后未能与左手及时合围，球在视频 2.52–2.76 s 向右下方落地，双手接住球与投篮命中均未完成（0/2）。断点为双手接住球，现象为动态时机，根因保留为控制；最终交付文件多次重放均确认未接住，未把会话中其他试验算作进度。没有状态日志且手部画面较小，不能细分具体控制子因或精确接触时长。", + "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 2.2–2.76 s 帧显示球飞到右手处被挡开后落地，未被双手持住；两节点均未达成、断点 ball_caught、动态时机 + 控制（停在“控制”）均有证据支持，接触标注保守合理。", + "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": "2026-10-29T10:06:39Z", + "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/embodied-olympics-fun-catch-and-shoot-i00-privileged--qvqbash/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-catch-and-shoot-i00-privileged--qvqbash/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "0f9dba51dbb2201097119c29c40842d815a465aac1625a667e4b3584c33e6f46", + "video_sha256": "83025b898e21d2d74141fa5516b1fa6e8f7d7af300cb97e6dc502e1553da168b", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/catch_and_shoot/trajectories/embodied-olympics-fun-catch-and-shoot-i00-privileged--sxevm8b.json b/phase2/embodied-olympics-fun/catch_and_shoot/trajectories/embodied-olympics-fun-catch-and-shoot-i00-privileged--sxevm8b.json new file mode 100644 index 0000000000000000000000000000000000000000..c658498eb274871d8372772c59c7161dda58c476 --- /dev/null +++ b/phase2/embodied-olympics-fun/catch_and_shoot/trajectories/embodied-olympics-fun-catch-and-shoot-i00-privileged--sxevm8b.json @@ -0,0 +1,194 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-catch-and-shoot-i00-privileged--sxevm8b", + "taskKey": "embodied-olympics-fun/catch_and_shoot", + "dag_sha256": "f6a6d84d578d7c1084b9432c694b8d43d8fa3421ae8d0142dbb9a316794fb326", + "outcome": { + "success": 1, + "reward": 1.0, + "source": "catalog", + "raw": { + "success": 1, + "replay_success": 1, + "deterministic": 1, + "n_actions": 634, + "sim_steps": 634, + "horizon": 1500, + "episode_over": 1, + "progress": 1.0, + "stages_reached": 3, + "stages_total": 3, + "fell": 0, + "replay_s": 6.53, + "bad_trajectory": 0, + "metric_ball_height_m": 2.78, + "grader_error": 0, + "video_rendered": 1, + "agent_sessions": 158, + "agent_resets": 476, + "agent_replays": 3, + "agent_cost_self_reported": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 3.779, + "peak_processes": 44, + "peak_disk_gb": 0.241, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0, + "session_log_tampered": 0, + "session_log_from_monitor": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-catch-and-shoot-i00-privileged--sxevm8b/replay.mp4", + "duration_s": 6.36, + "fps": 25.0, + "frames": 159, + "sha256": "34e870c1dae576f99656fa7a1fa3aefe4315499020cf42ea7928eaa5dfe41bd8" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 2, + "final": 2, + "nodes": { + "ball_caught": { + "method": "visual", + "intervals": [ + [ + 2.64, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 2.64, + "log_step": null, + "text_zh": "观察：视频 2.28–2.44 s，橙色篮球从机器人左侧飞向身前；视频 2.48–2.76 s，两手向球两侧收拢，球减速并留在双手之间，随后视频 2.80–4.40 s 持续随双手移动，未先落地。接球事件约标在视频 2.64 s，远景不能逐毫秒核实双手有效接触窗口。" + }, + { + "t": null, + "log_step": 56, + "text_zh": "观察：log step 56 对最终保存的 634 行动作文件运行官方重放器，返回 success=true、failed=null，stages 依次为 catch、release、through the ring；inputs/run.json 的评分字段也记录 replay_success=1、deterministic=1、stages_reached=3。该结果支持双手接球已通过原生裁判。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "其他", + "note_zh": "推断：视频 2.48–2.96 s 双手在篮球两侧相向收拢，随后共同持球；按可见姿态标候选“夹持”（DECISIONS.md D12）。单只手是否形成指掌对向包握在远景中不可辨，不据手指闭合命令认定抓握；双手相向夹持没有对应的现有 how 条目。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：视频 2.28–2.96 s 机器人以双脚接触地板，支撑站立接球。" + } + ] + }, + "shot_through_ring": { + "method": "visual", + "intervals": [ + [ + 6.32, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 5.04, + "log_step": null, + "text_zh": "观察：视频 3.00–4.40 s 机器人将球保持在双手之间，视频 4.48–4.88 s 双手持球下摆至髋前，随后向前上方送球；视频 5.04 s 球已与双手分离，向画面右侧篮筐飞行，机器人仍直立。" + }, + { + "t": 6.32, + "log_step": null, + "text_zh": "观察：视频 5.36–5.72 s 篮球升至篮圈上方，视频 5.76–6.24 s 下降并对准篮圈，视频 6.28–6.32 s 球进入圈口并向圈下落去，未见向外弹开。投篮完成约标在视频 6.32 s；球与篮网局部重叠，精确的圈下验收高度由成功评分支持。" + }, + { + "t": null, + "log_step": 56, + "text_zh": "观察：log step 56 官方重放器对交付文件返回 success=true、stages=[catch, release, through the ring]、executed_rows=634、ball_height_m=2.78；log step 70 重新生成文件前后 MD5 相同，log step 75 再次重放仍为 success=true 且末态哈希与 log step 56 相同。inputs/run.json 的 success=1、replay_success=1 与视频末段入圈一致。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "其他", + "note_zh": "推断：视频 3.00–4.96 s 双手从球两侧共同夹持、承托并作下手投篮，视频 5.04 s 已离手；候选“夹持”（DECISIONS.md D12），不能从远景确认单手包握，双掌相向夹持无对应的现有 how 条目。离手后的飞行与穿圈不再有手球接触。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：视频 3.00–6.32 s 机器人双脚留在地板上，站立支撑持球、摆臂和出手，末段仍直立。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [], + "failure": null, + "summary_zh": "机器人约在视频 2.64 s 双手接住左侧来球，持球调整后下摆蓄力，约在视频 5.0 s 出手，篮球约在视频 6.32 s 从上方穿入篮圈，两个事件节点均完成。官方交付重放与最终评分均为成功，未设失败断点；本次无状态日志，远景无法独立测量毫秒接触窗口、单手包握细节和精确穿圈几何，接触按双手相向夹持处理（DECISIONS.md D12），精确验收由原生成功结果支持。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "两个节点（接球约 2.64 s、投篮入圈约 6.32 s）与视频和成功评分一致，无失败断点，接触标注与 DAG plan 1 及实际相符。", + "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": "2026-10-29T10:06:32Z", + "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/embodied-olympics-fun-catch-and-shoot-i00-privileged--sxevm8b/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-catch-and-shoot-i00-privileged--sxevm8b/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "8ea96657bc579a1b9383dc46b38395085e4cf8b77aa7e028a777b9631d06afc8", + "video_sha256": "34e870c1dae576f99656fa7a1fa3aefe4315499020cf42ea7928eaa5dfe41bd8", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/clean_and_jerk/task_dag.json b/phase2/embodied-olympics-fun/clean_and_jerk/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..84d1450cd05c055164456a9dffe0c7d90fcacd48 --- /dev/null +++ b/phase2/embodied-olympics-fun/clean_and_jerk/task_dag.json @@ -0,0 +1,361 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "embodied-olympics-fun/clean_and_jerk", + "suite": "embodied-olympics-fun", + "base": "embodied-olympics-fun-clean-and-jerk-i00", + "title_zh": "挺举杠铃", + "summary_zh": "杠铃先在肩前稳定上肩，再由双手举至头顶上方，双臂双腿伸直、双脚并齐并保持稳定至下放信号，全程无裁判违规。", + "embodiment_zh": "两种模式均为Unitree H1-2双足人形机器人，27个本体关节；左右臂各装Inspire RH56DFX五指灵巧手，以接触和摩擦握持自由杠铃，双腿与躯干在举重台上支撑和平衡，无轮式底盘。", + "task_class": "compositional", + "analysis_level": "step", + "spec": { + "version": "1.2", + "task": "embodied-olympics-fun-clean-and-jerk-i00", + "bench": "embodied-olympics-fun", + "progress": "live", + "entities": {}, + "milestones": [ + { + "when": null, + "optional": false, + "verify_zh": null, + "freeze_held": false, + "step": { + "cat": "位姿", + "item": "拿起" + }, + "contact": [ + { + "who": "双臂协作", + "how": "抓握", + "plan": 1, + "note_zh": "左右灵巧手各自握住杠铃握杆的一侧，双臂同时承载并抬升杠铃。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "plan": 1, + "note_zh": "双脚支撑在木色举重台上，双腿与躯干承受机器人和杠铃的重量。" + } + ], + "conf": 6, + "id": "bar_at_shoulders", + "zh": "杠铃上肩", + "mode": "event", + "after": [], + "hold_s": 0.49, + "goal": true, + "criterion_zh": "场馆斜前方视角可见银色杠铃已离开两只棕色支架，停在胸肩前方，机器人双手持杠、直立伸膝、双脚在木台上并齐：杠轴距两侧肩关节均不超过0.15米，双膝距伸直不超过10度，双脚中心沿垂直杠轴的水平方向错位不超过0.05米（杠轴偏离水平超过60度不合格），躯干直立余弦至少0.94、杠中心距两肩中点的水平距离不超过0.25米、杠轴俯视方向与肩线夹角不超过20度，杠铃和骨盆线速度均不超过0.1米/秒，全部条件连续满足50个控制步（名义0.5秒）。此前必须双手离架、杠中心至少到过台面以上0.48米且未违规回落；杠仍在低位、仅经过肩高未停稳或未站直均未完成，完成后上举离肩不撤销本事件。", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:103-114", + "inputs/task/privileged/tests/rfo_task.py:186-201", + "inputs/task/privileged/tests/rfo_task.py:230-260", + "inputs/task/privileged/tests/rfo_task.py:274-289", + "inputs/task/privileged/tests/rfo_task.py:314-348", + "inputs/task/privileged/tests/rfo_task.py:350-385", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-184", + "inputs/task/privileged/instruction.md:4", + "inputs/oracle/replay.mp4:16.0-18.5s" + ], + "note_zh": "上肩是文本单列、可停留后继续的中间目标，使用事件模式；成功要求曾完成上肩，故为目标节点。离架和过膝合入该步验收，不按四个原生阶段拆成四步；双手接触允许最多0.05秒间断。按D31，50个0.01秒控制步的首末采样间隔为0.49秒；无日志时不执行此条件。" + }, + { + "when": null, + "optional": false, + "verify_zh": null, + "freeze_held": false, + "step": { + "cat": "位姿", + "item": "拿起" + }, + "contact": [ + { + "who": "双臂协作", + "how": "抓握", + "plan": 1, + "note_zh": "左右灵巧手各自握住杠铃握杆的一侧，双臂同时承载并抬升杠铃。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "plan": 1, + "note_zh": "双脚支撑在木色举重台上，双腿与躯干承受机器人和杠铃的重量。" + } + ], + "conf": 6, + "id": "bar_overhead", + "zh": "杠铃过顶", + "mode": "state", + "after": [ + "bar_at_shoulders" + ], + "hold_s": 1.99, + "goal": true, + "criterion_zh": "完成上肩后，场馆视角可见银色杠铃举至头部上方、两手分握杠杆、双臂伸直，双腿伸直且双脚在木台上并齐：杠轴高度至少比头部中心高0.25米，双肘和双膝距伸直均不超过10度，脚沿垂直杠轴的水平方向错位不超过0.05米（杠轴偏离水平超过60度不合格），躯干直立余弦至少0.94、杠中心距两肩中点的水平距离不超过0.25米、杠轴俯视方向与肩线夹角不超过20度，杠铃与骨盆线速度均不超过0.1米/秒，全部条件连续满足200个控制步（名义2秒）直至裁判记录下放信号。仅把杠举高而屈肘、屈膝、摇晃或尚未保持满时长均未完成；杠降低、脱手或任一终态条件失去即不再成立，下放信号立即结束成功回合，无需再放下杠铃。", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:103-114", + "inputs/task/privileged/tests/rfo_task.py:249-289", + "inputs/task/privileged/tests/rfo_task.py:314-318", + "inputs/task/privileged/tests/rfo_task.py:350-393", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-184", + "inputs/task/privileged/instruction.md:4", + "inputs/oracle/replay.mp4:25.0-27.4s" + ], + "note_zh": "上举与稳定保持共同构成第二个语义步骤；不另列过顶瞬间、静止或下放信号。边来自文本先上肩再上举及裁判的阶段顺序；state本身不以after门控，视频还须核验上肩历史与全部约束。200个0.01秒控制步的首末采样间隔为1.99秒（D31）；两步均允许且要求持杠，禁用持物冻结。" + } + ], + "constraints": [ + { + "id": "off_platform_contact", + "zh": "触及台外地面", + "never": null, + "criterion_zh": "机器人任一部位或杠铃任一部分接触木色举重台外的浅色地面；以实际接触判定，不按投影越界判定。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:323-326" + ] + }, + { + "id": "nonfoot_platform_contact", + "zh": "非脚部触台", + "never": null, + "criterion_zh": "除双脚外的任何机器人部位接触木色举重台；深蹲本身不违规，也没有额外骨盆高度或倾角跌倒规则。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:106,327-329" + ] + }, + { + "id": "robot_block_contact", + "zh": "身体触及支架", + "never": null, + "criterion_zh": "机器人任一部位接触两只棕色举重支架；双手抓杠时也不得碰到支架。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:330-331" + ] + }, + { + "id": "one_hand_liftoff", + "zh": "非双手离架", + "never": null, + "criterion_zh": "在尝试尚未过膝时，杠轴高于静置高度超过0.03米且不足两只手持杠；手持按握杆与各手手指接触、允许最多0.05秒间断计算。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:314-318,332-348" + ] + }, + { + "id": "bar_released_during_attempt", + "zh": "过膝后脱手", + "never": null, + "criterion_zh": "杠中心到过台面以上0.48米、尝试标记成立后，下放信号前任一只手脱离握杆超过0.05秒。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:314-318,337-348" + ] + }, + { + "id": "bar_lowered_before_signal", + "zh": "提前回落触台", + "never": null, + "criterion_zh": "尝试过膝后、下放信号前，杠铃任一部分接触举重台或任一支架；初始搁在支架上不违规。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:337-348" + ] + }, + { + "id": "overhead_without_clean", + "zh": "未上肩先过顶", + "never": null, + "criterion_zh": "上肩尚未通过时，杠轴达到头部中心以上至少0.25米；即使双臂已经伸直，也作为未完成上肩就过顶立即判负。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:350-385" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "clean_position_ok", + "level": "new_code", + "source": "inputs/task/privileged/tests/rfo_task.py:350-385；需新增日志字段dyn.clean_position_ok，记录lift-off、attempt、双手、膝、脚、静止、在杠下及肩距条件的合取，不含保持计数。" + }, + { + "atom": "clean_reached", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-158；需新增日志字段dyn.clean_reached，记录clean阶段锁存，用于核验上肩历史及上举先后。" + }, + { + "atom": "overhead_position_ok", + "level": "new_code", + "source": "inputs/task/privileged/tests/rfo_task.py:350-390；需新增日志字段dyn.overhead_position_ok，记录common、overhead、elbows的合取，不含保持计数。" + }, + { + "atom": "down_signal_reached", + "level": "new_code", + "source": "inputs/task/privileged/tests/rfo_task.py:390-393；需新增日志字段dyn.down_signal_reached，记录down signal阶段锁存。" + }, + { + "atom": "checker_success", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:174-178；需记录每帧裁判自身success和failed，不使用交付或复现有效性代替世界状态。" + }, + { + "atom": "off_platform_contact", + "level": "new_code", + "source": "inputs/task/privileged/tests/rfo_task.py:323-326；需新增日志字段dyn.off_platform_contact，记录本约束对应的接触或几何违规条件。" + }, + { + "atom": "nonfoot_platform_contact", + "level": "new_code", + "source": "inputs/task/privileged/tests/rfo_task.py:106,327-329；需新增日志字段dyn.nonfoot_platform_contact，记录本约束对应的接触或几何违规条件。" + }, + { + "atom": "robot_block_contact", + "level": "new_code", + "source": "inputs/task/privileged/tests/rfo_task.py:330-331；需新增日志字段dyn.robot_block_contact，记录本约束对应的接触或几何违规条件。" + }, + { + "atom": "one_hand_liftoff", + "level": "new_code", + "source": "inputs/task/privileged/tests/rfo_task.py:314-318,332-348；需新增日志字段dyn.one_hand_liftoff，记录本约束对应的接触或几何违规条件。" + }, + { + "atom": "bar_released_during_attempt", + "level": "new_code", + "source": "inputs/task/privileged/tests/rfo_task.py:314-318,337-348；需新增日志字段dyn.bar_released_during_attempt，记录本约束对应的接触或几何违规条件。" + }, + { + "atom": "bar_lowered_before_signal", + "level": "new_code", + "source": "inputs/task/privileged/tests/rfo_task.py:337-348；需新增日志字段dyn.bar_lowered_before_signal，记录本约束对应的接触或几何违规条件。" + }, + { + "atom": "overhead_without_clean", + "level": "new_code", + "source": "inputs/task/privileged/tests/rfo_task.py:350-385；需新增日志字段dyn.overhead_without_clean，记录本约束对应的接触或几何违规条件。" + } + ] + }, + "native_metrics": [ + { + "name": "bar_height_m", + "unit": "m", + "threshold": null, + "direction": "higher_better", + "note_zh": "相对台面的杠中心高度；过顶阈值随头部中心高度变化，非固定绝对高度。", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:372-374" + ] + }, + { + "name": "bar_max_height_m", + "unit": "m", + "threshold": null, + "direction": "higher_better", + "note_zh": "尝试中最大杠中心高度，不能单独表示上肩或成功。", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:373-375" + ] + }, + { + "name": "elbow_flex_deg", + "unit": "deg", + "threshold": 10.0, + "direction": "lower_better", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:112", + "inputs/task/privileged/tests/rfo_task.py:358,376" + ] + }, + { + "name": "knee_flex_deg", + "unit": "deg", + "threshold": 10.0, + "direction": "lower_better", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:112", + "inputs/task/privileged/tests/rfo_task.py:357,377" + ] + }, + { + "name": "time_s", + "unit": "s", + "threshold": 60.0, + "direction": "lower_better", + "note_zh": "成功时记录的回合耗时；60秒为回合时限，依P21不列约束。", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:103,392" + ] + } + ], + "open_questions_zh": [ + "inputs/states/index.json为空，成功参照也无状态日志；依P8所有when、never和success均为null，无法进行状态条件覆盖及success与评分器一致性的回放自检。atoms仅列出需补记的字段，不将源码中可计算的量冒充已记录字段；视频可判断主要姿态，但无法精确测出全部距离、速度和0.05秒接触宽限。", + "文本与评分器边界差异（N1）：说明写杠铃及骨盆速度低于0.1米/秒，实际final_checks接受小于等于0.1；说明对未上肩先过顶写高出头部中心超过0.25米才失败，实际step复用overhead检查，达到0.25米即失败。本文判据均按实际裁判边界。", + "两模式共享相同任务模块与参数，物理成功条件一致；privileged的grade.py:106-115还要求两次新进程回放均成功且终态哈希一致，standard由服务记录最后一次现场回合的裁判结论（standard/server.py:157-177；efe_standard-files/v2_1/grade.py:93-99）。这些是复现或交付验收，不增加世界状态节点，也不是不同的独立任务场景。", + "本任务关闭通用跌倒判据（rfo_task.py:106），只按列出的接触规则判负；60秒属于回合时限，依P21不列约束。物理数值发散也会使回合失败（rfo/world.py:872-883），属于模拟运行有效性，未增造运动步骤或几何跌倒阈值。" + ], + "conf": 6, + "conf_reason_zh": "两步及唯一依赖直接来自文本，所有验收阈值和接触禁例均有裁判源码依据，两模式参数及任务模块一致，成功视频展示了两步；缺少状态日志，细小几何阈值和保持计数不能独立核验。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "任务说明列出两步（挺到肩、举过顶并保持），DAG恰为这两个节点，边、阈值（0.15/0.25/0.94/0.1/10°/0.05/0.5s/2s）与说明一致，约束均为裁判失败条件且为立即终止；60秒为回合时限，按P21不列约束。无状态日志，when/success为null并已在open_questions_zh说明，文本与评分器边界差异也已列出。未发现需改内容的问题。", + "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：['bar_at_shoulders', 'bar_overhead']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:48:15Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__clean_and_jerk/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__clean_and_jerk/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "78e64800f07070e7ab8a1655f4015342035a2f6c6f03347ac63570590d0bfc89", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/embodied-olympics-fun/clean_and_jerk/trajectories/embodied-olympics-fun-clean-and-jerk-i00-privileged--ukpfjx2.json b/phase2/embodied-olympics-fun/clean_and_jerk/trajectories/embodied-olympics-fun-clean-and-jerk-i00-privileged--ukpfjx2.json new file mode 100644 index 0000000000000000000000000000000000000000..2dd915f799e58169f201a339706da50a2fca2e9d --- /dev/null +++ b/phase2/embodied-olympics-fun/clean_and_jerk/trajectories/embodied-olympics-fun-clean-and-jerk-i00-privileged--ukpfjx2.json @@ -0,0 +1,237 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-clean-and-jerk-i00-privileged--ukpfjx2", + "taskKey": "embodied-olympics-fun/clean_and_jerk", + "dag_sha256": "f530c1d45bf82223dabd69c9d30cf24e240db1d588782640954595693a7015ca", + "outcome": { + "success": 1, + "reward": 1.0, + "source": "catalog", + "raw": { + "success": 1, + "replay_success": 1, + "deterministic": 1, + "n_actions": 1878, + "sim_steps": 1878, + "horizon": 6000, + "episode_over": 1, + "progress": 1.0, + "stages_reached": 4, + "stages_total": 4, + "fell": 0, + "replay_s": 36.08, + "bad_trajectory": 0, + "metric_bar_height_m": 2.108, + "metric_bar_max_height_m": 2.112, + "metric_elbow_flex_deg": 1.1, + "metric_knee_flex_deg": 3.0, + "metric_time_s": 18.78, + "grader_error": 0, + "video_rendered": 1, + "agent_sessions": 193, + "agent_resets": 31, + "agent_replays": 2, + "agent_cost_self_reported": 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/embodied-olympics-fun-clean-and-jerk-i00-privileged--ukpfjx2/replay.mp4", + "duration_s": 18.8, + "fps": 25.0, + "frames": 470, + "sha256": "801b46eb0917a70f13a880efa0b347ee09931dc595571f636092124d7673ef93" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 2, + "final": 2, + "nodes": { + "bar_at_shoulders": { + "method": "visual", + "intervals": [ + [ + 9.6, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 2.8, + "log_step": null, + "text_zh": "观察：视频 1.8–2.4 s，机器人蹲身后两手分别合握银色杠杆的两侧；视频 2.8–5.0 s，杠铃随机器人起身离开两只棕色支架，原支承位置空出，双脚仍在木台上。" + }, + { + "t": 9.6, + "log_step": null, + "text_zh": "观察：视频 8.4–9.6 s，机器人直立伸膝，两手持杠停留于胸肩前方，双脚并列支撑；视频 9.8–10.4 s 才继续向头顶上举。" + }, + { + "t": 9.6, + "log_step": null, + "text_zh": "推断：结合胸肩前停留后继续上举的画面，以及评分器 success=1、replay_success=1、stages_reached=4/4，上肩事件已通过；达成时间估为视频 9.6 s，约有 0.5 s 不确定性。该事件完成后锁存，上举离肩不撤销；肩距、线速度及连续50个控制步无法由单视角画面精确复测。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "机器人左右手各合握杠杆的一侧，共同将杠铃从棕色支架上抬离，起身并提至胸肩前。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "机器人双脚持续接触木色举重台，双腿从屈膝蹲身转为伸膝直立，承载本体及手中杠铃。" + } + ] + }, + "bar_overhead": { + "method": "visual", + "intervals": [ + [ + 18.76, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 11.6, + "log_step": null, + "text_zh": "观察：视频 9.8–11.6 s，机器人将杠铃由胸肩前举至头顶上方，两手始终分握杠杆；视频 11.6–17.2 s，机器人继续持杠过顶，双臂和双腿接近伸直，双脚保持在木台上。" + }, + { + "t": 18.76, + "log_step": null, + "text_zh": "观察：视频 17.6、17.8、18.0、18.2、18.4、18.6 及末帧 18.76 s，杠铃仍在机器人头顶上方，双手未脱离杠杆，机器人直立、双脚并列支撑，未见放杠或倒地。评分器记录 success=1、replay_success=1、deterministic=1、stages_reached=4/4，终态杠高2.108米、肘屈曲1.1度、膝屈曲3.0度。" + }, + { + "t": 18.76, + "log_step": null, + "text_zh": "推断：评分器成功结论与末段过顶姿态一致，说明完整过顶保持及下放信号验收已通过。视频 11.6 s 已出现过顶姿态，但不能据此认定速度阈值及连续200个控制步已满足；完整节点取成功终点附近的末帧视频 18.76 s 为近似达成锚点，终态为真，不把首次举高时刻当作完整达成时刻。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "机器人左右手持续分别握住杠杆两侧，共同从胸肩前上举并在头顶上方承载杠铃至回合结束。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "机器人双脚在木色举重台上并列支撑，双腿及躯干直立承载举过头顶的杠铃。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "bar_height_m", + "unit": "m", + "threshold": null, + "direction": "higher_better", + "value": 2.108 + }, + { + "name": "bar_max_height_m", + "unit": "m", + "threshold": null, + "direction": "higher_better", + "value": 2.112 + }, + { + "name": "elbow_flex_deg", + "unit": "deg", + "threshold": 10.0, + "direction": "lower_better", + "value": 1.1 + }, + { + "name": "knee_flex_deg", + "unit": "deg", + "threshold": 10.0, + "direction": "lower_better", + "value": 3.0 + }, + { + "name": "time_s", + "unit": "s", + "threshold": 60.0, + "direction": "lower_better", + "value": 18.78 + } + ], + "failure": null, + "summary_zh": "机器人完成杠铃上肩及过顶两个目标：约视频 9.6 s 上肩停留后继续上举，视频 11.6 s 已见过顶姿态，并保持持杠至视频 18.76 s 末帧；评分器确认四个原生阶段全部通过，成功且可确定性复现，无失败断点。实际接触为双手分握杠杆、双脚支撑木台。无状态日志，单视角不能精确核验距离、速度及保持计数，因此上肩时间为近似值，过顶完整验收以成功终点为锚点；成功参照仅用于比较姿态，不用于推定本次接触或时间。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "成功运行，两节点均为真，终态与评分器（success=1，stages 4/4，time_s 18.78）一致；上肩约视频 9.6 s、过顶锚点取末帧，帧上相符，接触（双臂协作抓握+腿/躯干支撑）符合画面，无失败断点。", + "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": "2026-10-29T10:04:37Z", + "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/embodied-olympics-fun-clean-and-jerk-i00-privileged--ukpfjx2/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-clean-and-jerk-i00-privileged--ukpfjx2/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "f15cbe96e8b260128c602383c7a51ccbc0939c381eddf9ec7d060804010a8550", + "video_sha256": "801b46eb0917a70f13a880efa0b347ee09931dc595571f636092124d7673ef93", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/clean_and_jerk/trajectories/embodied-olympics-fun-clean-and-jerk-i00-standard--xs9w6sr.json b/phase2/embodied-olympics-fun/clean_and_jerk/trajectories/embodied-olympics-fun-clean-and-jerk-i00-standard--xs9w6sr.json new file mode 100644 index 0000000000000000000000000000000000000000..746315764fae6978821d212a82e8787c9f3133bd --- /dev/null +++ b/phase2/embodied-olympics-fun/clean_and_jerk/trajectories/embodied-olympics-fun-clean-and-jerk-i00-standard--xs9w6sr.json @@ -0,0 +1,230 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-clean-and-jerk-i00-standard--xs9w6sr", + "taskKey": "embodied-olympics-fun/clean_and_jerk", + "dag_sha256": "f530c1d45bf82223dabd69c9d30cf24e240db1d588782640954595693a7015ca", + "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": 197, + "sim_steps": 1478, + "calls": 28, + "rows": 1093, + "resets": 5, + "reset_budget": 5, + "sim_startups": 1, + "elapsed_s": 1726.7, + "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": 0.0, + "peak_memory_gb": 1.268, + "peak_processes": 10, + "peak_disk_gb": 0.369, + "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/embodied-olympics-fun-clean-and-jerk-i00-standard--xs9w6sr/replay.mp4", + "duration_s": 2.0, + "fps": 25.0, + "frames": 50, + "sha256": "9d56149eb906b2f39bb1d612f95b3feb0b91c59db36d51b9cbc3982adbacef11" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "bar_at_shoulders": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察：视频 0.00 s 机器人站在木台中央，双手空着，银色杠铃在两只棕色支架上，低于膝部。" + }, + { + "t": 1.52, + "log_step": 143, + "text_zh": "观察：视频 1.28–1.96 s 机器人由站立逐渐向前倾倒，末段躯干接近水平；杠铃由支架处落向台面，未见双手合握或将杠铃抬至胸肩。log step 143 的最终现场回合只输出 squat fold 和首段 sq[1]，随后返回 failed，没有进入抓杠、站起或上肩保持阶段。" + }, + { + "t": 1.96, + "log_step": null, + "text_zh": "观察：视频 1.96 s 机器人仍在前扑，杠铃处于台面低位；成功参考视频 17.00 s 则清楚显示机器人直立、双手持杠于肩前。本次完整视频未出现该上肩姿态。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：视频 0.00–1.52 s 双脚在木色举重台上承重；上肩准备过程中随后向前失去支撑并扑倒。全片未见双手闭合抓握杠铃。" + } + ] + }, + "bar_overhead": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.96, + "log_step": 143, + "text_zh": "观察：视频 0.00–1.96 s 杠铃始终在膝部附近或更低处，末段机器人扑倒，杠铃未被举到头顶上方；成功参考视频 27.36 s 可清楚看到双臂向上伸展持杠，本次未出现。log step 143 的最终回合在下蹲准备阶段结束，没有执行过顶阶段。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "bar_height_m", + "unit": "m", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "bar_max_height_m", + "unit": "m", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "elbow_flex_deg", + "unit": "deg", + "threshold": 10.0, + "direction": "lower_better", + "value": null + }, + { + "name": "knee_flex_deg", + "unit": "deg", + "threshold": 10.0, + "direction": "lower_better", + "value": null + }, + { + "name": "time_s", + "unit": "s", + "threshold": 60.0, + "direction": "lower_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "bar_at_shoulders", + "t": 1.5, + "log_step": 143, + "where": "操作中", + "evidence_zh": "观察：视频约 1.5 s 机器人在杠铃上肩的下蹲准备中明显向前倾倒，视频 1.88–1.96 s 躯干接近水平，杠铃落向台面，始终没有完成抓杠上肩。log step 143 显示 agent 的最终脚本进入 squat fold 后该回合即 failed；log step 144 的机器人服务状态为 episode=6、control_steps=197、done=true、reason=failed。" + }, + "unmet": { + "bar_at_shoulders": false, + "bar_overhead": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 1.5, + "log_step": 143, + "evidence_zh": "观察：视频 1.28–1.96 s 机器人上身持续前倾，脚部支撑姿态破坏，末段身体前扑至接近水平、双手落向台面，未恢复站立；杠铃未上肩。该现象来自本次最终回合画面，并非先前测试的跌倒。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志 / 视频", + "evidence_zh": "推断：log step 142 所写最终脚本以关节控制执行下蹲并依据骨盆 IMU 做姿态反馈；log step 143 显示该脚本在 squat fold 首段便结束，视频约 1.5–1.96 s 同时显示双脚未能维持直立支撑、机器人向前扑倒。因此本次上肩准备的本体稳定控制未奏效。没有足够记录区分轨迹、反馈律或参数的具体责任。" + } + ], + "gap_zh": "没有状态日志或逐步接触、关节力矩记录；log step 142 的脚本正文中间被截去，最终回合日志只保留首段下蹲输出及 failed，不能确定失稳的具体控制子因或裁判最后触发的接触条件。agent 在 log step 147、153 对 joints 模式自主倒塌或初始小腿碰杠的说法仍是待核实猜测，不据此归为任务或环境问题。评分器未报告 DAG 所列五项原生指标，其 value 均填 null。" + }, + "summary_zh": "本次评分视频对应最后一次现场回合（log step 143–144），机器人在上肩准备时向前失稳，视频约 1.5–1.96 s 扑倒；杠铃未被双手举至肩前，也未过顶，两节点均未达到。现象为全身平衡失败，日志和画面支持归到控制，但缺少状态及接触记录，无法定位具体控制子因或裁判终止接触；前几次会话测试不计入本次进度。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "视频仅 2.0 s：约 1.2 s 前机器人直立，随后前倾扑倒，杠铃留在台面，两节点均未达成；断点 bar_at_shoulders、现象 平衡、根因 控制（未细分）与画面、log id 143–144 及服务状态一致，未发现需修改的问题。", + "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": "2026-10-29T10:04:55Z", + "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/embodied-olympics-fun-clean-and-jerk-i00-standard--xs9w6sr/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-clean-and-jerk-i00-standard--xs9w6sr/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "c1dbe4236d14e9d53869b4471a7f250646cfdecd74776d87052fb6177036fdb9", + "video_sha256": "9d56149eb906b2f39bb1d612f95b3feb0b91c59db36d51b9cbc3982adbacef11", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/discus/task_dag.json b/phase2/embodied-olympics-fun/discus/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..4e82c434324daba1eff3f3d7775b705b777864ed --- /dev/null +++ b/phase2/embodied-olympics-fun/discus/task_dag.json @@ -0,0 +1,341 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "embodied-olympics-fun/discus", + "suite": "embodied-olympics-fun", + "base": "embodied-olympics-fun-discus-i00", + "title_zh": "掷铁饼并合法离圈", + "summary_zh": "铁饼从圈内右手掷出，首次落地在扇形界内且距圈内沿至少15米，机器人随后从后半圈合法离圈并站稳1秒，全程不犯规或跌倒。", + "embodiment_zh": "两种模式均为 Unitree H1-2 双足人形机器人，27个本体关节；右臂末端为 Inspire RH56DFX 五指灵巧手，六驱动、十二手指关节，以真实接触摩擦持饼；左臂末端为无工作手指的拳垫，不能持饼。腿和躯干支撑投掷、后退离圈与站立，无轮式底盘；privileged 可访问仿真状态，standard 通过服务获取相机与本体传感。", + "task_class": "compositional", + "analysis_level": "step", + "spec": { + "version": "1.2", + "task": "embodied-olympics-fun-discus-i00", + "bench": "embodied-olympics-fun", + "progress": "live", + "entities": {}, + "milestones": [ + { + "id": "discus_landed_qualified", + "zh": "铁饼落区达标", + "mode": "event", + "when": null, + "after": [], + "hold_s": 0, + "goal": true, + "criterion_zh": "右手投出的橙红色铁饼已首次触地，落点严格位于两条白色扇形边线之间（半角 17.46 度，压线无效），首次接触点距圈心的水平距离减圈内半径 1.25 米至少为 15.0 米；尚在空中、落在界外或首次落点过近都未完成。场地斜视画面可见铁饼脱离右手向扇形飞去，但远处首次触地点和精确距离可能出画，需结合扇形相机或裁判 landing 阶段及 distance_m；不能以滑滚后的停留位置代替落点。", + "step": { + "cat": "动态", + "item": "抛/投" + }, + "contact": [ + { + "who": "手指/夹爪", + "how": "抓握", + "note_zh": "右侧五指灵巧手以手指和拇指或手掌持住铁饼，摆臂投出；飞行阶段随后无接触。", + "plan": 1 + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚接触投掷圈内地面，腿和躯干支撑并带动右手投掷。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:4", + "inputs/task/privileged/environment/rfo_task.py:154-158", + "inputs/task/privileged/environment/rfo_task.py:220-227", + "inputs/task/privileged/environment/rfo_task.py:290-325", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-158" + ], + "conf": 6, + "note_zh": "按连续动作规则 S10，脱手与飞行不另建节点；release 是 landing 的原生前序阶段。结果取首次落地并锁存，即使铁饼后来滑出扇形也不撤回；接触手段限制及历史犯规保留在约束。" + }, + { + "id": "rear_exit_stood", + "zh": "后圈离圈站稳", + "mode": "final", + "when": null, + "after": [ + "discus_landed_qualified" + ], + "hold_s": 0.99, + "freeze_held": false, + "goal": true, + "criterion_zh": "有效落地之后，机器人首次触及圈沿顶面或圈外地面的接触点在圈心横线后方（圈心坐标 x<0），且此前没有同时出现前半圈接触；随后连续 100 个 0.01 秒控制步只有脚接地、至少一脚接地，躯干竖直方向余弦不小于 0.94，骨盆三维速度不大于 0.35 米/秒。场地相机可见机器人向扇形反方向跨到灰色圆圈后侧绿色地台并保持直立低速；只投出铁饼、尚未后圈接地、仍在明显行走或尚未满足站稳窗口都未完成，窗口内失稳则重新计时。", + "step": { + "cat": "本体", + "item": "行进" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚从圈内向后半圈的圈沿或圈外地台迈出，腿和躯干支撑身体，继而站稳。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:30-35", + "inputs/task/privileged/environment/rfo_task.py:248-275", + "inputs/task/privileged/environment/rfo_task.py:327-329", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:180-184", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:292-301", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:1103-1105" + ], + "conf": 7, + "note_zh": "按 DECISIONS.md D11 当前两步读法，站稳是合法离圈的验收，合并而不另列。final 在有效落地之后评估，初始站立不计分；hold_s=0.99 对应连续100次采样首末间隔，物理窗口仍为评分器所称1.0秒（D31）。exit 一旦锁存不要求全身出圈或持续在圈外，站稳可由单脚接地满足。" + } + ], + "constraints": [ + { + "id": "fall", + "zh": "跌倒", + "never": null, + "criterion_zh": "骨盆世界高度低于 0.50 米，或除左右脚外的机器人部位接触地面，立即判负；录像中可见躯干倒地、手膝撑地等。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:153-158", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:276-282", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/task.py:21-23" + ] + }, + { + "id": "circle_foul", + "zh": "圈沿越界犯规", + "never": null, + "criterion_zh": "铁饼首次落地前，机器人任一部位接触圈沿顶面或圈外地面；或尚未合法离圈时出现前半圈接触点，立即判负。以圈心为原点，圈外地面接触点水平半径至少 1.25 米；圈沿地台接触高度至少 0.017 米算顶面，前半圈含 x=0，首次合法离圈后不再检查此项。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:154-158", + "inputs/task/privileged/environment/rfo_task.py:248-273" + ] + }, + { + "id": "wrong_body_touch", + "zh": "非右手触饼", + "never": null, + "criterion_zh": "从开始到铁饼首次落地前，铁饼接触右手手指、手掌、手腕以外的机器人部位，立即判负；例如左手或前臂碰饼。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:276-280", + "inputs/task/privileged/environment/rfo_task.py:303-306", + "inputs/task/privileged/environment/rfo_task.py:274-275" + ] + }, + { + "id": "unheld_carry", + "zh": "未握而托饼", + "never": null, + "criterion_zh": "尚未确认脱手时，铁饼仍接触右手但不满足原生持握（手指与拇指或手掌同时接触），从最后持握起超过 0.20 秒，立即判负；不能把长时间仅托在掌上当作有效持握。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:154-158", + "inputs/task/privileged/environment/rfo_task.py:281-288" + ] + }, + { + "id": "regripped", + "zh": "脱手后再握", + "never": null, + "criterion_zh": "脱手确认后、首次落地前，右手再次形成对铁饼的持握即判负。脱手时间取最后一次持握时刻，需随后连续 0.03 秒完全不接触机器人才能确认。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:290-304" + ] + }, + { + "id": "late_retouch", + "zh": "迟延触饼", + "never": null, + "criterion_zh": "脱手确认后至首次落地前，右手在距最后持握时刻超过 0.20 秒后再接触铁饼即判负；短暂擦碰仅在这段宽限内允许。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:154-158", + "inputs/task/privileged/environment/rfo_task.py:303-306" + ] + }, + { + "id": "second_propulsion", + "zh": "脱手后再加速", + "never": null, + "criterion_zh": "脱手后允许的右手擦碰所产生接触冲量的模之和除以铁饼质量，累计大于 0.50 米/秒即判负；普通录像无法量出冲量，须看裁判接触记录，不能凭飞行外观推断。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:41-58", + "inputs/task/privileged/environment/rfo_task.py:307-312" + ] + }, + { + "id": "invalid_landing", + "zh": "落点不合格", + "never": null, + "criterion_zh": "铁饼第一次接地若在扇形界线外或界线上、距圈内沿不足 15.0 米，立即判负；扇形以初始朝向为中线、总角度 34.92 度，距离取首次地面接触点到圈心的水平距离减 1.25 米。后续弹跳或滑动达到距离不补救首次落点。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:154-158", + "inputs/task/privileged/environment/rfo_task.py:313-325" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "qualified_landing", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:290-325；需记录 ref.stages 的 release/landing、首次接地点和 ref.metrics.distance_m。" + }, + { + "atom": "legal_rear_exit", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:248-273；需记录 task.trackers.exit、ref.stages.exit 和首次圈外接触点。" + }, + { + "atom": "standing_after_exit", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:180-184,292-301；需记录 exit 后 ref.standing()、ref._hold.stand、脚部接地、躯干方向余弦与骨盆速度。" + }, + { + "atom": "episode_success", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:174-184；需逐帧记录原生 ref.success、ref.failed 和 ref.stages；不可用离线交付判定替代。" + }, + { + "atom": "fall", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:153-158；inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:276-282；inputs/grader/efe_embodied-olympics-fun-files/rfo/task.py:21-23；需记录此规则的逐帧原生接触/持握/落点判据或 ref.failed 的对应原因。" + }, + { + "atom": "circle_foul", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:154-158；inputs/task/privileged/environment/rfo_task.py:248-273；需记录此规则的逐帧原生接触/持握/落点判据或 ref.failed 的对应原因。" + }, + { + "atom": "wrong_body_touch", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:276-280；inputs/task/privileged/environment/rfo_task.py:303-306；inputs/task/privileged/environment/rfo_task.py:274-275；需记录此规则的逐帧原生接触/持握/落点判据或 ref.failed 的对应原因。" + }, + { + "atom": "unheld_carry", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:154-158；inputs/task/privileged/environment/rfo_task.py:281-288；需记录此规则的逐帧原生接触/持握/落点判据或 ref.failed 的对应原因。" + }, + { + "atom": "regripped", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:290-304；需记录此规则的逐帧原生接触/持握/落点判据或 ref.failed 的对应原因。" + }, + { + "atom": "late_retouch", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:154-158；inputs/task/privileged/environment/rfo_task.py:303-306；需记录此规则的逐帧原生接触/持握/落点判据或 ref.failed 的对应原因。" + }, + { + "atom": "second_propulsion", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:41-58；inputs/task/privileged/environment/rfo_task.py:307-312；需记录此规则的逐帧原生接触/持握/落点判据或 ref.failed 的对应原因。" + }, + { + "atom": "invalid_landing", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:154-158；inputs/task/privileged/environment/rfo_task.py:313-325；需记录此规则的逐帧原生接触/持握/落点判据或 ref.failed 的对应原因。" + } + ] + }, + "native_metrics": [ + { + "name": "distance_m", + "unit": "m", + "threshold": 15.0, + "direction": "higher_better", + "note_zh": "首次接地点到圈心的水平距离减1.25米；界内条件另验。", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:313-325" + ] + }, + { + "name": "retouch_dv_m_s", + "unit": "m/s", + "threshold": 0.5, + "direction": "lower_better", + "note_zh": "脱手后短时擦碰的累计冲量除以铁饼质量；超过阈值立即判负。", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:307-312" + ] + } + ], + "open_questions_zh": [ + "inputs/states/index.json 的 logs 为空，所有节点 when、约束 never 及 success 均为 null（暂行：DECISIONS.md P8）；所列 new_code 原子是所需补充日志，不能当作已存在字段。首次落点距离、0.03秒无接触间隙和接触冲量不能仅由此斜视录像精确量出。", + "投掷落地与离圈站稳的粒度仍属开放读法；按 DECISIONS.md D11 当前处理保留两步：铁饼有效落地达标 → 后半圈合法离圈并站稳，不拆脱手、飞行或另建站稳步骤。", + "文本与评分器边界：文本说离圈后站稳，代码只在落地后第一次后半圈圈沿或圈外接触时锁存 exit，不检查全身离圈，也不检查站稳时仍在圈外（rfo_task.py:248-275,327-329）；依评分器设置验收，不额外添加全身出圈条件。", + "文本与评分器边界：文本概括禁止脱手后再次持握或违规触饼，但任务 substep 在首次落地后直接返回，不再检查铁饼接触规则（rfo_task.py:274-275）；这些约束按实际检查窗口截至首次落地。", + "两种模式共用相同任务模块、冻结实例和原生裁判物理成功条件；privileged 另要求两个新进程重放均成功且末态哈希一致（grade.py:105-115），standard 直接评判最后一次实时试验（standard/server.py:165-180）。这些交付和复现条件不构成世界状态步骤，success 无日志时保持 null。", + "release_time_s、release_speed_m_s、release_height_m、release_tilt_deg、release_spin_rad_s 是原生诊断指标而无达标阈值（rfo_task.py:297-301）；不据成功录像的数值新增约束或调整投掷门槛。", + "成功参考录像长6.84秒；run.json 的最终生成/验证记录（steps id=193、200）与 reward 一致：首次落地16.267米、原生阶段齐全、最终683控制步成功。仅用于核对目标与画面，不将其摆臂姿势、步法或阶段时刻写成必需条件。" + ], + "conf": 6, + "conf_reason_zh": "任务文本、两种模式的同一任务代码和裁判支持两步及必要顺序，阈值均来自评分器；成功录像与最终记录相符。但没有状态日志，远处首次落点与接触细节不能逐帧独立核验，投掷项目的两步粒度仍按D11暂行。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "两步（铁饼落区达标 → 后半圈离圈站稳）与文本一致；成功条件、各项失败约束与评分器相符；无状态日志故 when 均为 null（暂行 P8），criterion_zh 足以判读。60 秒限时属任务 horizon，按 P21 不列约束。未发现需要改动的内容。", + "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：['discus_landed_qualified', 'rear_exit_stood']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:43:04Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__discus/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__discus/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "d32f59528d475f7809d97847ab97c1d2cfd38d04bc47b55269a9e15d645ff903", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/embodied-olympics-fun/discus/trajectories/embodied-olympics-fun-discus-i00-privileged--bk3sdpn.json b/phase2/embodied-olympics-fun/discus/trajectories/embodied-olympics-fun-discus-i00-privileged--bk3sdpn.json new file mode 100644 index 0000000000000000000000000000000000000000..87f617ba8331a85bff0ec713f232d3238293c443 --- /dev/null +++ b/phase2/embodied-olympics-fun/discus/trajectories/embodied-olympics-fun-discus-i00-privileged--bk3sdpn.json @@ -0,0 +1,241 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-discus-i00-privileged--bk3sdpn", + "taskKey": "embodied-olympics-fun/discus", + "dag_sha256": "ffb63c003bdbc581b50872ffe890e6ce579309876e24b407597d79050cb48e95", + "outcome": { + "success": 1, + "reward": 1.0, + "source": "catalog", + "raw": { + "success": 1, + "replay_success": 1, + "deterministic": 1, + "n_actions": 775, + "sim_steps": 775, + "horizon": 6000, + "episode_over": 1, + "progress": 1.0, + "stages_reached": 3, + "stages_total": 3, + "fell": 0, + "replay_s": 8.23, + "bad_trajectory": 0, + "metric_distance_m": 17.989, + "metric_release_height_m": 1.729, + "metric_release_speed_m_s": 13.51, + "metric_release_spin_rad_s": 14.8, + "metric_release_tilt_deg": 70.7, + "metric_release_time_s": 1.576, + "grader_error": 0, + "video_rendered": 1, + "agent_sessions": 51, + "agent_resets": 1670, + "agent_replays": 2, + "agent_cost_self_reported": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 2.198, + "peak_processes": 18, + "peak_disk_gb": 0.256, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0, + "session_log_tampered": 0, + "session_log_from_monitor": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-discus-i00-privileged--bk3sdpn/replay.mp4", + "duration_s": 7.76, + "fps": 25.0, + "frames": 194, + "sha256": "682b4086990dfb6e9b30e2a9eb53299d91357e91e7058093f98dd2b55e3a2eaf" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 2, + "final": 2, + "nodes": { + "discus_landed_qualified": { + "method": "visual", + "intervals": [ + [ + 3.5, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察：视频 0.00、0.20、0.60、1.00 s，橙红色铁饼随机器人的右手从身前摆至身体右后侧，双脚在灰色圈内支撑；右手始终贴住饼，左手未参与持饼。" + }, + { + "t": 1.68, + "log_step": null, + "text_zh": "观察：视频 1.52–1.68 s，机器人转体、右臂向扇形方向挥出，视频 1.68 s 铁饼与右手明显分离并升空；视频 1.84–2.00 s 铁饼继续远离机器人。" + }, + { + "t": 3.5, + "log_step": 52, + "text_zh": "观察：log step 52 的最终生成并保存记录中，模拟裁判报告 landing 事件在仿真 3.503 s、distance_m=17.989；log step 53 对该交付轨迹独立重放报告 success=true、release/landing/exit 齐全、775 个控制步、仿真 time_s=7.75，评分器 reward 的距离和阶段结果一致。视频 3.12–3.92 s 已检查，远处首次触地点太小，不能直接量出距离或判断是否压线。" + }, + { + "t": 3.5, + "log_step": 53, + "text_zh": "推断：log step 28 的控制循环用每步 0.01 s，交付轨迹 775 步×0.01 s=7.75 s，与视频总长 7.76 s 相符；log step 52 的仿真脱手 1.576 s 与视频 1.60–1.68 s 脱手画面、仿真离圈 6.628 s 与视频 6.60–6.68 s 后圈接地画面相符，因此将该交付轨迹的仿真 landing 3.503 s 映射为视频约 3.50 s。结合原生落地阶段、17.989 米和成功判定，首次落地达标并锁存；不是以滑滚后的停留位置判定。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "抓握", + "note_zh": "右手持住橙红色铁饼随摆臂运动，随后张手投出（视频 0.00–1.68 s）。结合初始手指扣住饼沿的任务初态和无犯规的原生 release 判定，右手为掌指持握；没有双手协作。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "投掷过程中双脚在圈内接地承重，腿和躯干支撑右臂摆动及转体，投出后仍直立（视频 0.00–3.92 s）。" + } + ] + }, + "rear_exit_stood": { + "method": "visual", + "intervals": [ + [ + 7.75, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 6.68, + "log_step": 52, + "text_zh": "观察：视频约 5.60–6.68 s，机器人朝扇形反方向小步后退，视频 6.60–6.68 s 后侧脚踩到灰色圆圈后边缘及绿色地台交界；log step 52 的最终保存记录中，模拟裁判在仿真 6.628 s 锁存 exit。" + }, + { + "t": 7.72, + "log_step": null, + "text_zh": "观察：已逐段检查视频 6.20–7.72 s，机器人在后圈边缘逐渐停住，视频 7.00–7.72 s 保持直立、脚部承重，未见手膝或躯干撑地；末帧仍在圈边，而非全身远离圆圈。成功参照同样展示向后离圈后站立，但其步幅不作为本次达标条件。" + }, + { + "t": 7.75, + "log_step": 54, + "text_zh": "推断：沿用交付轨迹控制步×0.01 s 与视频的映射（775 步对应仿真 7.75 s，视频长 7.76 s；已用视频脱手及后圈接地核对），log step 52、54 的 success 事件在仿真 7.75 s，对应视频约 7.75 s 的站稳验收。log step 54 重新生成的 775 行动作与已验证文件逐行一致，log step 55 两次重放均成功且末态哈希一致；结合 reward.success=1、replay_success=1、deterministic=1，后圈离圈并站稳节点最终为真。精确速度、躯干余弦和连续 100 步窗口依赖原生裁判结果，不能由斜视画面独立测量。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "机器人双脚交替接地向后小步移动，后侧脚触及圈后边缘及绿色地台，随后以脚部支撑保持直立（视频约 5.60–7.72 s）。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "distance_m", + "unit": "m", + "threshold": 15.0, + "direction": "higher_better", + "value": 17.989 + }, + { + "name": "retouch_dv_m_s", + "unit": "m/s", + "threshold": 0.5, + "direction": "lower_better", + "value": null + } + ], + "failure": null, + "summary_zh": "本次交付轨迹完成投掷有效落地与后圈离圈站稳两个节点：首次落地约在视频 3.50 s，原生距离为 17.989 米，约在视频 7.75 s 通过最终站稳验收。无状态日志，远处首次接地点、手指接触细节及精确站稳阈值不能仅凭画面独立核验，相关结论结合与交付文件一致的原生裁判记录及确定性成功重放。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 2, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "fix", + "conf": 7, + "summary_zh": "两个节点的值与时间（落地约 3.50 s，站稳约 7.75 s）与视频、评分器字段一致，接触标注合理；唯一问题是 summary_zh 中无依据的“任务设计问题”断言。", + "n_issues": 1, + "n_major": 0 + }, + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 2, + "verdict": "agree", + "conf": 8, + "summary_zh": "两个节点（落地约视频 3.50 s、后圈离圈站稳约 7.75 s）与评分器字段（3/3 阶段、distance 17.989、fell 0）及视频一致，接触标注合理，时钟映射有核对；第 1 轮的 summary 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": "2026-10-29T10:07:24Z", + "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/embodied-olympics-fun-discus-i00-privileged--bk3sdpn/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-discus-i00-privileged--bk3sdpn/04_review_r1" + }, + { + "stage": "reconcile", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-discus-i00-privileged--bk3sdpn/05_reconcile_r1" + }, + { + "stage": "traj_review", + "round": 2, + "ok": true, + "space": "work/runs/embodied-olympics-fun-discus-i00-privileged--bk3sdpn/04_review_r2" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "0c9f5162643bb92536ebc00147d37d4cfe449aad3b9a203e3e6e207b407a69d9", + "video_sha256": "682b4086990dfb6e9b30e2a9eb53299d91357e91e7058093f98dd2b55e3a2eaf", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/discus/trajectories/embodied-olympics-fun-discus-i00-privileged--p2mfyzi.json b/phase2/embodied-olympics-fun/discus/trajectories/embodied-olympics-fun-discus-i00-privileged--p2mfyzi.json new file mode 100644 index 0000000000000000000000000000000000000000..556f822d73e1d9eebe02a2659a671adba0d5c92e --- /dev/null +++ b/phase2/embodied-olympics-fun/discus/trajectories/embodied-olympics-fun-discus-i00-privileged--p2mfyzi.json @@ -0,0 +1,222 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-discus-i00-privileged--p2mfyzi", + "taskKey": "embodied-olympics-fun/discus", + "dag_sha256": "ffb63c003bdbc581b50872ffe890e6ce579309876e24b407597d79050cb48e95", + "outcome": { + "success": 0, + "reward": 0.3333, + "source": "catalog", + "raw": { + "success": 0, + "replay_success": 0, + "deterministic": 1, + "n_actions": 1998, + "sim_steps": 1998, + "horizon": 6000, + "episode_over": 1, + "progress": 0.3333, + "stages_reached": 1, + "stages_total": 3, + "fell": 0, + "replay_s": 14.03, + "bad_trajectory": 0, + "metric_distance_m": 10.729, + "metric_release_height_m": 1.832, + "metric_release_speed_m_s": 9.83, + "metric_release_spin_rad_s": 18.0, + "metric_release_tilt_deg": 83.0, + "metric_release_time_s": 18.592, + "grader_error": 0, + "video_rendered": 1, + "agent_sessions": 356, + "agent_resets": 863, + "agent_replays": 13, + "agent_cost_self_reported": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 1.48, + "peak_processes": 13, + "peak_disk_gb": 0.842, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0, + "session_log_tampered": 0, + "session_log_from_monitor": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-discus-i00-privileged--p2mfyzi/replay.mp4", + "duration_s": 20.0, + "fps": 25.0, + "frames": 500, + "sha256": "44e97a43dd31fd7e922f3f76a7d026d56138cc22a4b408e1e95c2420682638ff" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "discus_landed_qualified": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 18.72, + "log_step": null, + "text_zh": "观察：视频 0–18.48 s 机器人右手持橙红色铁饼在圈内转向、移动；视频 18.56–18.72 s 机器人摆动右臂，铁饼离开右手飞向扇形。视频 19.04–19.96 s 机器人仍直立留在圈内，远处铁饼逐渐下降；该斜视画面不能精确量出首次接地点距离。" + }, + { + "t": 19.96, + "log_step": 326, + "text_zh": "观察：log step 325 重新生成交付文件，共1998行动行；log step 326 对该文件的新进程重放由原生裁判报告 distance_m=10.729、stages=[release]，失败原因为投掷距离不足。log step 328 的第二次重放距离和末态哈希完全相同，与 run.json reward 的 metric_distance_m=10.729、deterministic=1 一致。" + }, + { + "t": 19.96, + "log_step": 328, + "text_zh": "推断：交付文件的两次重放均只达到脱手阶段，首次落地10.729米小于15.0米门槛，且视频中确有对应的唯一投掷，因此铁饼落区达标节点从未成立。原生 progress=0.3333 计入脱手阶段，不代表本DAG的两个语义节点已完成一个。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "抓握", + "note_zh": "推断：视频 0–18.48 s 橙红色铁饼一直随右手移动，结合任务初态所述右手手指扣住饼缘、手掌承托及裁判认可 release，判为右手持握后摆臂投出；视频 18.72 s 铁饼已与右手分离。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：视频 0–18.72 s 机器人双腿支撑圈内转身、挪步及投掷动作，双脚与圈内地面交替接触。" + } + ] + }, + "rear_exit_stood": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 19.96, + "log_step": 326, + "text_zh": "观察：视频 18.72–19.96 s 机器人投出后仍站在灰色圆圈内，未见向后跨到圈外绿色地台；log step 326、328 对交付文件的原生裁判输出均只有 release，没有 landing 或 exit 阶段。" + }, + { + "t": 19.96, + "log_step": 328, + "text_zh": "推断：有效落地前置从未完成，视频结束前也未发生后圈离圈，故初始及投后圈内直立不计为离圈站稳，该节点未成立且未开始尝试。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "distance_m", + "unit": "m", + "threshold": 15.0, + "direction": "higher_better", + "value": 10.729 + }, + { + "name": "retouch_dv_m_s", + "unit": "m/s", + "threshold": 0.5, + "direction": "lower_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "discus_landed_qualified", + "t": 18.7, + "log_step": 326, + "where": "操作中", + "evidence_zh": "观察：视频 18.56–18.72 s 机器人在圈内摆右臂投出铁饼，视频 19.96 s 结束时仍在圈内；log step 326、328 对实际交付文件的裁判重放均以首次落地仅10.729米、投掷距离不足结束。推断：该次投掷未把铁饼送至15米合格落区，首次落地后不可补救，断点为铁饼落区达标；视频 18.7 s 是出手动作的粗略导航锚点。" + }, + "unmet": { + "discus_landed_qualified": false, + "rear_exit_stood": false + }, + "phenomena": [ + { + "cat": "运动", + "item": "位姿不准", + "t": 18.7, + "log_step": 326, + "evidence_zh": "观察：视频 18.56–18.72 s 铁饼从机器人右手向扇形飞出，视频末秒铁饼在远处下降；log step 326、328 的原生裁判均报告首次落地距离10.729米、未达到15米合格线。推断：这次抛投落点过近，未命中合格落区，属于投掷结果的位置偏差。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志 / 视频 / 评分器字段", + "evidence_zh": "观察：log step 319 所示生成脚本用行走接近、短时转身及躯干和右臂的低层关节命令完成投掷；log step 324 将髋部扭矩脉冲更新为选中的候选，log step 325 重新生成交付文件。视频 18.56–18.72 s 机器人确实执行摆臂出手，log step 326、328 的文件重放均测得出手速度9.83米/秒、出手高度1.832米、首次落地10.729米。推断：所实现的投掷控制未产生足以到达合格落区的出手运动；证据可支持控制层面归因，但不能区分关节轨迹、跟踪、接触几何或某一参数的具体责任。" + } + ], + "gap_zh": "没有状态日志；远处铁饼的首次接地点和精确距离无法从该斜视视频独立量测，距离以交付文件两次一致重放的原生裁判结果及 reward 为据。手指接触细节分辨率有限，抓握结合初态、持饼连续画面和裁判认可脱手推断；缺少出手前逐步接触与关节跟踪记录，不能进一步定位控制子类。retouch_dv_m_s 未在评分字段报告，保留 null。" + }, + "summary_zh": "机器人在圈内持饼转向并于视频约18.6秒摆臂投出；交付文件的两次独立重放均判首次落地10.729米，低于15米门槛，因此本DAG最终完成0/2节点。断点为铁饼落区达标，现象为投掷落点过近，根因归于投掷控制；视频结束前机器人仍留在圈内，后圈离圈未开始。无状态日志，精确落地距离依靠一致的原生裁判重放结果，具体控制子类无法确定。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "标签与证据一致：视频 18.56–18.72 s 出手（与 release_time 18.592 对应），19.96 s 机器人仍在圈内，两节点均未达成，final=0，断点与现象有裁判数据（10.729 m < 15 m）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": "2026-10-29T10:00:10Z", + "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/embodied-olympics-fun-discus-i00-privileged--p2mfyzi/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-discus-i00-privileged--p2mfyzi/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "c259d9b37f6fb27719c8150dbc7bf3e3a63a73551d4a64b5fdbc47f38f1896ae", + "video_sha256": "44e97a43dd31fd7e922f3f76a7d026d56138cc22a4b408e1e95c2420682638ff", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/dunk_between_legs/task_dag.json b/phase2/embodied-olympics-fun/dunk_between_legs/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..6613f542ac99340802a66540104fbf8a9b33a9e4 --- /dev/null +++ b/phase2/embodied-olympics-fun/dunk_between_legs/task_dag.json @@ -0,0 +1,419 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "embodied-olympics-fun/dunk_between_legs", + "suite": "embodied-olympics-fun", + "base": "embodied-olympics-fun-dunk-between-legs-i00", + "title_zh": "胯下运球后扣篮", + "summary_zh": "先完成一次合法胯下反弹并双手收球，再以至多两步把受手控制的篮球自上方扣穿篮圈，松手后落地站稳一秒且无有效违例。", + "embodiment_zh": "两种模式均为Unitree H1-2双足人形机器人，27个身体关节；左右臂各7关节，末端均为Inspire RH56DFX五指灵巧手，每手6驱动，接触与摩擦持球，无焊接；双足行走与蹬地落地，无轮式底盘。", + "task_class": "compositional", + "analysis_level": "step", + "spec": { + "version": "1.2", + "task": "embodied-olympics-fun-dunk-between-legs-i00", + "bench": "embodied-olympics-fun", + "progress": "live", + "entities": {}, + "milestones": [ + { + "id": "between_legs_bounce", + "zh": "胯下运球", + "mode": "event", + "when": null, + "after": [], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "场馆侧视中篮球离开双手后触地反弹，球心在无手接触时穿过双脚足底与胯部构成且面积至少0.02平方米的三角门，穿越点不高于较低膝关节，球自身沿穿越方向的法向速度至少1.0米/秒；到下一次手触球时，球仍在上一次触球时所在平面的另一侧，且最后一次穿越须符合上述条件。仅在腿旁弹跳、手持球穿腿、机器人跨过原地球、穿过后又返回，或尚未再次触球确认，均未完成；合法运球的首次触地包含在本步内。", + "step": { + "cat": "动态", + "item": "运球" + }, + "contact": [ + { + "who": "手指/夹爪", + "how": "推/扫", + "note_zh": "任一只灵巧手的手掌或手指把篮球推向地面；双手不可在运球中同时触球，左右手的先后接触不是同时的双臂协作", + "plan": 1 + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双足支撑躯干，可站在原地或随篮球移动；与手掌推球同时出现", + "plan": 1 + }, + { + "who": "腿/躯干", + "how": "避开", + "note_zh": "双腿与篮球保持无接触，让篮球从两腿间穿过", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:4", + "inputs/task/privileged/environment/rfo_task.py:266-311", + "inputs/task/privileged/environment/rfo_task.py:373-377", + "inputs/task/privileged/environment/rfo_task.py:574-584", + "inputs/task/privileged/environment/rfo_task.py:612-668", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159" + ], + "conf": 6, + "note_zh": "开始运球是胯下运球的起始阶段，不另列球离手或首次触地节点；完成后可在原地继续合法运球，再择时双手收球。原生阶段“between the legs”锁存，后续不要求持续胯下运球。" + }, + { + "id": "ball_gathered", + "zh": "双手收球", + "mode": "event", + "when": null, + "after": [ + "between_legs_bounce" + ], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "至少一次胯下反弹确认后，场馆侧视及头部视角可见左右手同时直接接触并接住篮球，球不再自由反弹，裁判进入收球后的持球阶段。开局的胸前双手持球不算本步，只有单手接触或尚未完成胯下反弹即双手接球均不算；本步按当时双手实际同时接触验收，之后允许换为单手控球，但球心入圈前不得完全脱手。", + "step": { + "cat": "动态", + "item": "接住" + }, + "contact": [ + { + "who": "双臂协作", + "how": "其他", + "note_zh": "候选接触名：双掌夹持；左右张开的手掌从篮球两侧夹住球，单手未包住球，不宜标抓握，现有接触表没有双掌侧向夹持项", + "plan": 1 + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "典型方案在双足支撑躯干时，双掌接住篮球", + "plan": 1 + }, + { + "who": "双臂协作", + "how": "托/舀", + "note_zh": "替代方案为左右手从下方及侧下方共同承接反弹的篮球，两手同时触球", + "plan": 2 + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "典型托球方案中双足支撑躯干，与双手承接篮球同时出现", + "plan": 2 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:4", + "inputs/task/privileged/environment/rfo_task.py:656-662", + "inputs/task/privileged/environment/rfo_task.py:585-610" + ], + "conf": 6, + "note_zh": "双手收球是文字独立要求且裁判单独验收的接住状态，不是一般搬运前的抓取准备；可以持球停顿后再完成扣篮。用event锁存曾经合法收球，不能用开局持球或末帧持球代替。" + }, + { + "id": "dunk_landed", + "zh": "扣篮落地站稳", + "mode": "final", + "when": null, + "after": [ + "ball_gathered" + ], + "hold_s": 0.99, + "freeze_held": false, + "goal": true, + "criterion_zh": "场馆侧视中球由手控制从高2.39米、内半径0.225米的篮圈上方向下进入，入圈瞬间须双手在球上，或至少一手的最近接触点高于球心至少0.06米；入圈判定允许至多0.03秒接触间隙，球不向上退回并继续降至圈面下一个球直径0.239米处，那里球心距篮圈轴小于0.3445米，且在球心入圈后至站稳前曾出现完全离手记录。继而连续100个10毫秒控制步可见机器人已有脚着地、着地部位仅限双足、躯干倾斜余弦至少0.94、骨盆速度至多0.35米/秒，才算完成；只扣入尚未站稳、抛球入篮、仍未出现离手记录或落地跌倒均未完成，过程中必须遵守收球后步数及控球约束。", + "step": { + "cat": "关系", + "item": "放入" + }, + "contact": [ + { + "who": "双臂协作", + "how": "其他", + "note_zh": "候选接触名：双掌夹持；左右手从篮球两侧持续夹持并向下送球，单手没有包住球，现有表缺少双掌侧夹项", + "plan": 1 + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双足在双掌持球助跑和蹬地时支撑躯干，落地后的足底承重见步骤说明", + "plan": 1 + }, + { + "who": "手指/夹爪", + "how": "推/扫", + "note_zh": "替代方案用一只灵巧手在篮球上方接触并向下压送，最近接触点高于球心至少0.06米", + "plan": 2 + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双足在单手控球和起跳蹬地时支撑躯干，随后脚部完成落地承重", + "plan": 2 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:370-377", + "inputs/task/privileged/environment/rfo_task.py:585-610", + "inputs/task/privileged/environment/rfo_task.py:670-678", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:180-184", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:292-301", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:429-457" + ], + "conf": 6, + "note_zh": "按S10把扣篮、离手、落地与站稳视为一次连续动作的整体验收，不把空中入圈和落地拆开；腿脚着地支撑与松手后的站稳依次发生。球受手控制穿入圈而非离手后飞行命中，故按直接造成的容纳关系标关系·放入；双手或单手上压均是明文允许方案。final防止开局站立获分；100个控制采样的时间跨度为0.99秒，hold_s按D31填写，未来若直接记录原生保持计数则应比较计数而不重复保持。" + } + ], + "constraints": [ + { + "id": "fell", + "zh": "跌倒", + "never": null, + "criterion_zh": "任意时刻骨盆高度低于0.35米，或除左右脚以外的机器人部位接触地面即失败；裁判在10毫秒控制步末检查。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:17", + "inputs/task/privileged/environment/rfo_task.py:367-377", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:272-282", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/task.py:21-23" + ] + }, + { + "id": "out_of_bounds", + "zh": "出界", + "never": null, + "criterion_zh": "篮球或机器人任意部位与地面的接触点落在端线或边线上及外侧即违禁；本场端线内沿为x=6.0米，边线内沿为y=正负7.5米。代码在扣穿阶段以前执行此检查，扣穿后提前返回而漏查，见开放问题。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:26", + "inputs/task/privileged/environment/rfo_task.py:105-106", + "inputs/task/privileged/environment/rfo_task.py:370-372", + "inputs/task/privileged/environment/rfo_task.py:509-510", + "inputs/task/privileged/environment/rfo_task.py:553-561" + ] + }, + { + "id": "travelling", + "zh": "走步", + "never": null, + "criterion_zh": "开始运球前，未抬脚时任一脚或已选定的轴心脚不得滑动超过0.20米，球未脱手前轴心脚不得离地；收球后至球扣穿篮圈最多两步、同一脚不得连续落地两次且着地脚不得滑动超过0.20米。离地持续至少0.03秒后再触地算一步，两脚在0.03秒内落地合并为一步，扣穿之前第三次落地立即失败。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:112-213", + "inputs/task/privileged/environment/rfo_task.py:376", + "inputs/task/privileged/environment/rfo_task.py:551-571", + "inputs/task/privileged/environment/rfo_task.py:586-600", + "inputs/task/privileged/environment/rfo_task.py:659-660" + ] + }, + { + "id": "illegal_dribble", + "zh": "运球违例", + "never": null, + "criterion_zh": "开始运球前篮球仍在手中却触地即失败；运球期间两次分离至少0.03秒的手触球之间必须触地，一段单手触球不得超过0.30秒，篮球不得碰到手以外的机器人部位，球心与骨盆的水平距离不得超过2.0米。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:373-376", + "inputs/task/privileged/environment/rfo_task.py:567-575", + "inputs/task/privileged/environment/rfo_task.py:612-623", + "inputs/task/privileged/environment/rfo_task.py:640-668", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:304-347" + ] + }, + { + "id": "premature_gather", + "zh": "提前收球", + "never": null, + "criterion_zh": "尚未确认至少一次胯下反弹时左右手同时接触球、结束运球，立即失败。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:646-662" + ] + }, + { + "id": "lost_after_gather", + "zh": "收球后失控", + "never": null, + "criterion_zh": "收球后到球心从上方进入篮圈前不得完全离手，收球后扣穿前不得触地；入圈后允许离手，但离手球若未扣穿就触地仍失败。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:585-610" + ] + }, + { + "id": "illegal_basket", + "zh": "入篮方式违例", + "never": null, + "criterion_zh": "运球开始前进篮、运球中无手控球就进篮、球由下向上穿入篮圈，或扣穿时入圈记录显示既无双手触球也无从上方压球的单手，均失败；入圈单手最近接触点须至少高于球心0.06米。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:373", + "inputs/task/privileged/environment/rfo_task.py:562-569", + "inputs/task/privileged/environment/rfo_task.py:588-600", + "inputs/task/privileged/environment/rfo_task.py:614-615" + ] + }, + { + "id": "basket_hanging", + "zh": "过久触篮架", + "never": null, + "criterion_zh": "机器人任何部位接触篮圈、连接板或篮板的累计物理步时间超过0.30秒即失败；这是累计接触而非单次连续接触。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:376", + "inputs/task/privileged/environment/rfo_task.py:487", + "inputs/task/privileged/environment/rfo_task.py:546-550" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "between_legs_bounce", + "level": "new_code", + "source": "需记录裁判ref.stages中的between the legs及Task.trackers的cross、floor_since_hand、between；inputs/task/privileged/environment/rfo_task.py:624-651" + }, + { + "atom": "ball_gathered", + "level": "new_code", + "source": "需记录裁判ref.stages中的gather和当次左右手实际接触；inputs/task/privileged/environment/rfo_task.py:656-662" + }, + { + "atom": "dunk_landed", + "level": "new_code", + "source": "需记录裁判ref.stages中的dunk、Task.trackers.let_go、Referee._hold[stand]及standing布尔值；inputs/task/privileged/environment/rfo_task.py:670-678；inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:180-184" + }, + { + "atom": "task_success", + "level": "new_code", + "source": "需记录原生每帧Referee.success及Referee.failed，不能绑定外部交付或复现判决；inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:174-178" + }, + { + "atom": "fell", + "level": "new_code", + "source": "需记录原生每帧Referee.failed及违规事件或其物理谓词；inputs/task/privileged/environment/rfo_task.py:17；inputs/task/privileged/environment/rfo_task.py:367-377；inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:272-282；inputs/grader/efe_embodied-olympics-fun-files/rfo/task.py:21-23" + }, + { + "atom": "out_of_bounds", + "level": "new_code", + "source": "需记录原生每帧Referee.failed及违规事件或其物理谓词；inputs/task/privileged/environment/rfo_task.py:26；inputs/task/privileged/environment/rfo_task.py:105-106；inputs/task/privileged/environment/rfo_task.py:370-372；inputs/task/privileged/environment/rfo_task.py:509-510；inputs/task/privileged/environment/rfo_task.py:553-561" + }, + { + "atom": "travelling", + "level": "new_code", + "source": "需记录原生每帧Referee.failed及违规事件或其物理谓词；inputs/task/privileged/environment/rfo_task.py:112-213；inputs/task/privileged/environment/rfo_task.py:376；inputs/task/privileged/environment/rfo_task.py:551-571；inputs/task/privileged/environment/rfo_task.py:586-600；inputs/task/privileged/environment/rfo_task.py:659-660" + }, + { + "atom": "illegal_dribble", + "level": "new_code", + "source": "需记录原生每帧Referee.failed及违规事件或其物理谓词；inputs/task/privileged/environment/rfo_task.py:373-376；inputs/task/privileged/environment/rfo_task.py:567-575；inputs/task/privileged/environment/rfo_task.py:612-623；inputs/task/privileged/environment/rfo_task.py:640-668；inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:304-347" + }, + { + "atom": "premature_gather", + "level": "new_code", + "source": "需记录原生每帧Referee.failed及违规事件或其物理谓词；inputs/task/privileged/environment/rfo_task.py:646-662" + }, + { + "atom": "lost_after_gather", + "level": "new_code", + "source": "需记录原生每帧Referee.failed及违规事件或其物理谓词；inputs/task/privileged/environment/rfo_task.py:585-610" + }, + { + "atom": "illegal_basket", + "level": "new_code", + "source": "需记录原生每帧Referee.failed及违规事件或其物理谓词；inputs/task/privileged/environment/rfo_task.py:373；inputs/task/privileged/environment/rfo_task.py:562-569；inputs/task/privileged/environment/rfo_task.py:588-600；inputs/task/privileged/environment/rfo_task.py:614-615" + }, + { + "atom": "basket_hanging", + "level": "new_code", + "source": "需记录原生每帧Referee.failed及违规事件或其物理谓词；inputs/task/privileged/environment/rfo_task.py:376；inputs/task/privileged/environment/rfo_task.py:487；inputs/task/privileged/environment/rfo_task.py:546-550" + } + ] + }, + "native_metrics": [ + { + "name": "basket_touch_s", + "unit": "s", + "threshold": 0.3, + "direction": "lower_better" + } + ], + "open_questions_zh": [ + "没有成功参考轨迹：inputs/oracle_none.txt表明16次目录运行全失败；无录像可核验视觉判据，仅按文本、初态和裁判代码规划（N2、N8）。inputs/states/index.json为空，所有when、never及success置null，不假设任何字段已经写入日志（P8）；atoms列出需要补录的原生世界状态。", + "步骤粒度：开始运球与胯下反弹合为胯下运球，不单列离手或首次触地；完成后可原地继续运球再选择收球，双手收球可持球停顿，所以保留双手收球作为独立动态·接住步骤。扣篮与落地之间不可暂停，按S10合为一节点；原生五阶段不是五个语义节点。", + "类型边界：扣篮必须在球心入圈时保持手触球，不能以离手投篮替代；这里以受控篮球进入篮圈标关系·放入，而非飞行命中的动态·抛/投，落地站稳作为本步验收。未有成功录像核对，保留此类型判断供审阅。", + "接触词表缺口（D12）：文字及初态代码描述张开的左右手掌侧向挤持篮球，未包住球；此方案用双臂协作×其他，候选名双掌夹持。双手托接及单手上压是另列的接触方案，所有方案均不限定先用哪只手。", + "文本与评分器不一致：文本说扣篮后无手接触球并站立，但评分器只要求trackers.let_go曾被设置，重新接触球不清除此记录；最终criterion按曾完全离手加站稳描述，不额外发明持续松手步骤（rfo_task.py:553-556、603-605、674-678）。", + "文本与评分器不一致：出界写为全程禁止，代码在phase为dunked时提前返回，扣穿后直到站稳不再检查球或机器人出界（rfo_task.py:553-561）；保留文字禁止的出界约束并明确代码适用时段。", + "文本步骤未单独受评：文字要求扣篮后落地，评分器却没有起跳、双足离地或落地事件门控，只在扣穿后以standing检查脚接地与姿态；本节点不添加评分器没有的腾空门槛，落地与站稳仍留在同一文字步骤（rfo_task.py:670-678、referee.py:292-301）。", + "文本与评分器不一致：文本称骨盆速度小于0.35米/秒且在双脚上站稳，standing实际接受速度等于0.35米/秒，且只要求至少一只脚接地、所有接地部位属于左右脚，并不要求双脚同时接地；最终判据采用代码边界（referee.py:292-301）。", + "保持时长：原生hold在100个10毫秒控制采样上计满1.0秒；按D31，若未来绑定未计数的即时条件，应使用0.99秒跨度，现节点hold_s为0.99；若绑定原生保持计数或stand阶段则hold_s改为0，避免重复保持。", + "两模式任务模块与instance.json逐字一致，物理裁判相同；特权模式额外要求两次新进程回放均成功且末态哈希相同（grader/efe_embodied-olympics-fun-files/grade.py:106-115），标准模式评分最后一次实时尝试的原生裁判结果（grader/efe_embodied-olympics-fun-files/standard/server.py:157-177）。这些交付与复现要求不建语义节点，空日志也不能表达外部最终判决。" + ], + "conf": 6, + "conf_reason_zh": "描述、场景初态和两模式共用裁判支持三步与必要顺序，阈值均来自代码；没有状态日志或成功录像可复核绑定和可见性，扣篮类型及双掌夹持接触存在词表边界，且文本和代码有已列明差异。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "三步（胯下运球、双手收球、扣篮落地站稳）与文本一致，扣篮与落地按S10合并；成功条件与评分器阶段一致，约束均为评分器失败条件且立即终止；无状态日志，when/never置null并在open_questions_zh中说明，文本与评分器差异均已列出。未发现需修改项。", + "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：['between_legs_bounce', 'ball_gathered', 'dunk_landed']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:46:03Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__dunk_between_legs/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__dunk_between_legs/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "8e88e97a13b00ea9c5bb9b60b804a5b3e078183bffb08a8eb2d83ae6810ffcf9", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/embodied-olympics-fun/dunk_between_legs/trajectories/embodied-olympics-fun-dunk-between-legs-i00-privileged--4wisuyu.json b/phase2/embodied-olympics-fun/dunk_between_legs/trajectories/embodied-olympics-fun-dunk-between-legs-i00-privileged--4wisuyu.json new file mode 100644 index 0000000000000000000000000000000000000000..0f243333c0dcb5600cc5b790e49d1c6576c1245b --- /dev/null +++ b/phase2/embodied-olympics-fun/dunk_between_legs/trajectories/embodied-olympics-fun-dunk-between-legs-i00-privileged--4wisuyu.json @@ -0,0 +1,182 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-dunk-between-legs-i00-privileged--4wisuyu", + "taskKey": "embodied-olympics-fun/dunk_between_legs", + "dag_sha256": "b3a9022ee9aaa233f14f035ce0b249f6a6cd4ac0723730a08b64571a639c20ad", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "replay_success": 0, + "deterministic": 1, + "n_actions": 10, + "sim_steps": 10, + "horizon": 4000, + "episode_over": 0, + "progress": 0.0, + "stages_reached": 0, + "stages_total": 5, + "fell": 0, + "replay_s": 1.99, + "bad_trajectory": 0, + "metric_ball_height_m": 1.105, + "metric_basket_touch_s": 0.0, + "grader_error": 0, + "video_rendered": 1, + "agent_sessions": 2, + "agent_resets": 0, + "agent_replays": 0, + "agent_cost_self_reported": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.476, + "peak_processes": 6, + "peak_disk_gb": 0.248, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0, + "session_log_tampered": 0, + "session_log_from_monitor": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-dunk-between-legs-i00-privileged--4wisuyu/replay.mp4", + "duration_s": 0.12, + "fps": 25.0, + "frames": 3, + "sha256": "68e58a8e2fe8670b3138563742b6ad0cde523449876c02d4e9c6c1f6b0228c2c" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 3, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "between_legs_bounce": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.08, + "log_step": 9, + "text_zh": "观察：已检查视频 0.00、0.04、0.08 s 的全部三帧；机器人始终原地站立，篮球停留在身前手部，没有离手触地、反弹或穿过双腿。log step 9 的交付生成脚本仅执行 s.hold() 对应动作 10 步后保存；评分器 n_actions=10、stages_reached=0，与录像一致。" + } + ] + }, + "ball_gathered": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.08, + "log_step": 9, + "text_zh": "观察：视频 0.00、0.04、0.08 s 只显示开局身前持球，未出现胯下反弹及反弹后的双手接球；开局持球按该节点判据不计为双手收球。log step 9 保存的是保持开局姿态的基线。" + } + ] + }, + "dunk_landed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.08, + "log_step": null, + "text_zh": "观察：视频 0.00、0.04、0.08 s 中机器人一直位于远离篮圈的开局位置，球仍在身前，未见起跳、球入圈或扣篮后落地；评分器 stages_reached=0、metric_basket_touch_s=0.0。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "basket_touch_s", + "value": 0.0, + "unit": "s", + "threshold": 0.3, + "direction": "lower_better" + } + ], + "failure": { + "breakpoint": { + "ref": "unknown", + "t": null, + "log_step": 25, + "where": "首步前", + "evidence_zh": "观察：log step 25 的最终消息明确为 Gemini 提供方过滤拦截；此前 log step 9 只保存了 10 步保持姿态基线，视频 0.00、0.04、0.08 s 均未开始运球。log step 19–24 是另建 Episode 的开发测试，没有保存替换交付轨迹；评分器 n_actions=10、episode_over=0、agent_replays=0。" + }, + "phenomena": [], + "root_causes": [], + "gap_zh": "观察：本次有录像，已逐帧检查全部三帧，但没有状态日志或成功参考录像。log step 25 出现模型提供方过滤中断；中断前仅有 log step 9 的初始保持姿态基线，未见完成任务后的保存和回放验证记录，评分器 agent_replays=0。推断：这是开发尚未完成时被外部过滤切断，按外部因素规则不标“没尝试”或机器人根因；log step 19–24 的测试结果不能作为评分录像的进度或失败现象，无法定位评分尝试中的物理失败时刻。", + "unmet": { + "between_legs_bounce": false, + "ball_gathered": false, + "dunk_landed": false + } + }, + "summary_zh": "评分录像的全部三帧（视频 0.00、0.04、0.08 s）均为原地持球，胯下运球、合法双手收球和扣篮落地站稳三个节点均未达到。log step 9 仅保存了 10 步保持姿态基线，后续开发测试未替换该交付，会话最后在 log step 25 被 Gemini 提供方过滤中断。中断前无完成交付及回放验证记录，按外部因素处理，断点记 unknown／首步前，现象与根因留空。", + "conf": 9, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "三帧录像（视频 0–0.08 s）确认机器人原地持球，三个节点均未达成；log step 25 为 Gemini 提供方过滤中断，此前仅保存 10 步保持姿态基线、无回放验证，按外部因素留空现象与根因、断点 unknown/首步前，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": "2026-10-29T10:02:55Z", + "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/embodied-olympics-fun-dunk-between-legs-i00-privileged--4wisuyu/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-dunk-between-legs-i00-privileged--4wisuyu/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "e5b587e2133e11009605fd36902498f3b26894523c23e47cacd62b1459299ec5", + "video_sha256": "68e58a8e2fe8670b3138563742b6ad0cde523449876c02d4e9c6c1f6b0228c2c", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/dunk_between_legs/trajectories/embodied-olympics-fun-dunk-between-legs-i00-standard--wqrxnzk.json b/phase2/embodied-olympics-fun/dunk_between_legs/trajectories/embodied-olympics-fun-dunk-between-legs-i00-standard--wqrxnzk.json new file mode 100644 index 0000000000000000000000000000000000000000..95fd9587c839e4d5e1f8426a83e06e7541b9ebba --- /dev/null +++ b/phase2/embodied-olympics-fun/dunk_between_legs/trajectories/embodied-olympics-fun-dunk-between-legs-i00-standard--wqrxnzk.json @@ -0,0 +1,228 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-dunk-between-legs-i00-standard--wqrxnzk", + "taskKey": "embodied-olympics-fun/dunk_between_legs", + "dag_sha256": "b3a9022ee9aaa233f14f035ce0b249f6a6cd4ac0723730a08b64571a639c20ad", + "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": 2401, + "calls": 59, + "rows": 1716, + "resets": 5, + "reset_budget": 5, + "sim_startups": 1, + "elapsed_s": 1606.1, + "video_rendered": 1, + "first_success": 0, + "first_reason_code": 2, + "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/embodied-olympics-fun-dunk-between-legs-i00-standard--wqrxnzk/replay.mp4", + "duration_s": 1.88, + "fps": 25.0, + "frames": 47, + "sha256": "23b1d4c9da82406086ba62d7cbb084df20cb6f9fb7e5687bca170f355639ce5c" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 3, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "between_legs_bounce": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.72, + "log_step": 123, + "text_zh": "观察：视频 0.00–0.56 s 机器人在胸前双手持球；视频 0.60–0.92 s 双臂向下后方摆动并放球，篮球却沿机器人身前下降。" + }, + { + "t": 1.28, + "log_step": 125, + "text_zh": "观察：逐帧检查视频 0.92–1.84 s，篮球在脚前约视频 0.96 s 触地，随后在身前反弹至膝部附近再落下，未见无手接触穿过两腿并在另一侧被手触球确认；log step 125 的服务裁判报告第 6 局 stages_reached=1、stages_total=5，仅开始运球，未达胯下运球。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "其他", + "note_zh": "候选接触名：双掌夹持。视频 0.00–0.60 s 左右手从两侧共同持住篮球，然后同步摆臂放球；未见单手合拢包住球，不标抓握，现有词表没有双掌侧夹项。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "视频 0.00–1.84 s 机器人双足支撑地面，原地摆臂运球，没有轮式底盘移动。" + } + ] + }, + "ball_gathered": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.8, + "log_step": 123, + "text_zh": "观察：视频 1.72–1.84 s 机器人双臂在腰后继续摆动，篮球仍在足前低处，没有被双手接住；log step 123 显示脚本在 control_steps=172 后执行首次 gather sweep，到 control_steps=185 被裁判终止。开局视频 0.00 s 的持球属于初态，不是胯下运球后的收球。" + }, + { + "t": 1.84, + "log_step": 125, + "text_zh": "推断：视频全程没有完成前置胯下运球或后续双手收球，且 log step 125 的第 6 局裁判仅报告原生阶段 1/5，因此本事件未达成。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "视频 1.72–1.84 s 机器人原地以双足支撑，双臂执行后扫收球尝试；没有观察到手与篮球建立新的接球接触。" + } + ] + }, + "dunk_landed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.84, + "log_step": 125, + "text_zh": "观察：视频 0.00–1.84 s 机器人始终在罚球线附近，篮球只在身前低处反弹，没有接近篮圈、受手控制扣穿或扣篮后落地动作；log step 125 确认第 6 局已判失败，阶段仅 1/5。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "basket_touch_s", + "value": null, + "unit": "s", + "threshold": 0.3, + "direction": "lower_better" + } + ], + "failure": { + "breakpoint": { + "ref": "between_legs_bounce", + "t": 0.8, + "log_step": 123, + "where": "操作中", + "evidence_zh": "观察：视频 0.60–1.40 s 机器人摆臂放球后，篮球落在脚前并在身前反弹，没有进入两腿之间；视频 1.72–1.84 s 后扫双臂仍未补成胯下穿越。log step 123、125 确认这是最终第 6 局，服务裁判因“the ball touched a part of you other than a hand”终止，原生阶段仅 1/5；最先未修复的是胯下运球。" + }, + "unmet": { + "between_legs_bounce": false, + "ball_gathered": false, + "dunk_landed": false + }, + "phenomena": [ + { + "cat": "运动", + "item": "位姿不准", + "t": 0.8, + "log_step": 123, + "evidence_zh": "观察：视频 0.60–0.92 s 篮球离手后的下降球路在机器人身前，视频约 0.96 s 落在脚前，视频 1.08–1.44 s 又在身前反弹；球没有按胯下运球要求穿过腿间，后续也未纠正。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "log_step": 123, + "evidence_zh": "推断：log step 123 的最终脚本保持腿部初态，将双肩 pitch 由约 -0.387 rad 改至 0.70 rad、双肘 pitch 改至 0.20 rad，以固定关节目标执行标为向后下方放球的动作，随后定时后扫收球，没有依据球的位置修改动作；视频 0.60–1.44 s 显示实际球路留在身前，而手臂已经摆到腰后。log step 160 中 agent 也自述摆臂放球方向错误。由脚本和视频可归因于未能控制出所需球路，但不足以细分为逆解、跟踪或增益问题。", + "source_zh": "日志" + } + ], + "gap_zh": "没有状态日志或成功参考视频，侧视分辨率不足以确定最后一次非手部接触的具体机器人链节及精确时刻。log step 125 的服务裁判确认发生非手部触球；agent 在 log step 160 自述是前臂击球，但视频 1.72–1.84 s 球靠近足部，不能据其自述认定前臂接触。评分器没有报告 basket_touch_s，原生指标留空；根因只标控制大类。" + }, + "summary_zh": "最终评分对象为第 6 局：机器人原地摆臂放球，球在身前落地反弹，没有完成胯下运球；后续虽作收球后扫，仍未接住球，扣篮未启动，三个节点均未完成。断点是胯下运球，表现为球路位置不准；固定关节动作未能产生所需球路，随后裁判因非手部触球终止该局。没有状态日志，末次碰撞的具体链节不能从画面确认。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "三个节点均未达成、断点为胯下运球与视频、log step 123/125 一致：视频 0.6–1.44 s 球落在脚前身前反弹，1.84 s 球仍在足前，服务裁判 stages 1/5、因球碰到手以外部位终止。现象“位姿不准”与根因“控制·控制”（固定关节脚本、未据球位修正）有日志与视频支持，且未越权细分；gap_zh 如实说明碰撞链节无法确认。未发现需修改项。", + "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": "2026-10-29T10:04:35Z", + "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/embodied-olympics-fun-dunk-between-legs-i00-standard--wqrxnzk/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-dunk-between-legs-i00-standard--wqrxnzk/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "566ae7a85384a6a4ab44e4db0d675451acfda0d7190daa5653a05bea4c18178b", + "video_sha256": "23b1d4c9da82406086ba62d7cbb084df20cb6f9fb7e5687bca170f355639ce5c", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/free_throw/task_dag.json b/phase2/embodied-olympics-fun/free_throw/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..07316090436492dd424cac30751ea9da6b533da3 --- /dev/null +++ b/phase2/embodied-olympics-fun/free_throw/task_dag.json @@ -0,0 +1,336 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "embodied-olympics-fun/free_throw", + "suite": "embodied-olympics-fun", + "base": "embodied-olympics-fun-free-throw-i00", + "title_zh": "篮球罚球", + "summary_zh": "机器人从罚球线后合法投出开局持有的篮球，使球心从上方穿过篮圈并继续降至圈下一个球直径处，裁判判定命中。", + "embodiment_zh": "特权与标准模式均为 Unitree H1-2 双足人形机器人：27 个身体关节、双臂，每臂末端为 Inspire RH56DFX 五指灵巧手（每手六驱动、接触摩擦持球），双腿与双脚支撑，无轮式底盘；两种模式的观测权限与交付方式不同，物理具身一致。", + "task_class": "atomic", + "analysis_level": "step", + "spec": { + "version": "1.2", + "task": "embodied-olympics-fun-free-throw-i00", + "bench": "embodied-olympics-fun", + "progress": "live", + "entities": {}, + "milestones": [ + { + "id": "free_throw_scored", + "zh": "罚球命中", + "level": "semantic", + "mode": "event", + "when": null, + "after": [], + "hold_s": 0, + "goal": true, + "criterion_zh": "场地固定视角中，橙色篮球离开机器人双手后飞向白色篮板下的橙色篮圈，球心从上向下穿过圈内并继续降至圈下一个球直径，才算罚球命中；只出手、飞到圈旁、碰圈后弹出或从下穿圈均未完成。精确验收为：先记录离手（双手接触各自允许 0.03 s 间断），球心向下越过 z=3.05 m 平面时距篮圈中心 (4.575,0) 的水平距离 < 0.225 m，期间不再向上穿回该平面，随后向下越过 z=2.811 m 平面且交点水平距离 < 0.3445 m；全过程未触发所列犯规，出手须满足 5.0 s 期限，并在 12 s 回合时限内命中，无额外静置或保持窗口。", + "step": { + "cat": "动态", + "item": "抛/投" + }, + "contact": [ + { + "who": "双臂协作", + "how": "其他", + "plan": 1, + "note_zh": "候选：夹持；左、右张开的灵巧手以掌侧及拇指从篮球两侧相向夹持并共同送球，两手均未合拢包住球，抓握不适用，托/舀也不能完整描述相向夹紧。此为开局及典型双手投篮方案，离手后篮球自由飞行。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "plan": 1, + "note_zh": "机器人双脚在罚球线后半圆内支撑躯干，为双手送出篮球提供稳定站姿。" + }, + { + "who": "双臂协作", + "how": "托/舀", + "plan": 2, + "note_zh": "一只灵巧手从下方托住并送出篮球，另一只灵巧手在球侧辅助定向；左右主辅手均可互换，不能由成功录像限定。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "plan": 2, + "note_zh": "机器人双脚接触罚球区地面支撑躯干，配合一手托球、另一手辅助投篮。" + }, + { + "who": "手指/夹爪", + "how": "托/舀", + "plan": 3, + "note_zh": "一只灵巧手从下方托球后独立送出篮球；另一只手可先退出，规则允许任意手部投篮姿势，不要求出手前始终双手同时触球。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "plan": 3, + "note_zh": "机器人双脚在地面支撑躯干，配合单手将篮球投出。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:3-4", + "inputs/task/privileged/instruction.md:23-28", + "inputs/task/privileged/environment/rfo_task.py:119-148 开局球在两手之间", + "inputs/task/privileged/environment/rfo_task.py:178-186 原生阶段与参数", + "inputs/task/privileged/environment/rfo_task.py:229-239 场地与重置", + "inputs/task/privileged/environment/rfo_task.py:241-255 手部接触和穿圈跟踪", + "inputs/task/privileged/environment/rfo_task.py:270-308 离手、穿圈和违规", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:429-458 HoopPass", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178 阶段顺序与成功", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/scene.py:51 官方篮球半径 0.1195 m", + "inputs/oracle/replay.mp4 @ 0.00-1.12 s", + "inputs/oracle/run.json steps[id=428] 最终文件的两次重放均成功，release=0.17 s，through the ring=1.121 s（模拟时间）" + ], + "conf": 6, + "note_zh": "这是一次不可在离手与入篮之间暂停的投篮，依 S10 只建一个节点；release 与 through the ring 是同一步的原生验收阶段。命中是已发生并锁存的事件，不要求篮球在回合末静止留在网中；不另列拿球、抬球、松手或站立节点。双手初始相向夹持不是合指抓握，其他投篮姿势作为替代接触方案列出；同一方案中手部送球与双脚支撑同时成立，离手后的无接触飞行仅记于此。" + } + ], + "constraints": [ + { + "id": "line_fault", + "zh": "踩线越线", + "never": null, + "criterion_zh": "球尚未触及篮圈或进入篮筐时，机器人任一部位与地面的接触点 x >= 0.35 m，即触及罚球线近边或其前方地面；空中越线而未触地本身不触发此项。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:231-232 篮圈中心 x=0.35+4.225=4.575 m", + "inputs/task/privileged/environment/rfo_task.py:61-63 罚球线近边 x=0.35 m", + "inputs/task/privileged/environment/rfo_task.py:257-275", + "inputs/task/privileged/environment/rfo_task.py:293-294" + ] + }, + { + "id": "semicircle_fault", + "zh": "离开半圆", + "never": null, + "criterion_zh": "球尚未触及篮圈或进入篮筐时，机器人任一地面接触点到罚球线近边中点 (0.35,0) 的平面距离 > 1.80 m；外部场地视角可见脚或其他部位在白色罚球半圆外触地。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:183-186", + "inputs/task/privileged/environment/rfo_task.py:257-275", + "inputs/task/privileged/environment/rfo_task.py:293-294" + ] + }, + { + "id": "nonhand_ball_contact", + "zh": "非法部位触球", + "never": null, + "criterion_zh": "裁判记录离手前，篮球接触机器人手以外的任何部位即失败；手包括手指、拇指、掌和腕，前臂、上臂、躯干或腿均不允许接球或夹球。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:243-246", + "inputs/task/privileged/environment/rfo_task.py:276-282", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:25-43" + ] + }, + { + "id": "ball_retouch", + "zh": "离手后再触球", + "never": null, + "criterion_zh": "裁判记录离手后超过 0.05 s，篮球再次接触机器人任何部位即失败；手部接触间隔不超过 0.03 s 时仍按持续在手处理，离手不由通用 held 字段判断。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:185-186", + "inputs/task/privileged/environment/rfo_task.py:252-255", + "inputs/task/privileged/environment/rfo_task.py:276-288", + "inputs/task/privileged/environment/rfo_task.py:297-298" + ] + }, + { + "id": "ball_floor_contact", + "zh": "篮球触地", + "never": null, + "criterion_zh": "投篮成功前，篮球触及地面即失败，开局仍在手中时篮球触地也失败；场地视角可见橙球落地或反弹，不能算命中。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:283-285", + "inputs/task/privileged/environment/rfo_task.py:299-303" + ] + }, + { + "id": "ball_up_through_ring", + "zh": "从下穿圈", + "never": null, + "criterion_zh": "离手后，尚未从上方进入篮圈的篮球，球心向上越过 z=3.05 m 平面且交点距篮圈中心的水平距离 < 0.225 m，即从下方穿圈并失败。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:183-186", + "inputs/task/privileged/environment/rfo_task.py:248-249", + "inputs/task/privileged/environment/rfo_task.py:289-296", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:413-423" + ] + }, + { + "id": "robot_fell", + "zh": "机器人跌倒", + "never": null, + "criterion_zh": "命中前，在控制步末检查到骨盆高度 < 0.35 m，或除左右脚以外任一身体部位触地即失败；机器人明显倒伏、手膝撑地均属违规。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:180-186", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/task.py:21-23", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:272-289" + ] + }, + { + "id": "release_deadline", + "zh": "超时未出手", + "never": null, + "criterion_zh": "仍未被裁判记录为离手且控制步末时间超过 5.0 s（代码加 1e-9 数值容差）即失败；一直持球或只抬球不投出不能通过。此项是任务明确规定的出手犯规期限，不是 12 s 总时限。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:185-186", + "inputs/task/privileged/environment/rfo_task.py:305-308" + ] + }, + { + "id": "simulation_unstable", + "zh": "物理失稳", + "never": null, + "criterion_zh": "控制过程中物理数值发散即判失败：出现异常加速度警告、非有限位置或超过平台 MAX_SPEED 的平移速度；若本控制步刚命中也可撤销成功，先前控制步已锁存的成功除外。录像异常跳变只能提示，精确判定需要平台数值及裁判失稳记录。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:108", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:867-885", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:166-172" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "free_throw_scored", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:276-303 + inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:429-458；需逐帧记录裁判原生 release、through the ring 阶段及 HoopPass 跟踪状态；当前没有状态日志，不能绑定。" + }, + { + "atom": "line_fault", + "level": "new_code", + "source": "需在回放适配器中逐帧记录此违规的原生裁判事件及所用接触/位置/时间量；依据 inputs/task/privileged/environment/rfo_task.py:231-232 篮圈中心 x=0.35+4.225=4.575 m; inputs/task/privileged/environment/rfo_task.py:61-63 罚球线近边 x=0.35 m; inputs/task/privileged/environment/rfo_task.py:257-275; inputs/task/privileged/environment/rfo_task.py:293-294" + }, + { + "atom": "semicircle_fault", + "level": "new_code", + "source": "需在回放适配器中逐帧记录此违规的原生裁判事件及所用接触/位置/时间量；依据 inputs/task/privileged/environment/rfo_task.py:183-186; inputs/task/privileged/environment/rfo_task.py:257-275; inputs/task/privileged/environment/rfo_task.py:293-294" + }, + { + "atom": "nonhand_ball_contact", + "level": "new_code", + "source": "需在回放适配器中逐帧记录此违规的原生裁判事件及所用接触/位置/时间量；依据 inputs/task/privileged/environment/rfo_task.py:243-246; inputs/task/privileged/environment/rfo_task.py:276-282; inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:25-43" + }, + { + "atom": "ball_retouch", + "level": "new_code", + "source": "需在回放适配器中逐帧记录此违规的原生裁判事件及所用接触/位置/时间量；依据 inputs/task/privileged/environment/rfo_task.py:185-186; inputs/task/privileged/environment/rfo_task.py:252-255; inputs/task/privileged/environment/rfo_task.py:276-288; inputs/task/privileged/environment/rfo_task.py:297-298" + }, + { + "atom": "ball_floor_contact", + "level": "new_code", + "source": "需在回放适配器中逐帧记录此违规的原生裁判事件及所用接触/位置/时间量；依据 inputs/task/privileged/environment/rfo_task.py:283-285; inputs/task/privileged/environment/rfo_task.py:299-303" + }, + { + "atom": "ball_up_through_ring", + "level": "new_code", + "source": "需在回放适配器中逐帧记录此违规的原生裁判事件及所用接触/位置/时间量；依据 inputs/task/privileged/environment/rfo_task.py:183-186; inputs/task/privileged/environment/rfo_task.py:248-249; inputs/task/privileged/environment/rfo_task.py:289-296; inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:413-423" + }, + { + "atom": "robot_fell", + "level": "new_code", + "source": "需在回放适配器中逐帧记录此违规的原生裁判事件及所用接触/位置/时间量；依据 inputs/task/privileged/environment/rfo_task.py:180-186; inputs/grader/efe_embodied-olympics-fun-files/rfo/task.py:21-23; inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:272-289" + }, + { + "atom": "release_deadline", + "level": "new_code", + "source": "需在回放适配器中逐帧记录此违规的原生裁判事件及所用接触/位置/时间量；依据 inputs/task/privileged/environment/rfo_task.py:185-186; inputs/task/privileged/environment/rfo_task.py:305-308" + }, + { + "atom": "simulation_unstable", + "level": "new_code", + "source": "需在回放适配器中逐帧记录此违规的原生裁判事件及所用接触/位置/时间量；依据 inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:108; inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:867-885; inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:166-172" + }, + { + "atom": "episode_success", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:174-178；需记录原生逐帧 ref.success、ref.failed 与物理失稳状态；当前不绑定交付或重放有效性字段。" + } + ] + }, + "native_metrics": [ + { + "name": "ball_height_m", + "unit": "m", + "threshold": null, + "direction": null, + "note_zh": "inputs/task/privileged/environment/rfo_task.py:309 每控制步记录球心高度，精度 0.001 m；仅诊断，无单独合格阈值，也非越高越好，不能用终帧高度代替穿圈事件。" + } + ], + "open_questions_zh": [ + "inputs/states/index.json 的 logs 为空，成功参考也无状态日志（inputs/oracle/oracle.json）；按 P8，全部节点 when、约束 never 与 success 均为 null。视频可见投篮和入圈，但无法核验毫秒级接触、精确穿圈交点或所有违规，需补记原生裁判状态；不得将代理叙述或离线总评分当作逐帧绑定。", + "接触词表边界（D12）：开局两只张开的手从篮球两侧夹持，未以手指包络抓握；典型方案用双臂协作 × 其他，候选“夹持”，单手托投及双手主辅托投另列替代方案，均不规定必须左右手同步出手。", + "模式差异：特权评分还要求轨迹有效且两次新进程重放成功、终态哈希一致（inputs/grader/efe_embodied-olympics-fun-files/grade.py:70-125）；标准模式采用最后一次现场回合的裁判结果（inputs/grader/efe_embodied-olympics-fun-files/standard/server.py:157-178；inputs/grader/efe_standard-files/v2_1/grade.py:83-110）。这些交付与基础设施检查不是世界状态或独立评测场景，DAG 保留共同物理目标，success=null，不伪造可绑定的重放确定性条件。", + "文本与实现的采样差异：特权说明称裁判每 1 ms 判断，但接触及穿圈在物理子步判断，跌倒、5.0 s 出手期限和总时限在 10 ms 控制步末判断（inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:266-289；inputs/task/privileged/environment/rfo_task.py:305-308）。节点与约束按实现写，不从低帧率录像推断毫秒级先后。", + "按 S10 与 D19，离手和篮球飞行合为一次罚球命中，进度为 0/1；原生球心高度仅作诊断。成功参考录像约 1.16 s，开局双手持球、随后上抛并在末段穿圈，与最终重放输出的两个原生阶段一致；其参数搜索和左右手动作不形成必经步骤。" + ], + "conf": 6, + "conf_reason_zh": "任务文本、场景重置和裁判代码共同支持一个投篮命中事件；几何阈值与违规均有源码依据，成功录像可见穿圈。没有状态日志，无法绑定或验证所有验收条件，夹持接触词表仍有缺口，因此置信度取 6。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "单节点原子任务（罚球命中），无状态日志故 when/never/success 为 null，符合 P8；离手与穿圈按 S10/D19 合为一步；阈值 0.225/0.3445/3.05/2.811 与 referee.py HoopPass 一致；约束均为评分器失败条件且为立即终止；open_questions_zh 已列文本/实现差异。未发现需修改项。", + "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：['free_throw_scored']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:47:33Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__free_throw/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__free_throw/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "93cdd2fc632fd2f2887fa99abdf7563a855cf31e17693340798f0e2cb4561084", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/embodied-olympics-fun/free_throw/trajectories/embodied-olympics-fun-free-throw-i00-privileged--gg2r6zn.json b/phase2/embodied-olympics-fun/free_throw/trajectories/embodied-olympics-fun-free-throw-i00-privileged--gg2r6zn.json new file mode 100644 index 0000000000000000000000000000000000000000..3ff04488adc4deb0351ddc8a464e62a41a074671 --- /dev/null +++ b/phase2/embodied-olympics-fun/free_throw/trajectories/embodied-olympics-fun-free-throw-i00-privileged--gg2r6zn.json @@ -0,0 +1,173 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-free-throw-i00-privileged--gg2r6zn", + "taskKey": "embodied-olympics-fun/free_throw", + "dag_sha256": "c2d4b07a9ca0ff585b81cafb3eecf9f2484164969c066ba3cd5ef111abe016a5", + "outcome": { + "success": 1, + "reward": 1.0, + "source": "catalog", + "raw": { + "success": 1, + "replay_success": 1, + "deterministic": 1, + "n_actions": 157, + "sim_steps": 157, + "horizon": 1200, + "episode_over": 1, + "progress": 1.0, + "stages_reached": 2, + "stages_total": 2, + "fell": 0, + "replay_s": 3.68, + "bad_trajectory": 0, + "metric_ball_height_m": 2.801, + "grader_error": 0, + "video_rendered": 1, + "agent_sessions": 63, + "agent_resets": 1640, + "agent_replays": 7, + "agent_cost_self_reported": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.768, + "peak_processes": 9, + "peak_disk_gb": 0.252, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0, + "session_log_tampered": 0, + "session_log_from_monitor": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-free-throw-i00-privileged--gg2r6zn/replay.mp4", + "duration_s": 1.6, + "fps": 25.0, + "frames": 40, + "sha256": "23538015987da2ef2006cabd85c6dc74927fc143a8f28a2ca95b9509819533e8" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "free_throw_scored": { + "method": "visual", + "intervals": [ + [ + 1.56, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察：视频 0.00–0.32 s，机器人双脚着地，左右张开的手在橙色篮球两侧共同持球；机器人先屈膝俯身，随后双臂同时向上送球。" + }, + { + "t": 0.4, + "log_step": null, + "text_zh": "观察：视频 0.36 s 球仍紧邻双手，视频 0.40 s 球已在双手上方并有可见间隙，随后独立向篮筐飞行；机器人没有追球或再接球。" + }, + { + "t": 1.56, + "log_step": null, + "text_zh": "观察：逐帧检查视频 1.12–1.56 s，篮球从篮板上方持续下降；视频 1.48 s 球在篮圈正上方，视频 1.52 s 降到圈口，视频 1.56 s 球已进入圈下篮网。机器人仍双脚着地站立；罚球命中的视频锚点取末帧约 1.56 s。" + }, + { + "t": 1.56, + "log_step": 199, + "text_zh": "观察：log step 198 的最终生成脚本先重置、执行动作、保存轨迹再重放；log step 199 的生成结果与官方重放均报告 success=true、157 个控制步、release 和 through the ring 两阶段完成，球心终值 2.801 m。inputs/run.json 的 reward 同样记录 success=1、replay_success=1、deterministic=1、stages_reached=2，与视频末段入网一致。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "其他", + "note_zh": "观察：视频 0.00–0.36 s，机器人左右张开的灵巧手从篮球两侧共同夹持、抬送篮球，随后分开离球。候选：夹持；未见单手合拢包住球，抓握不适用，单纯托/舀不能完整表达两侧相向夹持。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：视频 0.00–1.56 s，机器人双脚接触罚球区地面，屈膝及伸腿配合双臂投球，离手后维持站姿。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "ball_height_m", + "value": 2.801, + "unit": "m", + "threshold": null, + "direction": null + } + ], + "failure": null, + "summary_zh": "机器人双手夹持并送出篮球，视频约 1.56 s 球由上向下穿圈进入篮网，完成罚球命中节点（1/1），全程保持站立。最终交付的官方重放与评分器均判成功，未见需要失败标签的事件。没有状态日志，25 fps 录像不能独立核验毫秒级接触及精确穿圈交点；节点按画面入网与评分器成功结果共同判定。", + "conf": 9, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "罚球命中节点在视频约 1.56 s 入网，与画面及评分器 success/stages 2/2 一致；接触标注合理，无失败标签。", + "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": "2026-10-29T10:04:02Z", + "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/embodied-olympics-fun-free-throw-i00-privileged--gg2r6zn/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-free-throw-i00-privileged--gg2r6zn/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "d092a10138df9fcbb38953e3c57a261ad2dc8768d03a911df2626845d2c8457f", + "video_sha256": "23538015987da2ef2006cabd85c6dc74927fc143a8f28a2ca95b9509819533e8", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/free_throw/trajectories/embodied-olympics-fun-free-throw-i00-standard--vdpxbwo.json b/phase2/embodied-olympics-fun/free_throw/trajectories/embodied-olympics-fun-free-throw-i00-standard--vdpxbwo.json new file mode 100644 index 0000000000000000000000000000000000000000..d3a7e2ce56d1c2c32106bde8ace9cc927f4d5662 --- /dev/null +++ b/phase2/embodied-olympics-fun/free_throw/trajectories/embodied-olympics-fun-free-throw-i00-standard--vdpxbwo.json @@ -0,0 +1,189 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-free-throw-i00-standard--vdpxbwo", + "taskKey": "embodied-olympics-fun/free_throw", + "dag_sha256": "c2d4b07a9ca0ff585b81cafb3eecf9f2484164969c066ba3cd5ef111abe016a5", + "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": 73, + "sim_steps": 537, + "calls": 15, + "rows": 537, + "resets": 5, + "reset_budget": 5, + "sim_startups": 1, + "elapsed_s": 1870.0, + "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": 0.0, + "peak_memory_gb": 0.287, + "peak_processes": 7, + "peak_disk_gb": 0.358, + "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/embodied-olympics-fun-free-throw-i00-standard--vdpxbwo/replay.mp4", + "duration_s": 0.76, + "fps": 25.0, + "frames": 19, + "sha256": "b247ae3badd280a78a43a3540dc64bf2aefcff753685b69ea329df1212c02563" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "free_throw_scored": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.2, + "log_step": 372, + "text_zh": "观察：视频 0.00–0.20 s，机器人站立，橙色篮球位于两只张开的手之间；视频 0.24–0.32 s 双臂上抬，球向前离开手部，但始终明显低于篮圈。" + }, + { + "t": 0.72, + "log_step": 372, + "text_zh": "观察：逐帧检查视频 0.32–0.72 s，篮球在机器人前方低空前移并下降，末帧已贴近地面，未到达篮圈，更未从上方穿圈；log step 372 的机器人服务返回 episode=6、control_steps=73、done=True、outcome=\"the ball touched the floor\"、stages_reached=1/2，与录像一致。这里的阶段数是裁判原生的离手/穿圈阶段，不是 DAG 节点完成数。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "其他", + "note_zh": "观察：视频 0.00–0.20 s，机器人左右张开的手从篮球两侧相向夹持，随后双臂共同送球并张开离球。候选：夹持；两手未各自合拢包住球，抓握不适用，单纯托/舀也不能完整描述相向夹紧。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：视频 0.00–0.72 s，机器人双脚接触球场地面、支撑躯干完成双臂送球，未见倒伏。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "ball_height_m", + "value": null, + "unit": "m", + "threshold": null, + "direction": null + } + ], + "failure": { + "breakpoint": { + "ref": "free_throw_scored", + "t": 0.28, + "log_step": 372, + "where": "操作中", + "evidence_zh": "观察：视频 0.24–0.32 s，机器人双臂抬起，球却从手部前下方低平飞出；视频 0.40–0.72 s 球持续下降，在远未抵达篮圈处落地。log step 372 确认这是最后一次重置后的第 6 回合，机器人服务的裁判结果为篮球触地，罚球未命中。" + }, + "unmet": { + "free_throw_scored": false + }, + "phenomena": [ + { + "cat": "运动", + "item": "位姿不准", + "t": 0.28, + "log_step": 372, + "evidence_zh": "观察：视频 0.24–0.72 s，篮球离开机器人双手后仅在低处向前飞行，始终低于篮圈并在篮筐前落地，没有形成能从上方进入篮圈的抛物线；log step 372 的裁判终止原因为篮球触地。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志", + "evidence_zh": "推断：agent 的投篮脚本采用固定时序的双臂加速及肩部外展松球（log step 367），实际执行为最后第 6 回合（log step 372）。该回合本体反馈表显示，仿真 0.21 s 左肩俯仰目标为 -0.992 rad、实际为 -0.607 rad，仿真 0.28 s 实际为 -1.210 rad，肩力矩已由 -70.327 N·m 变为 +66.067 N·m（log step 376）；结合视频 0.24–0.40 s 双臂上抬而球低平飞出的结果，送球与松球的运动控制未产生所需的出球状态。证据不足以将主因进一步确定为某个控制子类。" + } + ], + "gap_zh": "没有逐帧状态日志或篮球离手瞬间的速度、接触力记录，25 fps 外部视角也不能判定毫秒级手球接触先后；不能确定低平出球主要来自速度不足、方向偏差还是短暂滑移，因此控制根因不再细分。DAG 的 ball_height_m 未在评分器字段中报告，原生指标值留空。" + }, + "summary_zh": "观察：评分视频对应最后第 6 回合，机器人双手夹持后尝试送球，罚球命中节点最终为 0/1；篮球在视频约 0.28 s 低平离手，随后在篮筐前落地，log step 372 的裁判结果同为篮球触地。推断：固定时序的送球与松球控制未形成入篮所需的出球状态；缺少离手瞬间的球速和接触力，无法再细分控制原因。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "单节点罚球命中未达成（0/1）与视频一致：约 0.24–0.28 s 球离手后在机器人前方低处下落并触地，未接近篮圈；断点、接触、现象与证据可检验，根因已声明不细分，未发现需改项。", + "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": "2026-10-29T10:04:27Z", + "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/embodied-olympics-fun-free-throw-i00-standard--vdpxbwo/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-free-throw-i00-standard--vdpxbwo/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "7d9866ea18ce3bdd8fb7ad7e2bf85e5f3f832133dfd30a3402d19d49a8770984", + "video_sha256": "b247ae3badd280a78a43a3540dc64bf2aefcff753685b69ea329df1212c02563", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/freeski_360/task_dag.json b/phase2/embodied-olympics-fun/freeski_360/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..ec650b8d7ec16073ebc9eb1e72a4ee23b65a2143 --- /dev/null +++ b/phase2/embodied-olympics-fun/freeski_360/task_dag.json @@ -0,0 +1,259 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "embodied-olympics-fun/freeski_360", + "suite": "embodied-olympics-fun", + "base": "embodied-olympics-fun-freeski-360-i00", + "title_zh": "滑雪转体整周", + "summary_zh": "机器人从起点自行滑入跳台，在空中以雪板完成一次360度转体且无被计入的空翻，雪板先落地后受控滑过终点，整次尝试未触发规则失败。", + "embodiment_zh": "两种模式均为29关节Unitree G1双足人形机器人，双臂各有7关节，末端为无可动手指的拳垫，本任务没有可抓物，抓握指令无效；双靴分别刚性固定于1.35米双翘雪板，无雪杖和轮式底盘，由双腿、腰和躯干控制助滑、转体及落地。", + "task_class": "atomic", + "analysis_level": "step", + "spec": { + "version": "1.2", + "task": "embodied-olympics-fun-freeski-360-i00", + "bench": "embodied-olympics-fun", + "progress": "live", + "entities": {}, + "milestones": [ + { + "id": "spin_360_ridden_away", + "zh": "转体落地滑出", + "mode": "event", + "when": null, + "after": [], + "hold_s": 0, + "goal": true, + "criterion_zh": "在能辨认机器人和雪板的跳台或场地视角中，应看见机器人自行滑下起点坡、从跳台红线附近腾空，雪板在空中转过约一圈后先于身体落到台面末端蓝线之后，并靠雪板支撑继续朝坡下滑到绿色终点线之外；只转动上身、在雪上补转、尚未滑出终点或途中摔倒均未完成，而本次参考视频的远景无法可靠辨认机器人，不能据空景断定这些动作发生或未发生。精确验收为起滑、起跳、落地、滑出依次成立：双雪板绑定主体中心越过起点前沿后，连续至少30个1毫秒物理步无雪板、双靴或其他身体部位触雪，起飞时骨盆距跳台唇口的水平距离严格小于4.0米、速度仰角严格大于10度；仅在该次飞行中取两板连续展开航向净变化的平均值，绝对值须大于等于270度且严格小于450度（朝左或朝右均可，板轴距竖直不足10度时暂停航向更新），按骨盆横向角速度积分减去61度后的角度和是否倒置判定空翻数必须为零且不得倒滑起跳；首次触雪须有雪板接触且无其他身体部位触雪，骨盆横坐标不早于台面末端蓝线，两板朝前误差各不超过45度；随后遵守全部触雪、失衡、雪上转体、停滞和未出发约束，直到两板绑定主体中心都严格越过缓冲区起点下游15.0米的终点线，且此刻骨盆沿雪面法线净高至少0.4米、两板朝前误差各不超过30度，无额外静止保持窗口。", + "step": { + "cat": "本体", + "item": "翻腾" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双腿和双靴经固定雪板承重，雪板与坡面接触完成助滑、落地及滑出；空中无雪面接触是依次发生的阶段，记于节点说明。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:3-30", + "inputs/task/standard/instruction.md:3-30", + "inputs/task/privileged/tests/rfo_task.py:778-805", + "inputs/task/privileged/tests/rfo_task.py:813-820", + "inputs/task/privileged/tests/rfo_task.py:908-914", + "inputs/task/privileged/tests/rfo_task.py:978-981", + "inputs/task/privileged/tests/rfo_task.py:1047-1062", + "inputs/task/privileged/tests/rfo_task.py:1100-1207", + "inputs/task/privileged/tests/rfo_task.py:1209-1284", + "inputs/task/privileged/tests/rfo_task.py:1299-1305", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178", + "inputs/oracle/run.json:reward", + "inputs/oracle/replay.mp4 @ 0-17.60 s" + ], + "conf": 6, + "note_zh": "依S8、S10，把起滑、助滑、起跳、转体、雪板落地及受控滑出作为同一次连续大跳台动作；四个原生裁判阶段只用于验收，不拆成进度节点。event在完整滑出后锁存，开局不计分；没有额外落地静置或恢复平衡步骤。双脚已固定在雪板上，无拿取工具步骤；接触采用整步的支承交互，腾空时依次变为无接触，不把它另列为同时接触或替代方案。" + } + ], + "constraints": [ + { + "id": "non_boot_snow_contact", + "zh": "身体触雪", + "never": null, + "criterion_zh": "在整次尝试中，除双靴外的任一机器人部位（包括手、膝、骨盆、躯干）接触场景的雪面即失败；空中动作结束时身体先触雪也包括在内，双靴及其固定雪板触雪允许。裁判按接触距离不大于零的实际接触判定。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:833-836", + "inputs/task/privileged/tests/rfo_task.py:1133-1157", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:195-224" + ] + }, + { + "id": "down_too_long", + "zh": "持续失衡", + "never": null, + "criterion_zh": "起滑后，在非主跳腾空阶段且雪板或靴接触雪面时，骨盆沿雪面法线的净高小于0.4米，或躯干相对雪面法线倾斜超过70度，即累计低伏时间；按每0.01秒控制步增加低伏时间，恢复时以四分之一速率衰减至零，累计达到1.0秒即失败。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:1254-1279", + "inputs/task/privileged/tests/rfo_task.py:1065-1072", + "inputs/task/privileged/tests/rfo_task.py:912-914" + ] + }, + { + "id": "snow_spin", + "zh": "雪上转体", + "never": null, + "criterion_zh": "首次落地0.5秒后固定两块雪板各自的连续展开航向为基准，此后至终点，任一雪板航向相对自己的基准变化超过90度即失败；比较展开航向相对基准的净偏差，不把往返转动的绝对路程累加。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:1186-1197", + "inputs/task/privileged/tests/rfo_task.py:1109-1117", + "inputs/task/privileged/tests/rfo_task.py:912-914" + ] + }, + { + "id": "stalled", + "zh": "滑行停滞", + "never": null, + "criterion_zh": "起滑阶段成立后，骨盆三维线速度模长连续3.0秒小于0.3米每秒即失败；任一控制步速度达到0.3米每秒便清零该连续计数。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:1260-1263", + "inputs/task/privileged/tests/rfo_task.py:1280-1284", + "inputs/task/privileged/tests/rfo_task.py:912-914" + ] + }, + { + "id": "did_not_start", + "zh": "未及时起滑", + "never": null, + "criterion_zh": "到20.0秒仍处于起点平台阶段即判未出发；起滑要求两块雪板的绑定主体中心都越过平台前沿，即各自的世界横坐标均大于6.0米。此为发令后的未出发判罚，不是40秒总时限。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:1158-1161", + "inputs/task/privileged/tests/rfo_task.py:1260-1263", + "inputs/task/privileged/tests/rfo_task.py:909-914" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "dyn.freeski_360_ridden_away", + "level": "new_code", + "source": "inputs/task/privileged/tests/rfo_task.py:1133-1284; record per-frame native referee.stages, referee.success, referee.failed, task.trackers.result, task.trackers.phase and skis. Needed for the complete world event; no state log supplied." + }, + { + "atom": "dyn.non_boot_snow_contact", + "level": "new_code", + "source": "inputs/task/privileged/tests/rfo_task.py:1133-1157; log body_on and snow contact sets at each physics substep" + }, + { + "atom": "dyn.down_too_long", + "level": "new_code", + "source": "inputs/task/privileged/tests/rfo_task.py:1264-1279; log task.trackers.down_s, pelvis normal height and torso tilt" + }, + { + "atom": "dyn.snow_spin", + "level": "new_code", + "source": "inputs/task/privileged/tests/rfo_task.py:1186-1197; log task.trackers.head, head_ref, land_ph and ride_turn" + }, + { + "atom": "dyn.stalled", + "level": "new_code", + "source": "inputs/task/privileged/tests/rfo_task.py:1280-1284; log task.trackers.slow_k and pelvis linear speed" + }, + { + "atom": "dyn.did_not_start", + "level": "new_code", + "source": "inputs/task/privileged/tests/rfo_task.py:1158-1161,1260-1263; log phase, time and ski binding-center x coordinates" + } + ] + }, + "native_metrics": [ + { + "name": "trick_spin_deg", + "unit": "deg", + "threshold": 360, + "direction": "target", + "note_zh": "目标为整周；实际验收是两板空中净转角均值的绝对值属于[270,450)，原始带符号数值允许正负，按半周单位取最近整数。", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:778-805", + "inputs/task/privileged/tests/rfo_task.py:1214-1221", + "inputs/task/privileged/tests/rfo_task.py:1299-1305" + ] + }, + { + "name": "landing_heading_deg", + "unit": "deg", + "threshold": 45, + "direction": "lower_better", + "note_zh": "原生值是两块雪板首次触雪时的带符号航向误差列表；逐板比较绝对值，不超过45度。", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:1233-1240" + ] + }, + { + "name": "ride_turn_deg", + "unit": "deg", + "threshold": 90, + "direction": "lower_better", + "note_zh": "从落地0.5秒后的各自基准航向起，两板最大绝对连续航向偏差的历史最大值；超过90度立即失败。", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:1186-1199" + ] + } + ], + "open_questions_zh": [ + "inputs/states/index.json的logs为空；依P8全部节点when、约束never及success均为null，无法做状态回放的达标和评分一致性核验，atoms仅列出需要新增记录的世界状态，未把评分文件或复现散列绑定为节点条件。", + "视频可见性缺口：已查看17.64秒参考视频的全程接触表和原分辨率12秒、17.5秒帧，画面主要是被近景雪坡遮挡的固定场地远景，不能可靠辨认机器人、雪板转角、首次触雪或过线；因此判据描述预期可见动作并保留精确裁判标准，不伪称在此视频看到成功动作。成功由run.json的reward报告：四阶段完成、转体355.0度、空翻0、落点在圆顶之后3.17米、滑出最大转向13.6度。", + "评分器要求，文本未提：确认主跳需连续至少30个1毫秒物理步无雪面接触；文本只定义离雪至下次触雪，未给此确认窗口（rfo_task.py:1133-1176）。该窗口是同一跳跃的验收，不另建腾空节点，也不是节点hold_s。", + "文本与评分器边界差异：文本写270至450度，实际floor(abs(spin)/180+0.5)==2为[270,450)；文本写距唇口4米以内、仰角至少10度，代码分别严格小于4米、严格大于10度（rfo_task.py:785,1167）。本文件按代码给出精确验收。", + "两模式任务代码字节一致，起始状态散列及参数一致，世界成功标准相同；privileged外层另要求有限且格式正确的轨迹在两个新进程重放均成功且最终状态散列一致，standard评分最后一次实时服务尝试（privileged grade.py:90-115；standard server.py:157-178；standard grade.py:93-99）。这些交付与运行有效性要求不构造语义节点，无状态日志时也不能在success中绑定。", + "粒度采用S8、S10的连续动作阅读：整次360只有在受控滑过终点后被接受，起跳、空中转体、落地和滑出不拆开；不是按四个原生阶段建四步。起点站立属于全程不摔倒约束，雪板已安装，不增站立或拿板步骤。", + "幅度、风格、抓板不参与判分；air_s、height_m、from_knoll_m及躯干自身转角等原生记录不产生额外步骤或约束，其中落点本身仍须通过台面末端、雪板转角仍按专用测量验收。40秒总时限依P21不列约束；20秒未出发是显式DNS判罚，单独保留。" + ], + "conf": 6, + "conf_reason_zh": "连续动作粒度来自文本和S10，全部验收阈值及规则失败均有裁判代码依据，两模式代码与初态一致；但没有逐帧状态日志，参考视频远景无法可靠辨认动作，不能独立复核每个阶段的视觉完成时刻，仅有成功记录支持整体完成。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "文本的助滑、起跳、空中转360、雪板落地、滑过终点是一次连续动作，按S10合为一个本体·翻腾节点合理；约束（身体触雪、持续失衡、雪上转体、停滞、未出发）与文本规则相符，无状态日志故when/never为null并已在open_questions_zh说明；验证器通过。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：['spin_360_ridden_away']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:49:39Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__freeski_360/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__freeski_360/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "2019b2e2988504c7c953fb31c8367827d679ad895fe2109b4836a3994d07d4f9", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/embodied-olympics-fun/freeski_360/trajectories/embodied-olympics-fun-freeski-360-i00-privileged--pa2yunn.json b/phase2/embodied-olympics-fun/freeski_360/trajectories/embodied-olympics-fun-freeski-360-i00-privileged--pa2yunn.json new file mode 100644 index 0000000000000000000000000000000000000000..afb43c50f0c41f6d9e8d6a638d228325182c628f --- /dev/null +++ b/phase2/embodied-olympics-fun/freeski_360/trajectories/embodied-olympics-fun-freeski-360-i00-privileged--pa2yunn.json @@ -0,0 +1,190 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-freeski-360-i00-privileged--pa2yunn", + "taskKey": "embodied-olympics-fun/freeski_360", + "dag_sha256": "861efa8ce64a6c121798b9927345b65cbb761cb3b91579bd06a292aa528db572", + "outcome": { + "success": 1, + "reward": 1.0, + "source": "catalog", + "raw": { + "success": 1, + "replay_success": 1, + "deterministic": 1, + "n_actions": 1763, + "sim_steps": 1763, + "horizon": 4000, + "episode_over": 1, + "progress": 1.0, + "stages_reached": 4, + "stages_total": 4, + "fell": 0, + "replay_s": 19.7, + "bad_trajectory": 0, + "metric_air_s": 1.92, + "metric_from_knoll_m": 3.17, + "metric_height_m": 2.64, + "metric_ride_turn_deg": 13.6, + "metric_speed_m_s": 18.198, + "metric_trick_flip_deg": -74.8, + "metric_trick_flips": 0.0, + "metric_trick_half_turns": 2.0, + "metric_trick_pelvis_spin_deg": 324.6, + "metric_trick_short_deg": 5.0, + "metric_trick_spin_deg": 355.0, + "grader_error": 0, + "video_rendered": 1, + "agent_sessions": 136, + "agent_resets": 0, + "agent_replays": 55, + "agent_cost_self_reported": 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/embodied-olympics-fun-freeski-360-i00-privileged--pa2yunn/replay.mp4", + "duration_s": 17.64, + "fps": 25.0, + "frames": 441, + "sha256": "ac85c85d4e528032a2b390ff97ebcaf5a7b045f99f78980a7004730009bd8eb3" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "spin_360_ridden_away": { + "method": "visual", + "intervals": [ + [ + 17.6, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 12.0, + "log_step": null, + "text_zh": "观察：已检查视频全程每约0.5秒的帧及视频12.0 s原分辨率帧；场地远景中近侧雪坡占据大部分画面，无法可靠辨认起滑、空中雪板转角及首次触雪。" + }, + { + "t": 17.6, + "log_step": null, + "text_zh": "观察：视频15.6–17.6 s缓冲区局部放大及视频16.6–17.6 s密集帧中，末段有微小灰色人形轮廓向画面右侧终点方向移动；视频17.6 s原分辨率帧仍不足以分辨双雪板、身体部位与雪面的接触或双板是否均已越线。" + }, + { + "t": null, + "log_step": 273, + "text_zh": "观察：log step 273中，agent对最终封装轨迹的两次新进程回放检查输出均为success、执行1763行且最终状态散列相同。inputs/run.json的评分器reward另报告success=1、replay_success=1、deterministic=1、stages_reached=4/4、fell=0、雪板空中转角355.0度、空翻数0、落点越过圆顶3.17米、滑出转向13.6度；这些是评分与交付记录，不把会话中的较早试验当成本次动作。" + }, + { + "t": 17.6, + "log_step": null, + "text_zh": "推断：评分器四阶段成功决定唯一目标节点在终局为真；已检查的视频末段没有直接反证。区间起点以视频17.6 s末帧作完成的近似终局锚点，精确首次过线时刻无法从远景独立核实；未将agent工作时间或会话试验时刻换算成视频时间。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察：视频16.6–17.6 s密集帧和末段局部放大只能辨出微小轮廓移动，无法分辨腿、双靴及固定雪板的具体承重接触；没有逐帧状态接触日志，实际接触记为未知，不照抄任务DAG的典型支撑方案。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "trick_spin_deg", + "unit": "deg", + "threshold": 360, + "direction": "target", + "value": 355.0 + }, + { + "name": "landing_heading_deg", + "unit": "deg", + "threshold": 45, + "direction": "lower_better", + "value": null + }, + { + "name": "ride_turn_deg", + "unit": "deg", + "threshold": 90, + "direction": "lower_better", + "value": 13.6 + } + ], + "failure": null, + "summary_zh": "推断：依据评分器success=1及四阶段全部完成，本次交付完成唯一“转体落地滑出”节点；评分器记录雪板空中转角355.0度、空翻数0及滑出转向13.6度，无失败断点。观察：log step 273的最终文件两次新进程回放均成功且最终状态散列一致。观察：视频远景及末段放大无法独立核实转体、首次触雪与精确过线时刻，故视频17.6 s仅作终局近似锚点，实际接触记为未知；没有状态日志，评分器reward未报告landing_heading_deg，该原生指标留空。", + "conf": 6, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "评分器success=1、4/4阶段、fell=0，目标节点final=true、无失败断点，与证据一致；远景视频无法辨认机器人，故接触记未知、17.6 s仅作推断锚点，处理诚实。", + "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": "2026-10-29T10:06:44Z", + "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/embodied-olympics-fun-freeski-360-i00-privileged--pa2yunn/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-freeski-360-i00-privileged--pa2yunn/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "a0891200dbd3f16d86d46ae1900db69681fa95e3613702a80b8b38bdf79945b2", + "video_sha256": "ac85c85d4e528032a2b390ff97ebcaf5a7b045f99f78980a7004730009bd8eb3", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/freeski_360/trajectories/embodied-olympics-fun-freeski-360-i00-standard--gqg69cq.json b/phase2/embodied-olympics-fun/freeski_360/trajectories/embodied-olympics-fun-freeski-360-i00-standard--gqg69cq.json new file mode 100644 index 0000000000000000000000000000000000000000..c6b7e5dfa12552bfc0c05075cf97c8ebe06a86a0 --- /dev/null +++ b/phase2/embodied-olympics-fun/freeski_360/trajectories/embodied-olympics-fun-freeski-360-i00-standard--gqg69cq.json @@ -0,0 +1,198 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-freeski-360-i00-standard--gqg69cq", + "taskKey": "embodied-olympics-fun/freeski_360", + "dag_sha256": "861efa8ce64a6c121798b9927345b65cbb761cb3b91579bd06a292aa528db572", + "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": 1167, + "sim_steps": 5590, + "calls": 29, + "rows": 5590, + "resets": 5, + "reset_budget": 5, + "sim_startups": 1, + "elapsed_s": 642.7, + "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": 0.0, + "peak_memory_gb": 0.187, + "peak_processes": 6, + "peak_disk_gb": 0.252, + "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/embodied-olympics-fun-freeski-360-i00-standard--gqg69cq/replay.mp4", + "duration_s": 11.68, + "fps": 25.0, + "frames": 292, + "sha256": "9c6b29ba6c1b8c8d804c5f80ae514c312a14f79b65f624fff776739d1dbd9079" + }, + "state": null, + "progress": { + "method": "log", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "spin_360_ridden_away": { + "method": "log", + "intervals": [], + "final": false, + "evidence": [ + { + "t": null, + "log_step": 21, + "text_zh": "观察：log step 21 的机器人服务规范说明仅最后一次 episode 参与评分，结束后 meta.outcome 是裁判判决、stages_reached/stages_total 是裁判完成阶段数。log step 89 启动最终控制脚本；log step 90 显示重置进入 episode 6，助滑控制提交结束时裁判返回 outcome=fell、stages_reached=1/4、仿真 11.67 s；log step 92 确认该轮 done=True、reason=failed，之后没有新尝试。" + }, + { + "t": 11.64, + "log_step": 90, + "text_zh": "观察：已检查视频全程约每秒帧、助滑坡区域约每半秒放大帧、视频 10.0–11.6 s 约每0.2秒帧及视频 11.64 s 末帧，固定远景主要显示雪坡和护栏，无法可靠辨认机器人及雪板接触；参考视频全程也存在相同遮挡。推断：依据最终评分轮裁判仅1/4阶段且以摔倒终止，完整的转体落地滑出事件从未成立，非单凭总分为零判断。" + } + ], + "contact_actual": [ + { + "who": "未知", + "how": "未知", + "note_zh": "观察：视频全程远景及视频 10.0–11.64 s 的密集末段、放大坡面均不能可靠辨认双腿、雪板或身体与雪面的实际接触，且没有状态接触日志。log step 90 证明最终轮已进行起滑及助滑尝试，但不能据此把任务典型支撑链复制为已观察到的接触。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "trick_spin_deg", + "unit": "deg", + "threshold": 360, + "direction": "target", + "value": null + }, + { + "name": "landing_heading_deg", + "unit": "deg", + "threshold": 45, + "direction": "lower_better", + "value": null + }, + { + "name": "ride_turn_deg", + "unit": "deg", + "threshold": 90, + "direction": "lower_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "spin_360_ridden_away", + "t": null, + "log_step": 90, + "where": "操作中", + "evidence_zh": "观察：最终第6轮中，agent 的脚本正在提交助滑、缓冲及起跳准备的 lowcmd 关节控制；log step 90 打印该次服务返回 done=True、outcome=fell、stages_reached=1/4、仿真 11.67 s，尚未进入空中控制段。log step 92 确认该轮已结束且之后未重启；视频全程及末秒密集帧无法辨认失衡起始姿态，故不指定视频断点时刻。推断：失败发生在唯一节点的助滑操作中，完整动作未获修复。" + }, + "unmet": { + "spin_360_ridden_away": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": null, + "log_step": 90, + "evidence_zh": "观察：log step 90 所记录的最终第6轮机器人服务裁判返回 outcome=fell，仅完成1/4阶段；log step 21 说明该字段为原生裁判给出的终止原因，而非 agent 对姿态的自述。视频末段至视频 11.64 s 不能可靠看清倒地方式，因此这里只记录裁判明确报告的摔倒，不细分身体哪一部位先触雪。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志", + "evidence_zh": "推断：助滑姿态控制未能维持全身稳定。log step 88 的最终脚本为滑行配置对称腿部目标、躯干前倾、PD增益及踝前馈力矩，log step 89 实际执行，log step 90 在助滑控制段得到裁判 fell。会话辅助证据为 log step 59–60 的第3轮在仿真 11.1–11.64 s 出现俯仰快速变化并摔倒，log step 67、78 的第4、5轮修改控制后仍在起跳前以 fell、1/4结束；这些测试只支持控制层归因，不计作最终轮进度或现象。缺少最终轮逐步姿态与力矩，无法确定具体控制子类，也不能把第3轮的踝力矩饱和直接归给第6轮。" + } + ], + "gap_zh": "本次与参考视频均为受雪坡遮挡的固定远景，本次全程和末秒密集检查无法可靠辨认机器人、雪板、接触或失衡起始时刻；未提供状态日志。节点未完成及摔倒采用已确认属于最终第6轮的原生裁判字段，视频断点时刻留空、contact_actual记未知。log step 92 的1167个控制步与log step 7 的0.01秒周期相乘为仿真11.67秒，与视频11.68秒时长相符，仅用于核对评分轮身份，不把不可见的摔倒起始点硬映射到视频。最终控制脚本中段在所给记录中被截断，且缺少第6轮逐步遥测，根因仅推断到控制层；三个原生指标均未报告，记为null。" + }, + "summary_zh": "评分对应最后一次第6轮，裁判报告起滑后仅完成1/4原生阶段，在助滑控制段以 fell 结束，唯一的“转体落地滑出”节点未完成（log step 90、92）。结合最终脚本和此前姿态控制试验，推断助滑控制未维持平衡，具体机制未证实。全程视频及末段放大检查仍无法可靠辨认机器人和实际接触，也无状态日志，因此接触记未知、视频断点时刻留空。", + "conf": 6, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "标注与证据一致：评分轮为第6轮，裁判返回 outcome=fell、stages 1/4，唯一节点未达成；视频为遮挡远景，看不到机器人，因此 contact 记未知、断点时刻留空是合理的。现象“平衡”来自裁判 fell，根因“控制”只推断到控制层，已如实标为推断。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": "2026-10-29T10:07:37Z", + "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/embodied-olympics-fun-freeski-360-i00-standard--gqg69cq/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-freeski-360-i00-standard--gqg69cq/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "f2539890c3d3c25294070b165f1888a8bf52cc681bc9c69dcc9a3db2897c92be", + "video_sha256": "9c6b29ba6c1b8c8d804c5f80ae514c312a14f79b65f624fff776739d1dbd9079", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/golf_putt/task_dag.json b/phase2/embodied-olympics-fun/golf_putt/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..5006756039839bcba9b9ff4040841eedf524f237 --- /dev/null +++ b/phase2/embodied-olympics-fun/golf_putt/task_dag.json @@ -0,0 +1,310 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "embodied-olympics-fun/golf_putt", + "suite": "embodied-olympics-fun", + "base": "embodied-olympics-fun-golf-putt-i00", + "title_zh": "高尔夫一杆入洞", + "summary_zh": "右手始终持杆，以一次合法杆头击球将三米外的白球送入球洞，使全球低于果岭表面并在洞内停稳一秒，且全程没有规则违规。", + "embodiment_zh": "两模式均为 Unitree G1 双足人形机器人，躯干及双臂共 29 个本体关节；右臂末端为 Dex3-1 三指灵巧手，开局靠手指接触与摩擦握持独立自由体推杆，左臂末端为橡胶拳式手及抽象握持接口，可辅助握柄；双脚支撑，无轮式底盘。", + "task_class": "atomic", + "analysis_level": "step", + "spec": { + "version": "1.2", + "task": "embodied-olympics-fun-golf-putt-i00", + "bench": "embodied-olympics-fun", + "progress": "live", + "entities": {}, + "milestones": [ + { + "id": "ball_holed", + "zh": "一杆推球入洞", + "mode": "state", + "when": null, + "after": [], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "场地相机中，右手持推杆击出脚前白球，白球沿果岭滚向远处浅色条带上的圆形洞口并落下，在洞口及洞沿不再露出；有洞口近景时须看到全球落到果岭表面以下，洞内停稳 1.0 秒（球心相对静止参考点位移不超过 0.001 米，水平距洞心小于 0.054 米且球心高度加球半径 0.02135 米小于 0）。仅挥杆、白球滚向球洞、经过洞口或停在洞沿均未完成；裁判须已按顺序记录 stroke、holed 且没有规则违规，状态节点在球离洞或不再满足停稳条件时不成立。", + "step": { + "cat": "动态", + "item": "击打" + }, + "contact": [ + { + "who": "手持工具", + "how": "其他", + "plan": 1, + "note_zh": "右手握持的推杆杆头击打白球；候选接触名为“击打”，现有接触方式没有击打项。仅记录推杆至球的工具链，不另加握杆接触；规则允许左手同时辅助握柄。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "plan": 1, + "note_zh": "机器人双脚在果岭上支撑身体，同时右手挥动推杆击球。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:4-6", + "inputs/task/standard/instruction.md:3-5", + "inputs/task/privileged/environment/rfo_task.py:305-316", + "inputs/task/privileged/environment/rfo_task.py:358-395", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/scene.py:62", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178", + "inputs/oracle/replay.mp4 @ 0.00-6.64 s", + "inputs/oracle/run.json: steps[id=330], reward" + ], + "conf": 6, + "note_zh": "按 S1、S8、S10，击球、滚动与入洞停稳构成一个连续步骤；原生 stroke 是过程标志，不单列节点。初始右手已经持杆，失去持杆状态会立即失败，故不设取杆或保持持杆节点。停稳验收已包含完整 1.0 秒窗口，hold_s=0 避免重复计时；无日志时依据视频和裁判记录判断。" + } + ], + "constraints": [ + { + "id": "putter_released", + "zh": "推杆脱手", + "never": null, + "criterion_zh": "推杆握柄未同时接触右手食指或中指以及拇指或掌面，连续中断超过 0.05 秒即违规；左手辅助不能代替右手持杆。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:326-350" + ] + }, + { + "id": "club_anchored", + "zh": "球杆触身", + "never": null, + "criterion_zh": "推杆任一部位接触机器人双手以外的身体部位即违规；持杆时杆头接触果岭允许。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:351-352", + "inputs/task/privileged/environment/rfo_task.py:189" + ] + }, + { + "id": "illegal_ball_touch", + "zh": "违规触球", + "never": null, + "criterion_zh": "机器人身体任一部位，或推杆杆身、握柄接触白球即违规；仅推杆杆头及杆面允许触球。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:320-324", + "inputs/task/privileged/environment/rfo_task.py:353-356" + ] + }, + { + "id": "illegal_stance", + "zh": "跨线击球", + "never": null, + "criterion_zh": "首次杆头触球时，双脚与地面的接触点在球初始位置至洞心直线的两侧，或任何接触点在线上，即违规；代码检查接触点 x 坐标的最小值与最大值是否夹住 0.44 米。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:359-365", + "inputs/task/privileged/environment/rfo_task.py:192" + ] + }, + { + "id": "ball_pushed", + "zh": "推压白球", + "never": null, + "criterion_zh": "杆头与白球累计接触时间超过 0.03 秒即违规，包含推、刮、舀等长接触。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:368-371", + "inputs/task/privileged/environment/rfo_task.py:194" + ] + }, + { + "id": "second_stroke", + "zh": "二次击球", + "never": null, + "criterion_zh": "首次杆头触球超过 0.2 秒后再有杆头触球即违规；0.2 秒内的意外多次接触允许，但仍受累计 0.03 秒上限约束。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:358-371", + "inputs/task/privileged/environment/rfo_task.py:194" + ] + }, + { + "id": "ball_off_green", + "zh": "球离果岭", + "never": null, + "criterion_zh": "白球接触低于果岭表面的外围地面且球心水平距洞心大于 0.054 米时，裁判判定球离开果岭；视频可见白球越过矩形果岭边缘跌落。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:270-273", + "inputs/task/privileged/environment/rfo_task.py:372-374", + "inputs/task/privileged/environment/rfo_task.py:190" + ] + }, + { + "id": "ball_rest_outside", + "zh": "洞外停球", + "never": null, + "criterion_zh": "击球后白球不再接触杆头，球心相对静止参考点的位移不超过 0.001 米达 1.0 秒，但水平距洞心不小于 0.054 米或球顶未低于果岭表面，即洞外停球；停在洞沿也失败。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:385-395", + "inputs/task/privileged/environment/rfo_task.py:190-194" + ] + }, + { + "id": "robot_fallen", + "zh": "机器人跌倒", + "never": null, + "criterion_zh": "骨盆低于果岭表面 0.35 米，或除双脚以外的机器人身体部位接触地面即跌倒；视频中机器人跪地、手撑地或躺倒均违规，球杆本身触地允许。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:189-204", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:276-282", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/task.py:21-23" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "ball_holed_and_resting", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:385-393; log ball position, Task.trackers.first/still_at/still_since, ref.stages and ref.success; radius from inputs/grader/efe_embodied-olympics-fun-files/rfo/scene.py:62" + }, + { + "atom": "putter_released", + "level": "new_code", + "source": "Log the per-frame referee failure reason and relevant contact/pose tracker: inputs/task/privileged/environment/rfo_task.py:326-350" + }, + { + "atom": "club_anchored", + "level": "new_code", + "source": "Log the per-frame referee failure reason and relevant contact/pose tracker: inputs/task/privileged/environment/rfo_task.py:351-352, inputs/task/privileged/environment/rfo_task.py:189" + }, + { + "atom": "illegal_ball_touch", + "level": "new_code", + "source": "Log the per-frame referee failure reason and relevant contact/pose tracker: inputs/task/privileged/environment/rfo_task.py:320-324, inputs/task/privileged/environment/rfo_task.py:353-356" + }, + { + "atom": "illegal_stance", + "level": "new_code", + "source": "Log the per-frame referee failure reason and relevant contact/pose tracker: inputs/task/privileged/environment/rfo_task.py:359-365, inputs/task/privileged/environment/rfo_task.py:192" + }, + { + "atom": "ball_pushed", + "level": "new_code", + "source": "Log the per-frame referee failure reason and relevant contact/pose tracker: inputs/task/privileged/environment/rfo_task.py:368-371, inputs/task/privileged/environment/rfo_task.py:194" + }, + { + "atom": "second_stroke", + "level": "new_code", + "source": "Log the per-frame referee failure reason and relevant contact/pose tracker: inputs/task/privileged/environment/rfo_task.py:358-371, inputs/task/privileged/environment/rfo_task.py:194" + }, + { + "atom": "ball_off_green", + "level": "new_code", + "source": "Log the per-frame referee failure reason and relevant contact/pose tracker: inputs/task/privileged/environment/rfo_task.py:270-273, inputs/task/privileged/environment/rfo_task.py:372-374, inputs/task/privileged/environment/rfo_task.py:190" + }, + { + "atom": "ball_rest_outside", + "level": "new_code", + "source": "Log the per-frame referee failure reason and relevant contact/pose tracker: inputs/task/privileged/environment/rfo_task.py:385-395, inputs/task/privileged/environment/rfo_task.py:190-194" + }, + { + "atom": "robot_fallen", + "level": "new_code", + "source": "Log the per-frame referee failure reason and relevant contact/pose tracker: inputs/task/privileged/environment/rfo_task.py:189-204, inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:276-282, inputs/grader/efe_embodied-olympics-fun-files/rfo/task.py:21-23" + } + ] + }, + "native_metrics": [ + { + "name": "ball_to_hole_m", + "unit": "m", + "threshold": 0.054, + "direction": "lower_better" + }, + { + "name": "ball_speed_m_s", + "unit": "m/s", + "threshold": null, + "direction": "lower_better" + }, + { + "name": "head_contact_s", + "unit": "s", + "threshold": 0.03, + "direction": "lower_better" + }, + { + "name": "grip_break_max_s", + "unit": "s", + "threshold": 0.05, + "direction": "lower_better" + } + ], + "open_questions_zh": [ + "输入 inputs/states/index.json 的 logs 为空。按 P8，唯一节点 when、全部约束 never 及 success 均为 null；atoms 标明需补录的量。成功视频只有远景，毫米级静止、全球高度及毫秒级接触规则无法仅凭视频精确核验。", + "按 S10 将裁判的 stroke → holed 合为一个步骤：击球后的滚动及入洞不能暂停再续做；开局已持杆，Y17 工具取用例外不产生取杆节点。成功参考记录有 head_hits=2、head_contact_s=0.025，说明“一杆”并非仅允许一次接触。", + "文本与评分器边界差异：instruction.md 说 1.0 秒内球心移动小于 0.001 米；rfo_task.py:387-390 仅在位移大于 0.001 米时重置计时，因此恰好等于也接受。采用代码的固定参考点及不超过阈值；ball_speed_m_s 没有独立验收阈值，head_hits 只是诊断计数，不要求等于 1（rfo_task.py:380-395）。", + "两模式物理裁判相同。privileged 外层评分要求两次重放均成功且末态哈希一致（inputs/grader/efe_embodied-olympics-fun-files/grade.py:106-115）；standard 取服务末次现场回合的裁判结果（inputs/grader/efe_embodied-olympics-fun-files/standard/server.py:157-178；inputs/grader/efe_standard-files/v2_1/grade.py:93-99）。交付有效性及基础设施条件不能绑定语义节点；无状态日志也不能自动核对这些条件。" + ], + "conf": 6, + "conf_reason_zh": "一个入洞步骤由文本直接给出，起始持杆状态及所有验收阈值均有源码依据，成功参考视频与裁判记录相符；但没有状态日志，远景视频无法精确复核洞内停稳及细粒度接触，故置信度为 6。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "任务文本只描述一个步骤（一杆把球推入洞），单节点 ball_holed 与评分器的 stroke→holed 一致；约束与 rfo_task.py 的失败条件逐条对应，均为立即终止；无状态日志，when/never/success 为 null 并写入 open_questions_zh，处理合理。阈值（0.054、0.001、1.0 s、0.05 s、0.2 s、0.03 s）与代码一致。标签（动态·击打；手持工具×其他 + 腿/躯干×支撑）符合 S16。", + "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：['ball_holed']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:49:26Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__golf_putt/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__golf_putt/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "382a8cfb2ac61e2f7a20dd5814f07d84c988f5cd4a788b1596e6d61223f57513", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/embodied-olympics-fun/golf_putt/trajectories/embodied-olympics-fun-golf-putt-i00-privileged--ecpkzxa.json b/phase2/embodied-olympics-fun/golf_putt/trajectories/embodied-olympics-fun-golf-putt-i00-privileged--ecpkzxa.json new file mode 100644 index 0000000000000000000000000000000000000000..8632ae67c98d42cff0e78f2977c65347a9a69b46 --- /dev/null +++ b/phase2/embodied-olympics-fun/golf_putt/trajectories/embodied-olympics-fun-golf-putt-i00-privileged--ecpkzxa.json @@ -0,0 +1,196 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-golf-putt-i00-privileged--ecpkzxa", + "taskKey": "embodied-olympics-fun/golf_putt", + "dag_sha256": "74a1c00f893c3316873e4cc2017a91bda4a2926d26f9ae1a009d7dbb3a3e0854", + "outcome": { + "success": 1, + "reward": 1.0, + "source": "catalog", + "raw": { + "success": 1, + "replay_success": 1, + "deterministic": 1, + "n_actions": 534, + "sim_steps": 534, + "horizon": 3000, + "episode_over": 1, + "progress": 1.0, + "stages_reached": 2, + "stages_total": 2, + "fell": 0, + "replay_s": 7.18, + "bad_trajectory": 0, + "metric_ball_speed_m_s": 0.002, + "metric_ball_to_hole_m": 0.02, + "metric_grip_break_max_s": 0.002, + "metric_head_contact_s": 0.013, + "metric_head_hits": 1.0, + "grader_error": 0, + "video_rendered": 1, + "agent_sessions": 200, + "agent_resets": 2011, + "agent_replays": 11, + "agent_cost_self_reported": 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/embodied-olympics-fun-golf-putt-i00-privileged--ecpkzxa/replay.mp4", + "duration_s": 5.36, + "fps": 25.0, + "frames": 134, + "sha256": "28c6133367d6ff4d0086e4c9d7725ac5ea5ed5ed13eafb7ed49753d0becf4042" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "ball_holed": { + "method": "visual", + "intervals": [ + [ + 5.32, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 1.64, + "log_step": null, + "text_zh": "观察：视频 1.48–1.64 s，机器人右手持推杆向脚前白球挥动杆头，白球离开初始位置；视频 1.72–3.20 s，白球持续沿果岭滚向浅色条带上的洞口，机器人双脚保持着地支撑。" + }, + { + "t": 3.68, + "log_step": null, + "text_zh": "观察：视频 3.20–3.64 s 的洞口裁剪帧中，白球从近侧接近洞口，随后在洞口处消失；视频 3.68–5.32 s 未再见白球露出洞口或滚回果岭。成功参考仅用于识别画面线索：参考视频 4.44–5.16 s 同样显示白球接近洞口后消失，参考视频 5.28–6.60 s 洞口不再露球。" + }, + { + "t": 5.32, + "log_step": null, + "text_zh": "观察：视频 4.20–5.32 s 逐帧检查中，机器人仍站立持杆，洞口无白球露出。inputs/run.json 的评分器 reward 记录 success=1、replay_success=1、deterministic=1、stages_reached=2/2、fell=0、metric_head_hits=1，末态球心水平距洞心 metric_ball_to_hole_m=0.020 m。" + }, + { + "t": 5.32, + "log_step": null, + "text_zh": "推断：结合白球在洞口消失且未再露出，以及评分器对本次交付的成功重放判定，一杆推球入洞节点最终成立。完整验收还要求洞内停稳 1.0 秒；无状态日志，远景无法量出洞内毫米级运动，故首次完成时间近似记为视频末帧 5.32 s（约视频 5.3 s），精确时刻依赖裁判结果，不能仅由入洞画面确定。" + } + ], + "contact_actual": [ + { + "who": "手持工具", + "how": "其他", + "note_zh": "观察：视频 1.48–1.64 s，机器人以右手持推杆挥击白球，视频 1.64 s 后白球开始滚离；作用端为推杆杆头，对象为白球。候选接触方式“击打”，现有方式无对应击打项；同一工具链不重复记录握杆。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：视频 0.00–5.32 s，机器人双脚着地承重，右臂移杆和挥杆时身体保持站立，击球后继续站在原处。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "ball_to_hole_m", + "unit": "m", + "threshold": 0.054, + "direction": "lower_better", + "value": 0.02 + }, + { + "name": "ball_speed_m_s", + "unit": "m/s", + "threshold": null, + "direction": "lower_better", + "value": 0.002 + }, + { + "name": "head_contact_s", + "unit": "s", + "threshold": 0.03, + "direction": "lower_better", + "value": 0.013 + }, + { + "name": "grip_break_max_s", + "unit": "s", + "threshold": 0.05, + "direction": "lower_better", + "value": 0.002 + } + ], + "failure": null, + "summary_zh": "机器人右手持推杆完成一次击球，白球约在视频 3.6–3.7 s 到达洞口并消失，直到视频 5.32 s 末帧未再露出；结合评分器成功重放与阶段 2/2，唯一目标节点最终完成。完整停稳验收的完成时间近似标在视频 5.32 s；没有状态日志，洞内毫米级静止及毫秒级接触不能由远景独立核验，采用评分器结果与原生指标佐证。", + "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 ball_holed 成立，评分器 success=1、stages 2/2；视频 3.6 s 球到洞口消失，之后不再露出，末帧时间 5.32 s 与约 4.7 s 的停稳时刻相差不到一秒，接触标注与证据措辞合规，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": "2026-10-29T10:06:48Z", + "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/embodied-olympics-fun-golf-putt-i00-privileged--ecpkzxa/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-golf-putt-i00-privileged--ecpkzxa/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "dcae8ef5439586a050e57b5ffe628046a5b49974a9a3dc8831939d1c7d6ccb63", + "video_sha256": "28c6133367d6ff4d0086e4c9d7725ac5ea5ed5ed13eafb7ed49753d0becf4042", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/golf_putt/trajectories/embodied-olympics-fun-golf-putt-i00-privileged--hayytdm.json b/phase2/embodied-olympics-fun/golf_putt/trajectories/embodied-olympics-fun-golf-putt-i00-privileged--hayytdm.json new file mode 100644 index 0000000000000000000000000000000000000000..917c8ebca020800eebd410c2bfa74d069911cda9 --- /dev/null +++ b/phase2/embodied-olympics-fun/golf_putt/trajectories/embodied-olympics-fun-golf-putt-i00-privileged--hayytdm.json @@ -0,0 +1,205 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-golf-putt-i00-privileged--hayytdm", + "taskKey": "embodied-olympics-fun/golf_putt", + "dag_sha256": "74a1c00f893c3316873e4cc2017a91bda4a2926d26f9ae1a009d7dbb3a3e0854", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "replay_success": 0, + "deterministic": 1, + "n_actions": 1200, + "sim_steps": 1200, + "horizon": 3000, + "episode_over": 0, + "progress": 0.0, + "stages_reached": 0, + "stages_total": 2, + "fell": 0, + "replay_s": 9.59, + "bad_trajectory": 0, + "metric_ball_speed_m_s": 0.0, + "metric_ball_to_hole_m": 3.0, + "metric_grip_break_max_s": 0.002, + "metric_head_contact_s": 0.0, + "metric_head_hits": 0.0, + "grader_error": 0, + "video_rendered": 1, + "agent_sessions": 47, + "agent_resets": 28, + "agent_replays": 2, + "agent_cost_self_reported": 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/embodied-olympics-fun-golf-putt-i00-privileged--hayytdm/replay.mp4", + "duration_s": 12.04, + "fps": 25.0, + "frames": 301, + "sha256": "6269c1e4d44e4805c49aef1ab28a09488e43901187ec914dd87667020648332e" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "ball_holed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 2.8, + "log_step": null, + "text_zh": "观察：视频 0.0–2.2 s 机器人右手持杆站立，视频 2.4–3.2 s 右臂带动推杆前挥，杆头从白球旁边经过，白球没有离开脚前原位；未出现向远处球洞滚动或落洞。" + }, + { + "t": 12.0, + "log_step": 92, + "text_zh": "观察：视频 4.0–12.0 s 机器人持杆停留，逐帧检查视频 11.0–12.0 s，白球仍在脚前，洞口在远处浅色条带上。log step 90 保存最终 1200 行轨迹，log step 92 对该文件的重放返回 stages=[]、head_hits=0、head_contact_s=0.0、ball_to_hole_m=3.0；评分器 reward 同样报告这些数值及 success=0、deterministic=1，与视频一致。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：视频 0.0–12.0 s 机器人双脚接触果岭支撑身体，右手持推杆完成前挥。杆头未与白球建立接触（视频 2.4–3.2 s，评分器 head_hits=0、head_contact_s=0.0），因此不记录工具对白球的击打接触；握工具不另重复计入工具链。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "ball_to_hole_m", + "unit": "m", + "threshold": 0.054, + "direction": "lower_better", + "value": 3.0 + }, + { + "name": "ball_speed_m_s", + "unit": "m/s", + "threshold": null, + "direction": "lower_better", + "value": 0.0 + }, + { + "name": "head_contact_s", + "unit": "s", + "threshold": 0.03, + "direction": "lower_better", + "value": 0.0 + }, + { + "name": "grip_break_max_s", + "unit": "s", + "threshold": 0.05, + "direction": "lower_better", + "value": 0.002 + } + ], + "failure": { + "breakpoint": { + "ref": "ball_holed", + "t": 2.8, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察：视频 2.4–3.2 s 机器人正在挥动右手推杆，杆头与脚前白球保持分离，挥过后白球未动；视频 3.2–12.0 s 没有补成触球或入洞。最终评分器 head_hits=0、stages_reached=0，断在一杆推球入洞的击球尝试。" + }, + "phenomena": [ + { + "cat": "接触", + "item": "没建立接触", + "t": 2.8, + "log_step": 92, + "evidence_zh": "观察：视频 2.4–3.2 s 机器人持杆前挥，但杆头没有碰到静止白球；视频 11.0–12.0 s 白球仍留在起点。log step 92 最终交付文件重放及评分器 reward 均记录 head_hits=0、head_contact_s=0.0、ball_to_hole_m=3.0，所需杆头—球接触从未建立。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "视频 / 日志 / agent 自述 / 评分器字段", + "evidence_zh": "推断：机器人知道要使杆头穿过白球，失败在将挥杆指令落实为有效杆头路径。log step 80 的生成脚本以右手抓握位姿的逆解生成前挥，log step 84、90 调整并保存最终参数；视频 2.4–3.2 s 实际杆头仍错过白球，log step 92 对最终文件重放确认零触球。log step 93 的另一次诊断显示逆解残差及手、杆头运动差异，agent 在 log step 91、94 将其解释为逆解误差、PD 跟踪滞后及球杆滞后；这些诊断属于会话证据，支持控制失败，但不足以验证具体子类，故停在“控制”。" + } + ], + "gap_zh": "没有状态日志，远景视频不能精确量化握杆接触、关节跟踪或杆头与白球的三维间隙；因此无法独立区分逆解误差、反馈跟踪误差和握持中球杆滞后的贡献。log step 93 是保存并验证交付后的独立诊断，未将其仿真时间或数值当作本次视频的状态轨迹。", + "unmet": { + "ball_holed": false + } + }, + "summary_zh": "机器人右手持杆、双脚站立并作出前挥，但视频约 2.4–3.2 s 杆头未触及白球，视频 12.0 s 球仍在起点，“一杆推球入洞”最终为 0/1。最终文件重放与评分器均报告零触球、球距洞 3.0 米，断点在该步骤操作中，现象为没建立接触，根因保守归为控制。缺少状态日志，无法确定逆解、跟踪与球杆滞后各自的贡献。", + "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 ball_holed 操作中、没建立接触、根因保守归为控制，均与视频帧、评分器 head_hits=0 和 log step 92/93 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": "2026-10-29T10:06:29Z", + "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/embodied-olympics-fun-golf-putt-i00-privileged--hayytdm/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-golf-putt-i00-privileged--hayytdm/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "7fff0add04b231abfe629e842499fe33881cce541dda89ac47826fa679f8669c", + "video_sha256": "6269c1e4d44e4805c49aef1ab28a09488e43901187ec914dd87667020648332e", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/high_bar_swing/task_dag.json b/phase2/embodied-olympics-fun/high_bar_swing/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..ba1b0b1ee757dd6711bf8c78937324cd067f7cd8 --- /dev/null +++ b/phase2/embodied-olympics-fun/high_bar_swing/task_dag.json @@ -0,0 +1,330 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "embodied-olympics-fun/high_bar_swing", + "suite": "embodied-olympics-fun", + "base": "embodied-olympics-fun-high-bar-swing-i00", + "title_zh": "跃起抓杠并大回环", + "summary_zh": "机器人从起跳块跃起后双手抓杠，保持挂杠并在杠上伸展，完成一次有效的单向360°大回环且不触碰禁触物。", + "embodiment_zh": "两模式均为 Unitree G1 人形机器人，29个身体关节，双臂各装一只 Dex3-1 三指灵巧手（每手7个驱动），手部驱动力为原设备的5倍；双腿支撑和起跳，躯干、双腿及肩部驱动摆动，无轮式底盘。", + "task_class": "compositional", + "analysis_level": "step", + "spec": { + "version": "1.2", + "task": "embodied-olympics-fun-high-bar-swing-i00", + "bench": "embodied-olympics-fun", + "progress": "live", + "entities": {}, + "milestones": [ + { + "id": "bar_caught", + "zh": "跃起抓杠", + "mode": "event", + "when": null, + "after": [], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "外部视角中，机器人从棕色起跳块腾起，双脚离块后双手合拢抓住银色横杠，身体悬在杠下；验收要求曾从块上离开且当时双手均未抓杠，随后在不触块时两手各有食指或中指以及拇指或掌部同时接触横杠。仅伸手靠近、只有一手抓住，或仍接触起跳块时双手抓杠均未完成；不要求额外悬挂时长。", + "step": { + "cat": "本体", + "item": "跳跃" + }, + "contact": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "左灵巧手和右灵巧手各用手指与拇指或掌部环握固定横杠，同时建立双手抓杠。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:141-166 双手抓杠与挂杠的不同接触定义", + "inputs/task/privileged/tests/rfo_task.py:287-301 初态站块、双臂上举且未抓杠", + "inputs/task/privileged/tests/rfo_task.py:304-325 离块起跳后双手抓杠，原生阶段 jumped、caught the bar", + "inputs/task/privileged/instruction.md:4-11 任务句和起始场地", + "inputs/oracle/replay.mp4:0.00-0.90 起跳抓杠可见" + ], + "conf": 6, + "note_zh": "文本明确要求跃起后双手抓杠，评分器独立记录抓杠状态，因此它是本体目标而非操作物体前的抓取阶段。起跳并入本步；脚蹬起跳块的支撑先发生、腾空后才双手抓杠，不列为同时接触。抓杠后可悬挂蓄摆，与大回环之间可暂停；event 保留已抓杠的历史事实，后续脱手由约束处理。" + }, + { + "id": "giant_circle_completed", + "zh": "完成大回环", + "mode": "event", + "when": null, + "after": [ + "bar_caught" + ], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "外部连续画面中，双手保持挂在横杠上，身体从杠下摆到杠上方并继续沿同一方向绕回，完成以横杠至质心连线计的有效单向 360°，向前或向后均可；重心高于 2.80 米横杠时，到双腕中点的距离须至少 0.59 米，双手各自离杠连续间隔不得超过 0.05 秒，且不得发生禁触。只来回蓄摆、仅越过杠顶或尚未转足一圈均未完成；反向回退超过 5°从转折点重新累计，杠上伸展不足从当前点重计，达到阈值即完成而不要求下杠或落地。", + "step": { + "cat": "本体", + "item": "悬垂摆动" + }, + "contact": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "左灵巧手与右灵巧手同时环握横杠，双臂承载机器人，使身体绕固定横杠回环。", + "plan": 1 + }, + { + "who": "双臂协作", + "how": "勾", + "note_zh": "左灵巧手与右灵巧手在初次合格抓杠后以弯曲手指勾挂横杠回环；两手均持续接触即可，不要求拇指或掌部持续夹住横杠。", + "plan": 2 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:168-184 双手接触间隔和非手禁触", + "inputs/task/privileged/tests/rfo_task.py:231-235 阶段顺序及 360 度、0.59 米、5 度、0.05 秒参数", + "inputs/task/privileged/tests/rfo_task.py:335-370 有效角度、伸展重计、过顶阶段、整圈成功", + "inputs/task/privileged/tests/rfo_task.py:63-75 摆角与双腕半径定义", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178 阶段顺序和成功条件", + "inputs/oracle/replay.mp4:0.90-16.24 蓄摆后绕杠回环" + ], + "conf": 6, + "note_zh": "一整圈为一个累计角度目标，不拆上顶、四个象限或蓄摆次数。原生 over the top 阶段并入本步验收，且 full circle 只能在它之后登记；两步之间唯一边来自物理必需和原生阶段顺序。完成后裁判立即结束，采用 event。钩挂为规则明确接受的替代接触方案；腿和躯干在空中运动不等于与支撑物接触。" + } + ], + "constraints": [ + { + "id": "ground_touched", + "zh": "触地或垫", + "never": null, + "criterion_zh": "抓杠前后，机器人任一部位接触蓝色落地垫或地面均失败；抓杠前允许落回棕色起跳块后重新起跳。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:54 落地垫属于 ground", + "inputs/task/privileged/tests/rfo_task.py:310-311 抓杠前触地", + "inputs/task/privileged/tests/rfo_task.py:332-333 抓杠后触地" + ] + }, + { + "id": "bar_touched_by_nonhand", + "zh": "非手触杠", + "never": null, + "criterion_zh": "从起始至完成，机器人除裁判所定义的双手以外的部位接触横杠即失败；不能用前臂、腿或躯干搭杠代替挂杠。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:155-158 非手几何体选择", + "inputs/task/privileged/tests/rfo_task.py:182-184 抓杠后检查", + "inputs/task/privileged/tests/rfo_task.py:307-309 抓杠前检查" + ] + }, + { + "id": "bar_caught_without_jump", + "zh": "未跳先抓杠", + "never": null, + "criterion_zh": "双手均形成抓杠接触时，若此前未满足离块且双手均未抓杠的起跳事件，或此时机器人仍触碰起跳块，即失败。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:312-320" + ] + }, + { + "id": "bar_released", + "zh": "抓杠后脱手", + "never": null, + "criterion_zh": "合格抓杠后，任一手与横杠所有部位持续无接触超过 0.05 秒即失败，不能松手后重新抓回；不要求持续有拇指或掌部接触。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:168-181", + "inputs/task/privileged/tests/rfo_task.py:234-235 off_bar_s=0.05" + ] + }, + { + "id": "uprights_touched_after_catch", + "zh": "抓后触支架", + "never": null, + "criterion_zh": "合格抓杠后，机器人任一部位接触单杠两端立柱或其底脚即失败；该检查在抓杠前分支未执行。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:295-298 posts 为 post 前缀全部几何体", + "inputs/task/privileged/tests/rfo_task.py:328-329" + ] + }, + { + "id": "block_touched_after_catch", + "zh": "抓后触起跳块", + "never": null, + "criterion_zh": "合格抓杠后，机器人任一部位接触棕色起跳块即失败；画面投影重叠不能单独判定实际接触。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:330-331" + ] + }, + { + "id": "simulation_unstable", + "zh": "物理仿真发散", + "never": null, + "criterion_zh": "物理积分出现加速度发散警告、非有限位姿或任一受检线性自由度的速度分量绝对值超过150米/秒时，尝试以仿真不稳定立即结束；需仿真记录确认，视频不能精确判定。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:108-109 不稳定原因及150米每秒上限", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:872-885" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "bar_caught", + "level": "new_code", + "source": "Log w.ref.stages (jumped, caught the bar) and Task.trackers.caught; inputs/task/privileged/tests/rfo_task.py:312-324" + }, + { + "atom": "giant_circle_completed", + "level": "new_code", + "source": "Log w.ref.stages (over the top, full circle), w.ref.success, and Task.trackers.anchor/ext/sense; inputs/task/privileged/tests/rfo_task.py:335-370" + }, + { + "atom": "ground_touched", + "level": "new_code", + "source": "Log ref.touching(robot, ref.ground); inputs/task/privileged/tests/rfo_task.py:310-311,332-333" + }, + { + "atom": "bar_touched_by_nonhand", + "level": "new_code", + "source": "Log ref.touching(not_on_bar, bar_g); inputs/task/privileged/tests/rfo_task.py:182-184,307-309" + }, + { + "atom": "bar_caught_without_jump", + "level": "new_code", + "source": "Log Task.trackers.flown, on_block and holding(ref); inputs/task/privileged/tests/rfo_task.py:312-320" + }, + { + "atom": "bar_released", + "level": "new_code", + "source": "Log each hand on_bar contact and on_t gap at 1 ms resolution; inputs/task/privileged/tests/rfo_task.py:168-181" + }, + { + "atom": "uprights_touched_after_catch", + "level": "new_code", + "source": "Log Task.trackers.caught and ref.touching(robot, posts); inputs/task/privileged/tests/rfo_task.py:328-329" + }, + { + "atom": "block_touched_after_catch", + "level": "new_code", + "source": "Log Task.trackers.caught and ref.touching(robot, block); inputs/task/privileged/tests/rfo_task.py:330-331" + }, + { + "atom": "simulation_unstable", + "level": "new_code", + "source": "Log World.unstable and relevant integration diagnostics; inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:872-883" + } + ] + }, + "native_metrics": [ + { + "name": "travel_deg", + "unit": "deg", + "threshold": 360.0, + "direction": "higher_better" + }, + { + "name": "grip_break_max_s", + "unit": "s", + "threshold": 0.05, + "direction": "lower_better" + }, + { + "name": "wrist_radius_m", + "unit": "m", + "threshold": 0.59, + "direction": "higher_better" + }, + { + "name": "max_forward_deg", + "unit": "deg", + "threshold": null, + "direction": "higher_better" + }, + { + "name": "max_backward_deg", + "unit": "deg", + "threshold": null, + "direction": "higher_better" + }, + { + "name": "time_s", + "unit": "s", + "threshold": null, + "direction": "lower_better" + } + ], + "open_questions_zh": [ + "inputs/states/index.json 的 logs 为空；按 DECISIONS.md P8，所有节点 when、约束 never 和 success 均为 null。atoms 仅列今后需要记录的世界状态，不假定已有这些字段；无法自动核对成功回放的步骤覆盖和 success 与评分一致性。", + "按 S10 将跃起至双手抓杠合为一步，抓稳后可悬挂再蓄摆，因此抓杠与大回环分开；起跳和过顶两个原生阶段分别并入对应步骤。两步均是裁判成功所必需的独立文本步骤，以 event 记录；抓杠后的保持接触由脱手约束验收，不再建保持节点。", + "两模式使用相同任务模块和物理裁判；privileged 另要求有效轨迹在两次独立重放中成功且终态哈希相同（inputs/grader/efe_embodied-olympics-fun-files/grade.py:87-115），standard 判最后一次在线尝试（inputs/grader/efe_embodied-olympics-fun-files/standard/server.py:157-178）。文件合法性、重放确定性和基础设施有效性不是世界状态，未建语义节点，现有日志也不能表达完整交付判分。", + "文本称裁判每1毫秒判断事件，但代码在每1毫秒检查接触，而角度、伸展和回环成功在每10毫秒控制步检查（inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:867-885；referee.py:266-289）。视频为25帧每秒，无法直接分辨0.05秒接触容差或精确量出质心至双腕0.59米；判据保留代码阈值，不能用视频估计值替换。", + "travel_deg 是有效行程历史最大值，四舍五入到0.1度，并非当前累计量；实际成功比较未舍入 travel >= 2*pi 且顺序登记 full circle（rfo_task.py:359-370）。未来不能仅用 travel_deg >= 360 代替精确成功条件。wrist_radius_m 的0.59米阈值只在质心高于2.80米横杠时适用；反转或伸展不足仅重计，不是历史否决或立即失败。", + "成功视频为固定斜侧外部视角，上顶时部分身体被画面上沿裁切，精确接触和伸展存在可见性限制；该成功的评分记录为四阶段齐全、有效角度360.2度、最大脱杠间隔0.002秒，不能把这些单次结果变成任务阈值（inputs/oracle/run.json reward）。" + ], + "conf": 6, + "conf_reason_zh": "两步直接来自文本，阶段顺序、起始条件和所有阈值由同一裁判及两模式任务模块确认，成功视频支持抓杠后蓄摆再回环；没有任何状态日志，精确接触、伸展和有效角度无法自动核验，故降低置信度。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "文本步骤为跃起抓杠、完成大回环，DAG 两节点与之一致；起跳与抓杠是一次连续腾空动作故合并，过顶并入整圈。无状态日志，按 P8 所有 when/never/success 为 null 合规；约束覆盖评分器全部失败条件，类型与接触标注符合 taxonomy，开放问题已列出文本与评分差异。", + "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：['bar_caught', 'giant_circle_completed']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:48:33Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__high_bar_swing/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__high_bar_swing/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "b28324da8350ff2bc96f652f3174f0ad4286d7646dbfded0c6fcda7dd9fce6c0", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/embodied-olympics-fun/high_bar_swing/trajectories/embodied-olympics-fun-high-bar-swing-i00-privileged--qptjc3u.json b/phase2/embodied-olympics-fun/high_bar_swing/trajectories/embodied-olympics-fun-high-bar-swing-i00-privileged--qptjc3u.json new file mode 100644 index 0000000000000000000000000000000000000000..695e5d1de61357b09b3c062cbbda9405152c524a --- /dev/null +++ b/phase2/embodied-olympics-fun/high_bar_swing/trajectories/embodied-olympics-fun-high-bar-swing-i00-privileged--qptjc3u.json @@ -0,0 +1,224 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-high-bar-swing-i00-privileged--qptjc3u", + "taskKey": "embodied-olympics-fun/high_bar_swing", + "dag_sha256": "94fc49d886b4eb0d110e6e93e7d10de82eecae87b2952d2f2744f01b83806fc4", + "outcome": { + "success": 1, + "reward": 1.0, + "source": "catalog", + "raw": { + "success": 1, + "replay_success": 1, + "deterministic": 1, + "n_actions": 653, + "sim_steps": 653, + "horizon": 6000, + "episode_over": 1, + "progress": 1.0, + "stages_reached": 4, + "stages_total": 4, + "fell": 0, + "replay_s": 12.59, + "bad_trajectory": 0, + "metric_grip_break_max_s": 0.001, + "metric_max_backward_deg": 464.0, + "metric_max_forward_deg": 143.6, + "metric_time_s": 6.53, + "metric_travel_deg": 360.2, + "grader_error": 0, + "video_rendered": 1, + "agent_sessions": 929, + "agent_resets": 0, + "agent_replays": 13, + "agent_cost_self_reported": 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/embodied-olympics-fun-high-bar-swing-i00-privileged--qptjc3u/replay.mp4", + "duration_s": 6.56, + "fps": 25.0, + "frames": 164, + "sha256": "c6338aaca9761f059fcad6e996f0dbe80ce176ada6e4fa604463b5e79a857e35" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 2, + "final": 2, + "nodes": { + "bar_caught": { + "method": "visual", + "intervals": [ + [ + 0.68, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.68, + "log_step": null, + "text_zh": "观察：视频 0.28–0.48 s 机器人双脚在棕色起跳块上蹬伸，视频 0.52–0.64 s 双脚与块顶出现间隙、双手上举靠近横杠；视频 0.68–0.88 s 左右手合在银色横杠处，双脚悬空，屈膝身体由双臂悬挂并开始摆动。将双手抓杠首次可辨时刻记为视频约 0.68 s，存在数帧的不确定性。" + }, + { + "t": 0.88, + "log_step": 164, + "text_zh": "观察：log step 162 的候选结果列出 jumped、caught the bar、over the top、full circle 四阶段；log step 164 确认该 653 行候选在两次独立重放中均成功、终态哈希一致并已被选为交付，log step 169 确认最终生成器逐行复现交付。评分器 reward 同样报告 n_actions=653、stages_reached=4、replay_success=1、deterministic=1，与视频的腾空抓杠一致；其他会话试验不计入节点。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "观察：视频 0.68–0.88 s 左右灵巧手分别合拢在横杠两处，共同承载悬空身体；此前视频 0.28–0.48 s 双脚在起跳块上支撑蹬伸，脚的支撑先于空中双手抓握发生。手指的精确接触对在远景中不可分辨，合格双手抓杠另由评分器四阶段成功结果支持。" + } + ] + }, + "giant_circle_completed": { + "method": "visual", + "intervals": [ + [ + 6.52, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 5.4, + "log_step": null, + "text_zh": "观察：视频 0.88–4.80 s 机器人双手留在横杠处，反复摆腿、屈髋蓄摆，并有屈身越杠的准备动作；视频 5.20–5.56 s 身体在杠上倒立并逐渐伸展，随后视频 5.60–6.04 s 沿同一方向摆下，经横杠下方，再于视频 6.08–6.52 s 摆升回杠上另一侧。已逐帧检查视频 5.00–6.52 s，末帧仍双手挂杠，未见脱离后坠落。" + }, + { + "t": 6.52, + "log_step": 162, + "text_zh": "推断：视频末段连续回转与评分器 reward 的 success=1、replay_success=1、stages_reached=4、metric_travel_deg=360.2、metric_grip_break_max_s=0.001 一致，因此完整有效回环已达成，完成锚点取实际检查的末帧视频 6.52 s（约数）。log step 162 输出的同一 653 行候选包含 full circle 阶段，log step 164、169 确认其被交付且逐行复现；这里没有将 agent 工作时钟换算成视频时钟。视频远景不能独立测出 0.59 米伸展阈值或 0.05 秒离杠容差，合格性由成功裁定支持。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "观察：视频 0.88–6.52 s 左右手同时环合在横杠上，随身体绕杠回转，两臂共同承载；双腿和躯干在空中摆动。远景不能分清每根手指或拇指的持续接触，未据此另记仅用弯指勾挂的替代方案。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "travel_deg", + "unit": "deg", + "threshold": 360.0, + "direction": "higher_better", + "value": 360.2 + }, + { + "name": "grip_break_max_s", + "unit": "s", + "threshold": 0.05, + "direction": "lower_better", + "value": 0.001 + }, + { + "name": "wrist_radius_m", + "unit": "m", + "threshold": 0.59, + "direction": "higher_better", + "value": null + }, + { + "name": "max_forward_deg", + "unit": "deg", + "threshold": null, + "direction": "higher_better", + "value": 143.6 + }, + { + "name": "max_backward_deg", + "unit": "deg", + "threshold": null, + "direction": "higher_better", + "value": 464.0 + }, + { + "name": "time_s", + "unit": "s", + "threshold": null, + "direction": "lower_better", + "value": 6.53 + } + ], + "failure": null, + "summary_zh": "机器人在视频约 0.68 s 腾空后双手抓杠，随后蓄摆、屈伸身体，至视频末帧 6.52 s 完成有效大回环，两个节点均完成。评分器确认交付成功、两次重放确定一致，有效行程为 360.2°、最大离杠间隔为 0.001 秒；会话中其他候选的失败不属于本次评分轨迹。没有状态日志，远景及过顶时的画面上沿裁切限制了逐指接触、质心伸展距离和精确完成时刻的独立核验，相关阈值依据评分结果。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "成功运行，两节点均达成；视频帧（0.4、0.68、0.9、5.4、6.0、6.5 s）与标注及评分器 stages_reached=4、metric_time_s=6.53 一致，校验通过，未发现需修改的问题。", + "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": "2026-10-29T10:05:50Z", + "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/embodied-olympics-fun-high-bar-swing-i00-privileged--qptjc3u/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-high-bar-swing-i00-privileged--qptjc3u/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "0514e84dcc5b5fad413a6737b1bb9534b7c662ae01d3b8a3a339f5f6d219788f", + "video_sha256": "c6338aaca9761f059fcad6e996f0dbe80ce176ada6e4fa604463b5e79a857e35", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/high_bar_swing/trajectories/embodied-olympics-fun-high-bar-swing-i00-standard--jt2x2pc.json b/phase2/embodied-olympics-fun/high_bar_swing/trajectories/embodied-olympics-fun-high-bar-swing-i00-standard--jt2x2pc.json new file mode 100644 index 0000000000000000000000000000000000000000..2768f9518d9699d0e3a272a8aa37c4a18a5a8357 --- /dev/null +++ b/phase2/embodied-olympics-fun/high_bar_swing/trajectories/embodied-olympics-fun-high-bar-swing-i00-standard--jt2x2pc.json @@ -0,0 +1,257 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-high-bar-swing-i00-standard--jt2x2pc", + "taskKey": "embodied-olympics-fun/high_bar_swing", + "dag_sha256": "94fc49d886b4eb0d110e6e93e7d10de82eecae87b2952d2f2744f01b83806fc4", + "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": 6000, + "sim_steps": 12374, + "calls": 34, + "rows": 12374, + "resets": 3, + "reset_budget": 5, + "sim_startups": 1, + "elapsed_s": 6997.8, + "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": 0.0, + "peak_memory_gb": 8.1, + "peak_processes": 18, + "peak_disk_gb": 4.356, + "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/embodied-olympics-fun-high-bar-swing-i00-standard--jt2x2pc/replay.mp4", + "duration_s": 60.04, + "fps": 25.0, + "frames": 1501, + "sha256": "844662c3df4a4e2d567c9f0a8f5c8f241402f863eb7dfd29c766b07cf2ca3c38" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "bar_caught": { + "method": "visual", + "intervals": [ + [ + 0.88, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.88, + "log_step": null, + "text_zh": "观察：视频 0.68–0.76 s 机器人由屈膝转为蹬伸；视频 0.80–0.88 s 双脚离开棕色起跳块，两手到达横杠并闭合，随后身体悬在杠下。视频 0.92–1.32 s 双手位置固定在杠处，双脚与块顶有可见间隙；首次抓杠时间约为视频 0.88 s。" + }, + { + "t": 60.0, + "log_step": 173, + "text_zh": "推断：log step 173 最后一次在线尝试（episode 4）的服务裁判返回 stages_reached=2、stages_total=4，与视频中的离块及双手悬挂一致，支持按原生阶段顺序已完成 jumped 和 caught the bar。该节点是事件，抓杠事实保留至结束；未使用前三次尝试或本地副本的抓杠时间。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "观察：视频约 0.88 s 后，机器人左右灵巧手同时合拢在银色横杠上承重，双脚已离起跳块；细小手指的具体接触对由裁判已到抓杠阶段辅助确认。" + } + ] + }, + "giant_circle_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 2.4, + "log_step": 159, + "text_zh": "观察：视频 1.6–8.8 s 机器人双手留在横杠上，身体在杠下小幅往返摆动，摆幅逐渐减小；全片粗采样及视频 59.0–60.0 s 密集末段均未见身体绕至杠上方。log step 159 所示在线脚本在开场动作后，用 np.tile(pose, (n, 1)) 重复最后的关节目标并保持双手闭合。" + }, + { + "t": 60.0, + "log_step": 173, + "text_zh": "观察：视频 59.0、59.2、59.4、59.6、59.8 和 60.0 s 机器人仍近乎竖直悬在杠下，没有完成单向整圈。log step 173 最后一次尝试的服务裁判返回 outcome=time、stages_reached=2/4、reason=failed；这与视频未上杠、未完成回环一致。" + }, + { + "t": 60.0, + "log_step": 173, + "text_zh": "推断：最后一次尝试的控制步长为 0.01 s（log step 132 的 submit 调用），log step 173 的 episode 4 以 control_steps=6000、sim_time_s=60.0 结束；6000×0.01=视频 60.0 s，与已检查的末帧悬挂终态相符。这里没有把该日志内 agent 会话 time_s=6997.95 换算成视频时间。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "观察：视频 1.6–8.8 s 的往返摆动及视频 59.0–60.0 s 的近静止悬挂中，左右手都闭合在横杠处承载身体；两脚在空中，不把投影靠近起跳块记为脚部支撑。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "travel_deg", + "unit": "deg", + "threshold": 360.0, + "direction": "higher_better", + "value": null + }, + { + "name": "grip_break_max_s", + "unit": "s", + "threshold": 0.05, + "direction": "lower_better", + "value": null + }, + { + "name": "wrist_radius_m", + "unit": "m", + "threshold": 0.59, + "direction": "higher_better", + "value": null + }, + { + "name": "max_forward_deg", + "unit": "deg", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "max_backward_deg", + "unit": "deg", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "time_s", + "unit": "s", + "threshold": null, + "direction": "lower_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "giant_circle_completed", + "t": 2.4, + "log_step": 173, + "where": "操作中", + "evidence_zh": "观察：机器人已在视频约 0.88 s 跃起抓杠；视频 2.4 s 起的挂杠阶段只见小幅往返，随后摆动衰减，视频 59.0–60.0 s 仍几乎竖直悬挂，没有上杠或整圈。log step 173 的最终 episode 4 返回原生阶段 2/4、outcome=time。断点定位在完成大回环这一步的悬挂蓄摆阶段，视频 2.4 s 仅作导航锚点。" + }, + "phenomena": [ + { + "cat": "运动", + "item": "其他", + "t": 2.4, + "log_step": 173, + "note_zh": "候选：悬垂摆动衰减。现有步态用于行走、跑跳，位姿不准不能准确描述这里往返摆动逐渐消失的过程；没有看到脱手、跌倒或阻挡接触。", + "evidence_zh": "观察：视频 1.6、2.0、2.4、2.8、3.2 s 可见机器人在杠下交替偏向两侧，视频 4.0–8.8 s 摆幅缩小；后续全片粗采样及视频 59.0–60.0 s 密集帧中，身体都停留在杠下并趋于静止。可见问题是悬挂运动衰减、未持续增加摆幅。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "t": 2.4, + "log_step": 159, + "source_zh": "日志", + "evidence_zh": "推断：log step 132 创建、log step 159 再次显示的最终在线脚本，在开场动作后只把 q[-1] 作为固定 pose 重复发送，并闭合双手；摆角估计仅用于打印，没有用于改变关节目标。log step 173 最终执行仍是该脚本，仅修改起跳关键帧与合手时刻，随后连续运行 hang 循环至 episode 4 结束。这种控制输出未提供持续蓄摆的屈伸或肩部驱动，与视频 2.4–8.8 s 摆幅衰减及视频 59.0–60.0 s 近静止一致；证据足以定位在线运动控制缺少有效摆动驱动，不能再细分为某个已证实错误的增益、逆解或接触参数。" + } + ], + "gap_zh": "无状态日志，640×360、25 fps 的单一外部视角不能逐指核实 0.05 s 脱杠容差或精确测量质心至双腕的半径；抓杠节点结合本次服务裁判的 2/4 阶段确认。评分 reward 未报告 DAG 所列原生运动指标，故 native_metrics 均留 null；日志中的会话 time_s、elapsed_s 不充当原生运动耗时。最终控制代码可见，但没有证据进一步确定缺少有效摆动驱动的控制子型。", + "unmet": { + "bar_caught": true, + "giant_circle_completed": false + } + }, + "summary_zh": "最后一次在线尝试在视频约 0.88 s 跃起并双手抓杠，完成 1/2 个语义节点；之后机器人只在杠下小幅往返，最终近乎静止，至视频 60.0 s 未完成大回环。最终脚本反复发送固定挂杠姿态，未产生持续蓄摆驱动，服务裁判在 log step 173 返回原生阶段 2/4 并以 time 结束。无状态日志及原生运动指标，精细手指接触与有效角度无法量化。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "抓杠节点（视频约 0.88 s）、未完成大回环、断点 giant_circle_completed、服务读数 stages_reached=2/4 与控制步到视频时间的换算均与帧和日志一致；现象“运动·其他”与根因“控制”证据足够且未过度断言，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": "2026-10-29T10:08:36Z", + "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/embodied-olympics-fun-high-bar-swing-i00-standard--jt2x2pc/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-high-bar-swing-i00-standard--jt2x2pc/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "a581efb39846caa7d2a9dca573015c85deb27bb2a633c334856d25c8b72098eb", + "video_sha256": "844662c3df4a4e2d567c9f0a8f5c8f241402f863eb7dfd29c766b07cf2ca3c38", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/high_jump/task_dag.json b/phase2/embodied-olympics-fun/high_jump/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..f4b93283d27da0e5e7125d68517833e6733ebd12 --- /dev/null +++ b/phase2/embodied-olympics-fun/high_jump/task_dag.json @@ -0,0 +1,279 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "embodied-olympics-fun/high_jump", + "suite": "embodied-olympics-fun", + "base": "embodied-olympics-fun-high-jump-i00", + "title_zh": "跳高过杆落地", + "summary_zh": "机器人合法单脚起跳，在同一次腾空中全身越过两立柱之间的横杆后落地，且落地至少2秒后横杆仍静置于两支座上，期间无规则违例。", + "embodiment_zh": "两种模式均为Unitree H1-2人形机器人，27个关节、约1.78米高、67千克，双腿步行，无轮式底盘；左右臂各7个关节，末端均为固定拳垫，没有可工作的手指，本任务无可抓取物，夹持命令无效。跳高主要由腿和躯干完成，双臂可摆动配合。", + "task_class": "atomic", + "analysis_level": "step", + "spec": { + "version": "1.2", + "task": "embodied-olympics-fun-high-jump-i00", + "bench": "embodied-olympics-fun", + "progress": "live", + "entities": {}, + "milestones": [ + { + "id": "high_jump_landed", + "zh": "跳过杆并落地", + "mode": "event", + "when": null, + "after": [], + "hold_s": 0, + "goal": true, + "criterion_zh": "固定场地镜头中，机器人从红色跑道单脚起跳，骨盆和双踝在同一次腾空中从两根灰色立柱之间、高于红白横杆上沿的位置通过，随后全身越过横杆近侧平面并落到另一侧（成功录像落在蓝垫上）；通过平面x=-0.015米时骨盆和双踝均须满足|y|<2.01米、z>1.0米，所有机器人碰撞几何的最小x也须大于该平面，腾空至少0.05秒且骨盆比离地时上升至少0.08米，回穿的骨盆或踝须重新通过，未净空便触地或垫、仅上身过杆均未完成。合法单脚起跳须满足离地前0.15秒非起跳脚地面力小于体重5%或它在起跳脚最后完整接地期间一直离地，其他身体部分在相应期间无外部支撑（足部接触间断至多0.03秒仍合并）；首次落地不得有地面接触点位于起跳侧或平面上，且须通过所列横杆和接触约束，净空且落地至少2.0秒后，红白横杆仍同时接触两支座，中心沿跑道偏移绝对值≤0.02米、高度偏差绝对值≤0.005米、线速度≤0.05米/秒、角速度≤0.5弧度/秒，才完成此步。", + "step": { + "cat": "本体", + "item": "跳跃" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "左脚或右脚在红色跑道上承重蹬地，净空后脚、腿或躯干与蓝色落地垫或杆后地面接触承重；这些支撑依次发生，腾空期间无需外物接触。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:3-38", + "inputs/task/standard/instruction.md:3-38", + "inputs/task/privileged/tests/rfo_task.py:110-120", + "inputs/task/privileged/tests/rfo_task.py:200-217", + "inputs/task/privileged/tests/rfo_task.py:245-363", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:353-394", + "inputs/oracle/replay.mp4:15.5-18.72" + ], + "conf": 6, + "note_zh": "按S10将助跑后的起跳、净空、落地与延迟验收合为一次连续跳跃；三个原生阶段只是验收顺序，不分节点。横杆开局在支架上，失去支承会直接失败，故按约束处理；摔倒本身不失败，不要求站稳。hold_s=0：2秒是落地后等待至验收的延迟，不是要求所有几何与静止条件连续成立2秒。典型接触标起跳与落地支撑，不把过杆时允许的短暂擦碰标成必需接触；成功录像的背越式、左脚起跳和转身都不限制其他合法策略。" + } + ], + "constraints": [ + { + "id": "bar_off_supports", + "zh": "横杆离开支架", + "never": null, + "criterion_zh": "横杆中心相对首个物理步下降超过0.05米即失败；净空并落地至少2秒后的验收时，横杆若未同时接触两支座，或中心沿跑道偏移绝对值大于0.02米、高度偏差绝对值大于0.005米、线速度大于0.05米/秒、角速度大于0.5弧度/秒，也失败。画面中红白横杆掉落或离开灰色支架可直接确认，细小偏差和速度须裁判数据。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:249-252", + "inputs/task/privileged/tests/rfo_task.py:341-363" + ] + }, + { + "id": "bar_contact_excess", + "zh": "横杆支承超时", + "never": null, + "criterion_zh": "净空并首次落地之前，机器人与横杆累计接触超过0.3秒即失败；不能借横杆长时间托住身体。落地后代码不再累计此时长。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:258-263" + ] + }, + { + "id": "bar_impulse_excess", + "zh": "横杆冲量超限", + "never": null, + "criterion_zh": "整次尝试中横杆对机器人的累计法向冲量超过2.0牛顿秒即失败，包括净空落地后的等待期间；视频中是否借横杆明显撑起身体可作线索，准确冲量需物理步接触力。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:219-230", + "inputs/task/privileged/tests/rfo_task.py:253-259" + ] + }, + { + "id": "invalid_takeoff", + "zh": "非单脚起跳", + "never": null, + "criterion_zh": "跳跃离地前最后0.15秒，非起跳脚必须始终承受少于体重5%的地面法向力，或在起跳脚末次完整接地期间一直离地；其他身体部分也须在该窗口或该完整接地期间没有外部支撑，任一不满足即失败，双脚错开蹬地也可能违规。足部接触间断不超过0.03秒仍作同一接触。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:200-217", + "inputs/task/privileged/tests/rfo_task.py:271-279" + ] + }, + { + "id": "invalid_crossing", + "zh": "绕杆或钻杆", + "never": null, + "criterion_zh": "骨盆或任一踝关节由起跳侧穿过横杆近侧平面x=-0.015米时，若横向位置绝对值不小于2.01米，或高度不大于1.0米，即失败；画面中绕灰色立柱外侧或从红白横杆下方通过均不合法。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:24-40", + "inputs/task/privileged/tests/rfo_task.py:282-296" + ] + }, + { + "id": "premature_contact", + "zh": "未过杆先触地", + "never": null, + "criterion_zh": "尚未在同一次腾空中完成全身净空便触及杆后地面或蓝垫，或已认定为本次跳跃的飞行在净空前触及横杆以外的外物，即失败；先落垫再抬腿越杆不能补成成功。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:297-324" + ] + }, + { + "id": "invalid_flight", + "zh": "腾空不构成跳跃", + "never": null, + "criterion_zh": "过杆后的首次外物接触结束飞行时，若飞行短于0.05秒，或骨盆最高点相对离地时上升不足0.08米，即失败；单纯向杆后跌落不算完成跳高。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:114-120", + "inputs/task/privileged/tests/rfo_task.py:326-333", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:368-389" + ] + }, + { + "id": "takeoff_side_landing", + "zh": "落地触起跳侧", + "never": null, + "criterion_zh": "净空后首次落地时，只要机器人与地面或垫的任一接触点x不大于-0.015米，即失败；身体必须落在横杆近侧平面的另一侧，允许任意身体部位落地且不要求重新站立。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:309-310", + "inputs/task/privileged/tests/rfo_task.py:334-339" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "dyn.high_jump_accepted", + "level": "new_code", + "source": "需逐帧记录同一世界的Referee.success、stages和Task.trackers，源于inputs/task/privileged/tests/rfo_task.py:245-363及inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178；当前无状态日志，不绑定when。" + }, + { + "atom": "dyn.bar_off_supports", + "level": "new_code", + "source": "需逐帧记录对应规则的违例状态或Referee.failed与events；inputs/task/privileged/tests/rfo_task.py:249-252, inputs/task/privileged/tests/rfo_task.py:341-363" + }, + { + "atom": "dyn.bar_contact_excess", + "level": "new_code", + "source": "需逐帧记录对应规则的违例状态或Referee.failed与events；inputs/task/privileged/tests/rfo_task.py:258-263" + }, + { + "atom": "dyn.bar_impulse_excess", + "level": "new_code", + "source": "需逐帧记录对应规则的违例状态或Referee.failed与events；inputs/task/privileged/tests/rfo_task.py:219-230, inputs/task/privileged/tests/rfo_task.py:253-259" + }, + { + "atom": "dyn.invalid_takeoff", + "level": "new_code", + "source": "需逐帧记录对应规则的违例状态或Referee.failed与events；inputs/task/privileged/tests/rfo_task.py:200-217, inputs/task/privileged/tests/rfo_task.py:271-279" + }, + { + "atom": "dyn.invalid_crossing", + "level": "new_code", + "source": "需逐帧记录对应规则的违例状态或Referee.failed与events；inputs/task/privileged/tests/rfo_task.py:24-40, inputs/task/privileged/tests/rfo_task.py:282-296" + }, + { + "atom": "dyn.premature_contact", + "level": "new_code", + "source": "需逐帧记录对应规则的违例状态或Referee.failed与events；inputs/task/privileged/tests/rfo_task.py:297-324" + }, + { + "atom": "dyn.invalid_flight", + "level": "new_code", + "source": "需逐帧记录对应规则的违例状态或Referee.failed与events；inputs/task/privileged/tests/rfo_task.py:114-120, inputs/task/privileged/tests/rfo_task.py:326-333, inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:368-389" + }, + { + "atom": "dyn.takeoff_side_landing", + "level": "new_code", + "source": "需逐帧记录对应规则的违例状态或Referee.failed与events；inputs/task/privileged/tests/rfo_task.py:309-310, inputs/task/privileged/tests/rfo_task.py:334-339" + } + ] + }, + "native_metrics": [ + { + "name": "air_s", + "unit": "s", + "threshold": 0.05, + "direction": "higher_better" + }, + { + "name": "rise_m", + "unit": "m", + "threshold": 0.08, + "direction": "higher_better" + }, + { + "name": "bar_impulse_Ns", + "unit": "N s", + "threshold": 2.0, + "direction": "lower_better" + } + ], + "open_questions_zh": [ + "inputs/states/index.json中logs为空；按P8全部节点when、约束never和success为null，atoms仅说明需要补录的世界裁判状态，不使用未记录的评分结论。视频能判定过杆和落地，但无法精确测量起跳脚力、横杆冲量及毫米级位置和速度；这些须结合执行时裁判或未来状态日志核实。", + "任务设计问题：概述要求落在蓝色垫上（instruction.md:3及目录description），同一指令详细规则却明确允许垫外、任意身体部位落地；评分代码只记录on_the_mat而不要求它为真（rfo_task.py:326-339）。依N3采用详细规则，将落地位置作为同一次跳跃的验收，不另加可能排除合法成功的入垫步骤。", + "文本要求未完整受评：指令称机器人与横杆累计接触不得超过0.3秒，代码在净空且首次落地后提前返回，不再累计接触时长（rfo_task.py:258-263）；冲量及横杆下降仍检查。此处按实际评分器记录约束，并保留该差异。", + "模式交付差异：特权模式需有限且格式正确的轨迹在两次重放均成功、终态哈希一致（inputs/grader/efe_embodied-olympics-fun-files/grade.py:87-115）；标准模式直接评最后一次实时尝试，同一任务裁判没有附加运动规则（standard/server.py:157-178）。共同世界状态目标按一次合法跳高建模，文件格式、重放有效性和基础设施状态不建语义节点；无日志故success为null。" + ], + "conf": 6, + "conf_reason_zh": "一个连续跳高步骤直接来自文本，S10明确排除按裁判阶段拆分；两模式共用相同任务代码，初态与全部阈值有源码依据，成功录像显示过杆后落垫且横杆保留。但没有状态日志，精确力与接触、净空和最终横杆公差不能自动核验，文本内部的落地位置及接触时长也有差异。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "单一连续跳高步骤（起跳→过杆→落地→横杆静置）符合文本与S10；无状态日志故when/success为null合理；约束均为评分器失败条件且为立即终止；任务设计问题（落垫 vs 允许垫外）与0.3秒接触差异已列入open_questions_zh；步骤类型本体·跳跃与接触腿/躯干×支撑合理。", + "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：['high_jump_landed']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:43:38Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__high_jump/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__high_jump/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "b74ca34771ab66ea9d058990ccc47eb3a88ded5ad8c884848d0d58f8f5511238", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/embodied-olympics-fun/high_jump/trajectories/embodied-olympics-fun-high-jump-i00-privileged--d6jwkzp.json b/phase2/embodied-olympics-fun/high_jump/trajectories/embodied-olympics-fun-high-jump-i00-privileged--d6jwkzp.json new file mode 100644 index 0000000000000000000000000000000000000000..94f6ea0ac73bb1630e5fa1c3c8581a86b88276fa --- /dev/null +++ b/phase2/embodied-olympics-fun/high_jump/trajectories/embodied-olympics-fun-high-jump-i00-privileged--d6jwkzp.json @@ -0,0 +1,186 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-high-jump-i00-privileged--d6jwkzp", + "taskKey": "embodied-olympics-fun/high_jump", + "dag_sha256": "c300ffe03322943c62dc1bb110296f40a1e3ae3d2ab8a7501f467100e7abad26", + "outcome": { + "success": 1, + "reward": 1.0, + "source": "catalog", + "raw": { + "success": 1, + "replay_success": 1, + "deterministic": 1, + "n_actions": 1936, + "sim_steps": 1936, + "horizon": 6000, + "episode_over": 1, + "progress": 1.0, + "stages_reached": 3, + "stages_total": 3, + "fell": 0, + "replay_s": 11.62, + "bad_trajectory": 0, + "metric_air_s": 0.629, + "metric_rise_m": 0.367, + "grader_error": 0, + "video_rendered": 1, + "agent_sessions": 74, + "agent_resets": 0, + "agent_replays": 12, + "agent_cost_self_reported": 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/embodied-olympics-fun-high-jump-i00-privileged--d6jwkzp/replay.mp4", + "duration_s": 19.4, + "fps": 25.0, + "frames": 485, + "sha256": "70758d1027dbe0a10705535c5a88c5b156b78699916c2e5146902e0a3fdc2fae" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "high_jump_landed": { + "method": "visual", + "intervals": [ + [ + 19.36, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 15.6, + "log_step": null, + "text_zh": "观察：视频 0–15.4 s，机器人沿红色跑道走近横杆并停下；视频 15.6–16.6 s，机器人抬起一脚、降低骨盆，另一脚在跑道上支撑，随后伸腿起跳。" + }, + { + "t": 17.12, + "log_step": null, + "text_zh": "观察：逐帧检查视频 16.4–17.6 s；视频约 16.76 s 机器人离地，视频 16.84–17.12 s 身体上升、双腿屈曲越过两立柱之间的横杆，随后向杆后蓝垫下降。" + }, + { + "t": 17.36, + "log_step": null, + "text_zh": "观察：视频约 17.36 s，机器人骨盆及躯干落到蓝垫上，之后在垫上收拢身体；视频 18.36–19.36 s 横杆始终横置于两立柱支架上，最后一帧机器人仍在垫上。" + }, + { + "t": 19.36, + "log_step": null, + "text_zh": "推断：根据视频约 17.36 s 的落垫画面及 DAG 所列落地后等待 2.0 秒的验收规则，将本节点首次完成锚定在视频约 19.36 s（约有 0.1 秒不确定性）；inputs/run.json 的评分器字段 success=1、replay_success=1、deterministic=1、stages_reached=3/3 与过杆、落垫且横杆保留的画面一致，最终记为完成。脚力、精确净空和横杆公差由成功判定支持，不能从远景独立测得。" + }, + { + "t": null, + "log_step": 282, + "text_zh": "观察：log step 282 输出最终交付文件的两次新进程重放均成功，终态哈希相同；log step 283 的最终检查输出 rows=1936、success=true，与评分器 n_actions=1936 一致。本节点依据上述交付录像判断，不将会话早期搜索中的失败跳跃计入。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "机器人助跑时双脚交替接触红色跑道，杆前抬起一脚并由另一脚承重蹬地；过杆后骨盆及躯干落到蓝垫上承重，随后身体留在垫上。起跳前与落地后的支撑依次发生。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "air_s", + "unit": "s", + "threshold": 0.05, + "direction": "higher_better", + "value": 0.629 + }, + { + "name": "rise_m", + "unit": "m", + "threshold": 0.08, + "direction": "higher_better", + "value": 0.367 + }, + { + "name": "bar_impulse_Ns", + "unit": "N s", + "threshold": 2.0, + "direction": "lower_better", + "value": null + } + ], + "failure": null, + "summary_zh": "机器人沿跑道接近横杆，单脚支撑起跳，在同一次腾空中收腿过杆并于视频约 17.36 s 落到蓝垫上；计入落地后两秒验收，跳过杆并落地节点在视频约 19.36 s 完成，终态为 1/1。评分器判定成功且重放一致，片尾横杆仍在支架上；无状态日志，远景无法独立核验脚力、横杆冲量和毫米级验收公差，完成时间为结合画面与验收延迟的近似值，bar_impulse_Ns 因评分器未报告而留空。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "成功运行，单节点 high_jump_landed 在视频约 19.36 s 完成，final=true；画面显示越杆落垫、片尾横杆仍在支架上，与评分器字段一致，接触标注合理，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": "2026-10-29T10:06:41Z", + "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/embodied-olympics-fun-high-jump-i00-privileged--d6jwkzp/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-high-jump-i00-privileged--d6jwkzp/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "359e4d7240f14b723e682c9d3c139d901dcc76960e025c7f5059caed94565050", + "video_sha256": "70758d1027dbe0a10705535c5a88c5b156b78699916c2e5146902e0a3fdc2fae", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/high_jump/trajectories/embodied-olympics-fun-high-jump-i00-privileged--xtljosb.json b/phase2/embodied-olympics-fun/high_jump/trajectories/embodied-olympics-fun-high-jump-i00-privileged--xtljosb.json new file mode 100644 index 0000000000000000000000000000000000000000..8dac047267dbd9aeb22c191543b0ec0a9b4a8f41 --- /dev/null +++ b/phase2/embodied-olympics-fun/high_jump/trajectories/embodied-olympics-fun-high-jump-i00-privileged--xtljosb.json @@ -0,0 +1,184 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-high-jump-i00-privileged--xtljosb", + "taskKey": "embodied-olympics-fun/high_jump", + "dag_sha256": "c300ffe03322943c62dc1bb110296f40a1e3ae3d2ab8a7501f467100e7abad26", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "replay_success": 0, + "deterministic": 1, + "n_actions": 1, + "sim_steps": 1, + "horizon": 6000, + "episode_over": 0, + "progress": 0.0, + "stages_reached": 0, + "stages_total": 3, + "fell": 0, + "replay_s": 2.87, + "bad_trajectory": 0, + "grader_error": 0, + "video_rendered": 1, + "agent_sessions": 418, + "agent_resets": 11, + "agent_replays": 0, + "agent_cost_self_reported": 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/embodied-olympics-fun-high-jump-i00-privileged--xtljosb/replay.mp4", + "duration_s": 0.04, + "fps": 25.0, + "frames": 1, + "sha256": "66f440a1ee7a787f7f7640988e3e3242c776c61c1578bd5aff3ad2333e0c1a8e" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "high_jump_landed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察：本次评分视频共1帧、时长0.04秒；视频0.00 s的唯一一帧中，机器人双脚站在红色跑道起点区域，远离横杆与蓝色落地垫，没有腾空、越杆或落垫。该帧也是全片末帧，跳过杆并落地的判据未成立。" + }, + { + "t": null, + "log_step": 47, + "text_zh": "观察：log step 47中，agent的脚本创建新Session，只执行一次s.hold()返回的站立动作后调用s.save_trajectory()，工具报告执行成功并生成trajectory.npz；log step 48写入的生成脚本也仅包含这一个保持动作。" + }, + { + "t": null, + "log_step": 658, + "text_zh": "观察：log step 658回读的生成脚本仍只执行一次保持站姿；评分器字段n_actions=1、sim_steps=1、stages_reached=0、episode_over=0、fell=0，与视频0.00 s所示初态相符。会话中的助跑和跳跃测试不属于这条评分轨迹；本节点未开展实际跳跃操作，故不记录初始站立的背景支撑接触。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "air_s", + "unit": "s", + "threshold": 0.05, + "direction": "higher_better", + "value": null + }, + { + "name": "rise_m", + "unit": "m", + "threshold": 0.08, + "direction": "higher_better", + "value": null + }, + { + "name": "bar_impulse_Ns", + "unit": "N s", + "threshold": 2.0, + "direction": "lower_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "high_jump_landed", + "t": 0.0, + "log_step": 47, + "where": "首步前", + "evidence_zh": "观察：视频0.00 s是全片唯一一帧，机器人仍在跑道起点站立，跳高尚未启动；log step 47保存的交付轨迹只执行一个保持站姿动作。评分器n_actions=1、stages_reached=0、episode_over=0，说明本次交付在开始助跑、起跳前已结束。" + }, + "unmet": { + "high_jump_landed": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 47, + "evidence_zh": "观察：视频0.00 s中机器人站在横杆前方远处；log step 47明确只执行一次hold并保存轨迹，评分器也只重放1个动作。该基线未包含助跑、单脚蹬地、腾空或过杆动作，评分尝试未启动跳高。" + } + ], + "root_causes": [], + "gap_zh": "日志与单帧视频足以确定交付的是未启动跳高的站立基线，但不足以确定agent为何最终没有替换该基线；不能仅凭没尝试归因于规划、控制或交付操作错误，故根因留空。log step 1042的最后发言仍是等待独立测试结果，没有完成声明或解释交付选择的记录。本次无状态日志，评分器未报告air_s、rise_m、bar_impulse_Ns，原生指标均记null；会话测试不作为评分局的动作或失败现象。" + }, + "summary_zh": "评分轨迹仅含一个保持站姿动作，视频0.00 s的唯一一帧中机器人仍在跑道起点，跳过杆并落地未达成，节点最终为0/1。断点为high_jump_landed的首步前，现象为没尝试；log step 47的基线保存操作与评分器仅重放1个动作相互印证。会话虽有跳跃测试，但不足以确定最终保留基线的根因，且没有状态日志或原生指标值。", + "conf": 9, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "评分视频仅1帧（0.00 s站在跑道起点），评分器n_actions=1、stages_reached=0，节点未达成、断点首步前+没尝试与证据一致；log step 47/48/658存在且内容与引用相符，步骤885只把前缀轨迹存到/tmp，未覆盖交付文件，根因留空并有gap_zh，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": "2026-10-29T10:07:02Z", + "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/embodied-olympics-fun-high-jump-i00-privileged--xtljosb/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-high-jump-i00-privileged--xtljosb/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "f2027b9ea81628267fda441e1052d2a983fae01b5d19363bcacb6ed009ee0922", + "video_sha256": "66f440a1ee7a787f7f7640988e3e3242c776c61c1578bd5aff3ad2333e0c1a8e", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/juggling/task_dag.json b/phase2/embodied-olympics-fun/juggling/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..46b2b81fb020998d25e8f5968bd87cd28f6bf95e --- /dev/null +++ b/phase2/embodied-olympics-fun/juggling/task_dag.json @@ -0,0 +1,658 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "embodied-olympics-fun/juggling", + "suite": "embodied-olympics-fun", + "base": "embodied-olympics-fun-juggling-i00", + "title_zh": "连续颠球十次", + "summary_zh": "足球从双掌离开后，由脚、小腿、大腿或头累计完成十次符合回升高度和速度要求的有效颠球，期间不触地、不违规触球且机器人不跌倒，第十次计数成立即成功。", + "embodiment_zh": "两模式相同：Unitree G1 双足人形机器人，29 个关节，身高 1.32 米、重 33 千克；双臂末端均为无工作手指的平掌垫，非二指夹爪、灵巧手、吸盘或安装工具，靠双臂挤压与摩擦夹球，抓握指令无效；头、双脚、小腿及大腿可击球，双腿支撑和调整身体，无轮式底盘。", + "task_class": "compositional", + "analysis_level": "step", + "spec": { + "version": "1.2", + "task": "embodied-olympics-fun-juggling-i00", + "bench": "embodied-olympics-fun", + "progress": "live", + "entities": {}, + "milestones": [ + { + "id": "valid_touch_1", + "zh": "第1次颠球", + "mode": "event", + "when": null, + "after": [], + "hold_s": 0, + "goal": true, + "criterion_zh": "球先完全离开机器人至少 0.03 秒，再由脚、小腿、大腿或头击起；场边视角可见黑白球由下落转为离体上升，在下一次独立触球前球心较本次触球起点升高至少 0.22 米（一个球直径），累计有效次数达到 1，间隔不足 0.05 秒的接触合为一次、单次接触持续不超过 0.15 秒，且首次触球免速度比例要求。仅碰到球、停球或未升够高度均未完成；参考视频中球在头顶反复弹起，脚或腿击球也可，已计入次数保持锁存。", + "step": { + "cat": "动态", + "item": "击打" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "其他", + "plan": 1, + "note_zh": "候选：击打；机器人用脚、小腿、大腿或头短促击打足球使其离体上升，现有接触方式没有击打项，持续推、压或支撑均不能准确表达。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "task/privileged/instruction.md:18-27", + "task/privileged/environment/rfo_task.py:17-19", + "task/privileged/environment/rfo_task.py:139-140", + "task/privileged/environment/rfo_task.py:191-247", + "grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178", + "grader/efe_embodied-olympics-fun-files/rfo/referee.py:305-347" + ], + "conf": 6, + "note_zh": "按文本十次连续计数及 S5，每个有效次数单列，依原生 touch 阶段顺序锁存；已完成的计数不因后续失败而抹去。球离手、球飞行、身体调整及每次接触后的回升均并入该次验收，不另建步骤。双掌的初始夹持和释放先于首次击球，不是本节点同时发生的接触；不限定击球部位或左右次序。" + }, + { + "id": "valid_touch_2", + "zh": "第2次颠球", + "mode": "event", + "when": null, + "after": [ + "valid_touch_1" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "在前 1 次有效计数之后，脚、小腿、大腿或头再次击起足球；场边视角可见黑白球由下落转为离体上升，在下一次独立触球前球心较本次触球起点升高至少 0.22 米（一个球直径），累计有效次数达到 2，间隔不足 0.05 秒的接触合为一次、单次接触持续不超过 0.15 秒，且按评分器折算到触球起始高度的离球速率须至少为入球速率的 0.95 倍，低于此值的被动反弹不计数。仅碰到球、停球或未升够高度均未完成；参考视频中球在头顶反复弹起，脚或腿击球也可，已计入次数保持锁存。", + "step": { + "cat": "动态", + "item": "击打" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "其他", + "plan": 1, + "note_zh": "候选：击打；机器人用脚、小腿、大腿或头短促击打足球使其离体上升，现有接触方式没有击打项，持续推、压或支撑均不能准确表达。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "task/privileged/instruction.md:18-27", + "task/privileged/environment/rfo_task.py:17-19", + "task/privileged/environment/rfo_task.py:139-140", + "task/privileged/environment/rfo_task.py:191-247", + "grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178", + "grader/efe_embodied-olympics-fun-files/rfo/referee.py:305-347" + ], + "conf": 6, + "note_zh": "按文本十次连续计数及 S5，每个有效次数单列，依原生 touch 阶段顺序锁存；已完成的计数不因后续失败而抹去。球离手、球飞行、身体调整及每次接触后的回升均并入该次验收，不另建步骤。双掌的初始夹持和释放先于首次击球，不是本节点同时发生的接触；不限定击球部位或左右次序。" + }, + { + "id": "valid_touch_3", + "zh": "第3次颠球", + "mode": "event", + "when": null, + "after": [ + "valid_touch_2" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "在前 2 次有效计数之后，脚、小腿、大腿或头再次击起足球；场边视角可见黑白球由下落转为离体上升，在下一次独立触球前球心较本次触球起点升高至少 0.22 米（一个球直径），累计有效次数达到 3，间隔不足 0.05 秒的接触合为一次、单次接触持续不超过 0.15 秒，且按评分器折算到触球起始高度的离球速率须至少为入球速率的 0.95 倍，低于此值的被动反弹不计数。仅碰到球、停球或未升够高度均未完成；参考视频中球在头顶反复弹起，脚或腿击球也可，已计入次数保持锁存。", + "step": { + "cat": "动态", + "item": "击打" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "其他", + "plan": 1, + "note_zh": "候选：击打；机器人用脚、小腿、大腿或头短促击打足球使其离体上升，现有接触方式没有击打项，持续推、压或支撑均不能准确表达。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "task/privileged/instruction.md:18-27", + "task/privileged/environment/rfo_task.py:17-19", + "task/privileged/environment/rfo_task.py:139-140", + "task/privileged/environment/rfo_task.py:191-247", + "grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178", + "grader/efe_embodied-olympics-fun-files/rfo/referee.py:305-347" + ], + "conf": 6, + "note_zh": "按文本十次连续计数及 S5，每个有效次数单列，依原生 touch 阶段顺序锁存；已完成的计数不因后续失败而抹去。球离手、球飞行、身体调整及每次接触后的回升均并入该次验收，不另建步骤。双掌的初始夹持和释放先于首次击球，不是本节点同时发生的接触；不限定击球部位或左右次序。" + }, + { + "id": "valid_touch_4", + "zh": "第4次颠球", + "mode": "event", + "when": null, + "after": [ + "valid_touch_3" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "在前 3 次有效计数之后，脚、小腿、大腿或头再次击起足球；场边视角可见黑白球由下落转为离体上升，在下一次独立触球前球心较本次触球起点升高至少 0.22 米（一个球直径），累计有效次数达到 4，间隔不足 0.05 秒的接触合为一次、单次接触持续不超过 0.15 秒，且按评分器折算到触球起始高度的离球速率须至少为入球速率的 0.95 倍，低于此值的被动反弹不计数。仅碰到球、停球或未升够高度均未完成；参考视频中球在头顶反复弹起，脚或腿击球也可，已计入次数保持锁存。", + "step": { + "cat": "动态", + "item": "击打" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "其他", + "plan": 1, + "note_zh": "候选：击打；机器人用脚、小腿、大腿或头短促击打足球使其离体上升，现有接触方式没有击打项，持续推、压或支撑均不能准确表达。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "task/privileged/instruction.md:18-27", + "task/privileged/environment/rfo_task.py:17-19", + "task/privileged/environment/rfo_task.py:139-140", + "task/privileged/environment/rfo_task.py:191-247", + "grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178", + "grader/efe_embodied-olympics-fun-files/rfo/referee.py:305-347" + ], + "conf": 6, + "note_zh": "按文本十次连续计数及 S5，每个有效次数单列，依原生 touch 阶段顺序锁存；已完成的计数不因后续失败而抹去。球离手、球飞行、身体调整及每次接触后的回升均并入该次验收，不另建步骤。双掌的初始夹持和释放先于首次击球，不是本节点同时发生的接触；不限定击球部位或左右次序。" + }, + { + "id": "valid_touch_5", + "zh": "第5次颠球", + "mode": "event", + "when": null, + "after": [ + "valid_touch_4" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "在前 4 次有效计数之后，脚、小腿、大腿或头再次击起足球；场边视角可见黑白球由下落转为离体上升，在下一次独立触球前球心较本次触球起点升高至少 0.22 米（一个球直径），累计有效次数达到 5，间隔不足 0.05 秒的接触合为一次、单次接触持续不超过 0.15 秒，且按评分器折算到触球起始高度的离球速率须至少为入球速率的 0.95 倍，低于此值的被动反弹不计数。仅碰到球、停球或未升够高度均未完成；参考视频中球在头顶反复弹起，脚或腿击球也可，已计入次数保持锁存。", + "step": { + "cat": "动态", + "item": "击打" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "其他", + "plan": 1, + "note_zh": "候选：击打；机器人用脚、小腿、大腿或头短促击打足球使其离体上升，现有接触方式没有击打项，持续推、压或支撑均不能准确表达。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "task/privileged/instruction.md:18-27", + "task/privileged/environment/rfo_task.py:17-19", + "task/privileged/environment/rfo_task.py:139-140", + "task/privileged/environment/rfo_task.py:191-247", + "grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178", + "grader/efe_embodied-olympics-fun-files/rfo/referee.py:305-347" + ], + "conf": 6, + "note_zh": "按文本十次连续计数及 S5，每个有效次数单列，依原生 touch 阶段顺序锁存；已完成的计数不因后续失败而抹去。球离手、球飞行、身体调整及每次接触后的回升均并入该次验收，不另建步骤。双掌的初始夹持和释放先于首次击球，不是本节点同时发生的接触；不限定击球部位或左右次序。" + }, + { + "id": "valid_touch_6", + "zh": "第6次颠球", + "mode": "event", + "when": null, + "after": [ + "valid_touch_5" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "在前 5 次有效计数之后，脚、小腿、大腿或头再次击起足球；场边视角可见黑白球由下落转为离体上升，在下一次独立触球前球心较本次触球起点升高至少 0.22 米（一个球直径），累计有效次数达到 6，间隔不足 0.05 秒的接触合为一次、单次接触持续不超过 0.15 秒，且按评分器折算到触球起始高度的离球速率须至少为入球速率的 0.95 倍，低于此值的被动反弹不计数。仅碰到球、停球或未升够高度均未完成；参考视频中球在头顶反复弹起，脚或腿击球也可，已计入次数保持锁存。", + "step": { + "cat": "动态", + "item": "击打" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "其他", + "plan": 1, + "note_zh": "候选：击打；机器人用脚、小腿、大腿或头短促击打足球使其离体上升，现有接触方式没有击打项，持续推、压或支撑均不能准确表达。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "task/privileged/instruction.md:18-27", + "task/privileged/environment/rfo_task.py:17-19", + "task/privileged/environment/rfo_task.py:139-140", + "task/privileged/environment/rfo_task.py:191-247", + "grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178", + "grader/efe_embodied-olympics-fun-files/rfo/referee.py:305-347" + ], + "conf": 6, + "note_zh": "按文本十次连续计数及 S5，每个有效次数单列，依原生 touch 阶段顺序锁存；已完成的计数不因后续失败而抹去。球离手、球飞行、身体调整及每次接触后的回升均并入该次验收，不另建步骤。双掌的初始夹持和释放先于首次击球，不是本节点同时发生的接触；不限定击球部位或左右次序。" + }, + { + "id": "valid_touch_7", + "zh": "第7次颠球", + "mode": "event", + "when": null, + "after": [ + "valid_touch_6" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "在前 6 次有效计数之后，脚、小腿、大腿或头再次击起足球；场边视角可见黑白球由下落转为离体上升，在下一次独立触球前球心较本次触球起点升高至少 0.22 米（一个球直径），累计有效次数达到 7，间隔不足 0.05 秒的接触合为一次、单次接触持续不超过 0.15 秒，且按评分器折算到触球起始高度的离球速率须至少为入球速率的 0.95 倍，低于此值的被动反弹不计数。仅碰到球、停球或未升够高度均未完成；参考视频中球在头顶反复弹起，脚或腿击球也可，已计入次数保持锁存。", + "step": { + "cat": "动态", + "item": "击打" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "其他", + "plan": 1, + "note_zh": "候选：击打；机器人用脚、小腿、大腿或头短促击打足球使其离体上升，现有接触方式没有击打项，持续推、压或支撑均不能准确表达。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "task/privileged/instruction.md:18-27", + "task/privileged/environment/rfo_task.py:17-19", + "task/privileged/environment/rfo_task.py:139-140", + "task/privileged/environment/rfo_task.py:191-247", + "grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178", + "grader/efe_embodied-olympics-fun-files/rfo/referee.py:305-347" + ], + "conf": 6, + "note_zh": "按文本十次连续计数及 S5，每个有效次数单列，依原生 touch 阶段顺序锁存；已完成的计数不因后续失败而抹去。球离手、球飞行、身体调整及每次接触后的回升均并入该次验收，不另建步骤。双掌的初始夹持和释放先于首次击球，不是本节点同时发生的接触；不限定击球部位或左右次序。" + }, + { + "id": "valid_touch_8", + "zh": "第8次颠球", + "mode": "event", + "when": null, + "after": [ + "valid_touch_7" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "在前 7 次有效计数之后，脚、小腿、大腿或头再次击起足球；场边视角可见黑白球由下落转为离体上升，在下一次独立触球前球心较本次触球起点升高至少 0.22 米（一个球直径），累计有效次数达到 8，间隔不足 0.05 秒的接触合为一次、单次接触持续不超过 0.15 秒，且按评分器折算到触球起始高度的离球速率须至少为入球速率的 0.95 倍，低于此值的被动反弹不计数。仅碰到球、停球或未升够高度均未完成；参考视频中球在头顶反复弹起，脚或腿击球也可，已计入次数保持锁存。", + "step": { + "cat": "动态", + "item": "击打" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "其他", + "plan": 1, + "note_zh": "候选：击打；机器人用脚、小腿、大腿或头短促击打足球使其离体上升，现有接触方式没有击打项，持续推、压或支撑均不能准确表达。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "task/privileged/instruction.md:18-27", + "task/privileged/environment/rfo_task.py:17-19", + "task/privileged/environment/rfo_task.py:139-140", + "task/privileged/environment/rfo_task.py:191-247", + "grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178", + "grader/efe_embodied-olympics-fun-files/rfo/referee.py:305-347" + ], + "conf": 6, + "note_zh": "按文本十次连续计数及 S5，每个有效次数单列，依原生 touch 阶段顺序锁存；已完成的计数不因后续失败而抹去。球离手、球飞行、身体调整及每次接触后的回升均并入该次验收，不另建步骤。双掌的初始夹持和释放先于首次击球，不是本节点同时发生的接触；不限定击球部位或左右次序。" + }, + { + "id": "valid_touch_9", + "zh": "第9次颠球", + "mode": "event", + "when": null, + "after": [ + "valid_touch_8" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "在前 8 次有效计数之后，脚、小腿、大腿或头再次击起足球；场边视角可见黑白球由下落转为离体上升，在下一次独立触球前球心较本次触球起点升高至少 0.22 米（一个球直径），累计有效次数达到 9，间隔不足 0.05 秒的接触合为一次、单次接触持续不超过 0.15 秒，且按评分器折算到触球起始高度的离球速率须至少为入球速率的 0.95 倍，低于此值的被动反弹不计数。仅碰到球、停球或未升够高度均未完成；参考视频中球在头顶反复弹起，脚或腿击球也可，已计入次数保持锁存。", + "step": { + "cat": "动态", + "item": "击打" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "其他", + "plan": 1, + "note_zh": "候选：击打；机器人用脚、小腿、大腿或头短促击打足球使其离体上升，现有接触方式没有击打项，持续推、压或支撑均不能准确表达。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "task/privileged/instruction.md:18-27", + "task/privileged/environment/rfo_task.py:17-19", + "task/privileged/environment/rfo_task.py:139-140", + "task/privileged/environment/rfo_task.py:191-247", + "grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178", + "grader/efe_embodied-olympics-fun-files/rfo/referee.py:305-347" + ], + "conf": 6, + "note_zh": "按文本十次连续计数及 S5，每个有效次数单列，依原生 touch 阶段顺序锁存；已完成的计数不因后续失败而抹去。球离手、球飞行、身体调整及每次接触后的回升均并入该次验收，不另建步骤。双掌的初始夹持和释放先于首次击球，不是本节点同时发生的接触；不限定击球部位或左右次序。" + }, + { + "id": "valid_touch_10", + "zh": "第10次颠球", + "mode": "event", + "when": null, + "after": [ + "valid_touch_9" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "在前 9 次有效计数之后，脚、小腿、大腿或头再次击起足球；场边视角可见黑白球由下落转为离体上升，在下一次独立触球前球心较本次触球起点升高至少 0.22 米（一个球直径），累计有效次数达到 10，间隔不足 0.05 秒的接触合为一次、单次接触持续不超过 0.15 秒，且按评分器折算到触球起始高度的离球速率须至少为入球速率的 0.95 倍，低于此值的被动反弹不计数。仅碰到球、停球或未升够高度均未完成；参考视频中球在头顶反复弹起，脚或腿击球也可，已计入次数保持锁存，第十次成立即终止成功，无须另接球或站稳。", + "step": { + "cat": "动态", + "item": "击打" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "其他", + "plan": 1, + "note_zh": "候选：击打；机器人用脚、小腿、大腿或头短促击打足球使其离体上升，现有接触方式没有击打项，持续推、压或支撑均不能准确表达。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "task/privileged/instruction.md:18-27", + "task/privileged/environment/rfo_task.py:17-19", + "task/privileged/environment/rfo_task.py:139-140", + "task/privileged/environment/rfo_task.py:191-247", + "grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178", + "grader/efe_embodied-olympics-fun-files/rfo/referee.py:305-347" + ], + "conf": 6, + "note_zh": "按文本十次连续计数及 S5，每个有效次数单列，依原生 touch 阶段顺序锁存；已完成的计数不因后续失败而抹去。球离手、球飞行、身体调整及每次接触后的回升均并入该次验收，不另建步骤。双掌的初始夹持和释放先于首次击球，不是本节点同时发生的接触；不限定击球部位或左右次序。" + } + ], + "constraints": [ + { + "id": "ball_ground_contact", + "zh": "足球触地", + "never": null, + "criterion_zh": "无论球是否已离开双掌，足球与草地或地面发生接触即失败；场边视角可见球落到绿色场地表面。", + "consequence": "立即终止", + "sources": [ + "task/privileged/environment/rfo_task.py:203-205", + "task/privileged/environment/rfo_task.py:217-218" + ] + }, + { + "id": "illegal_pre_drop_contact", + "zh": "离手前违规触球", + "never": null, + "criterion_zh": "释放确认前，球接触机器人手掌垫及腕部所归属的双手以外任何部位即失败；例如双掌仍夹球时球已碰头、手臂或腿。释放确认是球连续 0.03 秒不碰机器人，释放时刻回记到最后一次接触。", + "consequence": "立即终止", + "sources": [ + "task/privileged/environment/rfo_task.py:17-19", + "task/privileged/environment/rfo_task.py:203-215" + ] + }, + { + "id": "handball_after_drop", + "zh": "离手后手球", + "never": null, + "criterion_zh": "自最后一次持球接触回记的释放时刻起超过 0.15 秒，任一手、前臂或上臂再碰足球即失败；视频中可见离手后的球再次擦到或压到手臂。", + "consequence": "立即终止", + "sources": [ + "task/privileged/environment/rfo_task.py:17", + "task/privileged/environment/rfo_task.py:211-220" + ] + }, + { + "id": "trunk_ball_contact", + "zh": "躯干触球", + "never": null, + "criterion_zh": "已释放的球接触胸、肩、骨盆等躯干部位即失败；头、脚、小腿、大腿为合法部位，头顶球不属此项。", + "consequence": "立即终止", + "sources": [ + "task/privileged/environment/rfo_task.py:18", + "task/privileged/environment/rfo_task.py:191-195", + "task/privileged/environment/rfo_task.py:221-223" + ] + }, + { + "id": "ball_carried", + "zh": "触球超时", + "never": null, + "criterion_zh": "同一次合法部位触球从首接触到末接触持续超过 0.15 秒即失败；不足 0.05 秒的短暂分离仍合为同一次，可见表现是球在头或腿上停住、滚动或被持续托着。", + "consequence": "立即终止", + "sources": [ + "task/privileged/environment/rfo_task.py:196", + "task/privileged/environment/rfo_task.py:232-233", + "grader/efe_embodied-olympics-fun-files/rfo/referee.py:325-347" + ] + }, + { + "id": "premature_retouch", + "zh": "回升不足再触", + "never": null, + "criterion_zh": "一次合法触球尚待确认上升 0.22 米时，又开始下一次独立合法触球，即失败；不是把不足 0.05 秒的接触抖动另算一次。视频中球尚未弹离足够高度便再次碰到头或腿。", + "consequence": "立即终止", + "sources": [ + "task/privileged/environment/rfo_task.py:227-239", + "grader/efe_embodied-olympics-fun-files/rfo/referee.py:325-347" + ] + }, + { + "id": "robot_fell", + "zh": "机器人跌倒", + "never": null, + "criterion_zh": "控制步末尚未成功时，骨盆高度低于 0.35 米，或任一脚以外的机器人部位碰到地面，即失败；可见身体塌低、手撑地或躯干倒地，单脚站立不算跌倒。", + "consequence": "立即终止", + "sources": [ + "task/privileged/environment/rfo_task.py:137-140", + "grader/efe_embodied-olympics-fun-files/rfo/referee.py:272-289" + ] + }, + { + "id": "simulation_unstable", + "zh": "仿真数值发散", + "never": null, + "criterion_zh": "物理步出现 BADQACC 警告、非有限位姿，或所监控线速度分量绝对值超过 150 米/秒，即数值发散并终止；同一控制步刚成立的成功也会被撤销。视频可能出现机器人或球突跳、消失，须以模拟器记录确认。", + "consequence": "立即终止", + "sources": [ + "grader/efe_embodied-olympics-fun-files/rfo/world.py:108-109", + "grader/efe_embodied-olympics-fun-files/rfo/world.py:866-883", + "grader/efe_embodied-olympics-fun-files/rfo/referee.py:166-172" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "valid_touches_ge_1", + "level": "new_code", + "source": "task/privileged/environment/rfo_task.py:245-256; log Task.trackers.count or Task.privileged().touches and ordered Referee.stages for count >= 1" + }, + { + "atom": "valid_touches_ge_2", + "level": "new_code", + "source": "task/privileged/environment/rfo_task.py:245-256; log Task.trackers.count or Task.privileged().touches and ordered Referee.stages for count >= 2" + }, + { + "atom": "valid_touches_ge_3", + "level": "new_code", + "source": "task/privileged/environment/rfo_task.py:245-256; log Task.trackers.count or Task.privileged().touches and ordered Referee.stages for count >= 3" + }, + { + "atom": "valid_touches_ge_4", + "level": "new_code", + "source": "task/privileged/environment/rfo_task.py:245-256; log Task.trackers.count or Task.privileged().touches and ordered Referee.stages for count >= 4" + }, + { + "atom": "valid_touches_ge_5", + "level": "new_code", + "source": "task/privileged/environment/rfo_task.py:245-256; log Task.trackers.count or Task.privileged().touches and ordered Referee.stages for count >= 5" + }, + { + "atom": "valid_touches_ge_6", + "level": "new_code", + "source": "task/privileged/environment/rfo_task.py:245-256; log Task.trackers.count or Task.privileged().touches and ordered Referee.stages for count >= 6" + }, + { + "atom": "valid_touches_ge_7", + "level": "new_code", + "source": "task/privileged/environment/rfo_task.py:245-256; log Task.trackers.count or Task.privileged().touches and ordered Referee.stages for count >= 7" + }, + { + "atom": "valid_touches_ge_8", + "level": "new_code", + "source": "task/privileged/environment/rfo_task.py:245-256; log Task.trackers.count or Task.privileged().touches and ordered Referee.stages for count >= 8" + }, + { + "atom": "valid_touches_ge_9", + "level": "new_code", + "source": "task/privileged/environment/rfo_task.py:245-256; log Task.trackers.count or Task.privileged().touches and ordered Referee.stages for count >= 9" + }, + { + "atom": "valid_touches_ge_10", + "level": "new_code", + "source": "task/privileged/environment/rfo_task.py:245-256; log Task.trackers.count or Task.privileged().touches and ordered Referee.stages for count >= 10" + }, + { + "atom": "ball_ground_contact", + "level": "new_code", + "source": "task/privileged/environment/rfo_task.py:203-205; task/privileged/environment/rfo_task.py:217-218; log per-step Referee.failed and event record" + }, + { + "atom": "illegal_pre_drop_contact", + "level": "new_code", + "source": "task/privileged/environment/rfo_task.py:17-19; task/privileged/environment/rfo_task.py:203-215; log per-step Referee.failed and event record" + }, + { + "atom": "handball_after_drop", + "level": "new_code", + "source": "task/privileged/environment/rfo_task.py:17; task/privileged/environment/rfo_task.py:211-220; log per-step Referee.failed and event record" + }, + { + "atom": "trunk_ball_contact", + "level": "new_code", + "source": "task/privileged/environment/rfo_task.py:18; task/privileged/environment/rfo_task.py:191-195; task/privileged/environment/rfo_task.py:221-223; log per-step Referee.failed and event record" + }, + { + "atom": "ball_carried", + "level": "new_code", + "source": "task/privileged/environment/rfo_task.py:196; task/privileged/environment/rfo_task.py:232-233; grader/efe_embodied-olympics-fun-files/rfo/referee.py:325-347; log per-step Referee.failed and event record" + }, + { + "atom": "premature_retouch", + "level": "new_code", + "source": "task/privileged/environment/rfo_task.py:227-239; grader/efe_embodied-olympics-fun-files/rfo/referee.py:325-347; log per-step Referee.failed and event record" + }, + { + "atom": "robot_fell", + "level": "new_code", + "source": "task/privileged/environment/rfo_task.py:137-140; grader/efe_embodied-olympics-fun-files/rfo/referee.py:272-289; log per-step Referee.failed and event record" + }, + { + "atom": "simulation_unstable", + "level": "new_code", + "source": "grader/efe_embodied-olympics-fun-files/rfo/world.py:108-109; grader/efe_embodied-olympics-fun-files/rfo/world.py:866-883; grader/efe_embodied-olympics-fun-files/rfo/referee.py:166-172; log per-step Referee.failed and event record" + }, + { + "atom": "task_success", + "level": "new_code", + "source": "grader/efe_embodied-olympics-fun-files/rfo/referee.py:174-178; grader/efe_embodied-olympics-fun-files/rfo/world.py:881-883; log benchmark Referee.success after each control step" + } + ] + }, + "native_metrics": [ + { + "name": "touches", + "unit": "count", + "threshold": 10, + "direction": "higher_better", + "sources": [ + "task/privileged/environment/rfo_task.py:245-252" + ] + }, + { + "name": "ball_height_m", + "unit": "m", + "threshold": null, + "direction": null, + "note_zh": "裁判记录球心绝对高度，供定位轨迹；没有独立的绝对高度成功阈值，0.22 米检查的是每次触球后的相对回升。", + "sources": [ + "task/privileged/environment/rfo_task.py:235-239", + "task/privileged/environment/rfo_task.py:249-252" + ] + } + ], + "open_questions_zh": [ + "inputs/states/index.json 的 logs 为空，无法绑定或复核毫秒级接触、回升高度及速度比例；按 DECISIONS.md P8，各节点 when、约束 never 和 success 均为 null，atoms 标出需要补记的裁判世界状态。视频可见逐次顶球，但精确阈值仍须原生裁判记录。", + "粒度采用 S5 的重复计数规则：十次颠球各一节点，即使连续飞行不能停住，也不按 S10 将文本明确计数的十次合为一步；球离手是首次颠球的中间事件，依据 S8 与规则 2.6 不另列释放步骤。", + "评分器与文本的采样差异：instruction.md:25 统称裁判在每个 1 毫秒物理步判定，实际球接触与计数在物理步判，跌倒只在 10 毫秒控制步末且未成功时判（referee.py:265-289）。按实际评分器写跌倒约束，不增设独立站稳节点。", + "两模式物理成功核心相同，但 privileged 还要求提交的有限数值轨迹在两个新进程中都成功且末状态哈希相同（grade.py:87-124）；standard 采纳最后一次现场尝试的裁判成功且无基础设施错误（standard/server.py:152-178；v2_1/grade.py:93-105）。这些交付、复现及基础设施检查不伪装成语义世界状态。", + "接触 how 缺少击打，依 DECISIONS.md S16、D16(c) 使用其他并注明候选“击打”；合法头球与脚、小腿、大腿击球均属腿/躯干，不把成功参考视频的头顶球方式限定为唯一路线。" + ], + "conf": 6, + "conf_reason_zh": "十次计数、边与阈值均由任务文本和两模式共用评分器直接支持，初态已核实，成功视频及原生评分结果显示十阶段全部通过；但没有状态日志，速度比例、接触时窗和各次精确计数时刻不能逐帧核验，所有条件仅提供视频判据与待补日志来源。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "十个顺序颠球节点与文本的 touch 1→…→touch 10 一致，无状态日志故 when/never/success 为 null 符合 P8；约束覆盖评分器全部失败条件，30 秒时限按 P21 不列；标签（动态·击打、腿/躯干×其他+击打）符合 S16。未发现需要修改的内容。", + "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：['valid_touch_1', 'valid_touch_2', 'valid_touch_3', 'valid_touch_4', 'valid_touch_5', 'valid_touch_6', 'valid_touch_7', 'valid_touch_8', 'valid_touch_9', 'valid_touch_10']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:49:44Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__juggling/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__juggling/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "fb80086a835431ce8f029e5920642822b776b2593b07984ec9b71fde6d7a2bc3", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/embodied-olympics-fun/juggling/trajectories/embodied-olympics-fun-juggling-i00-privileged--zjv9zmw.json b/phase2/embodied-olympics-fun/juggling/trajectories/embodied-olympics-fun-juggling-i00-privileged--zjv9zmw.json new file mode 100644 index 0000000000000000000000000000000000000000..b7ef1ba2835fe616c177ba45544b1d3421e89f01 --- /dev/null +++ b/phase2/embodied-olympics-fun/juggling/trajectories/embodied-olympics-fun-juggling-i00-privileged--zjv9zmw.json @@ -0,0 +1,437 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-juggling-i00-privileged--zjv9zmw", + "taskKey": "embodied-olympics-fun/juggling", + "dag_sha256": "25e819f9f1251e0a7ce9add4f150f9227df00d011426943d8c29867bd2eff282", + "outcome": { + "success": 1, + "reward": 1.0, + "source": "catalog", + "raw": { + "success": 1, + "replay_success": 1, + "deterministic": 1, + "n_actions": 639, + "sim_steps": 639, + "horizon": 3000, + "episode_over": 1, + "progress": 1.0, + "stages_reached": 10, + "stages_total": 10, + "fell": 0, + "replay_s": 4.38, + "bad_trajectory": 0, + "metric_ball_height_m": 1.631, + "metric_touches": 10.0, + "grader_error": 0, + "video_rendered": 1, + "agent_sessions": 51, + "agent_resets": 161, + "agent_replays": 1, + "agent_cost_self_reported": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.71, + "peak_processes": 9, + "peak_disk_gb": 0.052, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0, + "session_log_tampered": 0, + "session_log_from_monitor": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-juggling-i00-privileged--zjv9zmw/replay.mp4", + "duration_s": 6.4, + "fps": 25.0, + "frames": 160, + "sha256": "44e8c4f31adac97e99cb4d4d41db48bd53976a1c8286b356cbacf151e02c7042" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 10, + "base": 0, + "peak": 10, + "final": 10, + "nodes": { + "valid_touch_1": { + "method": "visual", + "intervals": [ + [ + 1.96, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 1.68, + "log_step": null, + "text_zh": "观察：视频 0.00–1.60 s，机器人以两侧平掌夹球并将球举至头上；视频 1.68–1.76 s 双掌向两侧分开，足球离掌下落，随后才发生首次头部击球。" + }, + { + "t": 1.96, + "log_step": null, + "text_zh": "观察：视频 1.80–1.96 s，足球在机器人头部附近由下落转为离体上升，为松手后第1个独立头球回升周期；双掌分开，机器人仍站立。" + }, + { + "t": 1.96, + "log_step": 104, + "text_zh": "观察：log step 104 是对已交付 trajectory.npz 的独立回放核验，裁判返回 stages 包含 touch 1，最终 touches=10、success=true、executed_rows=639；本节点的有效计数由该交付回放裁判结果确认，视频锚点取逐帧所见回升阶段，约有一帧误差，未将 agent 会话时间换算为视频时间。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "其他", + "note_zh": "候选：击打；视频 1.80–1.96 s，机器人以头部短促顶击足球，使球离体回升，为第1次颠球；不是持续托举或压擦，现有接触方式无击打项。" + } + ] + }, + "valid_touch_2": { + "method": "visual", + "intervals": [ + [ + 2.44, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 2.44, + "log_step": null, + "text_zh": "观察：视频 2.28–2.44 s，足球在机器人头部附近由下落转为离体上升，为松手后第2个独立头球回升周期；双掌分开，机器人仍站立。" + }, + { + "t": 2.44, + "log_step": 104, + "text_zh": "观察：log step 104 是对已交付 trajectory.npz 的独立回放核验，裁判返回 stages 包含 touch 2，最终 touches=10、success=true、executed_rows=639；本节点的有效计数由该交付回放裁判结果确认，视频锚点取逐帧所见回升阶段，约有一帧误差，未将 agent 会话时间换算为视频时间。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "其他", + "note_zh": "候选：击打；视频 2.28–2.44 s，机器人以头部短促顶击足球，使球离体回升，为第2次颠球；不是持续托举或压擦，现有接触方式无击打项。" + } + ] + }, + "valid_touch_3": { + "method": "visual", + "intervals": [ + [ + 2.92, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 2.92, + "log_step": null, + "text_zh": "观察：视频 2.80–2.92 s，足球在机器人头部附近由下落转为离体上升，为松手后第3个独立头球回升周期；双掌分开，机器人仍站立。" + }, + { + "t": 2.92, + "log_step": 104, + "text_zh": "观察：log step 104 是对已交付 trajectory.npz 的独立回放核验，裁判返回 stages 包含 touch 3，最终 touches=10、success=true、executed_rows=639；本节点的有效计数由该交付回放裁判结果确认，视频锚点取逐帧所见回升阶段，约有一帧误差，未将 agent 会话时间换算为视频时间。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "其他", + "note_zh": "候选：击打；视频 2.80–2.92 s，机器人以头部短促顶击足球，使球离体回升，为第3次颠球；不是持续托举或压擦，现有接触方式无击打项。" + } + ] + }, + "valid_touch_4": { + "method": "visual", + "intervals": [ + [ + 3.44, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 3.44, + "log_step": null, + "text_zh": "观察：视频 3.32–3.44 s，足球在机器人头部附近由下落转为离体上升，为松手后第4个独立头球回升周期；双掌分开，机器人仍站立。" + }, + { + "t": 3.44, + "log_step": 104, + "text_zh": "观察：log step 104 是对已交付 trajectory.npz 的独立回放核验，裁判返回 stages 包含 touch 4，最终 touches=10、success=true、executed_rows=639；本节点的有效计数由该交付回放裁判结果确认，视频锚点取逐帧所见回升阶段，约有一帧误差，未将 agent 会话时间换算为视频时间。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "其他", + "note_zh": "候选：击打；视频 3.32–3.44 s，机器人以头部短促顶击足球，使球离体回升，为第4次颠球；不是持续托举或压擦，现有接触方式无击打项。" + } + ] + }, + "valid_touch_5": { + "method": "visual", + "intervals": [ + [ + 3.96, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 3.96, + "log_step": null, + "text_zh": "观察：视频 3.80–3.96 s，足球在机器人头部附近由下落转为离体上升，为松手后第5个独立头球回升周期；双掌分开，机器人仍站立。" + }, + { + "t": 3.96, + "log_step": 104, + "text_zh": "观察：log step 104 是对已交付 trajectory.npz 的独立回放核验，裁判返回 stages 包含 touch 5，最终 touches=10、success=true、executed_rows=639；本节点的有效计数由该交付回放裁判结果确认，视频锚点取逐帧所见回升阶段，约有一帧误差，未将 agent 会话时间换算为视频时间。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "其他", + "note_zh": "候选：击打；视频 3.80–3.96 s，机器人以头部短促顶击足球，使球离体回升，为第5次颠球；不是持续托举或压擦，现有接触方式无击打项。" + } + ] + }, + "valid_touch_6": { + "method": "visual", + "intervals": [ + [ + 4.44, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 4.44, + "log_step": null, + "text_zh": "观察：视频 4.28–4.44 s，足球在机器人头部附近由下落转为离体上升，为松手后第6个独立头球回升周期；双掌分开，机器人仍站立。" + }, + { + "t": 4.44, + "log_step": 104, + "text_zh": "观察：log step 104 是对已交付 trajectory.npz 的独立回放核验，裁判返回 stages 包含 touch 6，最终 touches=10、success=true、executed_rows=639；本节点的有效计数由该交付回放裁判结果确认，视频锚点取逐帧所见回升阶段，约有一帧误差，未将 agent 会话时间换算为视频时间。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "其他", + "note_zh": "候选：击打；视频 4.28–4.44 s，机器人以头部短促顶击足球，使球离体回升，为第6次颠球；不是持续托举或压擦，现有接触方式无击打项。" + } + ] + }, + "valid_touch_7": { + "method": "visual", + "intervals": [ + [ + 4.92, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 4.92, + "log_step": null, + "text_zh": "观察：视频 4.80–4.92 s，足球在机器人头部附近由下落转为离体上升，为松手后第7个独立头球回升周期；双掌分开，机器人仍站立。" + }, + { + "t": 4.92, + "log_step": 104, + "text_zh": "观察：log step 104 是对已交付 trajectory.npz 的独立回放核验，裁判返回 stages 包含 touch 7，最终 touches=10、success=true、executed_rows=639；本节点的有效计数由该交付回放裁判结果确认，视频锚点取逐帧所见回升阶段，约有一帧误差，未将 agent 会话时间换算为视频时间。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "其他", + "note_zh": "候选：击打；视频 4.80–4.92 s，机器人以头部短促顶击足球，使球离体回升，为第7次颠球；不是持续托举或压擦，现有接触方式无击打项。" + } + ] + }, + "valid_touch_8": { + "method": "visual", + "intervals": [ + [ + 5.44, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 5.44, + "log_step": null, + "text_zh": "观察：视频 5.28–5.44 s，足球在机器人头部附近由下落转为离体上升，为松手后第8个独立头球回升周期；双掌分开，机器人仍站立。" + }, + { + "t": 5.44, + "log_step": 104, + "text_zh": "观察：log step 104 是对已交付 trajectory.npz 的独立回放核验，裁判返回 stages 包含 touch 8，最终 touches=10、success=true、executed_rows=639；本节点的有效计数由该交付回放裁判结果确认，视频锚点取逐帧所见回升阶段，约有一帧误差，未将 agent 会话时间换算为视频时间。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "其他", + "note_zh": "候选：击打；视频 5.28–5.44 s，机器人以头部短促顶击足球，使球离体回升，为第8次颠球；不是持续托举或压擦，现有接触方式无击打项。" + } + ] + }, + "valid_touch_9": { + "method": "visual", + "intervals": [ + [ + 5.92, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 5.92, + "log_step": null, + "text_zh": "观察：视频 5.76–5.92 s，足球在机器人头部附近由下落转为离体上升，为松手后第9个独立头球回升周期；双掌分开，机器人仍站立。" + }, + { + "t": 5.92, + "log_step": 104, + "text_zh": "观察：log step 104 是对已交付 trajectory.npz 的独立回放核验，裁判返回 stages 包含 touch 9，最终 touches=10、success=true、executed_rows=639；本节点的有效计数由该交付回放裁判结果确认，视频锚点取逐帧所见回升阶段，约有一帧误差，未将 agent 会话时间换算为视频时间。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "其他", + "note_zh": "候选：击打；视频 5.76–5.92 s，机器人以头部短促顶击足球，使球离体回升，为第9次颠球；不是持续托举或压擦，现有接触方式无击打项。" + } + ] + }, + "valid_touch_10": { + "method": "visual", + "intervals": [ + [ + 6.36, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 6.36, + "log_step": null, + "text_zh": "观察：视频 6.24–6.36 s，足球在机器人头部附近由下落转为离体上升，为松手后第10个独立头球回升周期；双掌分开，机器人仍站立。" + }, + { + "t": 6.36, + "log_step": 104, + "text_zh": "观察：log step 104 是对已交付 trajectory.npz 的独立回放核验，裁判返回 stages 包含 touch 10，最终 touches=10、success=true、executed_rows=639；本节点的有效计数由该交付回放裁判结果确认，视频锚点取逐帧所见回升阶段，约有一帧误差，未将 agent 会话时间换算为视频时间。" + }, + { + "t": 6.36, + "log_step": 104, + "text_zh": "推断：视频 6.24–6.36 s 可见第十次头球后足球继续上升，末帧位于视频 6.36 s；结合交付回放裁判最终十次均通过，将本节点定位在视频结尾约 6.4 s，以最近可见帧视频 6.36 s 作近似锚点。视频不单独分辨最后不足一帧内的验收阈值跨越。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "其他", + "note_zh": "候选：击打；视频 6.24–6.36 s，机器人以头部短促顶击足球，使球离体回升，为第10次颠球；不是持续托举或压擦，现有接触方式无击打项。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "touches", + "value": 10.0, + "unit": "count", + "threshold": 10, + "direction": "higher_better" + }, + { + "name": "ball_height_m", + "value": 1.631, + "unit": "m", + "threshold": null, + "direction": null + } + ], + "failure": null, + "summary_zh": "机器人将双掌夹持的足球举过头顶并松开，随后连续用头部顶球，十个颠球节点均完成；视频最后一秒仍可见连续回升，机器人保持站立、足球未触地。评分器 success=1、stages_reached=10，log step 104 对交付轨迹的回放也确认十次有效计数；log step 105 的过滤提示发生在成功交付核验之后，不改变结果。无状态日志，视频以每帧 0.04 s 核对接触及回升，节点时刻为近似锚点，毫秒级接触持续时间与速度比例依赖交付回放裁判判定，末次完成定位于视频结尾约 6.4 s。", + "conf": 9, + "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；视频 1.4–2.4 s 与 5.6–6.4 s 帧显示球举过头后双掌松开、连续头球回升、机器人站立、球未触地，与十个节点及终局值一致；log step 104 的交付回放（stages 含 touch 1–10）存在且未换算 agent 时间；无失败，failure=null 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": "2026-10-29T10:07:18Z", + "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/embodied-olympics-fun-juggling-i00-privileged--zjv9zmw/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-juggling-i00-privileged--zjv9zmw/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "63e0c98ab46eb2ab36c6bd5e5b9d3597668cc403a30939177c6dbbc672777ab9", + "video_sha256": "44e8c4f31adac97e99cb4d4d41db48bd53976a1c8286b356cbacf151e02c7042", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/juggling/trajectories/embodied-olympics-fun-juggling-i00-standard--nvkfgj7.json b/phase2/embodied-olympics-fun/juggling/trajectories/embodied-olympics-fun-juggling-i00-standard--nvkfgj7.json new file mode 100644 index 0000000000000000000000000000000000000000..bd64721a53ed75903dcb80c1db2553292d406c90 --- /dev/null +++ b/phase2/embodied-olympics-fun/juggling/trajectories/embodied-olympics-fun-juggling-i00-standard--nvkfgj7.json @@ -0,0 +1,313 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-juggling-i00-standard--nvkfgj7", + "taskKey": "embodied-olympics-fun/juggling", + "dag_sha256": "25e819f9f1251e0a7ce9add4f150f9227df00d011426943d8c29867bd2eff282", + "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": 107, + "sim_steps": 6126, + "calls": 831, + "rows": 4062, + "resets": 50, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 1689.8, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.391, + "peak_processes": 15, + "peak_disk_gb": 1.107, + "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/embodied-olympics-fun-juggling-i00-standard--nvkfgj7/replay.mp4", + "duration_s": 1.08, + "fps": 25.0, + "frames": 27, + "sha256": "ba687a96083cd68f7a9199140b220c551ee2a3d20e3aa466cfc58f4ce5d473e6" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 10, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "valid_touch_1": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.4, + "log_step": 134, + "text_zh": "观察：视频 0.00–0.76 s 中机器人双掌夹球，在抬右腿、屈支撑腿准备接球时逐渐侧倾；视频 0.80–1.04 s 中双臂分开、球向下移动，机器人继续向侧面倒下。逐帧检查全部 27 帧，未见合法部位击球后球离体回升一个球直径。" + }, + { + "t": 1.04, + "log_step": 136, + "text_zh": "观察：log step 134 的本回合控制器输出为 DONE failed fell stages 0；log step 136 的机器人服务返回 episode=51、control_steps=107、done=True、outcome=fell、stages_reached=0。log step 6 的服务 control_dt=0.01 s，107 个控制步对应仿真 1.07 s，与视频 1.04 s 末帧的倾倒及视频总长 1.08 s 相符；这些是最后评分回合的数据。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "其他", + "note_zh": "候选：双掌挤压夹持。视频 0.24–0.76 s 中，机器人左右平掌在首次颠球准备时共同从两侧夹持足球，随后双臂分开释放；平掌没有手指，任何一只手均未包住球，故不属抓握。此处记录首次尝试的实际持球准备，未观察到有效击球。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "视频 0.28–0.80 s 中，机器人抬起右腿准备以脚接球，左脚在草地上承担支撑，随后身体侧倾。此项是本体与地面的支撑接触，不表示脚已经击中足球。" + } + ] + }, + "valid_touch_2": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.04, + "log_step": 136, + "text_zh": "推断：视频 0.00–1.04 s 的全部帧未见首次有效击球回升，机器人在抬腿准备后侧倒；log step 136 的最后评分回合（episode=51）裁判累计计数 stages_reached=0。因此第2次颠球及其所需的前序计数均未达到，未进入本次重复动作。" + } + ] + }, + "valid_touch_3": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.04, + "log_step": 136, + "text_zh": "推断：视频 0.00–1.04 s 的全部帧未见首次有效击球回升，机器人在抬腿准备后侧倒；log step 136 的最后评分回合（episode=51）裁判累计计数 stages_reached=0。因此第3次颠球及其所需的前序计数均未达到，未进入本次重复动作。" + } + ] + }, + "valid_touch_4": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.04, + "log_step": 136, + "text_zh": "推断：视频 0.00–1.04 s 的全部帧未见首次有效击球回升，机器人在抬腿准备后侧倒；log step 136 的最后评分回合（episode=51）裁判累计计数 stages_reached=0。因此第4次颠球及其所需的前序计数均未达到，未进入本次重复动作。" + } + ] + }, + "valid_touch_5": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.04, + "log_step": 136, + "text_zh": "推断：视频 0.00–1.04 s 的全部帧未见首次有效击球回升，机器人在抬腿准备后侧倒；log step 136 的最后评分回合（episode=51）裁判累计计数 stages_reached=0。因此第5次颠球及其所需的前序计数均未达到，未进入本次重复动作。" + } + ] + }, + "valid_touch_6": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.04, + "log_step": 136, + "text_zh": "推断：视频 0.00–1.04 s 的全部帧未见首次有效击球回升，机器人在抬腿准备后侧倒；log step 136 的最后评分回合（episode=51）裁判累计计数 stages_reached=0。因此第6次颠球及其所需的前序计数均未达到，未进入本次重复动作。" + } + ] + }, + "valid_touch_7": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.04, + "log_step": 136, + "text_zh": "推断：视频 0.00–1.04 s 的全部帧未见首次有效击球回升，机器人在抬腿准备后侧倒；log step 136 的最后评分回合（episode=51）裁判累计计数 stages_reached=0。因此第7次颠球及其所需的前序计数均未达到，未进入本次重复动作。" + } + ] + }, + "valid_touch_8": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.04, + "log_step": 136, + "text_zh": "推断：视频 0.00–1.04 s 的全部帧未见首次有效击球回升，机器人在抬腿准备后侧倒；log step 136 的最后评分回合（episode=51）裁判累计计数 stages_reached=0。因此第8次颠球及其所需的前序计数均未达到，未进入本次重复动作。" + } + ] + }, + "valid_touch_9": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.04, + "log_step": 136, + "text_zh": "推断：视频 0.00–1.04 s 的全部帧未见首次有效击球回升，机器人在抬腿准备后侧倒；log step 136 的最后评分回合（episode=51）裁判累计计数 stages_reached=0。因此第9次颠球及其所需的前序计数均未达到，未进入本次重复动作。" + } + ] + }, + "valid_touch_10": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.04, + "log_step": 136, + "text_zh": "推断：视频 0.00–1.04 s 的全部帧未见首次有效击球回升，机器人在抬腿准备后侧倒；log step 136 的最后评分回合（episode=51）裁判累计计数 stages_reached=0。因此第10次颠球及其所需的前序计数均未达到，未进入本次重复动作。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "touches", + "value": 0, + "unit": "count", + "threshold": 10, + "direction": "higher_better" + }, + { + "name": "ball_height_m", + "value": null, + "unit": "m", + "threshold": null, + "direction": null + } + ], + "failure": { + "breakpoint": { + "ref": "valid_touch_1", + "t": 0.8, + "log_step": 134, + "where": "操作中", + "evidence_zh": "观察：视频 0.28–0.76 s 中机器人抬右腿并屈支撑腿，为第一次颠球建立姿势；视频约 0.80 s 起侧倾明显加剧，至视频 1.04 s 身体已向草地倒下，没有完成击球后的有效回升。log step 134、136 的最后评分回合均记录 fell、有效次数 0；失败发生在首次颠球的准备与释放过程中。" + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 0.84, + "log_step": 134, + "evidence_zh": "观察：视频 0.72–1.04 s 中，机器人在右腿抬起后持续向侧面倾倒，躯干越来越接近水平、身体高度下降，末段未恢复支撑。log step 134 记录姿态阶段仿真 0.75 s 横滚角 17.8°、结束时仿真 1.07 s 横滚角 62.6°，并输出裁判结果 fell；log step 136 的服务结果再次确认该回合因跌倒失败。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志 / 视频", + "t": 0.8, + "log_step": 134, + "evidence_zh": "推断：log step 126 的控制脚本将支撑腿设为屈膝姿态、右腿设为向前抬起的击球平台，姿态切换时长参数 t_ramp=0.6；log step 132、134 的实际运行版本还通过踝关节目标施加横滚反馈及右髋外展。执行时 log step 134 记录横滚角从姿态阶段仿真 0.75 s 的 17.8°增长到仿真 1.07 s 的 62.6°，视频 0.72–1.04 s 同样显示连续侧倒。因此该抬腿姿态切换中的控制未维持全身平衡，首次击球尚未完成便跌倒；现有证据不足以进一步区分承载布局、轨迹速度与反馈增益各自的作用。" + } + ], + "gap_zh": "没有状态日志，无法核定视频末段球与腿之间是否存在短暂擦碰及毫秒级接触时刻；但全片未见击球后的有效回升，且最后回合裁判计数为 0。根因可定位到抬腿阶段的平衡控制，不能验证某一个具体参数或接触布局是唯一原因。touches=0 取自 log step 134、136 的本回合裁判计数 stages_reached；评分器未报告 ball_height_m，log step 134 的球高是 agent 视觉脚本估计，未作为原生指标。", + "unmet": { + "valid_touch_1": false, + "valid_touch_2": false, + "valid_touch_3": false, + "valid_touch_4": false, + "valid_touch_5": false, + "valid_touch_6": false, + "valid_touch_7": false, + "valid_touch_8": false, + "valid_touch_9": false, + "valid_touch_10": false + } + }, + "summary_zh": "观察：评分采用最后的实时回合 episode=51，机器人抬腿准备首次颠球后侧倾倒下；视频 0.00–1.04 s 未见有效颠球回升，log step 134、136 的裁判结果为 fell、计数 0，十个节点均未达到。推断：断点为第1次颠球的操作中，抬腿姿态切换未能保持平衡，根因定位到控制；回合结束后的脚本修改未在本次评分视频中执行。无状态日志，末段瞬时擦碰及更细的控制原因无法确定。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "视频仅 1.08 s，机器人抬右腿后侧倒，全片无有效击球回升，十个节点均未达成、断点在第1次颠球操作中、平衡现象与控制根因均有 log 与视频支持；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": "2026-10-29T10:07:15Z", + "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/embodied-olympics-fun-juggling-i00-standard--nvkfgj7/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-juggling-i00-standard--nvkfgj7/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "e962931b42eca278c04f2123ae3c0ae587f9aff2013617ac5f778b6b7f74ae52", + "video_sha256": "ba687a96083cd68f7a9199140b220c551ee2a3d20e3aa466cfc58f4ce5d473e6", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/juggling_feet/task_dag.json b/phase2/embodied-olympics-fun/juggling_feet/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..d16ef76a60543c8bc6f091be193d58ab5021b82e --- /dev/null +++ b/phase2/embodied-olympics-fun/juggling_feet/task_dag.json @@ -0,0 +1,756 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "embodied-olympics-fun/juggling_feet", + "suite": "embodied-olympics-fun", + "base": "embodied-olympics-fun-juggling-feet-i00", + "title_zh": "足部颠球十次", + "summary_zh": "机器人从两掌间放落足球，在格内不跌倒且不犯规地以任意脚完成十次有效向上颠球，每次球心上升至少0.22米且球底达到离地0.36米，第十次计入时即成功。", + "embodiment_zh": "两种模式均为29关节宇树G1双足人形机器人，双臂末端是无可动手指的平掌垫及拳形外壳，没有二指夹爪、灵巧手、吸盘或安装工具，抓握指令不起作用；初态以双掌挤压及摩擦夹持足球，由腿脚完成击球和地面支撑，无轮式底盘。", + "task_class": "compositional", + "analysis_level": "step", + "spec": { + "version": "1.2", + "task": "embodied-olympics-fun-juggling-feet-i00", + "bench": "embodied-olympics-fun", + "progress": "live", + "entities": {}, + "milestones": [ + { + "id": "juggles_counted_1", + "zh": "第1次颠球", + "mode": "event", + "when": null, + "after": [], + "hold_s": 0, + "goal": true, + "criterion_zh": "固定场边视角可见黑白足球从任一脚短促击打后离脚上升，球心较本次触球开始上升至少0.22米且球底达到离地0.36米，累计第1次有效颠球即完成；尚在下落、贴脚或上升不足均未完成。同一触球按间隔不足0.05秒合并且接触历时不超过0.05秒，球在再次新触球前须达到高度条件并离开合法脚部接触，按评分器能量折算到触球开始高度的离球速度须至少为来球速度的0.95倍；低于速度门槛的反弹不计数也不单独判犯规，已计入的次数保留。", + "step": { + "cat": "动态", + "item": "击打" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "其他", + "plan": 1, + "note_zh": "任一脚的脚背、鞋底、侧面、脚尖或脚跟击打足球；候选接触名为“击打”，现有接触方式没有击打项，不能以推/扫或托/舀代替。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "plan": 1, + "note_zh": "支撑脚承重接触草地，使机器人在另一脚击打足球时保持站立；允许同一脚连续颠球，也允许换脚。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:4", + "inputs/task/standard/instruction.md:5", + "inputs/task/privileged/environment/rfo_task.py:155-162", + "inputs/task/privileged/environment/rfo_task.py:270-313", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:304-347", + "inputs/task/privileged/environment/instance.json:4-32", + "inputs/oracle/replay.mp4 @ 0.00-9.24 s", + "inputs/oracle/run.json: reward.stages_reached=10, reward.metric_touches=10" + ], + "conf": 6, + "note_zh": "按S5逐次计数，边仅表示裁判计数顺序；一次触球及其离脚上升合成一个完整步骤，不按接触、离脚、升高拆分。初始足球由双掌挤压，放开两掌及球下落是首次颠球的途经动作，不另建节点；颠球步骤无须手部接触。小腿碰撞体仅在脚坐标系前后位置介于踝后0.08米与踝前0.15米、且高于踝不超过0.12米的接触点算脚，其他小腿触球违规。录像没有逐帧裁判叠字，25帧每秒不能独立解析毫秒接触或精确速度，需补记裁判计数；不把成功样例使用左脚的方式限制为唯一方式。" + }, + { + "id": "juggles_counted_2", + "zh": "第2次颠球", + "mode": "event", + "when": null, + "after": [ + "juggles_counted_1" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "固定场边视角可见黑白足球从任一脚短促击打后离脚上升，球心较本次触球开始上升至少0.22米且球底达到离地0.36米，累计第2次有效颠球即完成；尚在下落、贴脚或上升不足均未完成。同一触球按间隔不足0.05秒合并且接触历时不超过0.05秒，球在再次新触球前须达到高度条件并离开合法脚部接触，按评分器能量折算到触球开始高度的离球速度须至少为来球速度的0.95倍；低于速度门槛的反弹不计数也不单独判犯规，已计入的次数保留。", + "step": { + "cat": "动态", + "item": "击打" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "其他", + "plan": 1, + "note_zh": "任一脚的脚背、鞋底、侧面、脚尖或脚跟击打足球；候选接触名为“击打”，现有接触方式没有击打项，不能以推/扫或托/舀代替。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "plan": 1, + "note_zh": "支撑脚承重接触草地，使机器人在另一脚击打足球时保持站立；允许同一脚连续颠球，也允许换脚。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:4", + "inputs/task/standard/instruction.md:5", + "inputs/task/privileged/environment/rfo_task.py:155-162", + "inputs/task/privileged/environment/rfo_task.py:270-313", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:304-347", + "inputs/task/privileged/environment/instance.json:4-32", + "inputs/oracle/replay.mp4 @ 0.00-9.24 s", + "inputs/oracle/run.json: reward.stages_reached=10, reward.metric_touches=10" + ], + "conf": 6, + "note_zh": "按S5逐次计数，边仅表示裁判计数顺序；一次触球及其离脚上升合成一个完整步骤，不按接触、离脚、升高拆分。初始足球由双掌挤压，放开两掌及球下落是首次颠球的途经动作，不另建节点；颠球步骤无须手部接触。小腿碰撞体仅在脚坐标系前后位置介于踝后0.08米与踝前0.15米、且高于踝不超过0.12米的接触点算脚，其他小腿触球违规。录像没有逐帧裁判叠字，25帧每秒不能独立解析毫秒接触或精确速度，需补记裁判计数；不把成功样例使用左脚的方式限制为唯一方式。" + }, + { + "id": "juggles_counted_3", + "zh": "第3次颠球", + "mode": "event", + "when": null, + "after": [ + "juggles_counted_2" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "固定场边视角可见黑白足球从任一脚短促击打后离脚上升，球心较本次触球开始上升至少0.22米且球底达到离地0.36米，累计第3次有效颠球即完成；尚在下落、贴脚或上升不足均未完成。同一触球按间隔不足0.05秒合并且接触历时不超过0.05秒，球在再次新触球前须达到高度条件并离开合法脚部接触，按评分器能量折算到触球开始高度的离球速度须至少为来球速度的0.95倍；低于速度门槛的反弹不计数也不单独判犯规，已计入的次数保留。", + "step": { + "cat": "动态", + "item": "击打" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "其他", + "plan": 1, + "note_zh": "任一脚的脚背、鞋底、侧面、脚尖或脚跟击打足球；候选接触名为“击打”，现有接触方式没有击打项，不能以推/扫或托/舀代替。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "plan": 1, + "note_zh": "支撑脚承重接触草地，使机器人在另一脚击打足球时保持站立；允许同一脚连续颠球，也允许换脚。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:4", + "inputs/task/standard/instruction.md:5", + "inputs/task/privileged/environment/rfo_task.py:155-162", + "inputs/task/privileged/environment/rfo_task.py:270-313", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:304-347", + "inputs/task/privileged/environment/instance.json:4-32", + "inputs/oracle/replay.mp4 @ 0.00-9.24 s", + "inputs/oracle/run.json: reward.stages_reached=10, reward.metric_touches=10" + ], + "conf": 6, + "note_zh": "按S5逐次计数，边仅表示裁判计数顺序；一次触球及其离脚上升合成一个完整步骤，不按接触、离脚、升高拆分。初始足球由双掌挤压，放开两掌及球下落是首次颠球的途经动作，不另建节点；颠球步骤无须手部接触。小腿碰撞体仅在脚坐标系前后位置介于踝后0.08米与踝前0.15米、且高于踝不超过0.12米的接触点算脚，其他小腿触球违规。录像没有逐帧裁判叠字，25帧每秒不能独立解析毫秒接触或精确速度，需补记裁判计数；不把成功样例使用左脚的方式限制为唯一方式。" + }, + { + "id": "juggles_counted_4", + "zh": "第4次颠球", + "mode": "event", + "when": null, + "after": [ + "juggles_counted_3" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "固定场边视角可见黑白足球从任一脚短促击打后离脚上升，球心较本次触球开始上升至少0.22米且球底达到离地0.36米，累计第4次有效颠球即完成；尚在下落、贴脚或上升不足均未完成。同一触球按间隔不足0.05秒合并且接触历时不超过0.05秒，球在再次新触球前须达到高度条件并离开合法脚部接触，按评分器能量折算到触球开始高度的离球速度须至少为来球速度的0.95倍；低于速度门槛的反弹不计数也不单独判犯规，已计入的次数保留。", + "step": { + "cat": "动态", + "item": "击打" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "其他", + "plan": 1, + "note_zh": "任一脚的脚背、鞋底、侧面、脚尖或脚跟击打足球；候选接触名为“击打”，现有接触方式没有击打项，不能以推/扫或托/舀代替。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "plan": 1, + "note_zh": "支撑脚承重接触草地，使机器人在另一脚击打足球时保持站立；允许同一脚连续颠球，也允许换脚。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:4", + "inputs/task/standard/instruction.md:5", + "inputs/task/privileged/environment/rfo_task.py:155-162", + "inputs/task/privileged/environment/rfo_task.py:270-313", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:304-347", + "inputs/task/privileged/environment/instance.json:4-32", + "inputs/oracle/replay.mp4 @ 0.00-9.24 s", + "inputs/oracle/run.json: reward.stages_reached=10, reward.metric_touches=10" + ], + "conf": 6, + "note_zh": "按S5逐次计数，边仅表示裁判计数顺序；一次触球及其离脚上升合成一个完整步骤，不按接触、离脚、升高拆分。初始足球由双掌挤压，放开两掌及球下落是首次颠球的途经动作，不另建节点；颠球步骤无须手部接触。小腿碰撞体仅在脚坐标系前后位置介于踝后0.08米与踝前0.15米、且高于踝不超过0.12米的接触点算脚，其他小腿触球违规。录像没有逐帧裁判叠字，25帧每秒不能独立解析毫秒接触或精确速度，需补记裁判计数；不把成功样例使用左脚的方式限制为唯一方式。" + }, + { + "id": "juggles_counted_5", + "zh": "第5次颠球", + "mode": "event", + "when": null, + "after": [ + "juggles_counted_4" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "固定场边视角可见黑白足球从任一脚短促击打后离脚上升，球心较本次触球开始上升至少0.22米且球底达到离地0.36米，累计第5次有效颠球即完成；尚在下落、贴脚或上升不足均未完成。同一触球按间隔不足0.05秒合并且接触历时不超过0.05秒，球在再次新触球前须达到高度条件并离开合法脚部接触，按评分器能量折算到触球开始高度的离球速度须至少为来球速度的0.95倍；低于速度门槛的反弹不计数也不单独判犯规，已计入的次数保留。", + "step": { + "cat": "动态", + "item": "击打" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "其他", + "plan": 1, + "note_zh": "任一脚的脚背、鞋底、侧面、脚尖或脚跟击打足球；候选接触名为“击打”，现有接触方式没有击打项，不能以推/扫或托/舀代替。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "plan": 1, + "note_zh": "支撑脚承重接触草地，使机器人在另一脚击打足球时保持站立；允许同一脚连续颠球，也允许换脚。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:4", + "inputs/task/standard/instruction.md:5", + "inputs/task/privileged/environment/rfo_task.py:155-162", + "inputs/task/privileged/environment/rfo_task.py:270-313", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:304-347", + "inputs/task/privileged/environment/instance.json:4-32", + "inputs/oracle/replay.mp4 @ 0.00-9.24 s", + "inputs/oracle/run.json: reward.stages_reached=10, reward.metric_touches=10" + ], + "conf": 6, + "note_zh": "按S5逐次计数，边仅表示裁判计数顺序；一次触球及其离脚上升合成一个完整步骤，不按接触、离脚、升高拆分。初始足球由双掌挤压，放开两掌及球下落是首次颠球的途经动作，不另建节点；颠球步骤无须手部接触。小腿碰撞体仅在脚坐标系前后位置介于踝后0.08米与踝前0.15米、且高于踝不超过0.12米的接触点算脚，其他小腿触球违规。录像没有逐帧裁判叠字，25帧每秒不能独立解析毫秒接触或精确速度，需补记裁判计数；不把成功样例使用左脚的方式限制为唯一方式。" + }, + { + "id": "juggles_counted_6", + "zh": "第6次颠球", + "mode": "event", + "when": null, + "after": [ + "juggles_counted_5" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "固定场边视角可见黑白足球从任一脚短促击打后离脚上升，球心较本次触球开始上升至少0.22米且球底达到离地0.36米，累计第6次有效颠球即完成；尚在下落、贴脚或上升不足均未完成。同一触球按间隔不足0.05秒合并且接触历时不超过0.05秒，球在再次新触球前须达到高度条件并离开合法脚部接触，按评分器能量折算到触球开始高度的离球速度须至少为来球速度的0.95倍；低于速度门槛的反弹不计数也不单独判犯规，已计入的次数保留。", + "step": { + "cat": "动态", + "item": "击打" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "其他", + "plan": 1, + "note_zh": "任一脚的脚背、鞋底、侧面、脚尖或脚跟击打足球；候选接触名为“击打”，现有接触方式没有击打项，不能以推/扫或托/舀代替。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "plan": 1, + "note_zh": "支撑脚承重接触草地，使机器人在另一脚击打足球时保持站立；允许同一脚连续颠球，也允许换脚。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:4", + "inputs/task/standard/instruction.md:5", + "inputs/task/privileged/environment/rfo_task.py:155-162", + "inputs/task/privileged/environment/rfo_task.py:270-313", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:304-347", + "inputs/task/privileged/environment/instance.json:4-32", + "inputs/oracle/replay.mp4 @ 0.00-9.24 s", + "inputs/oracle/run.json: reward.stages_reached=10, reward.metric_touches=10" + ], + "conf": 6, + "note_zh": "按S5逐次计数，边仅表示裁判计数顺序；一次触球及其离脚上升合成一个完整步骤，不按接触、离脚、升高拆分。初始足球由双掌挤压，放开两掌及球下落是首次颠球的途经动作，不另建节点；颠球步骤无须手部接触。小腿碰撞体仅在脚坐标系前后位置介于踝后0.08米与踝前0.15米、且高于踝不超过0.12米的接触点算脚，其他小腿触球违规。录像没有逐帧裁判叠字，25帧每秒不能独立解析毫秒接触或精确速度，需补记裁判计数；不把成功样例使用左脚的方式限制为唯一方式。" + }, + { + "id": "juggles_counted_7", + "zh": "第7次颠球", + "mode": "event", + "when": null, + "after": [ + "juggles_counted_6" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "固定场边视角可见黑白足球从任一脚短促击打后离脚上升，球心较本次触球开始上升至少0.22米且球底达到离地0.36米，累计第7次有效颠球即完成；尚在下落、贴脚或上升不足均未完成。同一触球按间隔不足0.05秒合并且接触历时不超过0.05秒，球在再次新触球前须达到高度条件并离开合法脚部接触，按评分器能量折算到触球开始高度的离球速度须至少为来球速度的0.95倍；低于速度门槛的反弹不计数也不单独判犯规，已计入的次数保留。", + "step": { + "cat": "动态", + "item": "击打" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "其他", + "plan": 1, + "note_zh": "任一脚的脚背、鞋底、侧面、脚尖或脚跟击打足球；候选接触名为“击打”，现有接触方式没有击打项，不能以推/扫或托/舀代替。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "plan": 1, + "note_zh": "支撑脚承重接触草地，使机器人在另一脚击打足球时保持站立；允许同一脚连续颠球，也允许换脚。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:4", + "inputs/task/standard/instruction.md:5", + "inputs/task/privileged/environment/rfo_task.py:155-162", + "inputs/task/privileged/environment/rfo_task.py:270-313", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:304-347", + "inputs/task/privileged/environment/instance.json:4-32", + "inputs/oracle/replay.mp4 @ 0.00-9.24 s", + "inputs/oracle/run.json: reward.stages_reached=10, reward.metric_touches=10" + ], + "conf": 6, + "note_zh": "按S5逐次计数，边仅表示裁判计数顺序；一次触球及其离脚上升合成一个完整步骤，不按接触、离脚、升高拆分。初始足球由双掌挤压，放开两掌及球下落是首次颠球的途经动作，不另建节点；颠球步骤无须手部接触。小腿碰撞体仅在脚坐标系前后位置介于踝后0.08米与踝前0.15米、且高于踝不超过0.12米的接触点算脚，其他小腿触球违规。录像没有逐帧裁判叠字，25帧每秒不能独立解析毫秒接触或精确速度，需补记裁判计数；不把成功样例使用左脚的方式限制为唯一方式。" + }, + { + "id": "juggles_counted_8", + "zh": "第8次颠球", + "mode": "event", + "when": null, + "after": [ + "juggles_counted_7" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "固定场边视角可见黑白足球从任一脚短促击打后离脚上升，球心较本次触球开始上升至少0.22米且球底达到离地0.36米，累计第8次有效颠球即完成；尚在下落、贴脚或上升不足均未完成。同一触球按间隔不足0.05秒合并且接触历时不超过0.05秒，球在再次新触球前须达到高度条件并离开合法脚部接触，按评分器能量折算到触球开始高度的离球速度须至少为来球速度的0.95倍；低于速度门槛的反弹不计数也不单独判犯规，已计入的次数保留。", + "step": { + "cat": "动态", + "item": "击打" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "其他", + "plan": 1, + "note_zh": "任一脚的脚背、鞋底、侧面、脚尖或脚跟击打足球；候选接触名为“击打”，现有接触方式没有击打项，不能以推/扫或托/舀代替。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "plan": 1, + "note_zh": "支撑脚承重接触草地，使机器人在另一脚击打足球时保持站立；允许同一脚连续颠球，也允许换脚。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:4", + "inputs/task/standard/instruction.md:5", + "inputs/task/privileged/environment/rfo_task.py:155-162", + "inputs/task/privileged/environment/rfo_task.py:270-313", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:304-347", + "inputs/task/privileged/environment/instance.json:4-32", + "inputs/oracle/replay.mp4 @ 0.00-9.24 s", + "inputs/oracle/run.json: reward.stages_reached=10, reward.metric_touches=10" + ], + "conf": 6, + "note_zh": "按S5逐次计数，边仅表示裁判计数顺序；一次触球及其离脚上升合成一个完整步骤，不按接触、离脚、升高拆分。初始足球由双掌挤压，放开两掌及球下落是首次颠球的途经动作，不另建节点；颠球步骤无须手部接触。小腿碰撞体仅在脚坐标系前后位置介于踝后0.08米与踝前0.15米、且高于踝不超过0.12米的接触点算脚，其他小腿触球违规。录像没有逐帧裁判叠字，25帧每秒不能独立解析毫秒接触或精确速度，需补记裁判计数；不把成功样例使用左脚的方式限制为唯一方式。" + }, + { + "id": "juggles_counted_9", + "zh": "第9次颠球", + "mode": "event", + "when": null, + "after": [ + "juggles_counted_8" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "固定场边视角可见黑白足球从任一脚短促击打后离脚上升，球心较本次触球开始上升至少0.22米且球底达到离地0.36米，累计第9次有效颠球即完成；尚在下落、贴脚或上升不足均未完成。同一触球按间隔不足0.05秒合并且接触历时不超过0.05秒，球在再次新触球前须达到高度条件并离开合法脚部接触，按评分器能量折算到触球开始高度的离球速度须至少为来球速度的0.95倍；低于速度门槛的反弹不计数也不单独判犯规，已计入的次数保留。", + "step": { + "cat": "动态", + "item": "击打" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "其他", + "plan": 1, + "note_zh": "任一脚的脚背、鞋底、侧面、脚尖或脚跟击打足球；候选接触名为“击打”，现有接触方式没有击打项，不能以推/扫或托/舀代替。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "plan": 1, + "note_zh": "支撑脚承重接触草地，使机器人在另一脚击打足球时保持站立；允许同一脚连续颠球，也允许换脚。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:4", + "inputs/task/standard/instruction.md:5", + "inputs/task/privileged/environment/rfo_task.py:155-162", + "inputs/task/privileged/environment/rfo_task.py:270-313", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:304-347", + "inputs/task/privileged/environment/instance.json:4-32", + "inputs/oracle/replay.mp4 @ 0.00-9.24 s", + "inputs/oracle/run.json: reward.stages_reached=10, reward.metric_touches=10" + ], + "conf": 6, + "note_zh": "按S5逐次计数，边仅表示裁判计数顺序；一次触球及其离脚上升合成一个完整步骤，不按接触、离脚、升高拆分。初始足球由双掌挤压，放开两掌及球下落是首次颠球的途经动作，不另建节点；颠球步骤无须手部接触。小腿碰撞体仅在脚坐标系前后位置介于踝后0.08米与踝前0.15米、且高于踝不超过0.12米的接触点算脚，其他小腿触球违规。录像没有逐帧裁判叠字，25帧每秒不能独立解析毫秒接触或精确速度，需补记裁判计数；不把成功样例使用左脚的方式限制为唯一方式。" + }, + { + "id": "juggles_counted_10", + "zh": "第10次颠球", + "mode": "event", + "when": null, + "after": [ + "juggles_counted_9" + ], + "hold_s": 0, + "goal": true, + "criterion_zh": "固定场边视角可见黑白足球从任一脚短促击打后离脚上升，球心较本次触球开始上升至少0.22米且球底达到离地0.36米，累计第10次有效颠球即完成；尚在下落、贴脚或上升不足均未完成。同一触球按间隔不足0.05秒合并且接触历时不超过0.05秒，球在再次新触球前须达到高度条件并离开合法脚部接触，按评分器能量折算到触球开始高度的离球速度须至少为来球速度的0.95倍；低于速度门槛的反弹不计数也不单独判犯规，已计入的次数保留。", + "step": { + "cat": "动态", + "item": "击打" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "其他", + "plan": 1, + "note_zh": "任一脚的脚背、鞋底、侧面、脚尖或脚跟击打足球；候选接触名为“击打”，现有接触方式没有击打项，不能以推/扫或托/舀代替。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "plan": 1, + "note_zh": "支撑脚承重接触草地，使机器人在另一脚击打足球时保持站立；允许同一脚连续颠球，也允许换脚。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:4", + "inputs/task/standard/instruction.md:5", + "inputs/task/privileged/environment/rfo_task.py:155-162", + "inputs/task/privileged/environment/rfo_task.py:270-313", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:304-347", + "inputs/task/privileged/environment/instance.json:4-32", + "inputs/oracle/replay.mp4 @ 0.00-9.24 s", + "inputs/oracle/run.json: reward.stages_reached=10, reward.metric_touches=10" + ], + "conf": 6, + "note_zh": "按S5逐次计数，边仅表示裁判计数顺序；一次触球及其离脚上升合成一个完整步骤，不按接触、离脚、升高拆分。初始足球由双掌挤压，放开两掌及球下落是首次颠球的途经动作，不另建节点；颠球步骤无须手部接触。小腿碰撞体仅在脚坐标系前后位置介于踝后0.08米与踝前0.15米、且高于踝不超过0.12米的接触点算脚，其他小腿触球违规。录像没有逐帧裁判叠字，25帧每秒不能独立解析毫秒接触或精确速度，需补记裁判计数；不把成功样例使用左脚的方式限制为唯一方式。" + } + ], + "constraints": [ + { + "id": "ball_ground_contact", + "zh": "足球触地", + "never": null, + "criterion_zh": "从初态到成功判定前，足球任何部位接触草地或评分器定义的地面即违规；包括放球之前，不能落地后继续累计。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:249-264" + ] + }, + { + "id": "illegal_predrop_contact", + "zh": "放球前违规触球", + "never": null, + "criterion_zh": "裁判确认放球前，足球只可接触双手掌垫及归为手部的腕部；若贴靠手臂、躯干或腿部则违规。放球须连续0.03秒不接触机器人，放球时刻回溯至最后一次手部接触结束。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:18-20", + "inputs/task/privileged/environment/rfo_task.py:249-261", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:25-43" + ] + }, + { + "id": "handball", + "zh": "手臂违规触球", + "never": null, + "criterion_zh": "放球后超过0.15秒，手或手臂再碰足球即违规；即使仍在宽限期，脚触球簇仍有效或本次球尚待高度和速度判定时，也不得同时碰手或手臂。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:265-266", + "inputs/task/privileged/environment/rfo_task.py:293-295", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:325-347" + ] + }, + { + "id": "nonfoot_ball_contact", + "zh": "非脚部触球", + "never": null, + "criterion_zh": "放球后足球触及头、躯干、大腿等非脚部即违规；小腿碰撞点只有在脚局部坐标纵向为踝后0.08米至踝前0.15米、竖向不高于踝0.12米时算合法脚部，越过任一边界即违规。手臂触球另按宽限期规则判断。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:229-239", + "inputs/task/privileged/environment/rfo_task.py:267-280" + ] + }, + { + "id": "ball_trapped_between_feet", + "zh": "双脚夹球", + "never": null, + "criterion_zh": "放球后两只脚同时接触足球即违规；合法脚背区域内的小腿碰撞体接触也纳入相应一侧脚的判断。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:270-284" + ] + }, + { + "id": "ball_carried_on_foot", + "zh": "脚部托球过久", + "never": null, + "criterion_zh": "一次触球簇从首次接触到最后接触的历时超过0.05秒即违规，表现为足球持续贴脚、滚在脚上或停在脚上；间隔不足0.05秒的接触合为同一触球簇。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:296-297", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:304-347" + ] + }, + { + "id": "premature_retouch", + "zh": "未判定便再触球", + "never": null, + "criterion_zh": "上一触球仍待判定时又开始新的脚部触球簇即违规；球尚未达到规定高度就再踢属于此情况。间隔不足0.05秒的接触仍属于原触球簇，不另加一次计数。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:285-308", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:325-342" + ] + }, + { + "id": "foot_outside_grid", + "zh": "脚部出界", + "never": null, + "criterion_zh": "任一脚的评分器足部标记点在水平任一轴的绝对坐标大于2.5米，即离开以初始机器人位置为中心、边长5米的格子而违规；脚抬起时也检查。场边视角可用草地白线辨认边界，但精确判定采用足部标记点。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:217-224", + "inputs/task/privileged/environment/rfo_task.py:315-321", + "inputs/task/privileged/environment/instance.json:13" + ] + }, + { + "id": "robot_fell", + "zh": "机器人跌倒", + "never": null, + "criterion_zh": "骨盆高度低于离地0.35米，或除双脚以外任一机器人部位触地即跌倒；单脚站立不违规，无额外躯干角度或静止时长要求。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:158-159", + "inputs/task/privileged/environment/rfo_task.py:215", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:276-282" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "juggles_counted_ge_1", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:309-313; log Task.trackers.count and Referee.stages; count >= 1" + }, + { + "atom": "juggles_counted_ge_2", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:309-313; log Task.trackers.count and Referee.stages; count >= 2" + }, + { + "atom": "juggles_counted_ge_3", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:309-313; log Task.trackers.count and Referee.stages; count >= 3" + }, + { + "atom": "juggles_counted_ge_4", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:309-313; log Task.trackers.count and Referee.stages; count >= 4" + }, + { + "atom": "juggles_counted_ge_5", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:309-313; log Task.trackers.count and Referee.stages; count >= 5" + }, + { + "atom": "juggles_counted_ge_6", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:309-313; log Task.trackers.count and Referee.stages; count >= 6" + }, + { + "atom": "juggles_counted_ge_7", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:309-313; log Task.trackers.count and Referee.stages; count >= 7" + }, + { + "atom": "juggles_counted_ge_8", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:309-313; log Task.trackers.count and Referee.stages; count >= 8" + }, + { + "atom": "juggles_counted_ge_9", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:309-313; log Task.trackers.count and Referee.stages; count >= 9" + }, + { + "atom": "juggles_counted_ge_10", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:309-313; log Task.trackers.count and Referee.stages; count >= 10" + }, + { + "atom": "ball_ground_contact", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:250-251,263-264; log ball-ground contact at each physics step or matching Referee.failed" + }, + { + "atom": "illegal_predrop_contact", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:249-261; log dropped, last_on, ball-contact parts and Referee.failed" + }, + { + "atom": "handball", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:265-266,293-295; log Referee.failed, dropped, touch.active, pending and ball-arm contacts" + }, + { + "atom": "nonfoot_ball_contact", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:267-280; log ball-contact parts and shin contact points in foot frame, or matching Referee.failed" + }, + { + "atom": "ball_trapped_between_feet", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:270-284; log per-side ball-foot contacts or matching Referee.failed" + }, + { + "atom": "ball_carried_on_foot", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:296-297; log Touch.active, t_start, t_end and duration at physics rate, or matching Referee.failed" + }, + { + "atom": "premature_retouch", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:287-290; log pending and new Touch.update event, or matching Referee.failed" + }, + { + "atom": "foot_outside_grid", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:315-321; log both robot.foot_site world positions or matching Referee.failed" + }, + { + "atom": "robot_fell", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:276-282; log pelvis height, ground-contact parts or Referee.failed=fell" + }, + { + "atom": "all_ten_juggles_without_fault", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:309-313; inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:174-178; log per-frame Referee.success and Referee.failed (native checker world-state record)" + } + ] + }, + "native_metrics": [ + { + "name": "touches", + "unit": "count", + "threshold": 10, + "direction": "higher_better" + }, + { + "name": "ball_height_m", + "unit": "m", + "threshold": null, + "direction": "higher_better" + } + ], + "open_questions_zh": [ + "inputs/states/index.json没有状态日志，按P8所有节点when、约束never和success均为null；atoms仅说明需要补记的裁判原生计数及接触数据，不代表已验证日志字段。成功录像为25帧每秒，不能精确核验0.03/0.05秒接触和0.95倍速度门槛，故置信度为6。", + "两模式任务模块逐字一致、实例规则参数一致，物理成功核心相同；特权评分器要求两次独立回放均成功且末态哈希一致（inputs/grader/efe_embodied-olympics-fun-files/grade.py:105-115），标准评分器采用最后一次实时回合成功记录（inputs/grader/efe_standard-files/v2_1/grade.py:93-99）。确定性、文件有效性和交付完整性不作为语义节点或世界状态原子。", + "文本与评分实现的边界：文本称球在本次触球判定前不能再次被碰；实现只在Touch.update返回“新触球簇”且pending为真时判再次触球，间隔不足0.05秒的再次接触仍合为原触球簇（任务模块:287-290；referee.py:325-342）。节点验收按原生计数，不能把每个视频接触都算作一次。", + "文本与评分实现的边界：宽限期内文本禁止手臂在脚触球时或触球待判定时触球，实现的touch.active在离脚后还保持至触球间隔达到0.05秒，即使已计数也可能继续禁止手臂触球（任务模块:293-295；referee.py:339-340）；约束按实现保留此触球簇窗口。", + "球底“高于膝盖”的文字在代码中以大于等于0.36米验收；离球速度以触球开始高度为基准，由当前速度和高度差按平方根能量式折算，并非直接读取分离瞬间速度（任务模块:300-305）。未以录像估计或调整阈值。", + "球高指标是当前球心高度，随每次颠球上下变化，没有单一终态门槛；不以不断增高代替有效计数，也不把被动反弹次数设为禁止条件。60秒为回合时限，按P21不建约束。" + ], + "conf": 6, + "conf_reason_zh": "十个计数步骤直接来自两模式文本，顺序和全部阈值来自相同任务模块及冻结实例；已查看初态、整段成功录像和运行记录。没有任何状态日志，且录像无法精确分辨毫秒接触与速度验收，故仅能给出视频判据和待补记字段，未宣称状态证据通过。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "文本只有\"放球后十次脚触球\"：十个第k次颠球计数步骤（链式after、event、when:null）与评分器stages touch1..10一致；放球属于第1次触球的前置动作（S8），其违规由放球前约束覆盖，不另建节点。成功条件仅为第10次触球，约束（落地、放球前违规、手臂触球、非脚部、夹球/托球、提前再触、出格、跌倒）均为评分器失败条件且立即终止，无状态日志故success为null合理。备注：放球一步若团队倾向单列为节点属可接受的另一读法，不要求修改。", + "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：['juggles_counted_1', 'juggles_counted_2', 'juggles_counted_3', 'juggles_counted_4', 'juggles_counted_5', 'juggles_counted_6', 'juggles_counted_7', 'juggles_counted_8', 'juggles_counted_9', 'juggles_counted_10']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:46:06Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__juggling_feet/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__juggling_feet/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "3754cfe829b3b1eac01a5cbd0b3ba94d02b09cc0477561e8b5a3a97ed6e7875c", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/embodied-olympics-fun/juggling_feet/trajectories/embodied-olympics-fun-juggling-feet-i00-privileged--snul9zt.json b/phase2/embodied-olympics-fun/juggling_feet/trajectories/embodied-olympics-fun-juggling-feet-i00-privileged--snul9zt.json new file mode 100644 index 0000000000000000000000000000000000000000..7d837f50cc425d3505ec9490713664510cfdab23 --- /dev/null +++ b/phase2/embodied-olympics-fun/juggling_feet/trajectories/embodied-olympics-fun-juggling-feet-i00-privileged--snul9zt.json @@ -0,0 +1,488 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-juggling-feet-i00-privileged--snul9zt", + "taskKey": "embodied-olympics-fun/juggling_feet", + "dag_sha256": "6fb5f868dd0ec1e7ac4f8d9035cd8456789adc45e6c5ceda5fec1ad9d3ba10f6", + "outcome": { + "success": 1, + "reward": 1.0, + "source": "catalog", + "raw": { + "success": 1, + "replay_success": 1, + "deterministic": 1, + "n_actions": 771, + "sim_steps": 771, + "horizon": 6000, + "episode_over": 1, + "progress": 1.0, + "stages_reached": 10, + "stages_total": 10, + "fell": 0, + "replay_s": 6.66, + "bad_trajectory": 0, + "metric_ball_height_m": 0.676, + "metric_bounces": 0.0, + "metric_touches": 10.0, + "grader_error": 0, + "video_rendered": 1, + "agent_sessions": 65, + "agent_resets": 46, + "agent_replays": 12, + "agent_cost_self_reported": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 7.408, + "peak_processes": 13, + "peak_disk_gb": 0.878, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0, + "session_log_tampered": 0, + "session_log_from_monitor": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-juggling-feet-i00-privileged--snul9zt/replay.mp4", + "duration_s": 7.72, + "fps": 25.0, + "frames": 193, + "sha256": "f25cb7071da058f968edfab8247b352da362a22a40fc80f721e5aa6281022a1d" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 10, + "base": 0, + "peak": 10, + "final": 10, + "nodes": { + "juggles_counted_1": { + "method": "visual", + "intervals": [ + [ + 1.84, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 1.84, + "log_step": null, + "text_zh": "观察：视频 1.68–1.76 s 足球落到机器人抬起的脚上方，脚向上迎球；视频 1.84–2.08 s 足球离脚上升，形成从放球后顺序观察到的第1次颠球，另一脚仍在草地上支撑。完成锚点约为视频 1.84 s，画面没有逐次裁判叠字。" + }, + { + "t": 1.84, + "log_step": null, + "text_zh": "推断：结合整段录像中依次出现的颠球过程与inputs/run.json评分器字段stages_reached=10、metric_touches=10、success=1及deterministic=1，第1次为有效计数并保持已完成；25 fps画面不能单独精测接触历时、上升距离和速度比门槛。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "其他", + "note_zh": "观察：视频 0.00–1.40 s 机器人两掌相对夹持足球并调整站姿，随后张开两臂放球，属于首次颠球的途经动作。候选接触名为“双掌挤压”；两只平掌从两侧对压，没有任何单手包住球，故不用抓握。" + }, + { + "who": "腿/躯干", + "how": "其他", + "note_zh": "观察：机器人抬起的脚在第1次颠球中短促向上击打足球，球随后离脚；该脚为左脚的判断结合视频连续动作及log step 216同一前缀的left_foot记录。候选接触名为“击打”，现有接触方式没有对应项。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：机器人在第1次颠球时以另一只脚（右脚）接触草地承重，保持站立。" + } + ] + }, + "juggles_counted_2": { + "method": "visual", + "intervals": [ + [ + 2.56, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 2.56, + "log_step": null, + "text_zh": "观察：视频 2.40–2.48 s 足球落到机器人抬起的脚上方，脚向上迎球；视频 2.56–2.80 s 足球离脚上升，形成从放球后顺序观察到的第2次颠球，另一脚仍在草地上支撑。完成锚点约为视频 2.56 s，画面没有逐次裁判叠字。" + }, + { + "t": 2.56, + "log_step": null, + "text_zh": "推断：结合整段录像中依次出现的颠球过程与inputs/run.json评分器字段stages_reached=10、metric_touches=10、success=1及deterministic=1，第2次为有效计数并保持已完成；25 fps画面不能单独精测接触历时、上升距离和速度比门槛。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "其他", + "note_zh": "观察：机器人抬起的脚在第2次颠球中短促向上击打足球，球随后离脚；该脚为左脚的判断结合视频连续动作及log step 216同一前缀的left_foot记录。候选接触名为“击打”，现有接触方式没有对应项。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：机器人在第2次颠球时以另一只脚（右脚）接触草地承重，保持站立。" + } + ] + }, + "juggles_counted_3": { + "method": "visual", + "intervals": [ + [ + 3.28, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 3.28, + "log_step": null, + "text_zh": "观察：视频 3.12–3.20 s 足球落到机器人抬起的脚上方，脚向上迎球；视频 3.28–3.52 s 足球离脚上升，形成从放球后顺序观察到的第3次颠球，另一脚仍在草地上支撑。完成锚点约为视频 3.28 s，画面没有逐次裁判叠字。" + }, + { + "t": 3.28, + "log_step": null, + "text_zh": "推断：结合整段录像中依次出现的颠球过程与inputs/run.json评分器字段stages_reached=10、metric_touches=10、success=1及deterministic=1，第3次为有效计数并保持已完成；25 fps画面不能单独精测接触历时、上升距离和速度比门槛。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "其他", + "note_zh": "观察：机器人抬起的脚在第3次颠球中短促向上击打足球，球随后离脚；该脚为左脚的判断结合视频连续动作及log step 216同一前缀的left_foot记录。候选接触名为“击打”，现有接触方式没有对应项。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：机器人在第3次颠球时以另一只脚（右脚）接触草地承重，保持站立。" + } + ] + }, + "juggles_counted_4": { + "method": "visual", + "intervals": [ + [ + 4.0, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 4.0, + "log_step": null, + "text_zh": "观察：视频 3.80–3.88 s 足球落到机器人抬起的脚上方，脚向上迎球；视频 4.00–4.20 s 足球离脚上升，形成从放球后顺序观察到的第4次颠球，另一脚仍在草地上支撑。完成锚点约为视频 4.00 s，画面没有逐次裁判叠字。" + }, + { + "t": 4.0, + "log_step": null, + "text_zh": "推断：结合整段录像中依次出现的颠球过程与inputs/run.json评分器字段stages_reached=10、metric_touches=10、success=1及deterministic=1，第4次为有效计数并保持已完成；25 fps画面不能单独精测接触历时、上升距离和速度比门槛。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "其他", + "note_zh": "观察：机器人抬起的脚在第4次颠球中短促向上击打足球，球随后离脚；该脚为左脚的判断结合视频连续动作及log step 216同一前缀的left_foot记录。候选接触名为“击打”，现有接触方式没有对应项。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：机器人在第4次颠球时以另一只脚（右脚）接触草地承重，保持站立。" + } + ] + }, + "juggles_counted_5": { + "method": "visual", + "intervals": [ + [ + 4.68, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 4.68, + "log_step": null, + "text_zh": "观察：视频 4.48–4.56 s 足球落到机器人抬起的脚上方，脚向上迎球；视频 4.68–4.88 s 足球离脚上升，形成从放球后顺序观察到的第5次颠球，另一脚仍在草地上支撑。完成锚点约为视频 4.68 s，画面没有逐次裁判叠字。" + }, + { + "t": 4.68, + "log_step": null, + "text_zh": "推断：结合整段录像中依次出现的颠球过程与inputs/run.json评分器字段stages_reached=10、metric_touches=10、success=1及deterministic=1，第5次为有效计数并保持已完成；25 fps画面不能单独精测接触历时、上升距离和速度比门槛。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "其他", + "note_zh": "观察：机器人抬起的脚在第5次颠球中短促向上击打足球，球随后离脚；该脚为左脚的判断结合视频连续动作及log step 216同一前缀的left_foot记录。候选接触名为“击打”，现有接触方式没有对应项。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：机器人在第5次颠球时以另一只脚（右脚）接触草地承重，保持站立。" + } + ] + }, + "juggles_counted_6": { + "method": "visual", + "intervals": [ + [ + 5.32, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 5.32, + "log_step": null, + "text_zh": "观察：视频 5.12–5.20 s 足球落到机器人抬起的脚上方，脚向上迎球；视频 5.32–5.48 s 足球离脚上升，形成从放球后顺序观察到的第6次颠球，另一脚仍在草地上支撑。完成锚点约为视频 5.32 s，画面没有逐次裁判叠字。" + }, + { + "t": 5.32, + "log_step": null, + "text_zh": "推断：结合整段录像中依次出现的颠球过程与inputs/run.json评分器字段stages_reached=10、metric_touches=10、success=1及deterministic=1，第6次为有效计数并保持已完成；25 fps画面不能单独精测接触历时、上升距离和速度比门槛。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "其他", + "note_zh": "观察：机器人抬起的脚在第6次颠球中短促向上击打足球，球随后离脚；该脚为左脚的判断结合视频连续动作及log step 216同一前缀的left_foot记录。候选接触名为“击打”，现有接触方式没有对应项。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：机器人在第6次颠球时以另一只脚（右脚）接触草地承重，保持站立。" + } + ] + }, + "juggles_counted_7": { + "method": "visual", + "intervals": [ + [ + 5.96, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 5.96, + "log_step": null, + "text_zh": "观察：视频 5.76–5.84 s 足球落到机器人抬起的脚上方，脚向上迎球；视频 5.96–6.16 s 足球离脚上升，形成从放球后顺序观察到的第7次颠球，另一脚仍在草地上支撑。完成锚点约为视频 5.96 s，画面没有逐次裁判叠字。" + }, + { + "t": 5.96, + "log_step": null, + "text_zh": "推断：结合整段录像中依次出现的颠球过程与inputs/run.json评分器字段stages_reached=10、metric_touches=10、success=1及deterministic=1，第7次为有效计数并保持已完成；25 fps画面不能单独精测接触历时、上升距离和速度比门槛。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "其他", + "note_zh": "观察：机器人抬起的脚在第7次颠球中短促向上击打足球，球随后离脚；该脚为左脚的判断结合视频连续动作及log step 216同一前缀的left_foot记录。候选接触名为“击打”，现有接触方式没有对应项。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：机器人在第7次颠球时以另一只脚（右脚）接触草地承重，保持站立。" + } + ] + }, + "juggles_counted_8": { + "method": "visual", + "intervals": [ + [ + 6.6, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 6.6, + "log_step": null, + "text_zh": "观察：视频 6.40–6.48 s 足球落到机器人抬起的脚上方，脚向上迎球；视频 6.60–6.76 s 足球离脚上升，形成从放球后顺序观察到的第8次颠球，另一脚仍在草地上支撑。完成锚点约为视频 6.60 s，画面没有逐次裁判叠字。" + }, + { + "t": 6.6, + "log_step": null, + "text_zh": "推断：结合整段录像中依次出现的颠球过程与inputs/run.json评分器字段stages_reached=10、metric_touches=10、success=1及deterministic=1，第8次为有效计数并保持已完成；25 fps画面不能单独精测接触历时、上升距离和速度比门槛。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "其他", + "note_zh": "观察：机器人抬起的脚在第8次颠球中短促向上击打足球，球随后离脚；该脚为左脚的判断结合视频连续动作及log step 216同一前缀的left_foot记录。候选接触名为“击打”，现有接触方式没有对应项。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：机器人在第8次颠球时以另一只脚（右脚）接触草地承重，保持站立。" + } + ] + }, + "juggles_counted_9": { + "method": "visual", + "intervals": [ + [ + 7.24, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 7.24, + "log_step": null, + "text_zh": "观察：视频 7.00–7.08 s 足球落到机器人抬起的脚上方，脚向上迎球；视频 7.24–7.36 s 足球离脚上升，形成从放球后顺序观察到的第9次颠球，另一脚仍在草地上支撑。完成锚点约为视频 7.24 s，画面没有逐次裁判叠字。" + }, + { + "t": 7.24, + "log_step": null, + "text_zh": "推断：结合整段录像中依次出现的颠球过程与inputs/run.json评分器字段stages_reached=10、metric_touches=10、success=1及deterministic=1，第9次为有效计数并保持已完成；25 fps画面不能单独精测接触历时、上升距离和速度比门槛。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "其他", + "note_zh": "观察：机器人抬起的脚在第9次颠球中短促向上击打足球，球随后离脚；该脚为左脚的判断结合视频连续动作及log step 216同一前缀的left_foot记录。候选接触名为“击打”，现有接触方式没有对应项。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：机器人在第9次颠球时以另一只脚（右脚）接触草地承重，保持站立。" + } + ] + }, + "juggles_counted_10": { + "method": "visual", + "intervals": [ + [ + 7.68, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 7.68, + "log_step": null, + "text_zh": "观察：视频 7.44–7.64 s 足球从第9次飞行的高点回落，机器人左脚再次快速抬起；视频 7.68 s 末帧可见足球在该脚上方向上运动，机器人仍由右脚支撑。第10次击球后的完整上升没有全部呈现在末帧之前。" + }, + { + "t": 7.68, + "log_step": 222, + "text_zh": "观察：inputs/run.json评分器字段success=1、replay_success=1、deterministic=1、stages_reached=10、metric_touches=10、metric_ball_height_m=0.676；log step 220启动对最终候选文件的两次新进程回放，log step 222检查两次成功且末态哈希一致后保存771行交付，log step 225输出该交付的成功和10次触球。" + }, + { + "t": 7.68, + "log_step": null, + "text_zh": "推断：评分器确认第10次有效颠球完成，录像末尾的再次击球与其一致，故终值为true；近似完成锚点取已检查的末帧视频 7.68 s，精确验收时刻在末帧附近，时间不确定约0.1 s。此时间依据画面定位，没有将agent工作时间或其他候选轨迹的完成时间换算为视频时间。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "其他", + "note_zh": "观察：机器人抬起的脚在第10次颠球中短促向上击打足球，球随后离脚；该脚为左脚的判断结合视频连续动作及log step 216同一前缀的left_foot记录。候选接触名为“击打”，现有接触方式没有对应项。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：机器人在第10次颠球时以另一只脚（右脚）接触草地承重，保持站立。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "touches", + "value": 10.0, + "unit": "count", + "threshold": 10, + "direction": "higher_better" + }, + { + "name": "ball_height_m", + "value": 0.676, + "unit": "m", + "threshold": null, + "direction": "higher_better" + } + ], + "failure": null, + "summary_zh": "观察：本次视频中机器人在双掌放球后以左脚连续颠球、右脚支撑，直至视频 7.68 s 末帧仍站立；评分器记录10次有效触球、零次被动反弹、success=1且确定性回放通过，十个累计节点全部完成，无失败断点。推断：第1至第9次按画面中的逐次离脚上升定位，第10次由末尾再次击球与评分器最终计数共同确认，完成锚点近似取视频 7.68 s。观察：输入没有状态日志，25 fps录像不能独立复核毫秒接触和精确高度、速度门槛，且第10次完整上升超出末帧可见范围，其精确验收时刻不确定约0.1 s；参考成功视频仅用于比较同一视角下的动作形态。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "成功运行，十个累计颠球节点均完成，无失败断点；抽查帧（0.0、1.0、1.4、1.84、7.0、7.24、7.48、7.68 s 视频时间）与标注一致，时间、接触（击打/支撑/双掌挤压）和推断标记无明显错误。", + "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": "2026-10-29T10:05:28Z", + "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/embodied-olympics-fun-juggling-feet-i00-privileged--snul9zt/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-juggling-feet-i00-privileged--snul9zt/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "83315b75ae6b1c4e4cecd34afc4f85191a28ff13005f1ac34fdf5adc4082f318", + "video_sha256": "f25cb7071da058f968edfab8247b352da362a22a40fc80f721e5aa6281022a1d", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/juggling_feet/trajectories/embodied-olympics-fun-juggling-feet-i00-standard--79hdmsa.json b/phase2/embodied-olympics-fun/juggling_feet/trajectories/embodied-olympics-fun-juggling-feet-i00-standard--79hdmsa.json new file mode 100644 index 0000000000000000000000000000000000000000..cfec916ca7de497510b3ee16cc5a81b86ad26d46 --- /dev/null +++ b/phase2/embodied-olympics-fun/juggling_feet/trajectories/embodied-olympics-fun-juggling-feet-i00-standard--79hdmsa.json @@ -0,0 +1,399 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-juggling-feet-i00-standard--79hdmsa", + "taskKey": "embodied-olympics-fun/juggling_feet", + "dag_sha256": "6fb5f868dd0ec1e7ac4f8d9035cd8456789adc45e6c5ceda5fec1ad9d3ba10f6", + "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": 0, + "sim_steps": 0, + "calls": 20, + "rows": 0, + "resets": 0, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3119.2, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.864, + "peak_processes": 7, + "peak_disk_gb": 0.318, + "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/embodied-olympics-fun-juggling-feet-i00-standard--79hdmsa/replay.mp4", + "duration_s": 0.04, + "fps": 25.0, + "frames": 1, + "sha256": "12d147d4a5f031b87ce4c65888a462965aa52f2b5691f991b4e3a59da41625a8" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 10, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "juggles_counted_1": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 5, + "text_zh": "观察：本次视频仅有视频 0.00 s 一帧：机器人双脚着地站立，足球位于胸腹前的两掌之间，没有放球、脚部击球或离脚上升；log step 5 的基准服务观测为 control_steps=0、sim_time_s=0，评分器最终字段 n_actions=0、sim_steps=0、rows=0。" + }, + { + "t": 0.0, + "log_step": 91, + "text_zh": "推断：评分回合始终停在未放球初态，故第1次有效颠球从未完成；log step 91 的模型提供方过滤器阻断发生在任何实时动作提交之前。日志内本地 MuJoCo 计算和试验不是评分回合。" + } + ], + "contact_actual": [ + { + "who": "不适用", + "how": "不适用", + "note_zh": "未交付动作轨迹；只有初态画面，评分回合没有开始本节点的操作，不将初始持球或站立接触当作已执行颠球。" + } + ] + }, + "juggles_counted_2": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 5, + "text_zh": "观察：本次视频仅有视频 0.00 s 一帧：机器人双脚着地站立，足球位于胸腹前的两掌之间，没有放球、脚部击球或离脚上升；log step 5 的基准服务观测为 control_steps=0、sim_time_s=0，评分器最终字段 n_actions=0、sim_steps=0、rows=0。" + }, + { + "t": 0.0, + "log_step": 91, + "text_zh": "推断：评分回合始终停在未放球初态，故第2次有效颠球从未完成；log step 91 的模型提供方过滤器阻断发生在任何实时动作提交之前。日志内本地 MuJoCo 计算和试验不是评分回合。" + } + ], + "contact_actual": [ + { + "who": "不适用", + "how": "不适用", + "note_zh": "未交付动作轨迹；只有初态画面，评分回合没有开始本节点的操作，不将初始持球或站立接触当作已执行颠球。" + } + ] + }, + "juggles_counted_3": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 5, + "text_zh": "观察：本次视频仅有视频 0.00 s 一帧：机器人双脚着地站立，足球位于胸腹前的两掌之间，没有放球、脚部击球或离脚上升；log step 5 的基准服务观测为 control_steps=0、sim_time_s=0，评分器最终字段 n_actions=0、sim_steps=0、rows=0。" + }, + { + "t": 0.0, + "log_step": 91, + "text_zh": "推断：评分回合始终停在未放球初态，故第3次有效颠球从未完成；log step 91 的模型提供方过滤器阻断发生在任何实时动作提交之前。日志内本地 MuJoCo 计算和试验不是评分回合。" + } + ], + "contact_actual": [ + { + "who": "不适用", + "how": "不适用", + "note_zh": "未交付动作轨迹；只有初态画面，评分回合没有开始本节点的操作，不将初始持球或站立接触当作已执行颠球。" + } + ] + }, + "juggles_counted_4": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 5, + "text_zh": "观察：本次视频仅有视频 0.00 s 一帧：机器人双脚着地站立，足球位于胸腹前的两掌之间，没有放球、脚部击球或离脚上升；log step 5 的基准服务观测为 control_steps=0、sim_time_s=0，评分器最终字段 n_actions=0、sim_steps=0、rows=0。" + }, + { + "t": 0.0, + "log_step": 91, + "text_zh": "推断：评分回合始终停在未放球初态，故第4次有效颠球从未完成；log step 91 的模型提供方过滤器阻断发生在任何实时动作提交之前。日志内本地 MuJoCo 计算和试验不是评分回合。" + } + ], + "contact_actual": [ + { + "who": "不适用", + "how": "不适用", + "note_zh": "未交付动作轨迹；只有初态画面，评分回合没有开始本节点的操作，不将初始持球或站立接触当作已执行颠球。" + } + ] + }, + "juggles_counted_5": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 5, + "text_zh": "观察：本次视频仅有视频 0.00 s 一帧：机器人双脚着地站立，足球位于胸腹前的两掌之间，没有放球、脚部击球或离脚上升；log step 5 的基准服务观测为 control_steps=0、sim_time_s=0，评分器最终字段 n_actions=0、sim_steps=0、rows=0。" + }, + { + "t": 0.0, + "log_step": 91, + "text_zh": "推断：评分回合始终停在未放球初态，故第5次有效颠球从未完成；log step 91 的模型提供方过滤器阻断发生在任何实时动作提交之前。日志内本地 MuJoCo 计算和试验不是评分回合。" + } + ], + "contact_actual": [ + { + "who": "不适用", + "how": "不适用", + "note_zh": "未交付动作轨迹；只有初态画面，评分回合没有开始本节点的操作，不将初始持球或站立接触当作已执行颠球。" + } + ] + }, + "juggles_counted_6": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 5, + "text_zh": "观察：本次视频仅有视频 0.00 s 一帧：机器人双脚着地站立，足球位于胸腹前的两掌之间，没有放球、脚部击球或离脚上升；log step 5 的基准服务观测为 control_steps=0、sim_time_s=0，评分器最终字段 n_actions=0、sim_steps=0、rows=0。" + }, + { + "t": 0.0, + "log_step": 91, + "text_zh": "推断：评分回合始终停在未放球初态，故第6次有效颠球从未完成；log step 91 的模型提供方过滤器阻断发生在任何实时动作提交之前。日志内本地 MuJoCo 计算和试验不是评分回合。" + } + ], + "contact_actual": [ + { + "who": "不适用", + "how": "不适用", + "note_zh": "未交付动作轨迹；只有初态画面，评分回合没有开始本节点的操作，不将初始持球或站立接触当作已执行颠球。" + } + ] + }, + "juggles_counted_7": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 5, + "text_zh": "观察：本次视频仅有视频 0.00 s 一帧：机器人双脚着地站立，足球位于胸腹前的两掌之间，没有放球、脚部击球或离脚上升；log step 5 的基准服务观测为 control_steps=0、sim_time_s=0，评分器最终字段 n_actions=0、sim_steps=0、rows=0。" + }, + { + "t": 0.0, + "log_step": 91, + "text_zh": "推断：评分回合始终停在未放球初态，故第7次有效颠球从未完成；log step 91 的模型提供方过滤器阻断发生在任何实时动作提交之前。日志内本地 MuJoCo 计算和试验不是评分回合。" + } + ], + "contact_actual": [ + { + "who": "不适用", + "how": "不适用", + "note_zh": "未交付动作轨迹；只有初态画面，评分回合没有开始本节点的操作，不将初始持球或站立接触当作已执行颠球。" + } + ] + }, + "juggles_counted_8": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 5, + "text_zh": "观察：本次视频仅有视频 0.00 s 一帧：机器人双脚着地站立，足球位于胸腹前的两掌之间，没有放球、脚部击球或离脚上升；log step 5 的基准服务观测为 control_steps=0、sim_time_s=0，评分器最终字段 n_actions=0、sim_steps=0、rows=0。" + }, + { + "t": 0.0, + "log_step": 91, + "text_zh": "推断：评分回合始终停在未放球初态，故第8次有效颠球从未完成；log step 91 的模型提供方过滤器阻断发生在任何实时动作提交之前。日志内本地 MuJoCo 计算和试验不是评分回合。" + } + ], + "contact_actual": [ + { + "who": "不适用", + "how": "不适用", + "note_zh": "未交付动作轨迹；只有初态画面，评分回合没有开始本节点的操作，不将初始持球或站立接触当作已执行颠球。" + } + ] + }, + "juggles_counted_9": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 5, + "text_zh": "观察：本次视频仅有视频 0.00 s 一帧：机器人双脚着地站立，足球位于胸腹前的两掌之间，没有放球、脚部击球或离脚上升；log step 5 的基准服务观测为 control_steps=0、sim_time_s=0，评分器最终字段 n_actions=0、sim_steps=0、rows=0。" + }, + { + "t": 0.0, + "log_step": 91, + "text_zh": "推断：评分回合始终停在未放球初态，故第9次有效颠球从未完成；log step 91 的模型提供方过滤器阻断发生在任何实时动作提交之前。日志内本地 MuJoCo 计算和试验不是评分回合。" + } + ], + "contact_actual": [ + { + "who": "不适用", + "how": "不适用", + "note_zh": "未交付动作轨迹；只有初态画面，评分回合没有开始本节点的操作，不将初始持球或站立接触当作已执行颠球。" + } + ] + }, + "juggles_counted_10": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 5, + "text_zh": "观察：本次视频仅有视频 0.00 s 一帧：机器人双脚着地站立，足球位于胸腹前的两掌之间，没有放球、脚部击球或离脚上升；log step 5 的基准服务观测为 control_steps=0、sim_time_s=0，评分器最终字段 n_actions=0、sim_steps=0、rows=0。" + }, + { + "t": 0.0, + "log_step": 91, + "text_zh": "推断：评分回合始终停在未放球初态，故第10次有效颠球从未完成；log step 91 的模型提供方过滤器阻断发生在任何实时动作提交之前。日志内本地 MuJoCo 计算和试验不是评分回合。" + } + ], + "contact_actual": [ + { + "who": "不适用", + "how": "不适用", + "note_zh": "未交付动作轨迹；只有初态画面，评分回合没有开始本节点的操作，不将初始持球或站立接触当作已执行颠球。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "touches", + "unit": "count", + "threshold": 10, + "direction": "higher_better", + "value": null + }, + { + "name": "ball_height_m", + "unit": "m", + "threshold": null, + "direction": "higher_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "unknown", + "t": null, + "log_step": 91, + "where": "首步前", + "evidence_zh": "观察：log step 91 最终消息为“This request was blocked by Gemini's filters”；此前 log step 87 仍在本地模型求逆解，log step 88–90 为读取规格的诊断。评分器最终 n_actions=0、sim_steps=0、rows=0，视频 0.00 s 唯一帧为双掌持球的站立初态；没有已启动的物理步骤或已确定的物理失败。" + }, + "unmet": { + "juggles_counted_1": false, + "juggles_counted_2": false, + "juggles_counted_3": false, + "juggles_counted_4": false, + "juggles_counted_5": false, + "juggles_counted_6": false, + "juggles_counted_7": false, + "juggles_counted_8": false, + "juggles_counted_9": false, + "juggles_counted_10": false + }, + "phenomena": [], + "root_causes": [], + "gap_zh": "观察：inputs/replay.mp4 存在，但仅含视频 0.00 s 的初态单帧（总长 0.04 s）；inputs/ 中没有状态日志，评分器未报告 touches 或 ball_height_m。log step 91 明确记录模型提供方过滤器阻断，且最终动作数、仿真步数及动作行数均为零。推断：这是交付任何实时动作前的外部中断，按 CASES.md FAIL-5 不归为机器人失败，断点为 unknown／首步前，现象与根因均留空；会话本地模型试验不能提供评分回合的接触或控制失败证据，应重测。" + }, + "summary_zh": "本次视频仅显示视频 0.00 s 的双掌持球站立初态；评分器记录 n_actions=0、sim_steps=0，十个颠球节点均未达到（0/10）。log step 91 明确报告 Gemini 过滤器阻断，发生在实时动作提交之前，按外部中断处理，保留原始 success=0 但不归因于机器人，应重测。没有动作片段、状态日志或裁判颠球计数，不能把会话中的本地模型试验当作评分回合的失败现象或根因。", + "conf": 10, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 9, + "summary_zh": "标签正确：视频仅有 0.00 s 初态一帧，评分器 n_actions=0、sim_steps=0，log step 91 为 Gemini 过滤器阻断，发生在任何实时动作之前，属外部中断（FAIL-5）；十个节点均未达成，断点 unknown/首步前，两层留空并有 gap_zh，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": "2026-10-29T10:03:05Z", + "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/embodied-olympics-fun-juggling-feet-i00-standard--79hdmsa/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-juggling-feet-i00-standard--79hdmsa/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "eb3c6360b4cc778ea90468d279072db83b08a5033335407c9f1043c1d1da1a60", + "video_sha256": "12d147d4a5f031b87ce4c65888a462965aa52f2b5691f991b4e3a59da41625a8", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/jump_rope/task_dag.json b/phase2/embodied-olympics-fun/jump_rope/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..c1e153e230b9ffec73dd9bde64ad5e9de170f41a --- /dev/null +++ b/phase2/embodied-olympics-fun/jump_rope/task_dag.json @@ -0,0 +1,965 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "embodied-olympics-fun/jump_rope", + "suite": "embodied-olympics-fun", + "base": "embodied-olympics-fun-jump-rope-i00", + "title_zh": "双脚正摇跳绳十七次", + "summary_zh": "一分钟内双手持柄完成17次有效正摇双脚跳绳，末次脚先落地并连续站稳1秒，且不发生绳柄脱手、绳柄触地或跌倒。", + "embodiment_zh": "两种模式均为Unitree G1人形机器人，29个本体关节、双臂，每只作用手均为Unitree Dex3-1三指灵巧手，靠接触与摩擦各握一个绳柄；双腿足式支撑和跳跃，无轮式底盘。特权模式可读仿真状态，标准模式使用相机、编码器与惯性传感器，硬件相同。", + "task_class": "compositional", + "analysis_level": "step", + "spec": { + "version": "1.2", + "task": "embodied-olympics-fun-jump-rope-i00", + "bench": "embodied-olympics-fun", + "progress": "live", + "entities": {}, + "milestones": [ + { + "id": "skip_1_completed", + "zh": "第1次过绳", + "mode": "event", + "when": null, + "after": [], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "累计完成第1次有效正摇双脚过绳；场地或侧面视角可见红白串珠绳先从头后越过头顶至头前，再从前向后穿过腾空的两脚下方：过脚时全身无部位触地，两脚离地时差不超过0.10秒，两脚过绳时差不超过0.15秒，绳在每只脚下的向后速度经评分器保留三位小数后至少为1.0米/秒，取样绳点低于该脚最低点加0.0095米。绳未先过头、仅过一脚、反摇、过脚时有身体部位触地或速度及同步条件不合格均不计数；额外空跳不计，一圈至多计一次，漏跳后可重新摇绳，已完成的有效次数不撤销。", + "step": { + "cat": "本体", + "item": "跳跃" + }, + "contact": [ + { + "who": "双臂协作", + "how": "抓握", + "plan": 1, + "note_zh": "左手灵巧手握左侧蓝色绳柄，右手灵巧手握右侧蓝色绳柄，两臂共同驱动红白串珠绳正向转动。" + }, + { + "who": "腿/躯干", + "how": "避开", + "plan": 1, + "note_zh": "双脚同时腾空，让转动的红白串珠绳从两脚下方通过；双手此时仍各握一个绳柄。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:3-34", + "inputs/task/standard/instruction.md:5-33", + "inputs/task/privileged/tests/rfo_task.py:121-125", + "inputs/task/privileged/tests/rfo_task.py:365-379", + "inputs/task/privileged/tests/rfo_task.py:383-451", + "inputs/task/privileged/tests/rfo_task.py:457-473", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/tests/rfo_task.py:21-22", + "inputs/task/privileged/tests/rfo_task.py:221-222" + ], + "conf": 6, + "note_zh": "按次数逐一建节点；该圈的过头、起跳、过脚不拆分。双脚起跳前和落地后的地板支撑先后发生，接触表列过绳时并存的双手持柄与双脚避绳。用原生有效次数达到该序号判定，不以任意一次腾空或绳的总圈数代替。" + }, + { + "id": "skip_2_completed", + "zh": "第2次过绳", + "mode": "event", + "when": null, + "after": [ + "skip_1_completed" + ], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "累计完成第2次有效正摇双脚过绳；场地或侧面视角可见红白串珠绳先从头后越过头顶至头前，再从前向后穿过腾空的两脚下方：过脚时全身无部位触地，两脚离地时差不超过0.10秒，两脚过绳时差不超过0.15秒，绳在每只脚下的向后速度经评分器保留三位小数后至少为1.0米/秒，取样绳点低于该脚最低点加0.0095米。绳未先过头、仅过一脚、反摇、过脚时有身体部位触地或速度及同步条件不合格均不计数；额外空跳不计，一圈至多计一次，漏跳后可重新摇绳，已完成的有效次数不撤销。", + "step": { + "cat": "本体", + "item": "跳跃" + }, + "contact": [ + { + "who": "双臂协作", + "how": "抓握", + "plan": 1, + "note_zh": "左手灵巧手握左侧蓝色绳柄，右手灵巧手握右侧蓝色绳柄，两臂共同驱动红白串珠绳正向转动。" + }, + { + "who": "腿/躯干", + "how": "避开", + "plan": 1, + "note_zh": "双脚同时腾空，让转动的红白串珠绳从两脚下方通过；双手此时仍各握一个绳柄。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:3-34", + "inputs/task/standard/instruction.md:5-33", + "inputs/task/privileged/tests/rfo_task.py:121-125", + "inputs/task/privileged/tests/rfo_task.py:365-379", + "inputs/task/privileged/tests/rfo_task.py:383-451", + "inputs/task/privileged/tests/rfo_task.py:457-473", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/tests/rfo_task.py:21-22", + "inputs/task/privileged/tests/rfo_task.py:221-222" + ], + "conf": 6, + "note_zh": "按次数逐一建节点；该圈的过头、起跳、过脚不拆分。双脚起跳前和落地后的地板支撑先后发生，接触表列过绳时并存的双手持柄与双脚避绳。用原生有效次数达到该序号判定，不以任意一次腾空或绳的总圈数代替。" + }, + { + "id": "skip_3_completed", + "zh": "第3次过绳", + "mode": "event", + "when": null, + "after": [ + "skip_2_completed" + ], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "累计完成第3次有效正摇双脚过绳；场地或侧面视角可见红白串珠绳先从头后越过头顶至头前，再从前向后穿过腾空的两脚下方：过脚时全身无部位触地，两脚离地时差不超过0.10秒，两脚过绳时差不超过0.15秒，绳在每只脚下的向后速度经评分器保留三位小数后至少为1.0米/秒，取样绳点低于该脚最低点加0.0095米。绳未先过头、仅过一脚、反摇、过脚时有身体部位触地或速度及同步条件不合格均不计数；额外空跳不计，一圈至多计一次，漏跳后可重新摇绳，已完成的有效次数不撤销。", + "step": { + "cat": "本体", + "item": "跳跃" + }, + "contact": [ + { + "who": "双臂协作", + "how": "抓握", + "plan": 1, + "note_zh": "左手灵巧手握左侧蓝色绳柄，右手灵巧手握右侧蓝色绳柄，两臂共同驱动红白串珠绳正向转动。" + }, + { + "who": "腿/躯干", + "how": "避开", + "plan": 1, + "note_zh": "双脚同时腾空，让转动的红白串珠绳从两脚下方通过；双手此时仍各握一个绳柄。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:3-34", + "inputs/task/standard/instruction.md:5-33", + "inputs/task/privileged/tests/rfo_task.py:121-125", + "inputs/task/privileged/tests/rfo_task.py:365-379", + "inputs/task/privileged/tests/rfo_task.py:383-451", + "inputs/task/privileged/tests/rfo_task.py:457-473", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/tests/rfo_task.py:21-22", + "inputs/task/privileged/tests/rfo_task.py:221-222" + ], + "conf": 6, + "note_zh": "按次数逐一建节点；该圈的过头、起跳、过脚不拆分。双脚起跳前和落地后的地板支撑先后发生，接触表列过绳时并存的双手持柄与双脚避绳。用原生有效次数达到该序号判定，不以任意一次腾空或绳的总圈数代替。" + }, + { + "id": "skip_4_completed", + "zh": "第4次过绳", + "mode": "event", + "when": null, + "after": [ + "skip_3_completed" + ], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "累计完成第4次有效正摇双脚过绳；场地或侧面视角可见红白串珠绳先从头后越过头顶至头前，再从前向后穿过腾空的两脚下方：过脚时全身无部位触地，两脚离地时差不超过0.10秒，两脚过绳时差不超过0.15秒，绳在每只脚下的向后速度经评分器保留三位小数后至少为1.0米/秒，取样绳点低于该脚最低点加0.0095米。绳未先过头、仅过一脚、反摇、过脚时有身体部位触地或速度及同步条件不合格均不计数；额外空跳不计，一圈至多计一次，漏跳后可重新摇绳，已完成的有效次数不撤销。", + "step": { + "cat": "本体", + "item": "跳跃" + }, + "contact": [ + { + "who": "双臂协作", + "how": "抓握", + "plan": 1, + "note_zh": "左手灵巧手握左侧蓝色绳柄，右手灵巧手握右侧蓝色绳柄，两臂共同驱动红白串珠绳正向转动。" + }, + { + "who": "腿/躯干", + "how": "避开", + "plan": 1, + "note_zh": "双脚同时腾空，让转动的红白串珠绳从两脚下方通过；双手此时仍各握一个绳柄。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:3-34", + "inputs/task/standard/instruction.md:5-33", + "inputs/task/privileged/tests/rfo_task.py:121-125", + "inputs/task/privileged/tests/rfo_task.py:365-379", + "inputs/task/privileged/tests/rfo_task.py:383-451", + "inputs/task/privileged/tests/rfo_task.py:457-473", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/tests/rfo_task.py:21-22", + "inputs/task/privileged/tests/rfo_task.py:221-222" + ], + "conf": 6, + "note_zh": "按次数逐一建节点；该圈的过头、起跳、过脚不拆分。双脚起跳前和落地后的地板支撑先后发生，接触表列过绳时并存的双手持柄与双脚避绳。用原生有效次数达到该序号判定，不以任意一次腾空或绳的总圈数代替。" + }, + { + "id": "skip_5_completed", + "zh": "第5次过绳", + "mode": "event", + "when": null, + "after": [ + "skip_4_completed" + ], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "累计完成第5次有效正摇双脚过绳；场地或侧面视角可见红白串珠绳先从头后越过头顶至头前，再从前向后穿过腾空的两脚下方：过脚时全身无部位触地，两脚离地时差不超过0.10秒，两脚过绳时差不超过0.15秒，绳在每只脚下的向后速度经评分器保留三位小数后至少为1.0米/秒，取样绳点低于该脚最低点加0.0095米。绳未先过头、仅过一脚、反摇、过脚时有身体部位触地或速度及同步条件不合格均不计数；额外空跳不计，一圈至多计一次，漏跳后可重新摇绳，已完成的有效次数不撤销。", + "step": { + "cat": "本体", + "item": "跳跃" + }, + "contact": [ + { + "who": "双臂协作", + "how": "抓握", + "plan": 1, + "note_zh": "左手灵巧手握左侧蓝色绳柄，右手灵巧手握右侧蓝色绳柄，两臂共同驱动红白串珠绳正向转动。" + }, + { + "who": "腿/躯干", + "how": "避开", + "plan": 1, + "note_zh": "双脚同时腾空，让转动的红白串珠绳从两脚下方通过；双手此时仍各握一个绳柄。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:3-34", + "inputs/task/standard/instruction.md:5-33", + "inputs/task/privileged/tests/rfo_task.py:121-125", + "inputs/task/privileged/tests/rfo_task.py:365-379", + "inputs/task/privileged/tests/rfo_task.py:383-451", + "inputs/task/privileged/tests/rfo_task.py:457-473", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/tests/rfo_task.py:21-22", + "inputs/task/privileged/tests/rfo_task.py:221-222" + ], + "conf": 6, + "note_zh": "按次数逐一建节点；该圈的过头、起跳、过脚不拆分。双脚起跳前和落地后的地板支撑先后发生，接触表列过绳时并存的双手持柄与双脚避绳。用原生有效次数达到该序号判定，不以任意一次腾空或绳的总圈数代替。" + }, + { + "id": "skip_6_completed", + "zh": "第6次过绳", + "mode": "event", + "when": null, + "after": [ + "skip_5_completed" + ], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "累计完成第6次有效正摇双脚过绳；场地或侧面视角可见红白串珠绳先从头后越过头顶至头前，再从前向后穿过腾空的两脚下方：过脚时全身无部位触地，两脚离地时差不超过0.10秒，两脚过绳时差不超过0.15秒，绳在每只脚下的向后速度经评分器保留三位小数后至少为1.0米/秒，取样绳点低于该脚最低点加0.0095米。绳未先过头、仅过一脚、反摇、过脚时有身体部位触地或速度及同步条件不合格均不计数；额外空跳不计，一圈至多计一次，漏跳后可重新摇绳，已完成的有效次数不撤销。", + "step": { + "cat": "本体", + "item": "跳跃" + }, + "contact": [ + { + "who": "双臂协作", + "how": "抓握", + "plan": 1, + "note_zh": "左手灵巧手握左侧蓝色绳柄，右手灵巧手握右侧蓝色绳柄，两臂共同驱动红白串珠绳正向转动。" + }, + { + "who": "腿/躯干", + "how": "避开", + "plan": 1, + "note_zh": "双脚同时腾空，让转动的红白串珠绳从两脚下方通过；双手此时仍各握一个绳柄。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:3-34", + "inputs/task/standard/instruction.md:5-33", + "inputs/task/privileged/tests/rfo_task.py:121-125", + "inputs/task/privileged/tests/rfo_task.py:365-379", + "inputs/task/privileged/tests/rfo_task.py:383-451", + "inputs/task/privileged/tests/rfo_task.py:457-473", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/tests/rfo_task.py:21-22", + "inputs/task/privileged/tests/rfo_task.py:221-222" + ], + "conf": 6, + "note_zh": "按次数逐一建节点；该圈的过头、起跳、过脚不拆分。双脚起跳前和落地后的地板支撑先后发生，接触表列过绳时并存的双手持柄与双脚避绳。用原生有效次数达到该序号判定，不以任意一次腾空或绳的总圈数代替。" + }, + { + "id": "skip_7_completed", + "zh": "第7次过绳", + "mode": "event", + "when": null, + "after": [ + "skip_6_completed" + ], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "累计完成第7次有效正摇双脚过绳；场地或侧面视角可见红白串珠绳先从头后越过头顶至头前，再从前向后穿过腾空的两脚下方：过脚时全身无部位触地，两脚离地时差不超过0.10秒，两脚过绳时差不超过0.15秒，绳在每只脚下的向后速度经评分器保留三位小数后至少为1.0米/秒，取样绳点低于该脚最低点加0.0095米。绳未先过头、仅过一脚、反摇、过脚时有身体部位触地或速度及同步条件不合格均不计数；额外空跳不计，一圈至多计一次，漏跳后可重新摇绳，已完成的有效次数不撤销。", + "step": { + "cat": "本体", + "item": "跳跃" + }, + "contact": [ + { + "who": "双臂协作", + "how": "抓握", + "plan": 1, + "note_zh": "左手灵巧手握左侧蓝色绳柄，右手灵巧手握右侧蓝色绳柄，两臂共同驱动红白串珠绳正向转动。" + }, + { + "who": "腿/躯干", + "how": "避开", + "plan": 1, + "note_zh": "双脚同时腾空，让转动的红白串珠绳从两脚下方通过；双手此时仍各握一个绳柄。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:3-34", + "inputs/task/standard/instruction.md:5-33", + "inputs/task/privileged/tests/rfo_task.py:121-125", + "inputs/task/privileged/tests/rfo_task.py:365-379", + "inputs/task/privileged/tests/rfo_task.py:383-451", + "inputs/task/privileged/tests/rfo_task.py:457-473", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/tests/rfo_task.py:21-22", + "inputs/task/privileged/tests/rfo_task.py:221-222" + ], + "conf": 6, + "note_zh": "按次数逐一建节点；该圈的过头、起跳、过脚不拆分。双脚起跳前和落地后的地板支撑先后发生，接触表列过绳时并存的双手持柄与双脚避绳。用原生有效次数达到该序号判定，不以任意一次腾空或绳的总圈数代替。" + }, + { + "id": "skip_8_completed", + "zh": "第8次过绳", + "mode": "event", + "when": null, + "after": [ + "skip_7_completed" + ], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "累计完成第8次有效正摇双脚过绳；场地或侧面视角可见红白串珠绳先从头后越过头顶至头前，再从前向后穿过腾空的两脚下方：过脚时全身无部位触地，两脚离地时差不超过0.10秒，两脚过绳时差不超过0.15秒，绳在每只脚下的向后速度经评分器保留三位小数后至少为1.0米/秒，取样绳点低于该脚最低点加0.0095米。绳未先过头、仅过一脚、反摇、过脚时有身体部位触地或速度及同步条件不合格均不计数；额外空跳不计，一圈至多计一次，漏跳后可重新摇绳，已完成的有效次数不撤销。", + "step": { + "cat": "本体", + "item": "跳跃" + }, + "contact": [ + { + "who": "双臂协作", + "how": "抓握", + "plan": 1, + "note_zh": "左手灵巧手握左侧蓝色绳柄，右手灵巧手握右侧蓝色绳柄，两臂共同驱动红白串珠绳正向转动。" + }, + { + "who": "腿/躯干", + "how": "避开", + "plan": 1, + "note_zh": "双脚同时腾空，让转动的红白串珠绳从两脚下方通过；双手此时仍各握一个绳柄。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:3-34", + "inputs/task/standard/instruction.md:5-33", + "inputs/task/privileged/tests/rfo_task.py:121-125", + "inputs/task/privileged/tests/rfo_task.py:365-379", + "inputs/task/privileged/tests/rfo_task.py:383-451", + "inputs/task/privileged/tests/rfo_task.py:457-473", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/tests/rfo_task.py:21-22", + "inputs/task/privileged/tests/rfo_task.py:221-222" + ], + "conf": 6, + "note_zh": "按次数逐一建节点；该圈的过头、起跳、过脚不拆分。双脚起跳前和落地后的地板支撑先后发生，接触表列过绳时并存的双手持柄与双脚避绳。用原生有效次数达到该序号判定，不以任意一次腾空或绳的总圈数代替。" + }, + { + "id": "skip_9_completed", + "zh": "第9次过绳", + "mode": "event", + "when": null, + "after": [ + "skip_8_completed" + ], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "累计完成第9次有效正摇双脚过绳；场地或侧面视角可见红白串珠绳先从头后越过头顶至头前，再从前向后穿过腾空的两脚下方：过脚时全身无部位触地，两脚离地时差不超过0.10秒，两脚过绳时差不超过0.15秒，绳在每只脚下的向后速度经评分器保留三位小数后至少为1.0米/秒，取样绳点低于该脚最低点加0.0095米。绳未先过头、仅过一脚、反摇、过脚时有身体部位触地或速度及同步条件不合格均不计数；额外空跳不计，一圈至多计一次，漏跳后可重新摇绳，已完成的有效次数不撤销。", + "step": { + "cat": "本体", + "item": "跳跃" + }, + "contact": [ + { + "who": "双臂协作", + "how": "抓握", + "plan": 1, + "note_zh": "左手灵巧手握左侧蓝色绳柄，右手灵巧手握右侧蓝色绳柄，两臂共同驱动红白串珠绳正向转动。" + }, + { + "who": "腿/躯干", + "how": "避开", + "plan": 1, + "note_zh": "双脚同时腾空，让转动的红白串珠绳从两脚下方通过；双手此时仍各握一个绳柄。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:3-34", + "inputs/task/standard/instruction.md:5-33", + "inputs/task/privileged/tests/rfo_task.py:121-125", + "inputs/task/privileged/tests/rfo_task.py:365-379", + "inputs/task/privileged/tests/rfo_task.py:383-451", + "inputs/task/privileged/tests/rfo_task.py:457-473", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/tests/rfo_task.py:21-22", + "inputs/task/privileged/tests/rfo_task.py:221-222" + ], + "conf": 6, + "note_zh": "按次数逐一建节点；该圈的过头、起跳、过脚不拆分。双脚起跳前和落地后的地板支撑先后发生，接触表列过绳时并存的双手持柄与双脚避绳。用原生有效次数达到该序号判定，不以任意一次腾空或绳的总圈数代替。" + }, + { + "id": "skip_10_completed", + "zh": "第10次过绳", + "mode": "event", + "when": null, + "after": [ + "skip_9_completed" + ], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "累计完成第10次有效正摇双脚过绳；场地或侧面视角可见红白串珠绳先从头后越过头顶至头前，再从前向后穿过腾空的两脚下方：过脚时全身无部位触地，两脚离地时差不超过0.10秒，两脚过绳时差不超过0.15秒，绳在每只脚下的向后速度经评分器保留三位小数后至少为1.0米/秒，取样绳点低于该脚最低点加0.0095米。绳未先过头、仅过一脚、反摇、过脚时有身体部位触地或速度及同步条件不合格均不计数；额外空跳不计，一圈至多计一次，漏跳后可重新摇绳，已完成的有效次数不撤销。", + "step": { + "cat": "本体", + "item": "跳跃" + }, + "contact": [ + { + "who": "双臂协作", + "how": "抓握", + "plan": 1, + "note_zh": "左手灵巧手握左侧蓝色绳柄，右手灵巧手握右侧蓝色绳柄，两臂共同驱动红白串珠绳正向转动。" + }, + { + "who": "腿/躯干", + "how": "避开", + "plan": 1, + "note_zh": "双脚同时腾空，让转动的红白串珠绳从两脚下方通过；双手此时仍各握一个绳柄。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:3-34", + "inputs/task/standard/instruction.md:5-33", + "inputs/task/privileged/tests/rfo_task.py:121-125", + "inputs/task/privileged/tests/rfo_task.py:365-379", + "inputs/task/privileged/tests/rfo_task.py:383-451", + "inputs/task/privileged/tests/rfo_task.py:457-473", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/tests/rfo_task.py:21-22", + "inputs/task/privileged/tests/rfo_task.py:221-222" + ], + "conf": 6, + "note_zh": "按次数逐一建节点；该圈的过头、起跳、过脚不拆分。双脚起跳前和落地后的地板支撑先后发生，接触表列过绳时并存的双手持柄与双脚避绳。用原生有效次数达到该序号判定，不以任意一次腾空或绳的总圈数代替。" + }, + { + "id": "skip_11_completed", + "zh": "第11次过绳", + "mode": "event", + "when": null, + "after": [ + "skip_10_completed" + ], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "累计完成第11次有效正摇双脚过绳；场地或侧面视角可见红白串珠绳先从头后越过头顶至头前，再从前向后穿过腾空的两脚下方：过脚时全身无部位触地，两脚离地时差不超过0.10秒，两脚过绳时差不超过0.15秒，绳在每只脚下的向后速度经评分器保留三位小数后至少为1.0米/秒，取样绳点低于该脚最低点加0.0095米。绳未先过头、仅过一脚、反摇、过脚时有身体部位触地或速度及同步条件不合格均不计数；额外空跳不计，一圈至多计一次，漏跳后可重新摇绳，已完成的有效次数不撤销。", + "step": { + "cat": "本体", + "item": "跳跃" + }, + "contact": [ + { + "who": "双臂协作", + "how": "抓握", + "plan": 1, + "note_zh": "左手灵巧手握左侧蓝色绳柄，右手灵巧手握右侧蓝色绳柄，两臂共同驱动红白串珠绳正向转动。" + }, + { + "who": "腿/躯干", + "how": "避开", + "plan": 1, + "note_zh": "双脚同时腾空，让转动的红白串珠绳从两脚下方通过；双手此时仍各握一个绳柄。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:3-34", + "inputs/task/standard/instruction.md:5-33", + "inputs/task/privileged/tests/rfo_task.py:121-125", + "inputs/task/privileged/tests/rfo_task.py:365-379", + "inputs/task/privileged/tests/rfo_task.py:383-451", + "inputs/task/privileged/tests/rfo_task.py:457-473", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/tests/rfo_task.py:21-22", + "inputs/task/privileged/tests/rfo_task.py:221-222" + ], + "conf": 6, + "note_zh": "按次数逐一建节点；该圈的过头、起跳、过脚不拆分。双脚起跳前和落地后的地板支撑先后发生，接触表列过绳时并存的双手持柄与双脚避绳。用原生有效次数达到该序号判定，不以任意一次腾空或绳的总圈数代替。" + }, + { + "id": "skip_12_completed", + "zh": "第12次过绳", + "mode": "event", + "when": null, + "after": [ + "skip_11_completed" + ], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "累计完成第12次有效正摇双脚过绳；场地或侧面视角可见红白串珠绳先从头后越过头顶至头前，再从前向后穿过腾空的两脚下方：过脚时全身无部位触地，两脚离地时差不超过0.10秒，两脚过绳时差不超过0.15秒，绳在每只脚下的向后速度经评分器保留三位小数后至少为1.0米/秒，取样绳点低于该脚最低点加0.0095米。绳未先过头、仅过一脚、反摇、过脚时有身体部位触地或速度及同步条件不合格均不计数；额外空跳不计，一圈至多计一次，漏跳后可重新摇绳，已完成的有效次数不撤销。", + "step": { + "cat": "本体", + "item": "跳跃" + }, + "contact": [ + { + "who": "双臂协作", + "how": "抓握", + "plan": 1, + "note_zh": "左手灵巧手握左侧蓝色绳柄，右手灵巧手握右侧蓝色绳柄，两臂共同驱动红白串珠绳正向转动。" + }, + { + "who": "腿/躯干", + "how": "避开", + "plan": 1, + "note_zh": "双脚同时腾空，让转动的红白串珠绳从两脚下方通过；双手此时仍各握一个绳柄。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:3-34", + "inputs/task/standard/instruction.md:5-33", + "inputs/task/privileged/tests/rfo_task.py:121-125", + "inputs/task/privileged/tests/rfo_task.py:365-379", + "inputs/task/privileged/tests/rfo_task.py:383-451", + "inputs/task/privileged/tests/rfo_task.py:457-473", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/tests/rfo_task.py:21-22", + "inputs/task/privileged/tests/rfo_task.py:221-222" + ], + "conf": 6, + "note_zh": "按次数逐一建节点；该圈的过头、起跳、过脚不拆分。双脚起跳前和落地后的地板支撑先后发生，接触表列过绳时并存的双手持柄与双脚避绳。用原生有效次数达到该序号判定，不以任意一次腾空或绳的总圈数代替。" + }, + { + "id": "skip_13_completed", + "zh": "第13次过绳", + "mode": "event", + "when": null, + "after": [ + "skip_12_completed" + ], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "累计完成第13次有效正摇双脚过绳；场地或侧面视角可见红白串珠绳先从头后越过头顶至头前，再从前向后穿过腾空的两脚下方：过脚时全身无部位触地，两脚离地时差不超过0.10秒，两脚过绳时差不超过0.15秒，绳在每只脚下的向后速度经评分器保留三位小数后至少为1.0米/秒，取样绳点低于该脚最低点加0.0095米。绳未先过头、仅过一脚、反摇、过脚时有身体部位触地或速度及同步条件不合格均不计数；额外空跳不计，一圈至多计一次，漏跳后可重新摇绳，已完成的有效次数不撤销。", + "step": { + "cat": "本体", + "item": "跳跃" + }, + "contact": [ + { + "who": "双臂协作", + "how": "抓握", + "plan": 1, + "note_zh": "左手灵巧手握左侧蓝色绳柄，右手灵巧手握右侧蓝色绳柄，两臂共同驱动红白串珠绳正向转动。" + }, + { + "who": "腿/躯干", + "how": "避开", + "plan": 1, + "note_zh": "双脚同时腾空，让转动的红白串珠绳从两脚下方通过；双手此时仍各握一个绳柄。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:3-34", + "inputs/task/standard/instruction.md:5-33", + "inputs/task/privileged/tests/rfo_task.py:121-125", + "inputs/task/privileged/tests/rfo_task.py:365-379", + "inputs/task/privileged/tests/rfo_task.py:383-451", + "inputs/task/privileged/tests/rfo_task.py:457-473", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/tests/rfo_task.py:21-22", + "inputs/task/privileged/tests/rfo_task.py:221-222" + ], + "conf": 6, + "note_zh": "按次数逐一建节点；该圈的过头、起跳、过脚不拆分。双脚起跳前和落地后的地板支撑先后发生，接触表列过绳时并存的双手持柄与双脚避绳。用原生有效次数达到该序号判定，不以任意一次腾空或绳的总圈数代替。" + }, + { + "id": "skip_14_completed", + "zh": "第14次过绳", + "mode": "event", + "when": null, + "after": [ + "skip_13_completed" + ], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "累计完成第14次有效正摇双脚过绳；场地或侧面视角可见红白串珠绳先从头后越过头顶至头前，再从前向后穿过腾空的两脚下方：过脚时全身无部位触地，两脚离地时差不超过0.10秒，两脚过绳时差不超过0.15秒，绳在每只脚下的向后速度经评分器保留三位小数后至少为1.0米/秒，取样绳点低于该脚最低点加0.0095米。绳未先过头、仅过一脚、反摇、过脚时有身体部位触地或速度及同步条件不合格均不计数；额外空跳不计，一圈至多计一次，漏跳后可重新摇绳，已完成的有效次数不撤销。", + "step": { + "cat": "本体", + "item": "跳跃" + }, + "contact": [ + { + "who": "双臂协作", + "how": "抓握", + "plan": 1, + "note_zh": "左手灵巧手握左侧蓝色绳柄，右手灵巧手握右侧蓝色绳柄，两臂共同驱动红白串珠绳正向转动。" + }, + { + "who": "腿/躯干", + "how": "避开", + "plan": 1, + "note_zh": "双脚同时腾空，让转动的红白串珠绳从两脚下方通过；双手此时仍各握一个绳柄。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:3-34", + "inputs/task/standard/instruction.md:5-33", + "inputs/task/privileged/tests/rfo_task.py:121-125", + "inputs/task/privileged/tests/rfo_task.py:365-379", + "inputs/task/privileged/tests/rfo_task.py:383-451", + "inputs/task/privileged/tests/rfo_task.py:457-473", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/tests/rfo_task.py:21-22", + "inputs/task/privileged/tests/rfo_task.py:221-222" + ], + "conf": 6, + "note_zh": "按次数逐一建节点；该圈的过头、起跳、过脚不拆分。双脚起跳前和落地后的地板支撑先后发生，接触表列过绳时并存的双手持柄与双脚避绳。用原生有效次数达到该序号判定，不以任意一次腾空或绳的总圈数代替。" + }, + { + "id": "skip_15_completed", + "zh": "第15次过绳", + "mode": "event", + "when": null, + "after": [ + "skip_14_completed" + ], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "累计完成第15次有效正摇双脚过绳；场地或侧面视角可见红白串珠绳先从头后越过头顶至头前，再从前向后穿过腾空的两脚下方：过脚时全身无部位触地，两脚离地时差不超过0.10秒，两脚过绳时差不超过0.15秒，绳在每只脚下的向后速度经评分器保留三位小数后至少为1.0米/秒，取样绳点低于该脚最低点加0.0095米。绳未先过头、仅过一脚、反摇、过脚时有身体部位触地或速度及同步条件不合格均不计数；额外空跳不计，一圈至多计一次，漏跳后可重新摇绳，已完成的有效次数不撤销。", + "step": { + "cat": "本体", + "item": "跳跃" + }, + "contact": [ + { + "who": "双臂协作", + "how": "抓握", + "plan": 1, + "note_zh": "左手灵巧手握左侧蓝色绳柄，右手灵巧手握右侧蓝色绳柄，两臂共同驱动红白串珠绳正向转动。" + }, + { + "who": "腿/躯干", + "how": "避开", + "plan": 1, + "note_zh": "双脚同时腾空，让转动的红白串珠绳从两脚下方通过；双手此时仍各握一个绳柄。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:3-34", + "inputs/task/standard/instruction.md:5-33", + "inputs/task/privileged/tests/rfo_task.py:121-125", + "inputs/task/privileged/tests/rfo_task.py:365-379", + "inputs/task/privileged/tests/rfo_task.py:383-451", + "inputs/task/privileged/tests/rfo_task.py:457-473", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/tests/rfo_task.py:21-22", + "inputs/task/privileged/tests/rfo_task.py:221-222" + ], + "conf": 6, + "note_zh": "按次数逐一建节点；该圈的过头、起跳、过脚不拆分。双脚起跳前和落地后的地板支撑先后发生，接触表列过绳时并存的双手持柄与双脚避绳。用原生有效次数达到该序号判定，不以任意一次腾空或绳的总圈数代替。" + }, + { + "id": "skip_16_completed", + "zh": "第16次过绳", + "mode": "event", + "when": null, + "after": [ + "skip_15_completed" + ], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "累计完成第16次有效正摇双脚过绳；场地或侧面视角可见红白串珠绳先从头后越过头顶至头前，再从前向后穿过腾空的两脚下方：过脚时全身无部位触地，两脚离地时差不超过0.10秒，两脚过绳时差不超过0.15秒，绳在每只脚下的向后速度经评分器保留三位小数后至少为1.0米/秒，取样绳点低于该脚最低点加0.0095米。绳未先过头、仅过一脚、反摇、过脚时有身体部位触地或速度及同步条件不合格均不计数；额外空跳不计，一圈至多计一次，漏跳后可重新摇绳，已完成的有效次数不撤销。", + "step": { + "cat": "本体", + "item": "跳跃" + }, + "contact": [ + { + "who": "双臂协作", + "how": "抓握", + "plan": 1, + "note_zh": "左手灵巧手握左侧蓝色绳柄，右手灵巧手握右侧蓝色绳柄，两臂共同驱动红白串珠绳正向转动。" + }, + { + "who": "腿/躯干", + "how": "避开", + "plan": 1, + "note_zh": "双脚同时腾空，让转动的红白串珠绳从两脚下方通过；双手此时仍各握一个绳柄。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:3-34", + "inputs/task/standard/instruction.md:5-33", + "inputs/task/privileged/tests/rfo_task.py:121-125", + "inputs/task/privileged/tests/rfo_task.py:365-379", + "inputs/task/privileged/tests/rfo_task.py:383-451", + "inputs/task/privileged/tests/rfo_task.py:457-473", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/tests/rfo_task.py:21-22", + "inputs/task/privileged/tests/rfo_task.py:221-222" + ], + "conf": 6, + "note_zh": "按次数逐一建节点；该圈的过头、起跳、过脚不拆分。双脚起跳前和落地后的地板支撑先后发生，接触表列过绳时并存的双手持柄与双脚避绳。用原生有效次数达到该序号判定，不以任意一次腾空或绳的总圈数代替。" + }, + { + "id": "skip_17_completed", + "zh": "第17跳站稳", + "mode": "final", + "when": null, + "after": [ + "skip_16_completed" + ], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "累计完成第17次有效正摇双脚过绳；场地或侧面视角可见红白串珠绳先从头后越过头顶至头前，再从前向后穿过腾空的两脚下方：过脚时全身无部位触地，两脚离地时差不超过0.10秒，两脚过绳时差不超过0.15秒，绳在每只脚下的向后速度经评分器保留三位小数后至少为1.0米/秒，取样绳点低于该脚最低点加0.0095米。第17次过绳后首次触地仅能是脚，随后连续100个10毫秒控制步仅脚触地、躯干竖直余弦至少0.94且骨盆三维速度不超过0.35米/秒；未落地或未站满该窗口均未完成。", + "step": { + "cat": "本体", + "item": "跳跃" + }, + "contact": [ + { + "who": "双臂协作", + "how": "抓握", + "plan": 1, + "note_zh": "左手灵巧手握左侧蓝色绳柄，右手灵巧手握右侧蓝色绳柄，两臂共同驱动红白串珠绳正向转动。" + }, + { + "who": "腿/躯干", + "how": "避开", + "plan": 1, + "note_zh": "双脚同时腾空，让转动的红白串珠绳从两脚下方通过；双手此时仍各握一个绳柄。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:3-34", + "inputs/task/standard/instruction.md:5-33", + "inputs/task/privileged/tests/rfo_task.py:121-125", + "inputs/task/privileged/tests/rfo_task.py:365-379", + "inputs/task/privileged/tests/rfo_task.py:383-451", + "inputs/task/privileged/tests/rfo_task.py:457-473", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159", + "inputs/task/privileged/tests/rfo_task.py:21-22", + "inputs/task/privileged/tests/rfo_task.py:221-222", + "inputs/task/privileged/tests/rfo_task.py:400-407", + "inputs/task/privileged/tests/rfo_task.py:495-497", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:174-184", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:292-301", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:1103-1105" + ], + "conf": 6, + "note_zh": "按次数逐一建节点；该圈的过头、起跳、过脚不拆分。双脚起跳前和落地后的地板支撑先后发生，接触表列过绳时并存的双手持柄与双脚避绳。按S10将最后过绳、脚先落地、站稳合为一次连续跳跃的完整验收，不能在空中暂停；末次采用final且不从初始站立计分。站稳使用原生100步保持计数，故hold_s不再叠加窗口；需补记该计数及落地标志。" + } + ], + "constraints": [ + { + "id": "handle_released", + "zh": "绳柄脱手", + "never": null, + "criterion_zh": "任一蓝色绳柄未被其对应手握持持续至少0.05秒即违规；握持按绳柄同时接触该手的手指以及拇指或手掌判定，瞬时不足0.05秒的缺失不直接终止。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:391-399", + "inputs/task/privileged/instruction.md:23", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/realhand.py:833-853" + ] + }, + { + "id": "handle_on_floor", + "zh": "绳柄触地", + "never": null, + "criterion_zh": "任一蓝色绳柄接触地面即违规，无脱手宽限；红白绳体触地本身不属于此规则。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:391-393" + ] + }, + { + "id": "robot_fell", + "zh": "机器人跌倒", + "never": null, + "criterion_zh": "骨盆高度低于0.35米，或除左右脚外任意机器人部位接触地面，即判跌倒；通用跌倒检查在每个10毫秒控制步执行。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:125", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:272-282", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/task.py:21-23" + ] + }, + { + "id": "last_landing_not_feet", + "zh": "末跳非脚落地", + "never": null, + "criterion_zh": "第17次有效过绳后，身体首次恢复地面接触时，接触部位包含脚以外的任何部位，即违规；脚接触与其他部位同时发生也不合格。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:400-407" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "valid_skip_count", + "level": "new_code", + "source": "Log Task.trackers.count and Referee.stages; inputs/task/privileged/tests/rfo_task.py:469-473; inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159" + }, + { + "atom": "last_landing_feet_first", + "level": "new_code", + "source": "Log Task.trackers.landing and Task.trackers.landed; inputs/task/privileged/tests/rfo_task.py:400-407" + }, + { + "atom": "standing_hold_steps", + "level": "new_code", + "source": "Log Referee._hold.stand, Referee.standing() and the stand stage; inputs/task/privileged/tests/rfo_task.py:495-497; inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:180-184,292-301" + }, + { + "atom": "handle_release_duration", + "level": "new_code", + "source": "Log each hand holding its own handle, Task.trackers.unheld and Referee.time; inputs/task/privileged/tests/rfo_task.py:391-399" + }, + { + "atom": "handle_ground_contact", + "level": "new_code", + "source": "Log each handle contact with Referee.ground; inputs/task/privileged/tests/rfo_task.py:391-393" + }, + { + "atom": "fall_state", + "level": "new_code", + "source": "Log pelvis height and Referee.parts_touching(Referee.ground) at control steps; inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:276-282" + }, + { + "atom": "task_success", + "level": "new_code", + "source": "Log the benchmark per-frame Referee.success flag and simulation time; inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:174-184,284-289" + } + ] + }, + "native_metrics": [ + { + "name": "count", + "unit": "次", + "threshold": 17, + "direction": "higher_better" + } + ], + "open_questions_zh": [ + "无成功示例：inputs/oracle_none.txt说明16条目录运行均失败；没有成功视频可核对可见线索，步骤据任务文本、初始设置与评分器规划（N2）。红白绳、蓝色绳柄及相机线索来自场景代码，并非已观察的成功画面。", + "inputs/states/index.json的logs为空，全部17个节点的when、全部约束的never及success均为null（P8）；atoms标记应补记的原生状态为new_code，不能据代码臆造已有日志字段。无重放可检查到达率或判分一致率，细绳过脚和毫秒级同步在普通视频中可能无法精确确认。", + "粒度采用S5及S10：每次有效跳绳一节点，共17个；第17次过绳后的脚先落地与连续站稳是该次跳跃的验收，不再建立第18个站立节点。此前漏跳只不计数，不清零已有有效次数，不要求17次无失误连续完成。", + "两模式任务代码逐字相同、instance.json内容相同，世界状态成功核心一致；特权评分还要求轨迹有效、两个新进程重放均成功且最终状态哈希一致（inputs/grader/efe_embodied-olympics-fun-files/grade.py:87-115），标准模式读取最终现场判定并验证交付状态（inputs/grader/efe_standard-files/v2_1/grade.py:89-100）。这些交付及重放有效性条件不是世界状态，不建节点，也不伪造success原子。", + "文本与实现的数值边界差异：instruction.md:23说脱手超过0.05秒失败，代码在至少0.05秒即失败（inputs/task/privileged/tests/rfo_task.py:398）；脚下绳速先保留三位小数再与1.0米/秒比较（同文件:443-465）。采用实现边界。通用跌倒与站稳按10毫秒控制步检查，过头、过脚和脱手按1毫秒物理步检查（inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:266-301），文本概括为逐物理步判定。", + "摇绳累计圈数rope_turns仅记录、不判成功（inputs/task/privileged/tests/rfo_task.py:475-497）；不将其作为步骤或约束，也不以总圈数替代有效跳数。60秒及6000控制步是时限，不是约束；成功须在截止检查前完成末次站稳，同一控制步先检查成功再检查超时（inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:283-287）。" + ], + "conf": 6, + "conf_reason_zh": "17次计数、依赖顺序、阈值、初始持柄和两模式公共规则均有源码依据；缺少成功视频及任何状态日志，所有状态绑定和实际可见性未验证，末次合并按S10解释。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "17个计数节点（S5）、链式 after、第17节点吸收末次落地与站稳1秒、四条立即终止约束均与说明和评分器（rfo_task.py:383-470）一致；无日志可用，when/never/success 为 null 并已在 open_questions_zh 说明。未发现需要修改内容的错误。", + "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：['skip_1_completed', 'skip_2_completed', 'skip_3_completed', 'skip_4_completed', 'skip_5_completed', 'skip_6_completed', 'skip_7_completed', 'skip_8_completed', 'skip_9_completed', 'skip_10_completed', 'skip_11_completed', 'skip_12_completed', 'skip_13_completed', 'skip_14_completed', 'skip_15_completed', 'skip_16_completed', 'skip_17_completed']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:48:10Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__jump_rope/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__jump_rope/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "0189814ade5f4a973221a22d6a2bc64f9680a8a5875992b3c361c4c711266977", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/embodied-olympics-fun/jump_rope/trajectories/embodied-olympics-fun-jump-rope-i00-privileged--dkyvnna.json b/phase2/embodied-olympics-fun/jump_rope/trajectories/embodied-olympics-fun-jump-rope-i00-privileged--dkyvnna.json new file mode 100644 index 0000000000000000000000000000000000000000..e4fbc12ad3200f9e84a3491d046722ca1e5a40b2 --- /dev/null +++ b/phase2/embodied-olympics-fun/jump_rope/trajectories/embodied-olympics-fun-jump-rope-i00-privileged--dkyvnna.json @@ -0,0 +1,379 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-jump-rope-i00-privileged--dkyvnna", + "taskKey": "embodied-olympics-fun/jump_rope", + "dag_sha256": "6defa81f79ee939b94ca5f1fd765d509ad9fae9b66e1b821a8e8b0890a4402d9", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "replay_success": 0, + "deterministic": 1, + "n_actions": 50, + "sim_steps": 50, + "horizon": 6000, + "episode_over": 0, + "progress": 0.0, + "stages_reached": 0, + "stages_total": 18, + "fell": 0, + "replay_s": 18.23, + "bad_trajectory": 0, + "metric_count": 0.0, + "metric_rope_turns": 0.006, + "grader_error": 0, + "video_rendered": 1, + "agent_sessions": 229, + "agent_resets": 0, + "agent_replays": 0, + "agent_cost_self_reported": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 3.705, + "peak_processes": 20, + "peak_disk_gb": 0.244, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0, + "session_log_tampered": 1, + "session_log_from_monitor": 1 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-jump-rope-i00-privileged--dkyvnna/replay.mp4", + "duration_s": 0.52, + "fps": 25.0, + "frames": 13, + "sha256": "d4563dc3586530ca47f97e6d26ad925a30447f576d09964ad544871a45bbf28b" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 17, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "skip_1_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.48, + "log_step": null, + "text_zh": "观察：逐帧检查视频0.00–0.48 s的全部13帧，机器人一直双脚着地站立，双臂保持低位，绳体未越过头顶，也没有双脚腾空过绳，故第1次有效过绳未发生；评分器reward.metric_count=0与画面一致。" + } + ] + }, + "skip_2_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.48, + "log_step": null, + "text_zh": "观察：逐帧检查视频0.00–0.48 s的全部13帧，机器人一直双脚着地站立，双臂保持低位，绳体未越过头顶，也没有双脚腾空过绳，故第2次有效过绳未发生；评分器reward.metric_count=0与画面一致。" + } + ] + }, + "skip_3_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.48, + "log_step": null, + "text_zh": "观察：逐帧检查视频0.00–0.48 s的全部13帧，机器人一直双脚着地站立，双臂保持低位，绳体未越过头顶，也没有双脚腾空过绳，故第3次有效过绳未发生；评分器reward.metric_count=0与画面一致。" + } + ] + }, + "skip_4_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.48, + "log_step": null, + "text_zh": "观察：逐帧检查视频0.00–0.48 s的全部13帧，机器人一直双脚着地站立，双臂保持低位，绳体未越过头顶，也没有双脚腾空过绳，故第4次有效过绳未发生；评分器reward.metric_count=0与画面一致。" + } + ] + }, + "skip_5_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.48, + "log_step": null, + "text_zh": "观察：逐帧检查视频0.00–0.48 s的全部13帧，机器人一直双脚着地站立，双臂保持低位，绳体未越过头顶，也没有双脚腾空过绳，故第5次有效过绳未发生；评分器reward.metric_count=0与画面一致。" + } + ] + }, + "skip_6_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.48, + "log_step": null, + "text_zh": "观察：逐帧检查视频0.00–0.48 s的全部13帧，机器人一直双脚着地站立，双臂保持低位，绳体未越过头顶，也没有双脚腾空过绳，故第6次有效过绳未发生；评分器reward.metric_count=0与画面一致。" + } + ] + }, + "skip_7_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.48, + "log_step": null, + "text_zh": "观察：逐帧检查视频0.00–0.48 s的全部13帧，机器人一直双脚着地站立，双臂保持低位，绳体未越过头顶，也没有双脚腾空过绳，故第7次有效过绳未发生；评分器reward.metric_count=0与画面一致。" + } + ] + }, + "skip_8_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.48, + "log_step": null, + "text_zh": "观察：逐帧检查视频0.00–0.48 s的全部13帧，机器人一直双脚着地站立，双臂保持低位，绳体未越过头顶，也没有双脚腾空过绳，故第8次有效过绳未发生；评分器reward.metric_count=0与画面一致。" + } + ] + }, + "skip_9_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.48, + "log_step": null, + "text_zh": "观察：逐帧检查视频0.00–0.48 s的全部13帧，机器人一直双脚着地站立，双臂保持低位，绳体未越过头顶，也没有双脚腾空过绳，故第9次有效过绳未发生；评分器reward.metric_count=0与画面一致。" + } + ] + }, + "skip_10_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.48, + "log_step": null, + "text_zh": "观察：逐帧检查视频0.00–0.48 s的全部13帧，机器人一直双脚着地站立，双臂保持低位，绳体未越过头顶，也没有双脚腾空过绳，故第10次有效过绳未发生；评分器reward.metric_count=0与画面一致。" + } + ] + }, + "skip_11_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.48, + "log_step": null, + "text_zh": "观察：逐帧检查视频0.00–0.48 s的全部13帧，机器人一直双脚着地站立，双臂保持低位，绳体未越过头顶，也没有双脚腾空过绳，故第11次有效过绳未发生；评分器reward.metric_count=0与画面一致。" + } + ] + }, + "skip_12_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.48, + "log_step": null, + "text_zh": "观察：逐帧检查视频0.00–0.48 s的全部13帧，机器人一直双脚着地站立，双臂保持低位，绳体未越过头顶，也没有双脚腾空过绳，故第12次有效过绳未发生；评分器reward.metric_count=0与画面一致。" + } + ] + }, + "skip_13_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.48, + "log_step": null, + "text_zh": "观察：逐帧检查视频0.00–0.48 s的全部13帧，机器人一直双脚着地站立，双臂保持低位，绳体未越过头顶，也没有双脚腾空过绳，故第13次有效过绳未发生；评分器reward.metric_count=0与画面一致。" + } + ] + }, + "skip_14_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.48, + "log_step": null, + "text_zh": "观察：逐帧检查视频0.00–0.48 s的全部13帧，机器人一直双脚着地站立，双臂保持低位，绳体未越过头顶，也没有双脚腾空过绳，故第14次有效过绳未发生；评分器reward.metric_count=0与画面一致。" + } + ] + }, + "skip_15_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.48, + "log_step": null, + "text_zh": "观察：逐帧检查视频0.00–0.48 s的全部13帧，机器人一直双脚着地站立，双臂保持低位，绳体未越过头顶，也没有双脚腾空过绳，故第15次有效过绳未发生；评分器reward.metric_count=0与画面一致。" + } + ] + }, + "skip_16_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.48, + "log_step": null, + "text_zh": "观察：逐帧检查视频0.00–0.48 s的全部13帧，机器人一直双脚着地站立，双臂保持低位，绳体未越过头顶，也没有双脚腾空过绳，故第16次有效过绳未发生；评分器reward.metric_count=0与画面一致。" + } + ] + }, + "skip_17_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.48, + "log_step": null, + "text_zh": "观察：逐帧检查视频0.00–0.48 s的全部13帧，机器人一直双脚着地站立，双臂保持低位，绳体未越过头顶，也没有双脚腾空过绳，故第17次有效过绳未发生；评分器reward.metric_count=0与画面一致。第17次过绳及其后落地站稳验收均未进入。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "count", + "unit": "次", + "threshold": 17, + "direction": "higher_better", + "value": 0.0 + } + ], + "failure": { + "breakpoint": { + "ref": "skip_1_completed", + "t": 0.0, + "log_step": 67, + "where": "首步前", + "evidence_zh": "观察：视频0.00–0.48 s全程机器人站立，未启动摇绳或起跳；log step 67的交付生成脚本以s.hold()取得保持姿态指令，重复50步后保存轨迹。log step 94检查交付文件actions形状为(50,151)，log step 108中agent明确自述交付文件是“0.5 s of standing still with no jumps”的占位轨迹，与视频及评分器n_actions=50、metric_count=0一致。" + }, + "unmet": { + "skip_1_completed": false, + "skip_2_completed": false, + "skip_3_completed": false, + "skip_4_completed": false, + "skip_5_completed": false, + "skip_6_completed": false, + "skip_7_completed": false, + "skip_8_completed": false, + "skip_9_completed": false, + "skip_10_completed": false, + "skip_11_completed": false, + "skip_12_completed": false, + "skip_13_completed": false, + "skip_14_completed": false, + "skip_15_completed": false, + "skip_16_completed": false, + "skip_17_completed": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 67, + "evidence_zh": "观察：视频0.00–0.48 s全部13帧中，机器人保持站姿，没有驱动绳体过头或双脚起跳；log step 67保存的脚本仅执行50个保持姿态控制步，log step 108中agent也承认提交静站占位轨迹。因此评分尝试中的第1次跳绳从未启动。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志", + "evidence_zh": "推断：会话中agent未实现能稳定摇绳并配合起跳的控制，因而最终保留静站占位轨迹。log step 71、73的多组手臂/手腕控制试验反复返回let go of a handle且HP为空；log step 86的摇绳与跳跃组合试验count始终为0，并返回过脚方向错误、绳柄触地或跌倒。log step 108中agent自述摇绳仍未解决，交付仍是log step 67生成的静站轨迹。这些会话证据支持一般控制能力缺口；具体接触设计、轨迹求解或反馈子类无法据此唯一确定。" + } + ], + "gap_zh": "本次没有状态日志或成功参考视频；交付视频已逐帧完整检查，足以确定未启动跳绳。根因依据会话试验日志作推断，试验中的脱手、碰撞和跌倒不属于本次静站交付的可见现象；这些试验没有本作业可查的视频及完整接触力记录，无法进一步确定控制子类。" + }, + "summary_zh": "评分轨迹仅含静站，视频0.00–0.48 s的全部13帧均未出现摇绳过头或起跳，17个节点全部未达成，有效跳数为0。断点为第1次过绳的首步前，现象为没尝试；log step 67、94、108确认交付的是50步静站占位文件，结合log step 71、73、86的失败试验，根因推断为未解决摇绳与起跳配合的控制问题。无状态日志和成功参考视频，具体控制子类缺乏充分证据。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "交付轨迹仅50步静站（视频0.52 s共13帧），17个节点均未达成、断点首步前、现象没尝试、根因停在控制，均有log step 67/94/108与评分器字段支持；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": "2026-10-29T10:05:06Z", + "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/embodied-olympics-fun-jump-rope-i00-privileged--dkyvnna/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-jump-rope-i00-privileged--dkyvnna/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "90c63aaedc1a46ca04d57eabe57e6f299b5f2a722066ea11c0a1caaabe3b0aaf", + "video_sha256": "d4563dc3586530ca47f97e6d26ad925a30447f576d09964ad544871a45bbf28b", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/jump_rope/trajectories/embodied-olympics-fun-jump-rope-i00-privileged--ruzgafv.json b/phase2/embodied-olympics-fun/jump_rope/trajectories/embodied-olympics-fun-jump-rope-i00-privileged--ruzgafv.json new file mode 100644 index 0000000000000000000000000000000000000000..c418e6d59f3f96d09d5ca6eee1bbc0839449cef9 --- /dev/null +++ b/phase2/embodied-olympics-fun/jump_rope/trajectories/embodied-olympics-fun-jump-rope-i00-privileged--ruzgafv.json @@ -0,0 +1,379 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-jump-rope-i00-privileged--ruzgafv", + "taskKey": "embodied-olympics-fun/jump_rope", + "dag_sha256": "6defa81f79ee939b94ca5f1fd765d509ad9fae9b66e1b821a8e8b0890a4402d9", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "replay_success": 0, + "deterministic": 1, + "n_actions": 1, + "sim_steps": 1, + "horizon": 6000, + "episode_over": 0, + "progress": 0.0, + "stages_reached": 0, + "stages_total": 18, + "fell": 0, + "replay_s": 11.81, + "bad_trajectory": 0, + "metric_count": 0.0, + "metric_rope_turns": 0.0, + "grader_error": 0, + "video_rendered": 1, + "agent_sessions": 89, + "agent_resets": 8, + "agent_replays": 0, + "agent_cost_self_reported": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.961, + "peak_processes": 9, + "peak_disk_gb": 0.837, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0, + "session_log_tampered": 0, + "session_log_from_monitor": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-jump-rope-i00-privileged--ruzgafv/replay.mp4", + "duration_s": 0.04, + "fps": 25.0, + "frames": 1, + "sha256": "194d8afc9eb0efa45b3ff1926b0c4bbaa95f05fd43d3b1f878c48692ecca0082" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 17, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "skip_1_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 329, + "text_zh": "观察：视频 0.00 s（全片唯一帧）中机器人双脚站在木地板上，双臂下垂，绳垂至地面；没有过头、起跳或过脚过程。log step 11 的生成脚本只执行一次 hold 后保存，log step 329 确认最终动作数组为 (1,151)；评分器 metric_count=0、metric_rope_turns=0，故未完成第1次有效过绳。" + } + ] + }, + "skip_2_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 329, + "text_zh": "观察：视频 0.00 s（全片唯一帧）中机器人双脚站在木地板上，双臂下垂，绳垂至地面；没有过头、起跳或过脚过程。log step 11 的生成脚本只执行一次 hold 后保存，log step 329 确认最终动作数组为 (1,151)；评分器 metric_count=0、metric_rope_turns=0，故未完成第2次有效过绳。" + } + ] + }, + "skip_3_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 329, + "text_zh": "观察：视频 0.00 s（全片唯一帧）中机器人双脚站在木地板上，双臂下垂，绳垂至地面；没有过头、起跳或过脚过程。log step 11 的生成脚本只执行一次 hold 后保存，log step 329 确认最终动作数组为 (1,151)；评分器 metric_count=0、metric_rope_turns=0，故未完成第3次有效过绳。" + } + ] + }, + "skip_4_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 329, + "text_zh": "观察：视频 0.00 s（全片唯一帧）中机器人双脚站在木地板上，双臂下垂，绳垂至地面；没有过头、起跳或过脚过程。log step 11 的生成脚本只执行一次 hold 后保存，log step 329 确认最终动作数组为 (1,151)；评分器 metric_count=0、metric_rope_turns=0，故未完成第4次有效过绳。" + } + ] + }, + "skip_5_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 329, + "text_zh": "观察：视频 0.00 s（全片唯一帧）中机器人双脚站在木地板上，双臂下垂，绳垂至地面；没有过头、起跳或过脚过程。log step 11 的生成脚本只执行一次 hold 后保存，log step 329 确认最终动作数组为 (1,151)；评分器 metric_count=0、metric_rope_turns=0，故未完成第5次有效过绳。" + } + ] + }, + "skip_6_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 329, + "text_zh": "观察：视频 0.00 s（全片唯一帧）中机器人双脚站在木地板上，双臂下垂，绳垂至地面；没有过头、起跳或过脚过程。log step 11 的生成脚本只执行一次 hold 后保存，log step 329 确认最终动作数组为 (1,151)；评分器 metric_count=0、metric_rope_turns=0，故未完成第6次有效过绳。" + } + ] + }, + "skip_7_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 329, + "text_zh": "观察：视频 0.00 s（全片唯一帧）中机器人双脚站在木地板上，双臂下垂，绳垂至地面；没有过头、起跳或过脚过程。log step 11 的生成脚本只执行一次 hold 后保存，log step 329 确认最终动作数组为 (1,151)；评分器 metric_count=0、metric_rope_turns=0，故未完成第7次有效过绳。" + } + ] + }, + "skip_8_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 329, + "text_zh": "观察：视频 0.00 s（全片唯一帧）中机器人双脚站在木地板上，双臂下垂，绳垂至地面；没有过头、起跳或过脚过程。log step 11 的生成脚本只执行一次 hold 后保存，log step 329 确认最终动作数组为 (1,151)；评分器 metric_count=0、metric_rope_turns=0，故未完成第8次有效过绳。" + } + ] + }, + "skip_9_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 329, + "text_zh": "观察：视频 0.00 s（全片唯一帧）中机器人双脚站在木地板上，双臂下垂，绳垂至地面；没有过头、起跳或过脚过程。log step 11 的生成脚本只执行一次 hold 后保存，log step 329 确认最终动作数组为 (1,151)；评分器 metric_count=0、metric_rope_turns=0，故未完成第9次有效过绳。" + } + ] + }, + "skip_10_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 329, + "text_zh": "观察：视频 0.00 s（全片唯一帧）中机器人双脚站在木地板上，双臂下垂，绳垂至地面；没有过头、起跳或过脚过程。log step 11 的生成脚本只执行一次 hold 后保存，log step 329 确认最终动作数组为 (1,151)；评分器 metric_count=0、metric_rope_turns=0，故未完成第10次有效过绳。" + } + ] + }, + "skip_11_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 329, + "text_zh": "观察：视频 0.00 s（全片唯一帧）中机器人双脚站在木地板上，双臂下垂，绳垂至地面；没有过头、起跳或过脚过程。log step 11 的生成脚本只执行一次 hold 后保存，log step 329 确认最终动作数组为 (1,151)；评分器 metric_count=0、metric_rope_turns=0，故未完成第11次有效过绳。" + } + ] + }, + "skip_12_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 329, + "text_zh": "观察：视频 0.00 s（全片唯一帧）中机器人双脚站在木地板上，双臂下垂，绳垂至地面；没有过头、起跳或过脚过程。log step 11 的生成脚本只执行一次 hold 后保存，log step 329 确认最终动作数组为 (1,151)；评分器 metric_count=0、metric_rope_turns=0，故未完成第12次有效过绳。" + } + ] + }, + "skip_13_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 329, + "text_zh": "观察：视频 0.00 s（全片唯一帧）中机器人双脚站在木地板上，双臂下垂，绳垂至地面；没有过头、起跳或过脚过程。log step 11 的生成脚本只执行一次 hold 后保存，log step 329 确认最终动作数组为 (1,151)；评分器 metric_count=0、metric_rope_turns=0，故未完成第13次有效过绳。" + } + ] + }, + "skip_14_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 329, + "text_zh": "观察：视频 0.00 s（全片唯一帧）中机器人双脚站在木地板上，双臂下垂，绳垂至地面；没有过头、起跳或过脚过程。log step 11 的生成脚本只执行一次 hold 后保存，log step 329 确认最终动作数组为 (1,151)；评分器 metric_count=0、metric_rope_turns=0，故未完成第14次有效过绳。" + } + ] + }, + "skip_15_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 329, + "text_zh": "观察：视频 0.00 s（全片唯一帧）中机器人双脚站在木地板上，双臂下垂，绳垂至地面；没有过头、起跳或过脚过程。log step 11 的生成脚本只执行一次 hold 后保存，log step 329 确认最终动作数组为 (1,151)；评分器 metric_count=0、metric_rope_turns=0，故未完成第15次有效过绳。" + } + ] + }, + "skip_16_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 329, + "text_zh": "观察：视频 0.00 s（全片唯一帧）中机器人双脚站在木地板上，双臂下垂，绳垂至地面；没有过头、起跳或过脚过程。log step 11 的生成脚本只执行一次 hold 后保存，log step 329 确认最终动作数组为 (1,151)；评分器 metric_count=0、metric_rope_turns=0，故未完成第16次有效过绳。" + } + ] + }, + "skip_17_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 329, + "text_zh": "观察：视频 0.00 s（全片唯一帧）中机器人双脚站在木地板上，双臂下垂，绳垂至地面；没有过头、起跳或过脚过程。log step 11 的生成脚本只执行一次 hold 后保存，log step 329 确认最终动作数组为 (1,151)；评分器 metric_count=0、metric_rope_turns=0，故未完成第17次有效过绳。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "count", + "value": 0.0, + "unit": "次", + "threshold": 17, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "skip_1_completed", + "t": 0.0, + "log_step": 11, + "where": "首步前", + "evidence_zh": "观察：视频 0.00 s 的唯一帧显示机器人仍双脚站立，绳垂在两侧至地面。log step 11 的交付生成脚本仅调用一次 hold、step 和 save_trajectory；log step 309 读回该脚本，log step 329 确认最终文件仅有一行动作，log step 330 也明确自述交付为 one-step hold。评分器 n_actions=1、sim_steps=1、metric_count=0、metric_rope_turns=0，首次摇绳起跳工作尚未启动。" + }, + "unmet": { + "skip_1_completed": false, + "skip_2_completed": false, + "skip_3_completed": false, + "skip_4_completed": false, + "skip_5_completed": false, + "skip_6_completed": false, + "skip_7_completed": false, + "skip_8_completed": false, + "skip_9_completed": false, + "skip_10_completed": false, + "skip_11_completed": false, + "skip_12_completed": false, + "skip_13_completed": false, + "skip_14_completed": false, + "skip_15_completed": false, + "skip_16_completed": false, + "skip_17_completed": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 329, + "evidence_zh": "观察：视频 0.00 s（全片唯一帧）中机器人保持起始站姿，未出现摇绳或腾空；log step 11 仅保存一次保持动作，log step 329 验证最终文件形状为 (1,151)，评分器记录只有一个控制步且绳圈数为0。评分尝试没有开始第1次过绳，会话中的运动测试不在交付内。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志", + "evidence_zh": "推断：会话测试未解决摇绳与双脚腾空的协调，最终未生成替换保持轨迹的有效跳跃。log step 95 的测试评分反馈 count=0，明确报告绳只过右脚、未过双脚，随后跌倒；log step 327 的最后一次测试反馈 count=0 并以 fell 结束。log step 317 的测试脚本仅在 count>0 时保存轨迹，log step 328 的交付说明承认始终未取得有效跳数，log step 329 验证仍交付单步保持。这些是会话控制失败及其与最终交付关系的证据，不把测试跌倒计作评分录像中的现象。" + } + ], + "gap_zh": "观察：inputs/ 仅提供0.04秒、1帧评分视频，没有状态日志或成功参考视频。评分尝试的首步未启动及零有效次数由画面、交付脚本和评分字段共同确定；控制根因依据会话测试反馈推断，不能从该单帧视频验证测试中的运动或区分接触设计、轨迹求解与反馈控制子类。全部过绳节点均未开始，故不为初始持柄、站立添加节点接触，也不把会话测试接触写入 contact_actual。" + }, + "summary_zh": "观察：评分视频仅有视频 0.00 s 一帧，机器人保持起始站姿；交付为一控制步保持轨迹，评分器有效跳数为0，17个节点全部未达成。推断：断点为第1次过绳的首步前，现象是没尝试；会话测试反复未能协调摇绳与双脚腾空，支持控制根因，最终仍交付原始保持轨迹。观察：无状态日志，单帧评分视频不能验证会话运动细节，根因不细分。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "标注与证据一致：评分视频仅一帧，交付为单步保持（log 11/309/329/330），评分字段 metric_count=0、n_actions=1，17个节点全未达成；断点 skip_1_completed/首步前/没尝试，根因控制来自会话测试证据（log 95、327），符合 CASES.md FAIL-22 的处理，且已声明为推断。", + "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": "2026-10-29T10:04:04Z", + "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/embodied-olympics-fun-jump-rope-i00-privileged--ruzgafv/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-jump-rope-i00-privileged--ruzgafv/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "aa3aa18006e229732580fa8fdddf06e23d9a16a81d0fbfd844b7d808e2302d26", + "video_sha256": "194d8afc9eb0efa45b3ff1926b0c4bbaa95f05fd43d3b1f878c48692ecca0082", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/pole_vault/task_dag.json b/phase2/embodied-olympics-fun/pole_vault/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..dd810a8b7490021d3001efd77a191fb30c162baf --- /dev/null +++ b/phase2/embodied-olympics-fun/pole_vault/task_dag.json @@ -0,0 +1,259 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "embodied-olympics-fun/pole_vault", + "suite": "embodied-olympics-fun", + "base": "embodied-olympics-fun-pole-vault-i00", + "title_zh": "撑杆跳高", + "summary_zh": "机器人用携带的杆完成合法助跑撑杆越过横杆并落地，文本要求落在垫上，裁判要求首次落地两秒后横杆仍未掉落且全程没有规则犯规。", + "embodiment_zh": "两种模式均为Unitree G1双足人形机器人，机身29个关节；左右臂各7个关节，两只末端均为Dex3-1三指灵巧手，以接触和摩擦持杆；双腿支撑并驱动身体，无轮式底盘。开局右手握住2.35米刚性杆中部、左手张开，骨盆位于零线前20米。", + "task_class": "atomic", + "analysis_level": "step", + "spec": { + "version": "1.2", + "task": "embodied-olympics-fun-pole-vault-i00", + "bench": "embodied-olympics-fun", + "progress": "live", + "entities": {}, + "milestones": [ + { + "id": "vault_completed", + "zh": "撑杆越杆落地", + "mode": "event", + "when": null, + "after": [], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "在场地视角中，应看到持杆机器人把杆尖送入跑道末端的深色插斗，助跑离地、手掌高于骨盆地悬挂在杆上，再于同一次腾空中使骨盆和两个踝关节均从两根灰色立柱之间且高于红白横杆通过、全部身体越过横杆近侧平面，随后落下，并在首次落地2.0秒后横杆仍未掉落（横杆中心相对首个物理步下降不超过0.05米）；文本要求落在蓝垫上，但代码实际以过杆后首次接触地面或垫子认定落地。验收还要求插杆时有手握杆、杆与插斗接触且杆尖场地坐标x在插斗前沿约-1.0797米至0.02米之间、z<0且|y|≤0.30米，起跳前0.5秒平均前向速度≥2.5米/秒、起跳时杆偏离竖直≥20度，同次腾空中杆承受≥半个体重且持杆手掌高于骨盆的合格悬挂累计≥0.15秒，三个过杆交点均|y|<2.15米且z>1.0米、全部身体范围x>0.185米，且无下列犯规；只插杆、腾空或过杆而尚未落地并等满两秒均未完成，画面不能证明承重和精确时间时须查裁判事件或补录状态。", + "step": { + "cat": "本体", + "item": "跳跃" + }, + "contact": [ + { + "who": "手指/夹爪", + "how": "抓握", + "plan": 1, + "note_zh": "单手方案：一只灵巧手握住撑杆并在腾空中承重，右手为开局持杆手；完成规则不限定哪只手。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "plan": 1, + "note_zh": "双脚在跑道上助跑、插杆和蹬地时承重，与持杆手的抓握同时发生；其后的离地、松杆和身体任意部位接触落地区域依次发生。" + }, + { + "who": "双臂协作", + "how": "抓握", + "plan": 2, + "note_zh": "双手方案：左右灵巧手分别握住撑杆的两个位置共同植杆并悬挂承重，不要求固定左右手高低。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "plan": 2, + "note_zh": "双脚在跑道上助跑、双手植杆及蹬地时承重，与双手抓杆同时发生；落地可用身体任意部位，不要求站稳。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:3-43", + "inputs/task/standard/instruction.md:5-40", + "inputs/task/privileged/environment/rfo_task.py:186-195", + "inputs/task/privileged/environment/rfo_task.py:292-310", + "inputs/task/privileged/environment/rfo_task.py:352-357", + "inputs/task/privileged/environment/rfo_task.py:372-471", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:236-259", + "docs/DECISIONS.md:135-144" + ], + "conf": 6, + "note_zh": "插杆、起跳悬挂、过杆与落地是同一次不可暂停的动作（S10），四个原生裁判阶段都属于单步验收；不另立助跑、悬挂、落地或横杆保持节点。support_s在同一次腾空内累计合格物理步，短暂失去悬挂条件不清零，非腾空时清零；落地后的两秒由t_land计时，hold_s=0避免再叠加等待。过杆后可持杆或放杆，摔倒本身不失败，横杆保持属于验收和掉杆约束。" + } + ], + "constraints": [ + { + "id": "bar_dropped", + "zh": "横杆掉落", + "never": null, + "criterion_zh": "从首个物理步记录的横杆中心高度起，中心下降严格超过0.05米即失败；落地后的两秒等待期间仍检查，身体或杆造成掉落均算。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:377-382" + ] + }, + { + "id": "early_far_ground_contact", + "zh": "提前触地", + "never": null, + "criterion_zh": "全部身体越过横杆平面前，身体或撑杆与地面、垫子的任一接触点在零线之后（场地坐标x>0）即失败；杆尖与插斗的合法接触不算。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:425-433" + ] + }, + { + "id": "invalid_bar_crossing", + "zh": "违规过杆", + "never": null, + "criterion_zh": "骨盆或任一踝关节从前向后跨越横杆近侧平面时，插值交点须满足横向绝对坐标严格小于2.15米且高度严格大于1.0米；绕过立柱或从杆下经过立即失败。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:434-451" + ] + }, + { + "id": "clearance_without_valid_vault", + "zh": "无效撑杆跳", + "never": null, + "criterion_zh": "全部身体过杆时若已不在同一腾空中，或该腾空中合格悬挂累计不足0.15秒即失败；合格悬挂要求持杆手掌高于骨盆、杆尖插斗且承受至少半个体重，起跳前0.5秒平均前向速度至少2.5米/秒、起跳杆倾角至少20度。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:452-463" + ] + }, + { + "id": "hand_climbed_pole", + "zh": "空中攀杆", + "never": null, + "criterion_zh": "过杆前的腾空中，任一持杆手沿杆距杆尖的距离，严格超过该腾空最初持杆时最高持杆手的距离加0.05米，即失败；手内滑移和松开后向上重新抓握都受此检查。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:401-424" + ] + }, + { + "id": "hand_steadied_bar", + "zh": "手扶横杆", + "never": null, + "criterion_zh": "首次落地前双手任一部位接触横杆的累计时间严格超过0.10秒即失败；两手同时接触的同一物理步只累计一次，非手部的碰杆本身允许。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:381-386" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "vault_completed", + "level": "new_code", + "source": "需补录每帧ref.stages、ref.success及Task.trackers的lift_off、support_s、t_clear、t_land和落地事件on_the_mat；逻辑来源inputs/task/privileged/environment/rfo_task.py:372-471；当前无状态日志，尚不绑定路径。" + }, + { + "atom": "bar_dropped", + "level": "new_code", + "source": "需在物理步级补录对应规则违规事件或原始接触与几何检查量；inputs/task/privileged/environment/rfo_task.py:377-382；当前无日志。" + }, + { + "atom": "early_far_ground_contact", + "level": "new_code", + "source": "需在物理步级补录对应规则违规事件或原始接触与几何检查量；inputs/task/privileged/environment/rfo_task.py:425-433；当前无日志。" + }, + { + "atom": "invalid_bar_crossing", + "level": "new_code", + "source": "需在物理步级补录对应规则违规事件或原始接触与几何检查量；inputs/task/privileged/environment/rfo_task.py:434-451；当前无日志。" + }, + { + "atom": "clearance_without_valid_vault", + "level": "new_code", + "source": "需在物理步级补录对应规则违规事件或原始接触与几何检查量；inputs/task/privileged/environment/rfo_task.py:452-463；当前无日志。" + }, + { + "atom": "hand_climbed_pole", + "level": "new_code", + "source": "需在物理步级补录对应规则违规事件或原始接触与几何检查量；inputs/task/privileged/environment/rfo_task.py:401-424；当前无日志。" + }, + { + "atom": "hand_steadied_bar", + "level": "new_code", + "source": "需在物理步级补录对应规则违规事件或原始接触与几何检查量；inputs/task/privileged/environment/rfo_task.py:381-386；当前无日志。" + } + ] + }, + "native_metrics": [ + { + "name": "support_s", + "unit": "s", + "threshold": 0.15, + "direction": "higher_better", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:413-421", + "inputs/task/privileged/environment/rfo_task.py:452-454" + ] + } + ], + "open_questions_zh": [ + "没有成功示例：inputs/oracle_none.txt说明16次目录运行全部失败，无法参照成功视频；本文件根据任务文本、固定初态及裁判代码制定（暂行：DECISIONS.md N2）。", + "inputs/states/index.json的logs为空；所有节点when、所有约束never及success均为null，补录需求记在atoms中，尚不能验证阈值绑定或与评分结果的一致性（暂行：DECISIONS.md P8）。", + "任务设计问题：文本要求落在垫上，但评分器过杆后只要身体首次接触ref.ground就设置t_land，on_the_mat仅写入landing事件，从未参与成功判断；所以落在垫外地面也可满足其落地检查（inputs/task/privileged/environment/rfo_task.py:387-393,467-471）。文本中的落垫要求未受评，作为这次完整跳跃的文本目标保留说明，不另拆连续动作节点（N1、S9、S10）。", + "采用S10以及D11对本任务的现行解释：plant、take-off、clearance、bar stays四个阶段合为一次撑杆跳步骤；落地两秒只是验收等待，不是额外步骤。", + "撑杆是每条合法成功路线都需要的工具：起跳阶段和过杆验收均要求持杆、植杆和承重悬挂；reset已将杆置于右手并闭合手指，故不增加拿起工具节点，也不要求成功结束时仍持杆（D2；rfo_task.py:292-310,352-357,413-421,452-463）。", + "两种模式的rfo_task.py与instance.json逐字节一致，物理成功条件相同；privileged还要求提交的有限动作序列在两个新进程回放中都成功且终态哈希一致，standard只评最后一次实时尝试的裁判结果。交付、回放有效性和基础设施条件不属于语义节点，当前success=null也不能表达这些外层差异（inputs/grader/efe_embodied-olympics-fun-files/grade.py:87-115；inputs/grader/efe_standard-files/v2_1/grade.py:93-99；inputs/grader/efe_embodied-olympics-fun-files/standard/server.py:157-178）。", + "代码中的合格悬挂0.15秒是在同一腾空内累计，并非必须连续；起跳速度先四舍五入到0.001米/秒、杆倾角到0.1度后比较，持杆位置先到0.0001米后比较。无状态日志时应以裁判原生数值为准，不用视频估算制造精度（rfo_task.py:396-424）。", + "文本所称横杆仍在托钉上，代码仅检查中心相对首个物理步下落是否严格超过0.05米，不直接检查两端托钉接触；手扶横杆累计检查在首次落地后停止，只有掉杆检查继续到成功。按代码验收，并明确记录此几何代理及检查时段（rfo_task.py:377-386,467-471）。", + "60秒为回合时限，未列为规则约束（P21）；fall_rule=False，未列摔倒约束。数值发散属于执行有效性失败，不据此增设世界状态步骤（rfo_task.py:186-189；inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:866-882）。" + ], + "conf": 6, + "conf_reason_zh": "单次连续撑杆跳的粒度有S10直接依据，初态及所有验收阈值均可由共同裁判代码确认；但没有成功视频和任何状态日志，不能验证实际可行轨迹、数值绑定或视频可见性，且落垫文本与评分器存在明确差异。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "单步（撑杆越杆落地）符合 DECISIONS.md 对 pole_vault 的现行决定（S10：植杆+起跳+过杆为一个连续动作）；六条立即终止约束与评分器失败条件对应，落垫未受评已作为任务设计问题写入 open_questions_zh。无状态日志，when 为 null 可接受。", + "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：['vault_completed']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:45:10Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__pole_vault/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__pole_vault/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "5961e3829f797ec7126a638ee884090f6a57b379697c11892ddbe152965f37cb", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/embodied-olympics-fun/pole_vault/trajectories/embodied-olympics-fun-pole-vault-i00-privileged--dtvmtvt.json b/phase2/embodied-olympics-fun/pole_vault/trajectories/embodied-olympics-fun-pole-vault-i00-privileged--dtvmtvt.json new file mode 100644 index 0000000000000000000000000000000000000000..5bbd1851806e74cd3c01829fe168b6be1180424d --- /dev/null +++ b/phase2/embodied-olympics-fun/pole_vault/trajectories/embodied-olympics-fun-pole-vault-i00-privileged--dtvmtvt.json @@ -0,0 +1,181 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-pole-vault-i00-privileged--dtvmtvt", + "taskKey": "embodied-olympics-fun/pole_vault", + "dag_sha256": "563d65428d5344656f3de560235015c833808d9cce8bab83d845705ceab114c3", + "outcome": { + "success": 0, + "reward": 0.25, + "source": "catalog", + "raw": { + "success": 0, + "replay_success": 0, + "deterministic": 1, + "n_actions": 2073, + "sim_steps": 2073, + "horizon": 6000, + "episode_over": 0, + "progress": 0.25, + "stages_reached": 1, + "stages_total": 4, + "fell": 0, + "replay_s": 26.55, + "bad_trajectory": 0, + "grader_error": 0, + "video_rendered": 1, + "agent_sessions": 123, + "agent_resets": 1053, + "agent_replays": 5, + "agent_cost_self_reported": 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/embodied-olympics-fun-pole-vault-i00-privileged--dtvmtvt/replay.mp4", + "duration_s": 20.76, + "fps": 25.0, + "frames": 519, + "sha256": "a6d3754d1c298163d164ed432ca37cf2ebd2ea77099c893bc8bcecc25796e84a" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "vault_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 18.0, + "log_step": null, + "text_zh": "观察：视频 18.0–20.4 s，机器人沿红色跑道持浅色杆接近蓝垫前的插斗，双脚交替支撑；杆随右手移动，未见左手共同持杆。" + }, + { + "t": 20.48, + "log_step": 363, + "text_zh": "观察：视频 20.4–20.72 s，机器人在蓝垫前俯身、抬臂，身体仍在横杆近侧，未见骨盆与双踝越过横杆或过杆后落地。log step 363 对最终2073行动行文件的验证回放只报告 stages=[plant]、success=false、done=false，plant事件为仿真 20.459 s；评分记录同样仅确认1/4原生阶段，不能视为完成这个完整跳跃节点。" + }, + { + "t": 20.72, + "log_step": 366, + "text_zh": "观察：log step 357 写出的生成脚本在 support_s 首次大于0时跳出并保存；log step 362 修改姿态后重新生成，log step 366 确认该终止条件仍在。视频 20.72 s 的末帧停在横杆前，交付未包含完整悬挂、越杆与落地验收。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "抓握", + "note_zh": "观察：视频 18.0–20.4 s，杆随机器人右手搬运并转向插斗；log step 363 的最终回放plant事件记录一个持杆位置 grip_from_tip_m=[1.1636]，支持单手握杆。未见双手共同抓杆。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：视频 17.0–20.4 s，机器人双脚在红色跑道上交替着地，承重前进到蓝垫前，配合右手持杆。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "support_s", + "value": null, + "unit": "s", + "threshold": 0.15, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "vault_completed", + "t": 20.5, + "log_step": 363, + "where": "操作中", + "evidence_zh": "观察：视频约20.5 s，机器人刚在蓝垫前进行植杆后的抬臂动作，至视频20.72 s结束仍未过横杆。log step 363 的最终交付回放只有plant阶段且done=false；log step 366 确认脚本在首次 support_s>0 后即保存，撑杆跳步骤从这里起未完成。" + }, + "unmet": { + "vault_completed": false + }, + "phenomena": [], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "推断：会话未形成能从植杆继续到持续承重悬挂的控制动作。log step 198 的独立测试输出显示臂姿过渡时杆角度与侧向位置偏离、持杆标志间歇丢失；agent在log step 293 自述固定手掌位置仍使杆在手内转动，log step 325 自述制动后杆尖仍滑动，log step 357 与372 自述最终仅获得0.001秒合格承重。最终脚本（log step 357、362、366）只保存至首次正承重，log step 363 回放也仅到plant；这些会话证据支持一般控制困难，不能确定具体控制子类。", + "source_zh": "日志 / agent 自述" + } + ], + "gap_zh": "视频早段机器人在画面外；末段视角较远，无法逐帧测定手指接触、杆载荷或毫秒级悬挂时间。inputs/没有状态日志，也没有成功参照视频；最终评分器和回放metrics均未给出support_s，0.001秒仅为agent自述，故原生指标值留空。交付在首次正承重附近结束，未显示后续滑杆、跌倒或完整起跳失败，现象层留空；log step 358 的说明中“随后杆滑出插斗”不能证明该事件包含在交付中。log step 6、136 所读源码注释声称本体无法完成任务，但可见日志截断了完整力学论证，不能据此排除所有合法方案或确定为任务设计问题。" + }, + "summary_zh": "机器人持杆助跑到插斗并达到plant阶段，但没有完成“撑杆越杆落地”节点，语义进度为0/1。交付脚本在首次出现正的悬挂承重时间后保存，视频20.72 s仍停在横杆前；会话测试支持植杆转悬挂的控制困难，但交付没有显示后续滑杆，故不标接触失败现象。缺少状态日志与评分器support_s数值，无法确认精细承重过程或进一步区分控制根因。", + "conf": 7, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 6, + "summary_zh": "单节点 vault_completed 未达成、final=false 与评分器 1/4 阶段（plant，sim 20.459 s）一致；视频 17–20.76 s 机器人持杆跑到插斗前，未见过杆。断点在该步骤、where=操作中合理；现象层留空并写明 gap_zh、根因『控制』标为推断，证据克制，未发现需改的错误。", + "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": "2026-10-29T10:03:06Z", + "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/embodied-olympics-fun-pole-vault-i00-privileged--dtvmtvt/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-pole-vault-i00-privileged--dtvmtvt/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "0d6c114c1950b7a04d0b1d683b0889a5fd806b7d4559a452920bd21fff5a0859", + "video_sha256": "a6d3754d1c298163d164ed432ca37cf2ebd2ea77099c893bc8bcecc25796e84a", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/pole_vault/trajectories/embodied-olympics-fun-pole-vault-i00-standard--yzuqd5x.json b/phase2/embodied-olympics-fun/pole_vault/trajectories/embodied-olympics-fun-pole-vault-i00-standard--yzuqd5x.json new file mode 100644 index 0000000000000000000000000000000000000000..9ab697f3c5c5178b362c9cf3e94423ea6d3c7f94 --- /dev/null +++ b/phase2/embodied-olympics-fun/pole_vault/trajectories/embodied-olympics-fun-pole-vault-i00-standard--yzuqd5x.json @@ -0,0 +1,192 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-pole-vault-i00-standard--yzuqd5x", + "taskKey": "embodied-olympics-fun/pole_vault", + "dag_sha256": "563d65428d5344656f3de560235015c833808d9cce8bab83d845705ceab114c3", + "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": 6000, + "sim_steps": 123146, + "calls": 5906, + "rows": 26919, + "resets": 37, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 3792.8, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 1.005, + "peak_processes": 11, + "peak_disk_gb": 0.519, + "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/embodied-olympics-fun-pole-vault-i00-standard--yzuqd5x/replay.mp4", + "duration_s": 60.04, + "fps": 25.0, + "frames": 1501, + "sha256": "42327718f2694f04f2a3f54a3239d26d78954efcc36e94ccd82370ea8fa2a58d" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "vault_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 22.0, + "log_step": null, + "text_zh": "观察：视频 17.0–24.8 s，机器人用右手携杆沿红色跑道步行进入蓝垫前端的狭槽，随后停止前进；视频 24.8–29.4 s 主要是右臂摆杆，未出现身体越过横杆的动作。" + }, + { + "t": 60.0, + "log_step": 112, + "text_zh": "观察：视频 29.4–60.0 s，机器人停在横杆近侧，杆保持倾斜，未腾空越过横杆，也没有过杆后的落地；已检查视频 59.0–60.0 s 的末秒。log step 112 运行最终脚本后，机器人服务返回 outcome=time、stages_reached=1、stages_total=4；这是仿真服务内裁判的阶段回执，表示仅达到插杆阶段。" + }, + { + "t": null, + "log_step": 114, + "text_zh": "观察：log step 114 的机器人服务状态为 episode=38、control_steps=6000、done=True、reason=failed，裁判回执仍为 time 1 4；最后评分回合与视频所示的持杆接近、摆杆后停留一致，完整撑杆越杆落地节点未完成。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "抓握", + "note_zh": "机器人右手握住杆身并携杆步行，随后右臂带动杆摆转（视频 17.0–29.4 s）；杆尖插斗由最后评分回合的裁判插杆阶段回执确认（log step 112、114）。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "机器人双脚交替在红色跑道上承重、向插斗附近步行；视频 22.0–24.0 s 可见脚与跑道的支撑接触，进入蓝垫前端狭槽后下肢被垫体遮挡。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "support_s", + "value": null, + "unit": "s", + "threshold": 0.15, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "vault_completed", + "t": 29.4, + "log_step": 112, + "where": "操作中", + "evidence_zh": "观察：视频 24.8–29.4 s，机器人在横杆近侧调整杆姿；视频 29.4 s 后停止推进，直到视频 60.0 s 仍未起跳越杆。log step 112、114 的最后回合裁判只认可插杆（1/4），未认可起跳悬挂及后续阶段；断点取摆杆结束后未衔接起跳的位置。" + }, + "unmet": { + "vault_completed": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "步态", + "t": 29.4, + "log_step": 112, + "evidence_zh": "观察：视频 22.0–24.0 s，机器人以步行方式接近插斗；视频 29.4–60.0 s，身体一直留在横杆近侧，没有撑杆腾空和过杆。最后回合裁判在 log step 112、114 仅报插杆阶段，起跳悬挂未达；可见问题是步行接近后没有形成所需跳跃，未见翻倒。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志", + "evidence_zh": "推断：agent 未实现从持杆行进、植杆到起跳悬挂的协调控制。会话测试中，log step 44、47、50、52、54 的脚本在运行策略中改变右臂植杆姿态，服务最终均报阶段 0；log step 105、108、110 又改变运行中的肩部摆动时机和幅度，仍均报阶段 0。最终脚本（log step 103 写入、log step 112 运行）仅回放步行，再以零移动指令摆右臂，随后持续取观测，没有起跳控制段，与视频 29.4–60.0 s 停在杆前相符。因此归于控制能力缺口，不能细分为逆解、反馈或具体抓握设计；前述会话测试仅用于根因，不作为评分回合的动作现象。" + } + ], + "gap_zh": "没有状态日志或成功参考视频；固定远景中，进入蓝垫狭槽后双脚、杆尖及手指接触细节被遮挡，不能从画面定量核验插斗接触力、悬挂承重或 support_s。最终回合的裁判阶段回执可确认只完成插杆，但未报告 support_s 数值，故原生指标留 null；会话测试不足以确定起跳控制失败的具体子类，也不能据 agent 自述断言杆长或握法使所有合法路线不可行。" + }, + "summary_zh": "机器人在最后评分回合持杆走到插斗附近并调整右臂，裁判只认可插杆阶段，完整撑杆越杆落地节点为 0/1。视频约 29.4 s 后机器人停在横杆近侧直至视频 60.0 s，没有完成起跳悬挂与过杆，失败现象为步态；会话脚本和多次测试支持尚未实现植杆与起跳衔接的控制。缺少状态日志和成功参考，无法判定具体控制子类或悬挂承重数值。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "节点 vault_completed 未达成、final=false、断点与步态现象、控制根因均有视频帧和 log step 112/114 的服务回执（time 1/4）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": "2026-10-29T10:03:20Z", + "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/embodied-olympics-fun-pole-vault-i00-standard--yzuqd5x/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-pole-vault-i00-standard--yzuqd5x/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "b3daf4a2c245592dd771a9135a49eff2e77ee1d31773faad79542b1725cabbe7", + "video_sha256": "42327718f2694f04f2a3f54a3239d26d78954efcc36e94ccd82370ea8fa2a58d", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/springboard_3m/task_dag.json b/phase2/embodied-olympics-fun/springboard_3m/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..7084bcbe255b45f54980db32b336d8197624be19 --- /dev/null +++ b/phase2/embodied-olympics-fun/springboard_3m/task_dag.json @@ -0,0 +1,436 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "embodied-olympics-fun/springboard_3m", + "suite": "embodied-olympics-fun", + "base": "embodied-olympics-fun-springboard-3m-i00", + "title_zh": "三米板前直体跳水", + "summary_zh": "机器人从三米弹性跳板双脚起跳，以合规直体向前翻转半周并头手先行垂直入水，直到全身入水且全过程无违规。", + "embodiment_zh": "两种模式均为 Unitree G1 人形机器人，29 个身体关节、双腿和双臂，无轮式底盘；左右腕末端均为固定拳垫，无可动手指，本任务没有可抓物体，抓握指令不起作用。脚在弹性板上支撑起跳，双臂用于空中及入水姿态。", + "task_class": "atomic", + "analysis_level": "step", + "spec": { + "version": "1.2", + "task": "embodied-olympics-fun-springboard-3m-i00", + "bench": "embodied-olympics-fun", + "progress": "live", + "entities": {}, + "milestones": [ + { + "id": "forward_straight_dive_completed", + "zh": "前直体跳水", + "mode": "event", + "when": null, + "after": [], + "hold_s": 0, + "goal": true, + "criterion_zh": "场地固定侧视角可见机器人从浅色板端双脚蹬离、向蓝色水面前翻，最后双臂伸过头顶、头朝下穿过水面且双脚也完全没入；只压板、跌落、腾空、翻至头朝下或仅头手入水均未完成，须按顺序通过起跳、入水、全身入水且无规则违规（全身最高碰撞包围盒点严格低于 z=0，完成后不再评水下动作）。验收为双脚离板不同步不超过 0.1 秒、连续腾空至少 0.1 秒且裁判记录的质心最高点较初始升高至少 0.1 米、无二次弹跳及违规器械接触，首次触水累计向前翻转 180±45 度且手或头先入水，双掌中点入水点位于静止板尖前方 0.5—3.0 米且横向偏差不超过 0.5 米，从累计前翻 90 度起至完成保持髋膝弯曲均不超过 20 度、双踝间距不超过 0.12 米、双踝跖屈均至少 0.3 弧度，骨盆过水面时身体线相对头朝下竖直偏差不超过 15 度，整个入水过程肩轴相对板宽方向偏差不超过 20 度、每掌沿身体线超过头中心至少 0.15 米、两掌相距不超过 0.20 米且每掌偏离身体线不超过 0.15 米；细小角度和接触需裁判记录辅助，远景消失本身不足以证明全过程合规。", + "step": { + "cat": "本体", + "item": "翻腾" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "plan": 1, + "note_zh": "双脚在弹性跳板上承重并同时蹬板起跳；这是完整跳水的必需支承接触。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:3-4,18-40", + "inputs/task/standard/instruction.md:3-5", + "inputs/task/privileged/environment/rfo_task.py:36-76,79-87,95-224", + "inputs/task/privileged/environment/rfo_task.py:328-340,390-402,452-476", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178,353-394", + "inputs/oracle/replay.mp4:视频 0.00、0.72、1.28、1.56、1.72、1.84 秒", + "inputs/oracle/run.json:reward（3 阶段通过、重放成功且确定性一致）" + ], + "conf": 6, + "note_zh": "按 DECISIONS.md S10，起跳、空中前翻与入水不能暂停后续做，合为一个连续步骤，三个裁判阶段仅作验收。event 在全身入水且合规时一次记成；不设单独站立、压板、腾空或直体节点。双脚蹬板后整段飞行及穿过绘制水面均无目标接触：水面无碰撞，双臂只调整姿态，没有抓握或扶板；这些依次发生的阶段不重复列成同时接触。" + } + ], + "constraints": [ + { + "id": "illegal_pre_takeoff_contact", + "zh": "起跳前违规接触", + "never": null, + "criterion_zh": "初始双脚落板稳定后、进入腾空阶段前，踝部外壳以上的身体部位接触跳板、支架、混凝土平台或塔体，或身体任何部位接触栏杆；例如手撑板或身体倒在板上即违规，脚和代码明确豁免的踝部外壳接触不违规。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:101-103,140-150,474-476", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:161-164,266-289" + ] + }, + { + "id": "asymmetric_takeoff", + "zh": "双脚离板不同步", + "never": null, + "criterion_zh": "仍有一脚触板时，另一脚自上次触板已超过 0.1 秒；画面可见单脚提前抬起，须结合逐物理步接触计时确认，不能把允许的双脚不离板压板当作违规。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:151-156,336", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:161-164,266-289" + ] + }, + { + "id": "double_bounce", + "zh": "二次弹跳", + "never": null, + "criterion_zh": "双脚离开支撑的飞行持续至少 0.1 秒后，又落回板或支撑面；进入腾空阶段后脚先回到板面上方的板面也会立即判为二次弹跳。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:157-169,336", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:161-164,266-289" + ] + }, + { + "id": "apparatus_contact_in_dive", + "zh": "跳水中触器械", + "never": null, + "criterion_zh": "进入腾空阶段后，身体任一部位再次接触跳板、支架、混凝土平台、塔体或栏杆；空中身体靠近但没有实际接触不触发。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:165-169,474-476", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:161-164,266-289" + ] + }, + { + "id": "insufficient_takeoff", + "zh": "无有效起跳", + "never": null, + "criterion_zh": "首次触及水面时尚未取得起跳阶段：腾空至少 0.1 秒后记录的最大质心高度相对初始高度未上升至少 0.1 米；直接向前跌落入水不能算完成跳水。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:161-184,336", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:161-164,266-289" + ] + }, + { + "id": "wrong_half_turn", + "zh": "前翻半周不足", + "never": null, + "criterion_zh": "首次触水时，身体线在世界 x-z 平面展开累计的向前转角不在 135 至 225 度内；不是仅凭身体某一帧头朝下就算合格。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:137-139,185-188,339", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:161-164,266-289" + ] + }, + { + "id": "wrong_first_entry_part", + "zh": "非头手先入水", + "never": null, + "criterion_zh": "身体碰撞几何包围盒的最低点首次到达 z=0 时，其部位不是左手、右手或头；脚、腿或躯干先入水都失败。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:79-87,173-190", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:161-164,266-289" + ] + }, + { + "id": "entry_outside_region", + "zh": "入水位置越界", + "never": null, + "criterion_zh": "首次触水时，以两掌中心中点的水平位置为入水点：相对静止板尖向前距离小于 0.5 米或大于 3.0 米，或偏离初始朝向中心线超过 0.5 米。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:191-200,226-229,335-339", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:161-164,266-289" + ] + }, + { + "id": "not_straight", + "zh": "直体姿势违规", + "never": null, + "criterion_zh": "累计前翻达到 90 度起及整个入水过程，任一髋或膝相对直线弯曲超过 20 度、双踝距离超过 0.12 米，或任一踝跖屈角小于 0.3 弧度即失败；空中后半程或入水时明显屈髋屈膝、分腿、勾脚均是可见提示。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:49-60,109-123,175-178,213-214,337", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:161-164,266-289" + ] + }, + { + "id": "entry_not_vertical", + "zh": "入水不垂直", + "never": null, + "criterion_zh": "骨盆第一次到达或低于 z=0 时，从双踝中点指向双肩中点的身体线相对头朝下竖直方向偏差超过 15 度；该角度只在骨盆过水面时取样。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:36-41,66,206-210,337", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:161-164,266-289" + ] + }, + { + "id": "entry_twisted", + "zh": "入水扭转", + "never": null, + "criterion_zh": "从首次触水至全身入水的任一物理步，左右肩轴相对跳板宽度正方向的夹角超过 20 度；正侧方相机未必能精确看出轻微扭转。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:67-69,203-212,338", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:161-164,266-289" + ] + }, + { + "id": "arms_not_extended", + "zh": "入水手臂违规", + "never": null, + "criterion_zh": "整个入水过程，任一掌沿身体线超过头中心的距离小于 0.15 米、两掌距离大于 0.20 米，或任一掌偏离身体线超过 0.15 米；可见手臂未伸过头顶或双手散开时不合格。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:61-65,203-218,338", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:161-164,266-289" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "dive_completed", + "level": "new_code", + "source": "需记录每帧原生裁判 ref.success、ref.stages、ref.failed 与 task.trackers.phase；inputs/task/privileged/environment/rfo_task.py:133-224；inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178。目前没有状态日志，未绑定任何路径。" + }, + { + "atom": "illegal_pre_takeoff_contact", + "level": "new_code", + "source": "需新增逐物理步记录：stand 阶段 settled 后 not_feet 与 platform 的接触，以及 robot_mask 与 rails 的接触；inputs/task/privileged/environment/rfo_task.py:101-103,140-150,474-476" + }, + { + "atom": "asymmetric_takeoff", + "level": "new_code", + "source": "需新增逐物理步记录：两脚 on、last_on、ref.time、phase；inputs/task/privileged/environment/rfo_task.py:151-156,336" + }, + { + "atom": "double_bounce", + "level": "new_code", + "source": "需新增逐物理步记录：Flights.cur、Flights.flights、phase 与落回 platform 的接触；Flights.update 见 inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:368-389；inputs/task/privileged/environment/rfo_task.py:157-169,336" + }, + { + "atom": "apparatus_contact_in_dive", + "level": "new_code", + "source": "需新增逐物理步记录：phase 属于 air 或 entry 时 robot_mask 与 platform 的接触；inputs/task/privileged/environment/rfo_task.py:165-169,474-476" + }, + { + "atom": "insufficient_takeoff", + "level": "new_code", + "source": "需新增逐物理步记录：phase、com0、com_max、flights 与 ref.stages 中 take-off；inputs/task/privileged/environment/rfo_task.py:161-184,336" + }, + { + "atom": "wrong_half_turn", + "level": "new_code", + "source": "需新增逐物理步记录：trackers.turn 与 metrics.somersault_deg；inputs/task/privileged/environment/rfo_task.py:137-139,185-188,339" + }, + { + "atom": "wrong_first_entry_part", + "level": "new_code", + "source": "需新增逐物理步记录：lowest_part 的 low、part 与 phase；inputs/task/privileged/environment/rfo_task.py:79-87,173-190" + }, + { + "atom": "entry_outside_region", + "level": "new_code", + "source": "需新增逐物理步记录：first_point、geo.front_x、start[1] 与 metrics.entry_distance_m；inputs/task/privileged/environment/rfo_task.py:191-200,226-229,335-339" + }, + { + "atom": "not_straight", + "level": "new_code", + "source": "需新增逐物理步记录：Body.shape 的 hip、knee、feet、toes 以及 phase、trackers.turn；inputs/task/privileged/environment/rfo_task.py:49-60,109-123,175-178,213-214,337" + }, + { + "atom": "entry_not_vertical", + "level": "new_code", + "source": "需新增逐物理步记录：trackers.hips 与 metrics.entry_tilt_deg；inputs/task/privileged/environment/rfo_task.py:36-41,66,206-210,337" + }, + { + "atom": "entry_twisted", + "level": "new_code", + "source": "需新增逐物理步记录：Body.shape.twist 与入水阶段；inputs/task/privileged/environment/rfo_task.py:67-69,203-212,338" + }, + { + "atom": "arms_not_extended", + "level": "new_code", + "source": "需新增逐物理步记录：Body.shape 的 beyond、apart、off 与入水阶段；inputs/task/privileged/environment/rfo_task.py:61-65,203-218,338" + } + ] + }, + "native_metrics": [ + { + "name": "somersault_deg", + "unit": "deg", + "threshold": 180, + "direction": "target", + "note_zh": "首次触水时累计前翻角，容差 ±45 度，即 [135,225] 度。", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:185-188,339" + ] + }, + { + "name": "entry_distance_m", + "unit": "m", + "threshold": null, + "direction": "target", + "note_zh": "首次触水双掌中点到静止板尖的前向距离，闭区间 [0.5,3.0] 米。", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:191-198,226-229,339" + ] + }, + { + "name": "entry_tilt_deg", + "unit": "deg", + "threshold": 15, + "direction": "lower_better", + "note_zh": "骨盆首次到水面时相对头朝下竖直方向的角度，≤15 度。", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:206-210,337" + ] + }, + { + "name": "entry_hip", + "unit": "deg", + "threshold": 160, + "direction": "higher_better", + "note_zh": "入水全程两髋几何夹角的最小值，180 度为伸直，须≥160 度；空中后半程另逐步检查。", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:49-58,109-115,125-131,175-178,221-222,337" + ] + }, + { + "name": "entry_knee", + "unit": "deg", + "threshold": 160, + "direction": "higher_better", + "note_zh": "入水全程两膝几何夹角的最小值，须≥160 度；空中后半程另逐步检查。", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:49-58,109-115,125-131,221-222,337" + ] + }, + { + "name": "entry_feet", + "unit": "m", + "threshold": 0.12, + "direction": "lower_better", + "note_zh": "入水全程双踝间距的最大值，须≤0.12 米。", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:59,116-117,125-131,221-222,337" + ] + }, + { + "name": "entry_toes", + "unit": "rad", + "threshold": 0.3, + "direction": "higher_better", + "note_zh": "入水全程两踝跖屈角的最小值，须≥0.3 弧度。", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:60,118-119,125-131,221-222,337" + ] + }, + { + "name": "entry_beyond", + "unit": "m", + "threshold": 0.15, + "direction": "higher_better", + "note_zh": "入水全程双掌沿身体线超过头中心距离的最小值，须≥0.15 米。", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:61-65,125-131,215-222,338" + ] + }, + { + "name": "entry_apart", + "unit": "m", + "threshold": 0.2, + "direction": "lower_better", + "note_zh": "入水全程两掌间距的最大值，须≤0.20 米。", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:64,125-131,215-222,338" + ] + }, + { + "name": "entry_off", + "unit": "m", + "threshold": 0.15, + "direction": "lower_better", + "note_zh": "入水全程掌到身体线距离的最大值，须≤0.15 米。", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:65,125-131,215-222,338" + ] + }, + { + "name": "entry_twist", + "unit": "deg", + "threshold": 20, + "direction": "lower_better", + "note_zh": "入水全程肩轴相对板宽正方向夹角的最大值，须≤20 度。", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:67-69,125-131,211-222,338" + ] + } + ], + "open_questions_zh": [ + "inputs/states/index.json 的 logs 为空；依 P8，唯一节点 when、所有约束 never 及 success 均为 null。atoms 的 new_code 仅说明将来需要记录的逐物理步原生裁判状态，不假定日志路径存在；目前无法以状态日志检验达成或 success 与评分一致性。", + "录像仅 1.88 秒、25 帧/秒、固定侧方远景：可见站在板端、蹬离、前翻、头向下入水，末帧尚有很小的露出形状；裁判成功发生在第 185 个控制步，而渲染每 4 步拍一帧，末帧只到第 184 步，故视频未独立展示全身入水瞬间（rcb_rfo.py:215-233）。成功样例的 reward 确认 3 阶段完成，但录像不能独立核实毫米级接触、0.1 秒窗口及全部姿态阈值；不可用样例的 1.84 秒作为通用完成阈值。", + "文本与评分器边界差异：说明写髋膝弯曲小于 20 度、部分规则写扭转小于 20 度；rfo_task.py:112-119,211 只在超过 20 度时判负，故判据按评分器使用≤20 度。说明说起跳前仅脚可触板，代码 :102-103 还豁免踝部外壳；:142-146 在初始双脚均触板稳定前暂不作站立阶段接触判负。保留这些差异，不增加步骤。", + "两模式跳水几何判据一致，rfo_task.py 文件逐字相同；特权模式还要求轨迹在两个独立进程均成功且最终状态哈希一致（grade.py:106-115），标准模式采现场裁判结果而不重放评分（standard/server.py:6-12,157-178；efe_standard-files/v2_1/grade.py:93-99）。这些交付与复现条件不作为世界状态节点，且当前无状态日志，success 留 null。", + "板尖最大下挠 board_tip_lowest_m 由 rfo_task.py:479-482 记录但没有通过阈值，不是约束；20 秒回合上限不作为约束（P21）。fall_rule=False（:331），不能添加通用低骨盆或身体倾倒约束；有意倒置入水是目标动作。" + ], + "conf": 6, + "conf_reason_zh": "单节点粒度由任务文本及 S10 支持；初态、两模式共同裁判、全部阈值与阶段顺序均有源码依据，成功录像和原生评分记录支持动作结果，但无状态日志，远景和末帧采样无法独立检查全过程细小接触及姿态，故机器条件全部留空并降低置信度。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "任务文本只描述一个连续动作（双脚起跳、前翻半周、头手先入水垂直入水），单节点 atomic 合理（S10）；十二条约束覆盖评分器全部失败分支（rfo_task.py:133-218），均为立即终止；20 秒上限属 horizon（P21）；无状态日志故 when/never/success 为 null 且已在 open_questions_zh 说明；标签 本体·翻腾 与 腿/躯干·支撑 合理。无需修改。", + "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：['forward_straight_dive_completed']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:44:15Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__springboard_3m/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__springboard_3m/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "6d88cf65562fd493de510ac6f7ca0eb16ba5904299ddd630c8580dbd960332d0", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/embodied-olympics-fun/springboard_3m/trajectories/embodied-olympics-fun-springboard-3m-i00-privileged--4kpacdj.json b/phase2/embodied-olympics-fun/springboard_3m/trajectories/embodied-olympics-fun-springboard-3m-i00-privileged--4kpacdj.json new file mode 100644 index 0000000000000000000000000000000000000000..03673efacf84b0ad1301015f12c020ff2266c42a --- /dev/null +++ b/phase2/embodied-olympics-fun/springboard_3m/trajectories/embodied-olympics-fun-springboard-3m-i00-privileged--4kpacdj.json @@ -0,0 +1,252 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-springboard-3m-i00-privileged--4kpacdj", + "taskKey": "embodied-olympics-fun/springboard_3m", + "dag_sha256": "da2d305d2df479b38c8f6f8ea05ea8ab8944e53746c08b499d4a5b65dee06ba4", + "outcome": { + "success": 0, + "reward": 0.3333, + "source": "catalog", + "raw": { + "success": 0, + "replay_success": 0, + "deterministic": 1, + "n_actions": 150, + "sim_steps": 150, + "horizon": 2000, + "episode_over": 1, + "progress": 0.3333, + "stages_reached": 1, + "stages_total": 3, + "fell": 0, + "replay_s": 2.87, + "bad_trajectory": 0, + "metric_board_tip_lowest_m": 0.1637, + "grader_error": 0, + "video_rendered": 1, + "agent_sessions": 10, + "agent_resets": 37, + "agent_replays": 12, + "agent_cost_self_reported": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.706, + "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, + "session_log_tampered": 0, + "session_log_from_monitor": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-springboard-3m-i00-privileged--4kpacdj/replay.mp4", + "duration_s": 1.52, + "fps": 25.0, + "frames": 38, + "sha256": "4ff026c4183aa4dc88b246880e90a2db36db567fac0d2eb3e7b614033b122ce4" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "forward_straight_dive_completed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.72, + "log_step": null, + "text_zh": "观察：已检查本次视频全部帧。视频 0.00—0.76 s，机器人在浅色板端由站立转为屈膝下蹲，随后伸腿并举起双臂；视频约 0.96—1.12 s 双脚离板腾空。" + }, + { + "t": 1.48, + "log_step": 117, + "text_zh": "观察：视频 1.12—1.48 s，机器人头和上身向画面左侧（跳板后方）倾倒，逐渐接近水平并落回板端附近，始终未向前翻至头朝下入水，更未全身入水。log step 117 明确恢复最终交付文件并两次独立回放，均为 150 行、同一终态哈希且仅到 take-off 阶段，重新生成后仍报告空中触跳板、支架或栏杆而终止；评分记录 stages_reached=1/3、success=0、deterministic=1 与画面相符。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：视频 0.00—约 0.92 s，机器人的双脚在弹性跳板端承重、下蹲压板并蹬离；双臂摆至头上方，未见手扶器械。视频 1.40—1.48 s 身体落回板端附近，log step 117 的最终文件回放确认发生空中触器械；具体碰撞身体部位未记录，不将其当作新的主动支撑方式。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "somersault_deg", + "unit": "deg", + "threshold": 180, + "direction": "target", + "value": null + }, + { + "name": "entry_distance_m", + "unit": "m", + "threshold": null, + "direction": "target", + "value": null + }, + { + "name": "entry_tilt_deg", + "unit": "deg", + "threshold": 15, + "direction": "lower_better", + "value": null + }, + { + "name": "entry_hip", + "unit": "deg", + "threshold": 160, + "direction": "higher_better", + "value": null + }, + { + "name": "entry_knee", + "unit": "deg", + "threshold": 160, + "direction": "higher_better", + "value": null + }, + { + "name": "entry_feet", + "unit": "m", + "threshold": 0.12, + "direction": "lower_better", + "value": null + }, + { + "name": "entry_toes", + "unit": "rad", + "threshold": 0.3, + "direction": "higher_better", + "value": null + }, + { + "name": "entry_beyond", + "unit": "m", + "threshold": 0.15, + "direction": "higher_better", + "value": null + }, + { + "name": "entry_apart", + "unit": "m", + "threshold": 0.2, + "direction": "lower_better", + "value": null + }, + { + "name": "entry_off", + "unit": "m", + "threshold": 0.15, + "direction": "lower_better", + "value": null + }, + { + "name": "entry_twist", + "unit": "deg", + "threshold": 20, + "direction": "lower_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "forward_straight_dive_completed", + "t": 1.4, + "log_step": 117, + "where": "操作中", + "evidence_zh": "观察：视频约 0.96 s 机器人已蹬板腾空，视频 1.12—1.48 s 却持续向后旋转并落回板端附近，没有完成向前半周翻转或入水。log step 117 的最终交付文件两次独立回放仅通过 take-off，随后裁判报告 touched the springboard, its stand or the railings during the dive (D 8.5.3)，回合终止；故失败步骤为前直体跳水。" + }, + "unmet": { + "forward_straight_dive_completed": false + }, + "phenomena": [ + { + "cat": "运动", + "item": "被挡/碰撞", + "t": 1.48, + "log_step": 117, + "evidence_zh": "观察：视频 1.28—1.48 s 机器人在板端上方由后仰变成近水平，向器械方向下落，未进入水面区域。log step 117 对最终交付轨迹的裁判回放明确报告跳水中触及跳板、支架或栏杆并终止；画面支持落回板端，精确碰撞对由远景无法分辨。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "evidence_zh": "推断：起跳控制未产生任务所需的向前飞行和前翻运动。log step 103 的交付生成脚本明确以双脚蹬板、摆臂和空中直体为目标；log step 112 对之后保存且在 log step 117 恢复验证的轨迹，脚本诊断输出向后水平速度及负向转角变化率；视频 1.00—1.48 s 独立显示上身向后转、身体回落到板端。控制输出与前翻目标不符，但现有证据不足以区分接触布局、轨迹生成和反馈控制的具体责任。", + "source_zh": "日志" + } + ], + "gap_zh": "没有状态日志或逐物理步接触对；固定远景不能确认末段具体哪个身体部位碰到哪件器械，碰撞事实由最终交付回放裁判记录辅助确认。log step 112 的质心与角速度数值是 agent 脚本诊断，不能独立核实精确起跳力矩机制，因此根因只标控制，不细分；入水前已终止，DAG 所列入水原生指标均未报告。" + }, + "summary_zh": "机器人完成下蹲压板和腾空，裁判仅通过起跳阶段；完整前直体跳水节点未达成，最终为 0/1。视频 1.12—1.48 s 显示机器人向后旋转并回落板端，最终文件回放确认空中触器械而终止，支持被挡/碰撞现象及控制根因。无状态日志，精确碰撞部位及起跳控制失配的具体机制无法确认。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "节点未达成、断点、现象与根因均与视频和日志相符：视频 0.96 s 腾空后向后旋转回落板端，裁判回放报告触器械，仅通过起跳阶段。", + "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": "2026-10-29T10:07:18Z", + "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/embodied-olympics-fun-springboard-3m-i00-privileged--4kpacdj/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-springboard-3m-i00-privileged--4kpacdj/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "c0923c2f2505e7f516c1d0bc8307d430e60131acc72bee4ba4639cf58fde87b2", + "video_sha256": "4ff026c4183aa4dc88b246880e90a2db36db567fac0d2eb3e7b614033b122ce4", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/springboard_3m/trajectories/embodied-olympics-fun-springboard-3m-i00-privileged--vjaxtwo.json b/phase2/embodied-olympics-fun/springboard_3m/trajectories/embodied-olympics-fun-springboard-3m-i00-privileged--vjaxtwo.json new file mode 100644 index 0000000000000000000000000000000000000000..dd851bd95ea9f3a2690a1ffab2dbf58ec01acbcf --- /dev/null +++ b/phase2/embodied-olympics-fun/springboard_3m/trajectories/embodied-olympics-fun-springboard-3m-i00-privileged--vjaxtwo.json @@ -0,0 +1,252 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-springboard-3m-i00-privileged--vjaxtwo", + "taskKey": "embodied-olympics-fun/springboard_3m", + "dag_sha256": "da2d305d2df479b38c8f6f8ea05ea8ab8944e53746c08b499d4a5b65dee06ba4", + "outcome": { + "success": 1, + "reward": 1.0, + "source": "catalog", + "raw": { + "success": 1, + "replay_success": 1, + "deterministic": 1, + "n_actions": 184, + "sim_steps": 184, + "horizon": 2000, + "episode_over": 1, + "progress": 1.0, + "stages_reached": 3, + "stages_total": 3, + "fell": 0, + "replay_s": 5.5, + "bad_trajectory": 0, + "metric_board_tip_lowest_m": 0.1758, + "metric_entry_apart": 0.1099, + "metric_entry_beyond": 0.3688, + "metric_entry_distance_m": 1.739, + "metric_entry_feet": 0.0964, + "metric_entry_hip": 171.6261, + "metric_entry_knee": 170.1238, + "metric_entry_off": 0.0557, + "metric_entry_tilt_deg": 0.0, + "metric_entry_toes": 0.4995, + "metric_entry_twist": 0.0572, + "metric_somersault_deg": 166.9, + "grader_error": 0, + "video_rendered": 1, + "agent_sessions": 110, + "agent_resets": 30997, + "agent_replays": 6, + "agent_cost_self_reported": 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/embodied-olympics-fun-springboard-3m-i00-privileged--vjaxtwo/replay.mp4", + "duration_s": 1.88, + "fps": 25.0, + "frames": 47, + "sha256": "9cda11f07e9d0cddeb1e0f528143a53f94f5fd28bd2eb48e81c96b7b1e22bd04" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "forward_straight_dive_completed": { + "method": "visual", + "intervals": [ + [ + 1.84, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.32, + "log_step": null, + "text_zh": "观察：视频 0.00—0.40 s，机器人站在浅色跳板端，随后屈膝压低身体；视频 0.44—0.68 s 伸腿、抬臂，视频 0.72 s 双脚已离开板端，身体向前上方腾空。双脚对跳板的承重和蹬离是本次实际支承接触。" + }, + { + "t": 1.28, + "log_step": null, + "text_zh": "观察：视频 0.80—1.60 s，机器人离板后连续向前翻转，身体由头在上经过近水平转至头在下，双臂沿身体方向伸过头顶；未见落回跳板。已逐帧查看全片，并对视频 0.40—0.84 s 的蹬离过程及视频 1.56—1.84 s 的入水过程放大核对。" + }, + { + "t": 1.84, + "log_step": null, + "text_zh": "观察：视频 1.64—1.84 s，机器人以头手朝下、身体近竖直的姿态穿过蓝色水面；视频 1.76 s 尚有腿脚露出，视频 1.80 s 只剩小段脚部，视频 1.84 s 末帧已无露出身体。故完成时刻以视频 1.84 s 为近似锚点，精度受 25 fps 采样限制。" + }, + { + "t": 1.84, + "log_step": 155, + "text_zh": "观察：inputs/run.json 的 reward 给出 success=1、replay_success=1、deterministic=1、stages_reached=3/3、sim_steps=184，且无评分器错误；log step 155 的最终封装检查报告交付文件为 184 行、生成文件逐字节一致、两次独立重放均成功且末状态一致。该交付结果与视频中的完整跳水相符，不采用早期搜索试次的成功或失败作为本次进度。" + }, + { + "t": 1.84, + "log_step": null, + "text_zh": "观察：本次评分器报告首次入水累计前翻角 166.9 度、入水点距板尖 1.739 m、骨盆过水面倾角 0.0 度；入水髋角最低 171.6261 度、膝角最低 170.1238 度、双踝间距最大 0.0964 m、跖屈角最低 0.4995 rad，双掌间距最大 0.1099 m、超头距离最低 0.3688 m、偏轴距离最大 0.0557 m、肩轴扭转最大 0.0572 度。推断：这些本次裁判数值和三阶段成功共同支持节点的精细验收；单凭侧方远景不能独立量清全部阈值。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：视频 0.00—0.68 s，机器人双脚在浅色弹性跳板端承重，屈膝后伸腿蹬板；视频 0.72 s 双脚均已离板。后续双臂抬过头顶用于姿态调整，未见手扶板或抓握物体；空中翻转及入水阶段不另列目标接触。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "somersault_deg", + "unit": "deg", + "threshold": 180, + "direction": "target", + "value": 166.9 + }, + { + "name": "entry_distance_m", + "unit": "m", + "threshold": null, + "direction": "target", + "value": 1.739 + }, + { + "name": "entry_tilt_deg", + "unit": "deg", + "threshold": 15, + "direction": "lower_better", + "value": 0.0 + }, + { + "name": "entry_hip", + "unit": "deg", + "threshold": 160, + "direction": "higher_better", + "value": 171.6261 + }, + { + "name": "entry_knee", + "unit": "deg", + "threshold": 160, + "direction": "higher_better", + "value": 170.1238 + }, + { + "name": "entry_feet", + "unit": "m", + "threshold": 0.12, + "direction": "lower_better", + "value": 0.0964 + }, + { + "name": "entry_toes", + "unit": "rad", + "threshold": 0.3, + "direction": "higher_better", + "value": 0.4995 + }, + { + "name": "entry_beyond", + "unit": "m", + "threshold": 0.15, + "direction": "higher_better", + "value": 0.3688 + }, + { + "name": "entry_apart", + "unit": "m", + "threshold": 0.2, + "direction": "lower_better", + "value": 0.1099 + }, + { + "name": "entry_off", + "unit": "m", + "threshold": 0.15, + "direction": "lower_better", + "value": 0.0557 + }, + { + "name": "entry_twist", + "unit": "deg", + "threshold": 20, + "direction": "lower_better", + "value": 0.0572 + } + ], + "failure": null, + "summary_zh": "本次交付完成唯一节点“前直体跳水”（最终 1/1）：机器人由双脚蹬板腾空、连续前翻，头手朝下近竖直入水，于视频约 1.84 s 全身没入。评分器确认三个阶段全部完成，success、replay_success、deterministic 均为 1，记录的入水位置、前翻角及各项入水姿态指标均满足任务判据，无失败断点。已对照成功参考并逐帧检查本次全片；无状态日志，双脚离板的毫秒级同步性、细小接触及姿态阈值无法仅凭远景独立复核，合规结论结合本次裁判记录。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "单节点“前直体跳水”达成、final=1、无失败断点均与视频及评分器结果（success=1，3/3 阶段，somersault 166.9 度在 135–225 内，各入水指标均过阈值）一致；视频时间 1.84 s 位于视频内且已说明末帧精度；contact_actual（腿/躯干·支撑）符合实际；无状态日志，未见时钟转换问题。", + "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": "2026-10-29T10:07:24Z", + "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/embodied-olympics-fun-springboard-3m-i00-privileged--vjaxtwo/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-springboard-3m-i00-privileged--vjaxtwo/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "9c652947b9aa6a347ee4dbe47cbcec5b83cfbf8a42e57e97ac5af73480d4bbfc", + "video_sha256": "9cda11f07e9d0cddeb1e0f528143a53f94f5fd28bd2eb48e81c96b7b1e22bd04", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/steeplechase/task_dag.json b/phase2/embodied-olympics-fun/steeplechase/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..11ffcf1a36179f3c6de8fc52b8dadd9466b0f6c5 --- /dev/null +++ b/phase2/embodied-olympics-fun/steeplechase/task_dag.json @@ -0,0 +1,271 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "embodied-olympics-fun/steeplechase", + "suite": "embodied-olympics-fun", + "base": "embodied-olympics-fun-steeplechase-i00", + "title_zh": "障碍栏水坑越障", + "summary_zh": "机器人合规越过固定障碍栏，通过或越过栏后水坑，全身到达远岸并仅以脚部支撑连续站稳1秒。", + "embodiment_zh": "两种模式均为Unitree H1-2双足人形机器人，27个关节、约1.78米高、67千克；两臂末端为固定拳垫，无工作手指、不能抓握，本任务无可抓物；依靠双腿及躯干运动，无轮式底盘。", + "task_class": "compositional", + "analysis_level": "step", + "spec": { + "version": "1.2", + "task": "embodied-olympics-fun-steeplechase-i00", + "bench": "embodied-olympics-fun", + "progress": "live", + "entities": {}, + "milestones": [ + { + "id": "hurdle_cleared", + "zh": "越过障碍栏", + "mode": "event", + "when": null, + "after": [], + "hold_s": 0, + "goal": true, + "criterion_zh": "固定全景中，机器人从白色横杆前方越到蓝色坑一侧，骨盆及两脚中心各自向前穿过栏中面（x=10.0635 m）时高度都严格大于0.914 m，骨盆过面时横向位置满足|y|≤1.83 m，且全身碰撞几何的最后端已越过横杆后缘（rear_x_m>10.127 m），才算越栏完成；只过了骨盆、仍有脚或身体拖在栏前均未完成。允许用手或脚撑栏、踩栏或攀越，不要求腾空，完成后作为已发生事件保留。", + "step": { + "cat": "本体", + "item": "跳跃" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "双脚在栏前跑道蹬地支撑身体，以跳跨方式越过固定横杆。", + "plan": 1 + }, + { + "who": "双臂协作", + "how": "支撑", + "note_zh": "替代攀越方案中，左右手的固定拳垫同时压撑白色横杆，承托身体越栏。", + "plan": 2 + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "替代攀越方案中，脚部与手撑栏同时在栏前跑道或横杆上支撑身体；不要求抓握横杆。", + "plan": 2 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:4,10,20-26", + "inputs/task/privileged/environment/rfo_task.py:38-43,75-79,121-145,154-161", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159,250-259", + "inputs/oracle/replay.mp4:视频17-19秒，白色横杆前后的位置变化", + "inputs/oracle/run.json:steps[id=152].out.text，交付轨迹的hurdle阶段在模拟18.29秒触发" + ], + "conf": 6, + "note_zh": "三处过面记录及全身清栏共同验收一个越栏步骤，不拆骨盆和两脚。类型采用本体·跳跃表示跨越障碍，包含规则允许的攀越，接触方案不强制跳跨或手撑栏。越栏后可停在坑底再继续，故与后续出坑分开。" + }, + { + "id": "pit_crossed_standing", + "zh": "出坑并站稳", + "mode": "final", + "when": null, + "after": [ + "hurdle_cleared" + ], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "全景中，机器人从蓝色坑区域到达其远端右侧的红色跑道，全身（包括后脚和伸出的手臂）均在坑远边缘之后：碰撞几何最后端rear_x_m>13.66 m，并在此位置连续100个10 ms控制步满足站立条件（评分器记为1.0 s）：有脚接地且接地部位仅限左脚或右脚，躯干偏离竖直≤20°、骨盆三维速度≤0.35 m/s。还留有身体部分在坑上方、出坑后仍用手或膝着地、尚在明显移动或站稳时长不足均未完成；完成后若退回坑内或失去上述站稳验收条件，则该终态不再保持。", + "step": { + "cat": "本体", + "item": "行进" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "脚部在坑底、斜坡或远岸跑道支撑并推进身体，最终由脚部支撑在远岸站稳；允许穿坑行走或跳到远岸。", + "plan": 1 + }, + { + "who": "腿/躯干", + "how": "避开", + "note_zh": "机器人身体与水坑两侧红色跑道上表面保持无接触，通过坑宽范围到达远岸；坑底和低处侧壁允许接触。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:4,21-27", + "inputs/task/privileged/environment/rfo_task.py:39-43,75-79,146-169", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-184,250-259,292-301", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:1103-1105", + "inputs/oracle/replay.mp4:视频约19-26.94秒，落入蓝色坑后沿斜坡走到远岸并站稳", + "inputs/oracle/run.json:steps[id=152].out.text，交付轨迹模拟25.35秒出坑、26.94秒站稳成功" + ], + "conf": 6, + "note_zh": "文本的越过或穿过水坑以全身出坑并站稳为完成状态；原生out of the water jump和stand两阶段是此步的位置与稳定性验收，不另建落坑、起身或终态静置节点。stand原生累计100个控制步，直接采用该验收，不再增加hold_s，避免重复等待；未来可记录该计数或stand阶段。final在越栏之后验收，不给开局站立计分；若未能一直保持合格站立，原生站立计数归零。" + } + ], + "constraints": [ + { + "id": "invalid_hurdle_crossing", + "zh": "违规越栏", + "never": null, + "criterion_zh": "在原生出坑阶段锁存前，每次骨盆或任一脚中心向前穿过x=10.0635 m的栏中面时，该点z≤0.914 m，或骨盆过面时|y|>1.83 m，立即判违规；反复向前过面也逐次检查，手脚触栏本身不违规。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:75-79,131-145" + ] + }, + { + "id": "pelvis_outside_pit", + "zh": "骨盆绕坑", + "never": null, + "criterion_zh": "在原生出坑阶段锁存前，只要骨盆x>10.0635 m而|y|>1.83 m，即认定从水坑侧面绕行；完成全身出坑阶段后不再检查此项。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:131-149,162-164" + ] + }, + { + "id": "side_track_touched", + "zh": "触碰坑侧跑道", + "never": null, + "criterion_zh": "在原生出坑阶段锁存前且骨盆x>10.0635 m时，身体任一部位与左右坑侧跑道发生接触，且接触点z>-0.05 m，即违规；更低的坑侧壁或坑底接触允许，跌倒本身不违规。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:40-43,123,131-152" + ] + }, + { + "id": "physics_unstable", + "zh": "仿真失稳", + "never": null, + "criterion_zh": "模拟器出现坏加速度警告、非有限位置值，或被监测线速度自由度任一分量绝对值超过150 m/s，记录the simulation became unstable并立即结束为失败；不能仅凭姿态摔倒认定仿真失稳。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:108-109,867-885", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:166-172" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "hurdle_cleared", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:128-145,160-161; log ref.stages and task.trackers.over for hurdle completion" + }, + { + "atom": "pit_crossed_standing", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:158-169; log rear_x_m, standing, ref._hold[stand], and ref.stages" + }, + { + "atom": "invalid_hurdle_crossing", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:135-143; log interpolated forward-crossing point and ref.failed/failed event details" + }, + { + "atom": "pelvis_outside_pit", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:133-149; log pelvis position and out-of-water-jump stage" + }, + { + "atom": "side_track_touched", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:150-152; log side-track contact-point heights and active checking interval" + }, + { + "atom": "physics_unstable", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:872-883; log world.unstable and native failure reason" + }, + { + "atom": "referee_success", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:174-178; log per-frame ref.success, not external replay validity" + } + ] + }, + "native_metrics": [ + { + "name": "rear_x_m", + "unit": "m", + "threshold": 13.66, + "direction": "higher_better", + "note_zh": "全身碰撞几何包围盒最后端的x坐标，出坑严格要求大于13.66；越栏阶段另要求大于10.127。" + }, + { + "name": "time_s", + "unit": "s", + "threshold": null, + "direction": "lower_better", + "note_zh": "仅成功时记录完成时间，成功不要求额外的速度成绩；60秒为回合时限，不列为约束。" + } + ], + "open_questions_zh": [ + "inputs/states/index.json为空，无逐帧状态日志；依DECISIONS.md P8，所有节点when、约束never及success均为null。atoms列出需要新增记录的原生字段，不能把run.json中的最终汇总或交付校验伪装成逐帧条件。", + "粒度判读（DECISIONS.md S2、S8、S10及D11）：保留越栏和出坑并站稳两个文本步骤；可在坑底暂停，故不强合成一次连续跳跃；最后站稳是出坑步骤的验收，不因评分器有三个阶段再建独立站稳节点。", + "模式差异：两种模式的rfo_task.py字节相同，几何、原生阶段和站稳条件一致；privileged评分另要求有效有限轨迹、两次新进程回放均成功且终态哈希一致（inputs/grader/efe_embodied-olympics-fun-files/grade.py:87-115），standard评分最后实时回合并校验服务终结记录（inputs/grader/efe_standard-files/v2_1/grade.py:93-99；inputs/grader/efe_embodied-olympics-fun-files/standard/server.py:157-178）。这些交付与有效性条件不建世界状态步骤，success在无日志时保留null。", + "视频仅有固定全景，精确的三点过栏高度、瞬间接触点、躯干角度和骨盆速度不宜从小幅画面单独量化；阈值均直接取评分器，成功录像及交付轨迹的裁判输出用于确认可见结果，未据此调整阈值。" + ], + "conf": 6, + "conf_reason_zh": "文本、两种模式的任务代码及成功录像支持两步和唯一必要顺序，全部几何与站稳阈值有源码依据；无状态日志，无法动态绑定或自动核验，且视频对精确越栏和接触细节的分辨率有限。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "两个文本步骤（越栏、出坑并站稳）与说明及评分器阶段一致；越栏阈值（0.914、|y|≤1.83、x=10.0635、rear>10.127）、出坑rear>13.66、站稳1.0s/20°/0.35 m/s均与rfo_task.py相符；四个约束为评分器失败条件且为立即终止；无状态日志，when/never/success为null符合P8，疑点已写入open_questions_zh。未发现需要改内容的错误。", + "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：['hurdle_cleared', 'pit_crossed_standing']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:43:34Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__steeplechase/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__steeplechase/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "41fa838f86d02441d49a754b09577f27f1303e6f8cf9aa84ef59f05bac173487", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/embodied-olympics-fun/steeplechase/trajectories/embodied-olympics-fun-steeplechase-i00-privileged--8fyjxjt.json b/phase2/embodied-olympics-fun/steeplechase/trajectories/embodied-olympics-fun-steeplechase-i00-privileged--8fyjxjt.json new file mode 100644 index 0000000000000000000000000000000000000000..0b34d2b4c3d35a45a7b154d09134629f3295fe59 --- /dev/null +++ b/phase2/embodied-olympics-fun/steeplechase/trajectories/embodied-olympics-fun-steeplechase-i00-privileged--8fyjxjt.json @@ -0,0 +1,203 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-steeplechase-i00-privileged--8fyjxjt", + "taskKey": "embodied-olympics-fun/steeplechase", + "dag_sha256": "d1654f75d2327bdc38995638268d9415b735f4a2b26d39708a7baeb8fa7d5b71", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "replay_success": 0, + "deterministic": 1, + "n_actions": 3490, + "sim_steps": 3490, + "horizon": 6000, + "episode_over": 0, + "progress": 0.0, + "stages_reached": 0, + "stages_total": 3, + "fell": 0, + "replay_s": 24.16, + "bad_trajectory": 0, + "metric_rear_x_m": 6.967, + "grader_error": 0, + "video_rendered": 1, + "agent_sessions": 145, + "agent_resets": 113, + "agent_replays": 3, + "agent_cost_self_reported": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.71, + "peak_processes": 13, + "peak_disk_gb": 0.246, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0, + "session_log_tampered": 0, + "session_log_from_monitor": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-steeplechase-i00-privileged--8fyjxjt/replay.mp4", + "duration_s": 34.92, + "fps": 25.0, + "frames": 873, + "sha256": "6aebff870215231e21f88ee44ebfe4cafd47620b716e281b884b10828e664bda" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "hurdle_cleared": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 16.7, + "log_step": null, + "text_zh": "观察：视频 0–16.7 s，机器人以双脚支撑沿红色跑道走到白色横杆前，骨盆及双脚仍在横杆的来向一侧；视频 17.6–18.5 s 机器人后仰倒在栏前跑道上。" + }, + { + "t": 34.88, + "log_step": 221, + "text_zh": "观察：视频 20–32 s 的间隔画面及视频 33.92–34.88 s 的密集末段画面中，机器人在栏前地面扭动，未见全身到达蓝色坑一侧；log step 221 对最终恢复的交付文件回放确认执行3490步、success=false，评分器 reward.stages_reached=0。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：视频 0–17.3 s，机器人双脚在栏前红色跑道交替支撑行走并停步；视频 18.5 s 起躯干及腿部倒在跑道地面承重，随后在原侧地面扭动至视频 34.88 s。" + } + ] + }, + "pit_crossed_standing": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 34.88, + "log_step": 221, + "text_zh": "观察：视频 0–34.88 s 所检画面中，机器人始终未越过白色横杆进入蓝色坑，也未到达坑远端红色跑道；视频 34.88 s 仍倒在栏前。评分器 reward.metric_rear_x_m=6.967，log step 221 最终回放success=false；没有出现出坑并以脚站稳的结果。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "rear_x_m", + "value": 6.967, + "unit": "m", + "threshold": 13.66, + "direction": "higher_better" + }, + { + "name": "time_s", + "value": null, + "unit": "s", + "threshold": null, + "direction": "lower_better" + } + ], + "failure": { + "breakpoint": { + "ref": "hurdle_cleared", + "t": 17.6, + "log_step": 211, + "where": "操作中", + "evidence_zh": "观察：视频 16.7–17.3 s 机器人在栏前站立，视频 17.6–18.2 s 躯干逐渐后仰并失去足部支撑，视频 18.5 s 倒在来向跑道；直到视频 34.88 s 仍未越栏。log step 211 的交付生成过程先执行栏前规划器保持，再执行移重心、抬脚上栏等指令，但机器人在首个保持阶段已倒地；该步随后始终未完成。" + }, + "unmet": { + "hurdle_cleared": false, + "pit_crossed_standing": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 17.6, + "log_step": null, + "evidence_zh": "观察：视频 17.6 s 机器人开始明显后仰，视频 18.2 s 身体向栏前地面翻倒、腿部抬起，视频 18.5 s 躯干已落在红色跑道上；视频 19.2 s 以及视频 33.92–34.88 s 仍是倒地姿态，没有恢复越栏。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "log_step": 211, + "source_zh": "日志 / 视频", + "evidence_zh": "推断：log step 211 显示 agent 的控制脚本在栏前由行走切换到规划器驱动的保持和上栏序列，首个settled输出骨盆高度已降至0.234 m、up=0.045，之后仍继续发出上栏动作而未稳定身体；视频 16.7–18.5 s 对应地从直立变为后仰倒地。log step 215 重新生成得到相同动作哈希，log step 220 恢复此3490步文件，log step 221 再回放确认失败，支持交付控制序列无法维持上栏所需平衡；没有足够证据区分轨迹求解与反馈跟踪的具体缺陷。" + } + ], + "gap_zh": "无逐帧状态日志；固定全景分辨率不足以确认拳垫是否短暂触栏及各支撑接触力，故不认定双臂撑栏或具体力控机制。log step 218–219 的恒定关节指令保持实验不属交付：log step 220 已恢复此前文件；log step 222 自述的稳定保持和移重心失败混入该撤回实验，不能据此替代交付视频。原生time_s仅以成功完成时间列于DAG，本次无成功完成时间，留null。" + }, + "summary_zh": "机器人走到栏前，但在视频约 17.6 s 的上栏准备阶段后仰，视频 18.5 s 已倒地，此后在栏前地面扭动至视频 34.88 s，越栏与出坑并站稳均未达到，最终为0/2。视频与最终恢复并回放验证的3490步交付一致，支持控制序列未能维持平衡；末尾较稳定的停步实验已被撤回。无逐帧状态日志，无法进一步确认瞬时手部触栏或区分逆解与反馈控制缺陷。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "两个节点均未达成与视频、评分器 stages_reached=0、rear_x_m=6.967 一致；断点 hurdle_cleared（视频约17.6 s 后仰倒地）、平衡现象和控制根因有 log step 211/215/220/221 支撑，撤回实验已在 gap_zh 说明，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": "2026-10-29T10:05:08Z", + "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/embodied-olympics-fun-steeplechase-i00-privileged--8fyjxjt/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-steeplechase-i00-privileged--8fyjxjt/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "c44410b4b23466b2aff171090c6e9be1d70ec9c293949d072e9a269644b7ffb1", + "video_sha256": "6aebff870215231e21f88ee44ebfe4cafd47620b716e281b884b10828e664bda", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/steeplechase/trajectories/embodied-olympics-fun-steeplechase-i00-privileged--u7eakr4.json b/phase2/embodied-olympics-fun/steeplechase/trajectories/embodied-olympics-fun-steeplechase-i00-privileged--u7eakr4.json new file mode 100644 index 0000000000000000000000000000000000000000..b04b398c27004412b4d562057b19e33b7e1fe8c4 --- /dev/null +++ b/phase2/embodied-olympics-fun/steeplechase/trajectories/embodied-olympics-fun-steeplechase-i00-privileged--u7eakr4.json @@ -0,0 +1,220 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-steeplechase-i00-privileged--u7eakr4", + "taskKey": "embodied-olympics-fun/steeplechase", + "dag_sha256": "d1654f75d2327bdc38995638268d9415b735f4a2b26d39708a7baeb8fa7d5b71", + "outcome": { + "success": 1, + "reward": 1.0, + "source": "catalog", + "raw": { + "success": 1, + "replay_success": 1, + "deterministic": 1, + "n_actions": 4251, + "sim_steps": 4251, + "horizon": 6000, + "episode_over": 1, + "progress": 1.0, + "stages_reached": 3, + "stages_total": 3, + "fell": 0, + "replay_s": 18.53, + "bad_trajectory": 0, + "metric_rear_x_m": 14.925, + "metric_time_s": 42.51, + "grader_error": 0, + "video_rendered": 1, + "agent_sessions": 126, + "agent_resets": 54, + "agent_replays": 3, + "agent_cost_self_reported": 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": 26, + "peak_disk_gb": 0.285, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0, + "session_log_tampered": 0, + "session_log_from_monitor": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-steeplechase-i00-privileged--u7eakr4/replay.mp4", + "duration_s": 42.52, + "fps": 25.0, + "frames": 1063, + "sha256": "079c18f1bc2589a65a448d988953cc07a4318b3b1e1343b5677dc2b2fe3ffd2f" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 2, + "base": 0, + "peak": 2, + "final": 2, + "nodes": { + "hurdle_cleared": { + "method": "visual", + "intervals": [ + [ + 27.1, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 21.5, + "log_step": null, + "text_zh": "观察：视频 14–21 s 机器人从栏前接近白色横杆并前倾，视频 21.5–24.25 s 双臂留在横杆附近承托身体，躯干翻转、两腿抬高后向蓝色坑一侧翻下；视频 24.5–26.5 s 身体低坐在坑近端，视频 27.1–28.25 s 起身，已无腿留在栏前。" + }, + { + "t": 27.1, + "log_step": 132, + "text_zh": "观察：log step 132 的模拟器裁判输出记录骨盆过面在仿真 21.95 s、高度 1.010 m，左脚在仿真 24.144 s、高度 1.505 m，右脚在仿真 24.150 s、高度 1.488 m；全身清栏阶段 hurdle 到仿真 27.14 s 才触发。该步骤保存交付轨迹，终态哈希 934ecc84… 与 log step 133、140 的官方回放相同。" + }, + { + "t": 27.1, + "log_step": 133, + "text_zh": "推断：交付文件的 4251 个控制步对应仿真 42.51 s，与视频总长 42.52 s 吻合，且视频约 24.0–24.25 s 的双脚翻过横杆、约 27.1 s 的起身清栏与裁判事件一致，故将越栏完成标在视频约 27.1 s，而不在骨盆或双脚刚过栏时提前计分；精度约半秒。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "机器人双脚在栏前红色跑道支撑并推进；翻身过栏后在坑近端低位停靠，脚落入坑底并支撑起身。" + }, + { + "who": "双臂协作", + "how": "支撑", + "note_zh": "机器人左右臂末端固定拳垫在白色横杆处共同压撑，配合躯干前翻和双腿抬起完成攀越；无手指包握。视频 21.5–24.25 s 可见撑栏翻转，左右具体接触点受固定远景分辨率限制。" + } + ] + }, + "pit_crossed_standing": { + "method": "visual", + "intervals": [ + [ + 42.5, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 33.5, + "log_step": null, + "text_zh": "观察：视频 28.5–30.5 s 机器人在坑内起身并由脚部支撑，视频 31.5–38.5 s 沿蓝色坑底斜坡向远岸行走，身体始终从坑宽范围内通过，未走上坑两侧跑道。" + }, + { + "t": 42.48, + "log_step": null, + "text_zh": "观察：视频 39.0–41.0 s 机器人继续向远岸红色跑道推进，视频 41.5、41.75、42.0、42.25 s 及最后一帧视频 42.48 s 均保持直立，脚在跑道上，手臂悬空，全身已离开蓝色坑区。" + }, + { + "t": 42.5, + "log_step": 132, + "text_zh": "观察：log step 132 的交付轨迹裁判输出记录 out of the water jump 在仿真 39.11 s 触发，stand 与 success 在仿真 42.51 s 触发；log step 133、140 官方回放再次成功且终态哈希一致。inputs/run.json 的评分器字段 success=1、replay_success=1、deterministic=1、stages_reached=3/3，metric_rear_x_m=14.925。" + }, + { + "t": 42.5, + "log_step": 133, + "text_zh": "推断：远岸位置和末尾连续直立画面与交付回放的成功验收一致，故出坑并站稳在视频约 42.5 s 达成并保持到结束。视频 39 s 左右刚出坑时仍明显行走，不提前计为站稳；精确躯干角度、骨盆速度及连续100步由原生裁判成功结果确认，视频时间取近似值。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "机器人双脚在坑底及上坡面交替接地推进，最终脚部在远岸红色跑道支撑直立；起身后未见手臂或膝部着地。" + }, + { + "who": "腿/躯干", + "how": "避开", + "note_zh": "机器人从蓝色坑宽范围内穿行并沿斜坡到达远岸，身体未接触坑左右两侧的红色跑道上表面；远岸跑道上的脚部支撑另记。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "rear_x_m", + "value": 14.925, + "unit": "m", + "threshold": 13.66, + "direction": "higher_better" + }, + { + "name": "time_s", + "value": 42.51, + "unit": "s", + "threshold": null, + "direction": "lower_better" + } + ], + "failure": null, + "summary_zh": "机器人撑栏翻身后在坑近端起身，视频约 27.1 s 完成全身越栏；随后沿坑底斜坡步行到远岸，视频约 42.5 s 完成出坑并站稳，两个节点最终均达成。交付轨迹的官方回放及评分器均成功，未设失败断点。无逐帧状态日志，固定远景不能精确量化手部接触点、过栏高度与站稳速度阈值，这些细节以同一交付轨迹的裁判事件和最终验收佐证。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "两节点均达成、成功运行无失败断点；视频 27.1 s 起身已越栏、42.4 s 在远岸直立，与裁判事件和评分器字段一致，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": "2026-10-29T10:06:07Z", + "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/embodied-olympics-fun-steeplechase-i00-privileged--u7eakr4/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-steeplechase-i00-privileged--u7eakr4/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "09f3ca383a100992f6c78a3cc4e7284da5c288966fa9c754cc8b52a68cb2cae3", + "video_sha256": "079c18f1bc2589a65a448d988953cc07a4318b3b1e1343b5677dc2b2fe3ffd2f", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/table_tennis_return/task_dag.json b/phase2/embodied-olympics-fun/table_tennis_return/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..08a3df878cd47fca86dcfa91f05fe89452e0f765 --- /dev/null +++ b/phase2/embodied-olympics-fun/table_tennis_return/task_dag.json @@ -0,0 +1,392 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "embodied-olympics-fun/table_tennis_return", + "suite": "embodied-olympics-fun", + "base": "embodied-olympics-fun-table-tennis-return-i00", + "title_zh": "乒乓球接发回球", + "summary_zh": "球在己方台面弹跳一次后被合法击回并触及对方台面，期间右手持续持拍且不触发任何判负规则。", + "embodiment_zh": "三模式均为 Unitree G1 双足人形机器人，29 个身体关节、双臂各 7 自由度；右端为 Dex3-1 三指灵巧手（7 个手指驱动），靠接触和摩擦握持独立球拍；左端为无活动手指的拳垫，双足支撑与移动，无轮式底盘。", + "task_class": "atomic", + "analysis_level": "step", + "spec": { + "version": "1.2", + "task": "embodied-olympics-fun-table-tennis-return-i00", + "bench": "embodied-olympics-fun", + "progress": "live", + "entities": {}, + "milestones": [ + { + "id": "return_landed", + "zh": "回球落对台", + "mode": "event", + "when": null, + "after": [], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "场馆斜侧视角中，黄绿色小球在机器人一侧蓝色台面完成一次弹跳后，由右手所持红色球拍拍面或右持拍手击回，越过或绕过白色球网（允许擦网）并触及发球机一侧蓝色台面上表面或上沿：触台时球心 x≥1.87 米、z≥0.76 米，裁判依次锁存 bounce、strike、opponent's court 且接受成功；持拍不松脱并满足全部约束，首次击球后 0.05 秒内的接触算同一击，无额外保持窗口。仅挥拍、击中球或球飞过网仍未完成；落己方、桌侧、地面或触发犯规也未完成，合法落对台即完成并锁存，无须等待画面出现再次反弹。", + "step": { + "cat": "动态", + "item": "击打" + }, + "contact": [ + { + "who": "手持工具", + "how": "其他", + "note_zh": "候选：击打；右手握持的球拍拍面瞬时击打乒乓球，现有抓握、推/扫、压/擦等接触值均不表示击打；只记录球拍到球这一工具链，不重复标右手抓拍。", + "plan": 1 + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "机器人双脚接触地面支撑身体，同时右手持拍击回乒乓球。", + "plan": 1 + }, + { + "who": "手指/夹爪", + "how": "其他", + "note_zh": "候选：击打；右持拍手按裁判的 right_hand 部位范围直接击打乒乓球，同时仍握住球拍；规则明确允许此替代方案，现有接触值没有击打。", + "plan": 2 + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "机器人双脚接触地面支撑身体，同时右持拍手直接击回乒乓球。", + "plan": 2 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:4-30", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:907-923", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:991-995", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1020-1048", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1100-1157", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178", + "inputs/oracle/replay.mp4:0.00-4.96 video_s；近台弹跳、击球、远台落点", + "inputs/oracle/run.json:reward；steps[415] 最终文件两次重放成功" + ], + "conf": 6, + "note_zh": "一次合法回球是连续动态步骤（S8、S10），己方弹跳与击球仅是验收时序，不拆节点或添加等待发球步骤；event 表示这一落台事件，无需球留在台面。球拍 reset 已在右手，不建立拿拍步骤；右持拍手直接击球也是合法策略。场馆摄像机中对方半台位于白网后发球机一侧；录像只作外观参照，不把示例落点、出手角度或行走设为要求。" + } + ], + "constraints": [ + { + "id": "paddle_released", + "zh": "球拍离手", + "never": null, + "criterion_zh": "拍柄不再同时接触右手手指组及拇指或手掌组，连续间断超过 0.05 秒（代码另加 1e-9 秒数值容差）即判负；可见拍柄滑离或飞离右手，短于容差的接触间断不算释放。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:918-921", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1043", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1081-1093" + ] + }, + { + "id": "paddle_wrong_body", + "zh": "球拍触身体", + "never": null, + "criterion_zh": "球拍任一部分接触右手之外的机器人部位即判负；例如球拍靠在左手、前臂或躯干上，右手自身触拍合法。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1035-1044", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1094-1095" + ] + }, + { + "id": "robot_net_contact", + "zh": "机器人触网", + "never": null, + "criterion_zh": "机器人任何部位或球拍触及白色球网组件即判负；回球本身在合法己方弹跳后擦网允许，不能将球擦网误判成机器人触网。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1096-1097", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1128-1133" + ] + }, + { + "id": "paddle_floor_contact", + "zh": "球拍触地", + "never": null, + "criterion_zh": "球拍拍面或拍柄触及地面即判负，右手仍握着也不豁免。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1098-1099" + ] + }, + { + "id": "premature_ball_contact", + "zh": "提前触球", + "never": null, + "criterion_zh": "球尚未离开发球机轮组，或尚未在机器人半台完成第一次合法弹跳时，机器人或球拍接触球即判负；须让球先在己方蓝色台面弹起再击回。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1100-1106", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1113-1116" + ] + }, + { + "id": "illegal_ball_contact", + "zh": "非法部位触球", + "never": null, + "criterion_zh": "发球后球接触拍柄或右持拍手以外的机器人身体即判负；只有拍面和裁判标作 right_hand 的部位可合法击球。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1039-1041", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1113-1119", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:25-43" + ] + }, + { + "id": "double_bounce", + "zh": "己方二跳", + "never": null, + "criterion_zh": "击球前球在己方蓝色台面出现第二个独立弹跳即判负；须离开上次台面接触至少 0.02 秒才计新接触，不能把同一次接触的连续物理步算成多跳。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:918", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1047-1048", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1136-1148", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:304-342" + ] + }, + { + "id": "double_strike", + "zh": "重复击球", + "never": null, + "criterion_zh": "球在首次合法击球后超过 0.05 秒又与拍面或右持拍手接触即判负，持续接触跨越此窗也算；窗内多次接触归为同一次击球。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:918", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1120-1126" + ] + }, + { + "id": "return_wrong_table", + "zh": "回球落错台面", + "never": null, + "criterion_zh": "击球后、触及对方台面前，球先接触己方半台（球心 x<1.87 米）或对侧桌面侧壁（触台时球心 z<0.76 米）即判负；蓝色台面的上沿可算上表面。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:911-912", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:991-995", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1136-1152" + ] + }, + { + "id": "ball_floor_contact", + "zh": "球触地", + "never": null, + "criterion_zh": "发球机出球后，球在成功前任意时刻触地即判负；包括漏接后落地及回球先落地。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1100-1106", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1134-1135" + ] + }, + { + "id": "free_hand_table", + "zh": "空闲手触台", + "never": null, + "criterion_zh": "球离开发球机轮组后，左手触及台面即判负；代码以左手与 table 的接触点最高 z≥0.758 米（台高 0.76 米减 0.002 米容差）为准，其他部位触台本身不构成此项犯规。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1044-1045", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1100-1110" + ] + }, + { + "id": "robot_fell", + "zh": "机器人摔倒", + "never": null, + "criterion_zh": "每个 0.01 秒控制步末，成功锁存前骨盆高度低于 0.35 米，或双脚以外任一身体部位接触地面，即判摔倒；可见身体倒下或非脚部位着地。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:918", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:981-987", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:276-282", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/task.py:21-23" + ] + }, + { + "id": "illegal_machine_service", + "zh": "机器发球不合法", + "never": null, + "criterion_zh": "机器出球后、己方首次合法弹跳前，球触网，或触台没有依次发生在发球机侧上表面、机器人侧上表面，即直接判发球不合法并失败；这是环境发球的裁判规则，不是机器人要另做的一步。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1128-1133", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1136-1146" + ] + }, + { + "id": "simulation_unstable", + "zh": "模拟数值不稳定", + "never": null, + "criterion_zh": "物理引擎出现坏加速度警告、位姿含非有限值，或被检查的平移自由度任一速度分量绝对值超过 150 米/秒，尝试失败并立即终止；可撤销同一控制步内刚记录的成功，不能仅凭画面静止判此项。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:108-109", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:866-885", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:166-172" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "legal_return_landed", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1100-1157 与 inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178；需新增逐帧记录裁判 ref.stages、ref.failed、ref.success、球位置/台面接触、strike_t；当前没有任何状态日志，尚无可绑定字段。" + }, + { + "atom": "paddle_released", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1081-1093；需逐物理步记录 in_hand、trackers.in_hand_t、ref.time 或原生 grip_break_max_s，当前无状态日志。" + }, + { + "atom": "paddle_wrong_body", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1094-1095；需记录球拍与非 right_hand 机器人几何体的原生接触判据。" + }, + { + "atom": "robot_net_contact", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1096-1097；需记录 robot_mask | paddle 与 netg 的接触。" + }, + { + "atom": "paddle_floor_contact", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1098-1099；需记录 paddle 与 ref.ground 的接触。" + }, + { + "atom": "premature_ball_contact", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1100-1116；需记录 trackers.served、ref.stages 及 ball 对 robot_mask | paddle 的原生接触。" + }, + { + "atom": "illegal_ball_contact", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1039-1041,1117-1119；需记录 ball 对 self.other 的原生接触及接触部位。" + }, + { + "atom": "double_bounce", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1136-1148；需记录 Touch.update 新接触事件、球侧别及原生 bounce 阶段，含 0.02 s 消抖。" + }, + { + "atom": "double_strike", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1120-1126；需记录 trackers.strike_t、ref.time 及 ball 与 striker 的接触。" + }, + { + "atom": "return_wrong_table", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1136-1152；需记录击球后新 table 接触及球心 x/z，阈值由 near_x+length/2 和 table_height 给定。" + }, + { + "atom": "ball_floor_contact", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1134-1135；需记录 served 后 ball 与 ref.ground 的接触。" + }, + { + "atom": "free_hand_table", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1107-1110；需记录 served 和左手对 table 的接触点高度。" + }, + { + "atom": "robot_fell", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:276-282；需记录骨盆高度与 ref.parts_touching(ground)，不能套用通用 fallen 的角度或高度比例。" + }, + { + "atom": "illegal_machine_service", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/table_tennis_return.py:1128-1146；需记录 served、server_bounce、bounce 阶段及球与球网/台面的接触、球心侧别及高度。" + }, + { + "atom": "simulation_unstable", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:866-885；需记录 world.unstable 及 ref.failed 的原生数值不稳定原因；视频通常无法独立判明。" + } + ] + }, + "native_metrics": [ + { + "name": "grip_break_max_s", + "unit": "s", + "threshold": 0.05, + "direction": "lower_better" + } + ], + "open_questions_zh": [ + "inputs/states/index.json 的 logs 为空，不能核验任何逐帧字段；按 DECISIONS.md P8，全部 when、never 与 success 为 null，atoms 仅说明需要补记的数据。成功录像与原生重放输出可核对这一例，但 evidence reach-all 与 success==grader 都只能报告 0/0。", + "目录摘要仅写拍面击球，三模式 instruction 的详细规则和评分器均允许右持拍手直接击球；按 DECISIONS.md N3 使用详细规则，保留两种接触方案。球拍在 reset 时已握于右手（任务模块 1020-1031），离手超过 0.05 秒立即判负（1081-1093），故不需 Y17 拿工具节点（D2）。", + "评分器要求，文本未提：己方首次弹跳前机器发球擦网，或未依次触及发球机侧与机器人侧上表面，直接以 BAD_SERVICE 判失败（任务模块 1128-1146），不重发；保留为约束，不添加机器人语义步骤。", + "执行时序细节：privileged 文本称裁判每 1 毫秒判定，但摔倒只在每 10 毫秒控制步末、尚未成功时检查（referee.py:276-282）；接触规则逐物理步检查。成功在落台物理步锁存，数值不稳定仍可撤销同一控制步的成功（world.py:866-885）。", + "三模式共享同一冻结初态、任务参数和世界状态裁判；privileged 另要求两次重放成功且最终哈希一致（grade.py:106-125），standard 与 standard-realtime 使用实时裁判（standard/server.py:157-178）。交付与重放有效性不冒充世界状态节点或条件，success 因无状态日志保持 null。", + "原生 landing_x_m、landing_y_m 是落点记录，没有额外居中或瞄准点阈值；net_touch 只记录球擦网，合法回球允许擦网，不是必须为零的约束（任务模块 1128-1133、1154-1155）。移动固定球台不受评（instruction.md:24），不建立约束。" + ], + "conf": 6, + "conf_reason_zh": "单个步骤来自三模式文本，连续运动粒度、阶段顺序、几何阈值与判负规则均有代码依据，成功录像及两次重放输出支持该读法；但没有状态日志，25 帧每秒的远景也无法逐毫秒核验持拍与接触消抖，所以条件未绑定，置信度保守取 6。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "单步任务（己方弹跳后用右手持拍击球并落对台）与文本和评分器一致；无状态日志故 when/success 为 null 符合 P8；约束均为评分器的失败条件且为立即终止；落台阈值 x≥1.87、z≥0.76 与评分器一致；已列出文本/评分器差异（直接用手击球、发球不合法）。", + "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：['return_landed']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:45:04Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__table_tennis_return/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__table_tennis_return/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "818398b7c5a52ea8c471f85f4812cb4df6f1f79624a336cff3a974eda13c693c", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/embodied-olympics-fun/table_tennis_return/trajectories/embodied-olympics-fun-table-tennis-return-i00-standard--lt4nskc.json b/phase2/embodied-olympics-fun/table_tennis_return/trajectories/embodied-olympics-fun-table-tennis-return-i00-standard--lt4nskc.json new file mode 100644 index 0000000000000000000000000000000000000000..df6abb18fd32106ccbe8a3bb93e86c232c7f5515 --- /dev/null +++ b/phase2/embodied-olympics-fun/table_tennis_return/trajectories/embodied-olympics-fun-table-tennis-return-i00-standard--lt4nskc.json @@ -0,0 +1,180 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-table-tennis-return-i00-standard--lt4nskc", + "taskKey": "embodied-olympics-fun/table_tennis_return", + "dag_sha256": "7c99ad814d8cc0948a54e1e64bae8599f94aff87d6cc23f6cd5219a93aa00bbb", + "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": 494, + "sim_steps": 1811, + "calls": 36, + "rows": 102, + "resets": 3, + "reset_budget": 5, + "sim_startups": 1, + "elapsed_s": 1568.4, + "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": 0.0, + "peak_memory_gb": 0.291, + "peak_processes": 9, + "peak_disk_gb": 1.178, + "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/embodied-olympics-fun-table-tennis-return-i00-standard--lt4nskc/replay.mp4", + "duration_s": 4.96, + "fps": 25.0, + "frames": 124, + "sha256": "c8b8d99141192f6d862a851c72844c86eef79baf838d45fd3256c98b3475799a" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "return_landed": { + "method": "visual", + "intervals": [ + [ + 4.94, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.0, + "log_step": 77, + "text_zh": "观察：视频 0.00–3.60 s 中机器人右手持红色球拍，从端线后向球台靠近并抬起右臂，左臂未参与击球；log step 77 的机器人服务返回 episode=4、control_steps=0，log step 78 的该局靠近和抬臂动作与视频顺序一致。本视频对应最后一次第4局，前三局的测试与失败不计入此轨迹。" + }, + { + "t": 3.96, + "log_step": null, + "text_zh": "观察：视频 3.84–4.12 s 连续帧中，小球向机器人侧蓝色台面下降，在视频约 3.92–3.96 s 贴近台面，随后向上弹起并接近右手球拍；这是回球前的己方弹跳，尚未完成落对台节点。" + }, + { + "t": 4.32, + "log_step": null, + "text_zh": "观察：视频 4.28–4.36 s 中，机器人右手的红色球拍向上挥动，小球在视频 4.32 s 紧邻拍面，随后在视频 4.36–4.60 s 离开球拍向发球机侧上升；机器人双脚支撑身体，左手垂在身体侧后方，球拍仍随右手运动。" + }, + { + "t": 4.92, + "log_step": 93, + "text_zh": "观察：视频 4.60–4.92 s 的逐帧画面中，小球越过球网后下降到发球机侧蓝色台面的上方，最后一帧已接近该侧台面上沿；机器人仍站立并持拍。log step 93 中机器人服务的实时裁判返回 episode=4、control_steps=494、sim_time_s=4.94、done=true、reason=success、stages_reached=3、stages_total=3；run.json 的 reward.live_success=1、graded_by_live=1，确认该局合法回球落对台。" + }, + { + "t": 4.94, + "log_step": 93, + "text_zh": "推断：依据 log step 85–93 的第4局控制行 dt=0.01 s，log step 93 control step 494 对应仿真 4.94 s，映射到视频约 4.94 s；已用视频约 4.32 s 的击球及 log step 87、89 的 control_steps=431、435 前后动作核对，视频总长 4.96 s 也与该局终止吻合。落对台瞬间处在视频末帧 4.92 s 附近的采样间隙，以裁判接受时刻约视频 4.94 s 记节点首次达成并锁存；不要求画面再出现反弹。" + } + ], + "contact_actual": [ + { + "who": "手持工具", + "how": "其他", + "note_zh": "候选：击打；推断：视频 4.28–4.36 s 中球在右手所持红色球拍拍面附近变向并向对方半台飞出，结合 log step 93 的合法击球阶段接受，判为球拍拍面击打乒乓球。击打不属于抓握、推/扫或压/擦；工具链仅记拍到球，不重复记右手抓拍。25 fps 远景不能解析瞬时接触点。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：视频 0.00–3.60 s 中机器人以双脚交替着地向台前移动；视频 3.96–4.92 s 中双脚着地支撑前倾躯干，配合右臂完成击球及收势，未见摔倒。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "grip_break_max_s", + "unit": "s", + "threshold": 0.05, + "direction": "lower_better", + "value": null + } + ], + "failure": null, + "summary_zh": "第4局评分轨迹完成“回球落对台”1/1：机器人持拍靠近球台，球在视频约 3.95 s 于己方弹跳，视频约 4.32 s 被击回，实时裁判在对应视频约 4.94 s 接受全部三阶段成功。视频末段与成功参考均呈现回球过网后下降至发球机侧台面的过程；本次成功有 log step 93 的服务裁判结果及 reward.live_success=1 支持，无失败断点。没有状态日志，25 fps 远景无法逐毫秒核验拍球接触、持拍中断或精确落台瞬间，且评分字段未报告 grip_break_max_s，故该原生指标值留空。", + "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 return_landed 在视频末尾 4.94 s 达成，有 log step 93 的实时裁判 success 与 reward.live_success=1 支持；接触标注（击打=其他、腿/躯干支撑）合理，无失败断点。", + "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": "2026-10-29T10:02:35Z", + "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/embodied-olympics-fun-table-tennis-return-i00-standard--lt4nskc/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-table-tennis-return-i00-standard--lt4nskc/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "760df3022dba49444e347a928c310181e3ae9d203c78d7152a0e03b027056d22", + "video_sha256": "c8b8d99141192f6d862a851c72844c86eef79baf838d45fd3256c98b3475799a", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/table_tennis_return/trajectories/embodied-olympics-fun-table-tennis-return-i00-standard-realtime--23dcjny.json b/phase2/embodied-olympics-fun/table_tennis_return/trajectories/embodied-olympics-fun-table-tennis-return-i00-standard-realtime--23dcjny.json new file mode 100644 index 0000000000000000000000000000000000000000..d203ea57e27caf922f62cbf428858962515fb81e --- /dev/null +++ b/phase2/embodied-olympics-fun/table_tennis_return/trajectories/embodied-olympics-fun-table-tennis-return-i00-standard-realtime--23dcjny.json @@ -0,0 +1,198 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-table-tennis-return-i00-standard-realtime--23dcjny", + "taskKey": "embodied-olympics-fun/table_tennis_return", + "dag_sha256": "7c99ad814d8cc0948a54e1e64bae8599f94aff87d6cc23f6cd5219a93aa00bbb", + "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": 439, + "sim_steps": 4685, + "calls": 34, + "rows": 8800, + "resets": 10, + "reset_budget": 10, + "sim_startups": 1, + "elapsed_s": 2623.7, + "video_rendered": 1, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 55.764, + "max_lag_s": 0.5079, + "stalls": 12, + "stalled_s": 6.029, + "stalled_share": 0.1081, + "held_steps": 0, + "late_steps": 0, + "requests_refused": 0, + "request_busy_s": 4.056, + "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/embodied-olympics-fun-table-tennis-return-i00-standard-realtime--23dcjny/replay.mp4", + "duration_s": 4.4, + "fps": 25.0, + "frames": 110, + "sha256": "6c1d45e4604da5c7a21e48a51c66fe3c64200b7babca3706edb0a2c2e6856c92" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "return_landed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 2.4, + "log_step": null, + "text_zh": "观察：视频 0.8–1.6 s 机器人向球台靠近，视频 2.0–3.6 s 抬起并伸出持红色球拍的右臂；已经开始接发回球尝试。" + }, + { + "t": 4.36, + "log_step": 208, + "text_zh": "观察：逐帧检查视频 3.68–4.36 s，来球进入己方半台、弹起并再次下降，机器人右手及球拍未迎上来球；末帧球仍在己方半台，没有击回对台的反向飞行。log step 207 启动 episode 11；log step 208 的机器人服务裁判返回该 episode 在 control_steps=439 结束，outcome 为“the ball bounced twice on your court”，stages_reached=1/3，仅完成己方弹跳阶段，未记录击球或对台落点。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "抓握", + "note_zh": "观察：视频 0.0–4.36 s 右手握着拍柄，持拍抬臂并调整拍面；实际接触为右手与球拍。未观察到球拍或右手击中球，故不记工具到球的击打接触。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：视频 0.0–4.36 s 机器人双脚接触地面承重，靠近球台后保持站立进行接发尝试。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "grip_break_max_s", + "unit": "s", + "threshold": 0.05, + "direction": "lower_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "return_landed", + "t": 4.3, + "log_step": 208, + "where": "操作中", + "evidence_zh": "观察：视频 4.00–4.36 s 来球在己方半台上方继续飞行并下降，机器人右臂持拍伸出，却未产生回击；log step 208 的 episode 11 服务裁判以己方二跳判负，未到击球及对台落点阶段。失败发生在回球落对台这一步的接发操作中。" + }, + "unmet": { + "return_landed": false + }, + "phenomena": [ + { + "cat": "运动", + "item": "动态时机", + "t": 4.3, + "log_step": 208, + "evidence_zh": "观察：视频 3.88–4.36 s 来球在己方台面弹起后向机器人移动并下落，拍面停留在伸出的右手附近，没有随来球完成拦截，球从拍前方继续下降；log step 208 裁判报告己方二跳且仅完成 1/3 阶段。对运动中来球的接发未跟上，未形成回击。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "视频", + "evidence_zh": "推断：视频 2.0–3.6 s 机器人已将球拍抬到台面上方，但视频 4.00–4.36 s 球接近并下降时，右臂和拍面没有移到与球相交的位置，未完成接发动作的空间与时间配合。此处只能归到控制，无法仅据远景确定是接触几何、轨迹求解还是跟踪问题。" + } + ], + "gap_zh": "没有逐帧状态日志，25 fps 场馆远景不能核验毫秒级接触及持拍间断；己方二跳判负由最后 episode 的服务裁判（log step 208）确认。最终控制脚本的修改内容在 log step 205–206 未完整展开，无法区分估计误差、逆解误差与执行误差；grip_break_max_s 未由评分器报告。" + }, + "summary_zh": "评分采用最后的 episode 11（log step 207–208）：机器人靠近球台并持拍伸臂接发，但来球在己方二跳前未被击回，“回球落对台”最终完成 0/1。视频 4.00–4.36 s 的漏接与服务裁判仅完成 1/3 内部阶段一致，失败现象为动态时机，根因保守归为控制。没有状态日志及完整的最终脚本修改记录，无法进一步细分控制原因。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "未发现需要修改的错误：回球落对台未达成（final false），服务裁判 log step 208（control_steps=439，约合视频 4.39 s，与视频 4.4 s 吻合）判己方二跳、stages 1/3，视频末段球落己方半台、未被击回；断点、动态时机现象与保守的控制根因均有证据且标注为推断。", + "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": "2026-10-29T10:02:37Z", + "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/embodied-olympics-fun-table-tennis-return-i00-standard-realtime--23dcjny/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-table-tennis-return-i00-standard-realtime--23dcjny/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "67902a5e63553e4208832aba895541e0f2b8bf7111a6867ab8b662a4da511fbe", + "video_sha256": "6c1d45e4604da5c7a21e48a51c66fe3c64200b7babca3706edb0a2c2e6856c92", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/taekwondo_kick/task_dag.json b/phase2/embodied-olympics-fun/taekwondo_kick/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..e14cef29c51589d4fd6cfe0d66f8a4244a7a0561 --- /dev/null +++ b/phase2/embodied-olympics-fun/taekwondo_kick/task_dag.json @@ -0,0 +1,243 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "embodied-olympics-fun/taekwondo_kick", + "suite": "embodied-olympics-fun", + "base": "embodied-olympics-fun-taekwondo-kick-i00", + "title_zh": "跆拳道踢靶", + "summary_zh": "以脚短促踢中假人躯干护具且冲量达到7.2 N·s，落地后脱离假人连续双脚站稳2.0秒，全程不触犯接触或跌倒规则。", + "embodiment_zh": "两模式均为Unitree G1人形机器人，29关节、约1.32 m高、33 kg，双腿与躯干参与踢击和地面支撑，浮动骨盆；两条七关节机械臂末端均为固定拳垫，无可动手指或有效抓握，无轮式底盘和手持工具。", + "task_class": "atomic", + "analysis_level": "step", + "spec": { + "version": "1.2", + "task": "embodied-olympics-fun-taekwondo-kick-i00", + "bench": "embodied-olympics-fun", + "progress": "live", + "entities": {}, + "milestones": [ + { + "id": "protector_kicked_and_stood", + "zh": "踢中护具并站稳", + "mode": "event", + "when": null, + "after": [], + "hold_s": 0, + "goal": true, + "freeze_held": false, + "criterion_zh": "场边视角可见银色机器人抬脚短促击中黑色底座假人胸腹前的蓝色白圆标护具，再收腿落到红色垫面；首次接触须为脚或踝下足部范围内的小腿碰撞壳（踝局部坐标 z≤0、x 在 −0.08 至 0.15 m），仅击中护具，接触跨度≤0.05 s、分离≥0.03 s后确认，护具累计法向冲量≥7.2 N·s。随后双脚且仅双脚触地、全身离开假人，躯干竖直余弦≥0.94、骨盆三维速率≤0.35 m/s，连续满足200个10 ms控制步才完成（裁判记为2.0 s）；未踢中、只有踢击尚未落地站稳、仍单脚支撑、倚靠假人或站稳窗口中断均未完成。", + "step": { + "cat": "动态", + "item": "击打" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "其他", + "plan": 1, + "note_zh": "任一脚以踝下足部击打假人的蓝色躯干护具；候选名称“击打”，现有接触方式没有踢击对应值，按S16用其他。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "plan": 1, + "note_zh": "典型原地踢法中，另一只脚同时支撑红色垫面；左右腿可互换。" + }, + { + "who": "腿/躯干", + "how": "其他", + "plan": 2, + "note_zh": "腾空踢法中，任一脚击打假人的蓝色躯干护具，击打瞬间没有另一只脚的地面支撑；候选名称“击打”，接触方式表没有此值。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:3-30", + "inputs/task/standard/instruction.md:5-28", + "inputs/task/privileged/environment/rfo_task.py:25-36 法向接触冲量积分", + "inputs/task/privileged/environment/rfo_task.py:78-81 实例阈值", + "inputs/task/privileged/environment/rfo_task.py:158-215 脚部、护具、接触时长、冲量与kick阶段", + "inputs/task/privileged/environment/rfo_task.py:217-224 kick后双脚落地、脱离假人并站稳", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178 阶段顺序与成功", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:180-184 连续控制步计时", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:292-301 站立速度与朝向判据", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:1103-1105 躯干朝上余弦", + "inputs/oracle/replay.mp4 @ 0.00-4.04 s", + "inputs/oracle/run.json:1 reward与steps[id=174,180]" + ], + "conf": 6, + "note_zh": "按S10将有效踢击、收腿落地和连续站稳视为一个完整动态动作，kick与stand是其验收阶段；不以抬腿、碰靶、收腿或双脚刚着地分别计分。接触方案1为单脚支撑踢，方案2为腾空踢，均不限定左右脚；随后双脚支撑地面的顺序接触在此说明，不另并列为同时接触。hold_s为0，因为完整判据已包含裁判200控制步的站稳计数，未来记录原生计数后不再额外等待2秒。" + } + ], + "constraints": [ + { + "id": "fell", + "zh": "摔倒", + "never": null, + "criterion_zh": "控制步检查时骨盆高度<0.35 m，或红蓝垫面上出现双脚以外的机器人部位接触（例如手、膝或躯干撑地）。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:77-81", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:272-282", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/task.py:21-23" + ] + }, + { + "id": "illegal_dummy_contact", + "zh": "非法部位碰靶", + "never": null, + "criterion_zh": "首次踢击判定前，除脚和满足踝下足部范围的小腿壳外任一机器人部位碰假人即判负；判定后，评分代码仅对排除脚及整段小腿后的其余部位碰假人判负，文本要求则更严格（见待确认问题）。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:158-195" + ] + }, + { + "id": "kick_outside_protector", + "zh": "踢中护具外", + "never": null, + "criterion_zh": "首次踢击尚未判定时，机器人接触假人的头、裸露躯干、支杆或黑色底座等护具以外的几何体。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:195-197" + ] + }, + { + "id": "protector_pushed", + "zh": "推压护具", + "never": null, + "criterion_zh": "首次触靶开始后，在仍接触假人的某个物理步，距首次接触时间已>0.05 s；短于0.03 s的分离不会重置首次触靶计时，不能用持续推压替代踢击。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:198-208" + ] + }, + { + "id": "kick_too_weak", + "zh": "踢击过弱", + "never": null, + "criterion_zh": "首次触靶结束并与假人分离至少0.03 s时，护具在该次触靶中获得的累计法向冲量<7.2 N·s。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:201-215" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "dyn.protector_kicked_and_stood", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:158-224; inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-184,292-301; 需在逐帧回放适配器中记录ref.stages、ref.failed、ref._hold[stand]及合法踢击的原生追踪量，以kick和stand均完成且无判负表达完整动作；当前没有状态日志。" + }, + { + "atom": "dyn.fell", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:276-282; 需记录骨盆高度、地面接触部位与原生跌倒判定。" + }, + { + "atom": "dyn.illegal_dummy_contact", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:158-195; 需记录首次踢击判定状态、接触部位、踝坐标接触点及原生非法部位判定。" + }, + { + "atom": "dyn.kick_outside_protector", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:195-197; 需记录判定前机器人与非hogu假人几何体的接触。" + }, + { + "atom": "dyn.protector_pushed", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:198-208; 需记录首次接触时刻、末次接触时刻、当前接触及原生推压判负。" + }, + { + "atom": "dyn.kick_too_weak", + "level": "new_code", + "source": "inputs/task/privileged/environment/rfo_task.py:201-215; 需记录未四舍五入的累计冲量和触靶结束判定。" + } + ] + }, + "native_metrics": [ + { + "name": "impulse_n_s", + "unit": "N s", + "threshold": 7.2, + "direction": "higher_better" + }, + { + "name": "contact_s", + "unit": "s", + "threshold": 0.05, + "direction": "lower_better" + } + ], + "open_questions_zh": [ + "inputs/states/index.json的logs为空，无法确认任何逐帧字段或执行状态绑定；按P8全部when、never及success均为null，atoms列出需新增的记录，冲量和毫秒级接触不能仅凭25帧/秒视频精确量出。成功参考的最终评分记录提供7.68 N·s、0.026 s；它不代替所有轨迹的状态日志。", + "粒度解释：按S10，踢中护具、收腿落地和连续站稳构成一次完整踢击，合为一个节点，未采用原生kick→stand的两节点划分；落地站稳前不提前记完成。", + "任务设计问题：文本要求踢后除脚外任何部位接触假人均失败，但rfo_task.py:165将双侧小腿整体排除在other之外，184-187在踢后直接返回而不再调用192处的_foot_ok，因此踢后踝上小腿碰假人未必判负；约束按实际代码记录此差异，不新增节点或假定它已经修复。", + "文本与评分时钟差异：说明称每1 ms物理步决定全部事件，实际触靶在物理步检查，跌倒、站稳及成功在10 ms控制步检查（world.py:878-885、referee.py:272-301）；站稳2.0秒是200个连续真采样，首末采样相隔1.99秒，使用原生计数，不另加完整2秒持续条件。", + "模式差异：任务脚本逐字相同，参数及冻结初态哈希相同；privileged由grade.py:106-115要求两个新进程重放均成功且终态哈希相同，standard由v2_1/grade.py:93-99取服务最后一次现场事件的成功并要求有效终结。共同DAG描述物理任务，不将文件合法性、确定性、交付或基础设施字段当作语义步骤；无状态日志时success保持null。", + "peak_force_n仅在rfo_task.py:203,210记录峰值，无通过阈值，不建约束；15秒/1500控制步仅为时限，按P21不建约束。", + "README称无参考恢复控制器且使用过注入恢复的测试（inputs/task/privileged/README.md:10,22），本次采用的是inputs/oracle中另一个目录提供的真实成功记录与视频；未把历史测试中的注入恢复设为合法动作或必经步骤。" + ], + "conf": 6, + "conf_reason_zh": "文本、两模式同一规则代码、初态及成功视频支持一个完整踢击节点，数值均来自评分器；但没有状态日志，冲量、接触时长和接触点不能由视频精确核验，且存在踢后小腿接触规则差异，故降低置信度。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "单节点（踢中护具并站稳）符合S10（踢-落地-站稳为连续动作，类比后空翻1步）；无状态日志故when/never为null符合P8；约束对应评分代码的失败条件，标签（动态·击打，腿/躯干×其他+支撑）符合S16；待确认问题已列出文本/评分差异。无需修改。", + "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：['protector_kicked_and_stood']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:46:19Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__taekwondo_kick/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__taekwondo_kick/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "966bb269b84e02a5e925ea8c924b6719dfbff1e118c32c5bee9ebe4d6c244111", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/embodied-olympics-fun/taekwondo_kick/trajectories/embodied-olympics-fun-taekwondo-kick-i00-privileged--hpdmveu.json b/phase2/embodied-olympics-fun/taekwondo_kick/trajectories/embodied-olympics-fun-taekwondo-kick-i00-privileged--hpdmveu.json new file mode 100644 index 0000000000000000000000000000000000000000..be45737413f57f427e44e0d1ce6bfc03792defda --- /dev/null +++ b/phase2/embodied-olympics-fun/taekwondo_kick/trajectories/embodied-olympics-fun-taekwondo-kick-i00-privileged--hpdmveu.json @@ -0,0 +1,179 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-taekwondo-kick-i00-privileged--hpdmveu", + "taskKey": "embodied-olympics-fun/taekwondo_kick", + "dag_sha256": "63ae944f698e6469950e0e47b5b44e99a40b4ebb86df1180bd7baa564cd727b9", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "replay_success": 0, + "deterministic": 1, + "n_actions": 10, + "sim_steps": 10, + "horizon": 1500, + "episode_over": 0, + "progress": 0.0, + "stages_reached": 0, + "stages_total": 2, + "fell": 0, + "replay_s": 2.42, + "bad_trajectory": 0, + "grader_error": 0, + "video_rendered": 1, + "agent_sessions": 2, + "agent_resets": 0, + "agent_replays": 1, + "agent_cost_self_reported": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.766, + "peak_processes": 8, + "peak_disk_gb": 0.251, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0, + "session_log_tampered": 0, + "session_log_from_monitor": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-taekwondo-kick-i00-privileged--hpdmveu/replay.mp4", + "duration_s": 0.12, + "fps": 25.0, + "frames": 3, + "sha256": "f4763c4e8d54fe7bb613afefeaa0274684b72246e7c9a8969d881212dbc30758" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "protector_kicked_and_stood": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察：视频 0.00 s 中，银色机器人双脚位于红垫上，站在假人前；双腿未抬起，脚与假人蓝色护具之间有明显间隔。" + }, + { + "t": 0.08, + "log_step": null, + "text_zh": "观察：已检查本次视频全部三帧（视频 0.00、0.04、0.08 s）；机器人始终保持初始站姿，未抬脚击中护具，也没有踢后收腿落地动作，完整踢击并站稳节点从未成立。" + }, + { + "t": null, + "log_step": 21, + "text_zh": "观察：log step 21 中，agent 的生成脚本仅取得 s.hold() 并执行10步，再调用 s.save_trajectory()；该交付文件的回放输出为 steps=10、stages=[]、success=false、done=false、metrics={}。评分器 reward 同样记录 n_actions=10、stages_reached=0、episode_over=0，与本次短视频的未踢击站姿一致。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "impulse_n_s", + "unit": "N s", + "threshold": 7.2, + "direction": "higher_better", + "value": null + }, + { + "name": "contact_s", + "unit": "s", + "threshold": 0.05, + "direction": "lower_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "protector_kicked_and_stood", + "t": 0.0, + "log_step": 21, + "where": "首步前", + "evidence_zh": "观察：视频 0.00–0.08 s 的全部三帧中，机器人保持双脚站立、与护具分离，没有启动踢击；log step 21 保存并回放验证的是10步 hold 基线，stages=[]。推断：评分轨迹止于踢击动作开始之前，断点为“踢中护具并站稳”的首步前。" + }, + "unmet": { + "protector_kicked_and_stood": false + }, + "phenomena": [ + { + "cat": "没尝试", + "item": "没尝试", + "t": 0.0, + "log_step": 21, + "evidence_zh": "观察：视频 0.00、0.04、0.08 s 中机器人均未抬腿或接触护具；log step 21 的交付生成脚本只执行10步 hold 并保存，随后回放报告 stages=[]。本现象针对已交付并验证的静止基线，后续会话调试中的踢击不属于该评分轨迹。" + } + ], + "root_causes": [], + "gap_zh": "观察：本目录没有状态日志，评分器未报告 impulse_n_s 或 contact_s，故两项原生指标保留 null，不能把未触靶换算成已测得的零冲量。log step 46 读取的会话测试输出曾报告 success=True、stages=[kick, stand]，但该测试未见保存为交付轨迹；log step 50 的最后消息是模型提供方过滤中断。log step 21 已在中断前保存并回放验证静止基线，按D37仍标注该评分轨迹；现有证据不足以认定保留基线源于错误规划、控制能力或交付操作错误，根因留空，不把会话测试的摔倒或成功移入评分轨迹。" + }, + "summary_zh": "本次评分轨迹只包含10步静止基线；视频 0.00–0.08 s 的全部三帧均未见机器人抬脚触靶，“踢中护具并站稳”最终为0/1。断点位于该节点首步前，现象为没尝试；log step 21 已保存并回放验证基线，log step 50 的提供方过滤发生在其后。后续会话测试的成功不属于交付轨迹，缺少足以解释为何未替换基线的根因证据，且无状态日志或已报告的接触指标。", + "conf": 9, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "标注与证据一致：交付轨迹仅为10步 hold 基线（log step 21 回放 stages=[]、steps=10，与 reward n_actions=10 吻合），视频仅3帧均无踢击，节点 false、断点首步前、没尝试符合 D37/FAIL-23；log 46 的会话成功未计入评分轨迹，处理正确；根因留空并有 gap_zh。", + "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": "2026-10-29T10:03:46Z", + "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/embodied-olympics-fun-taekwondo-kick-i00-privileged--hpdmveu/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-taekwondo-kick-i00-privileged--hpdmveu/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "dfdba197c08e9125f71c00623a1271a16bf21dff49e6398f302414c2a79d9610", + "video_sha256": "f4763c4e8d54fe7bb613afefeaa0274684b72246e7c9a8969d881212dbc30758", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/taekwondo_kick/trajectories/embodied-olympics-fun-taekwondo-kick-i00-privileged--wfo3maq.json b/phase2/embodied-olympics-fun/taekwondo_kick/trajectories/embodied-olympics-fun-taekwondo-kick-i00-privileged--wfo3maq.json new file mode 100644 index 0000000000000000000000000000000000000000..5337e8eb4e2332944c3d742e8b15d6e410be79bb --- /dev/null +++ b/phase2/embodied-olympics-fun/taekwondo_kick/trajectories/embodied-olympics-fun-taekwondo-kick-i00-privileged--wfo3maq.json @@ -0,0 +1,182 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-taekwondo-kick-i00-privileged--wfo3maq", + "taskKey": "embodied-olympics-fun/taekwondo_kick", + "dag_sha256": "63ae944f698e6469950e0e47b5b44e99a40b4ebb86df1180bd7baa564cd727b9", + "outcome": { + "success": 1, + "reward": 1.0, + "source": "catalog", + "raw": { + "success": 1, + "replay_success": 1, + "deterministic": 1, + "n_actions": 805, + "sim_steps": 505, + "horizon": 1500, + "episode_over": 1, + "progress": 1.0, + "stages_reached": 2, + "stages_total": 2, + "fell": 0, + "replay_s": 3.89, + "bad_trajectory": 0, + "metric_contact_s": 0.024, + "metric_impulse_n_s": 30.57, + "metric_peak_force_n": 4941.3, + "grader_error": 0, + "video_rendered": 1, + "agent_sessions": 3, + "agent_resets": 0, + "agent_replays": 9, + "agent_cost_self_reported": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 3.9, + "peak_processes": 27, + "peak_disk_gb": 0.242, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0, + "session_log_tampered": 0, + "session_log_from_monitor": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-taekwondo-kick-i00-privileged--wfo3maq/replay.mp4", + "duration_s": 5.08, + "fps": 25.0, + "frames": 127, + "sha256": "8547dff9e1b132a3718fe47a8df5aea4dceb500cfb5580cd35bca68e4190de34" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "protector_kicked_and_stood": { + "method": "visual", + "intervals": [ + [ + 5.04, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 1.88, + "log_step": null, + "text_zh": "观察：视频 1.52–1.80 s，机器人单脚支撑并抬起另一腿；视频 1.84–1.88 s，抬起的脚快速伸向假人胸腹前的蓝色白圆标护具，随后视频 1.92–2.12 s 屈膝收回。右脚留在红色垫面支撑，踢击腿为左腿（左右身份另由 log step 110 的动作说明交叉核对）。" + }, + { + "t": 2.28, + "log_step": null, + "text_zh": "观察：视频 2.16–2.28 s，机器人将踢击脚降回红色垫面；视频 2.32–3.40 s 逐帧可见双脚落地、身体与假人分离并逐渐稳定。视频 4.04–5.04 s 逐帧可见机器人继续直立、双脚支撑，未见手、膝或躯干撑地。" + }, + { + "t": 5.04, + "log_step": 87, + "text_zh": "观察：评分器 reward 记录 success=1、replay_success=1、deterministic=1、stages_reached=2/2、fell=0，护具冲量为 30.57 N·s，接触时长为 0.024 s；log step 87 对最终保存轨迹的两次验证回放均成功，报告 kick、stand 两阶段完成、505 个控制步及仿真 5.05 s。log step 108 再次回放同一交付文件，报告相同指标和 505 个执行行。" + }, + { + "t": 5.04, + "log_step": 87, + "text_zh": "推断：log step 8 的模拟器说明给出每控制步 10 ms，最终交付回放的 505 步对应仿真 5.05 s，与视频 5.08 s 总长相符，并已在视频末帧 5.04 s 的双脚站立终态核对。结合评分器确认 stand 已完成，将完整节点首次达成近似锚定在视频 5.04 s；未将踢中或刚落地提前算作完成。无逐帧状态日志，25 fps 视频不能独立测量踝下接触点、0.024 s 接触跨度、法向冲量或200个控制步的站稳计数，这些细节采用本次评分器结果。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "其他", + "note_zh": "观察：视频 1.84–1.88 s，机器人抬起的左脚短促击打假人的蓝色躯干护具并迅速收回；候选名称“击打”，接触方式表没有踢击对应值，按规则使用“其他”。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：踢击时机器人右脚支撑红色垫面；视频 2.28 s 左脚落地，随后至视频 5.04 s 由双脚支撑垫面完成站稳。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "impulse_n_s", + "value": 30.57, + "unit": "N s", + "threshold": 7.2, + "direction": "higher_better" + }, + { + "name": "contact_s", + "value": 0.024, + "unit": "s", + "threshold": 0.05, + "direction": "lower_better" + } + ], + "failure": null, + "summary_zh": "本次评分轨迹完成“踢中护具并站稳”1/1节点：机器人在视频约 1.88 s 以左脚短促踢中蓝色护具，随后收腿落地并保持双脚站立，完整动作在视频末尾约 5.04 s 达成。评分器记录冲量30.57 N·s、接触时长0.024 s、踢击与站稳阶段均通过，最终交付文件的验证回放与这些结果一致，无失败断点。没有逐帧状态日志，毫秒级接触细节及精确站稳窗口依赖本次评分器与交付回放记录，视频用于核对动作和终态。", + "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 protector_kicked_and_stood 达成、final=true 与评分器 success/stages 2/2/fell=0 及视频一致；踢击约视频 1.4–2.0 s，落地后站稳至片尾；contact_actual（腿/躯干·其他击打 + 支撑）正确，无失败断点，时钟均标明视频时间。", + "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": "2026-10-29T10:03:59Z", + "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/embodied-olympics-fun-taekwondo-kick-i00-privileged--wfo3maq/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-taekwondo-kick-i00-privileged--wfo3maq/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "a2a66462d1e6bf66c98b90efe04acff833a9401f121442fb7f5716e05f89a6cc", + "video_sha256": "8547dff9e1b132a3718fe47a8df5aea4dceb500cfb5580cd35bca68e4190de34", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/tennis_return/task_dag.json b/phase2/embodied-olympics-fun/tennis_return/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..aeb99b0260f07bcddb8e70a33ccbc6bc2481d60c --- /dev/null +++ b/phase2/embodied-olympics-fun/tennis_return/task_dag.json @@ -0,0 +1,342 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "embodied-olympics-fun/tennis_return", + "suite": "embodied-olympics-fun", + "base": "embodied-olympics-fun-tennis-return-i00", + "title_zh": "网球接发回球", + "summary_zh": "来球在己方反弹一次后，由始终握在右手的球拍合法回击越网，首次落入对面单打场地内且此前无违规。", + "embodiment_zh": "三模式均为Unitree H1-2双足人形机器人（27个关节）；右臂末端为Inspire五指灵巧手，初始以接触和摩擦握住独立球拍，左臂末端为无活动手指的拳垫；双腿与躯干承担站立或步行支撑，无轮式底盘。", + "task_class": "atomic", + "analysis_level": "step", + "spec": { + "version": "1.2", + "task": "embodied-olympics-fun-tennis-return-i00", + "bench": "embodied-olympics-fun", + "progress": "live", + "entities": {}, + "milestones": [ + { + "id": "return_landed_in", + "zh": "回球落入对场", + "mode": "event", + "when": null, + "after": [], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "固定场地视角中，来球先在近侧蓝色场地反弹一次，随后被右手握持的黄绿色球拍回击；球心从己方向远侧越过黑色球网平面（网柱之间须高于该处网顶，或从柱外绕过），并在远侧蓝色单打场地含白线内首次触地才完成，落点须满足12.385 < x <= 24.27米、-1.365 <= y <= 6.865米，且此前未触发任一违规。裁判须按bounce、racket、over the net、in依次确认且立即成功，无额外保持窗口；只击中、仍在飞行、落在近侧或出界均未完成，文字要求首次反弹在己方右发球区，而代码只验己方半场（差异见待决问题）。", + "step": { + "cat": "动态", + "item": "击打" + }, + "contact": [ + { + "who": "手持工具", + "how": "其他", + "plan": 1, + "note_zh": "候选：击打。右手灵巧手握持球拍，以拍弦、拍框或球拍其他部位短时撞击网球；现有怎么碰取值没有击打，按S16使用其他，持拍抓握不重复标注。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "plan": 1, + "note_zh": "机器人双脚接触球场地面，双腿和躯干支撑持拍回击时的身体。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:3-36", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:1004-1018", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:1117-1130", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:1150-1216", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178", + "inputs/oracle/replay.mp4 @ 0.00, 3.68, 4.08, 5.64 s", + "inputs/oracle/run.json: reward, steps[id=88]" + ], + "conf": 6, + "note_zh": "按S10，一次接发回击及随后的越网、落地是不可暂停后续做的完整动态步骤，原生四阶段只用于验收，不拆节点。初始已握拍，持拍是全程约束，不建拾拍或保持持拍节点；oracle的侧移、摆拍和挥拍时刻均不规定为必经策略。" + } + ], + "constraints": [ + { + "id": "racket_released", + "zh": "球拍脱手", + "never": null, + "criterion_zh": "右手手指与拇指或手掌未同时接触拍柄，连续中断超过0.05秒即脱手；掉拍、扔拍或滑出手均失败，短暂中断不超过0.05秒允许。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:1150-1163", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:1017" + ] + }, + { + "id": "net_touched", + "zh": "身体或拍触网", + "never": null, + "criterion_zh": "机器人任何身体部位或球拍接触球网及网柱即失败；只有球本身擦网或碰柱不触发此禁令。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:1138", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:1164-1167", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:817-825" + ] + }, + { + "id": "far_ground_touched", + "zh": "触及对场地面", + "never": null, + "criterion_zh": "机器人或球拍与地面的实际接触点中，任何一点的x坐标大于12.385米（球网平面）即失败；未触地的伸越不是这一项的判据。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:1008", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:1168-1170" + ] + }, + { + "id": "ball_touched_before_serve", + "zh": "发球前触球", + "never": null, + "criterion_zh": "球尚未离开发球机轮组时，身体或球拍触球即失败；发球机自身对球的接触不属机器人违规。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:1171-1178" + ] + }, + { + "id": "body_touched_ball", + "zh": "身体触球", + "never": null, + "criterion_zh": "发球后网球接触机器人身体任何部位（包括右手）即失败；须由球拍接触来球。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:1137", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:1181-1182" + ] + }, + { + "id": "volley_before_bounce", + "zh": "反弹前击球", + "never": null, + "criterion_zh": "己方首次弹地阶段尚未完成就由球拍击球，裁判拒绝racket阶段并判失败。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:1183-1188", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-159" + ] + }, + { + "id": "racket_double_hit", + "zh": "球拍二次击球", + "never": null, + "criterion_zh": "球拍与球的第二次独立接触即失败；两段接触间隔至少0.05秒才计为两次，短于该间隔按同一次计。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:1018", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:1183-1186", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:325-342" + ] + }, + { + "id": "ball_carried", + "zh": "球拍持球", + "never": null, + "criterion_zh": "同一次球拍接触事件从首次到末次接触的持续时间超过0.05秒即持球失败；该事件内不足0.05秒的短暂分离不另起一次接触。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:1018", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:1190-1191", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:325-347" + ] + }, + { + "id": "ball_crossed_below_net", + "zh": "球从网下穿过", + "never": null, + "criterion_zh": "击球后球心从己方向对方穿过x=12.385米平面时，若距中线不超过5.029米且球心不高于该处网顶，即失败；绕柱外通过可低于网顶。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:785-798", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:1192-1199" + ] + }, + { + "id": "second_bounce_before_hit", + "zh": "击球前二次弹地", + "never": null, + "criterion_zh": "首次己方弹地后、球拍击球前，球在己方（x不大于12.385米）再次产生独立地面接触即失败；普通地面接触间隔至少0.02秒才分为两次。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:1018", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:1200-1205", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:325-342" + ] + }, + { + "id": "ball_landed_own_side", + "zh": "回球落在己方", + "never": null, + "criterion_zh": "击球后球在己方（球心x不大于12.385米）首次触地即失败；球仅被击中尚不能计作完成。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:1189", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:1200-1207" + ] + }, + { + "id": "ball_landed_out", + "zh": "回球落出界", + "never": null, + "criterion_zh": "合法越网后的首次地面接触若球心x大于24.27米，或球心y偏离中线2.75米超过4.115米，即出界失败；压线算界内。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:785", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:1008", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:1211-1216" + ] + }, + { + "id": "serve_fell_short", + "zh": "发球未到己方", + "never": null, + "criterion_zh": "尚未完成合法回球越网阶段，球就在网的远侧首次触地，裁判以the serve did not reach your side结束失败。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:1208-1209" + ] + }, + { + "id": "robot_fell", + "zh": "机器人跌倒", + "never": null, + "criterion_zh": "控制步检查时骨盆高度低于0.35米，或除双脚外的身体部位接触地面即失败；球拍在己方触地且始终在手中允许。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:1002", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:272-282", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/task.py:21-23" + ] + }, + { + "id": "simulation_unstable", + "zh": "仿真失稳", + "never": null, + "criterion_zh": "发生非有限物理状态、数值发散警告或代码规定的异常线速度，裁判以the simulation became unstable作废本次尝试；这一控制步内刚取得的成功也可能被撤回。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:872-885", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:166-172" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "return_landed_in", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:1183-1214及inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178；需逐帧记录原生ref.stages、ref.success、ref.failed，并保留球的位置、接触及穿网事件；当前无状态日志，未绑定任何假定字段。" + }, + { + "atom": "rule_violations", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:1153-1216、inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:272-289及inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:872-885；需逐帧记录原生裁判失败原因和接触/时间跟踪器，支持各约束；当前never均为null。" + } + ] + }, + "native_metrics": [ + { + "name": "grip_break_max_s", + "unit": "s", + "threshold": 0.05, + "direction": "lower_better", + "note_zh": "最大连续脱离手部有效接触时间；超过阈值失败。来源：inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:1156-1163。" + }, + { + "name": "landing_x_m", + "unit": "m", + "threshold": [ + 12.385, + 24.27 + ], + "direction": "target", + "note_zh": "首次远侧落地球心x，合法区间左开右闭；需同时满足y边界和先前阶段。来源：inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:1200-1216。" + }, + { + "name": "landing_y_m", + "unit": "m", + "threshold": [ + -1.365, + 6.865 + ], + "direction": "target", + "note_zh": "首次远侧落地球心y，合法闭区间由2.75±4.115米得到；不是越接近区间中心越好。来源：inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:785、1008、1211-1216。" + } + ], + "open_questions_zh": [ + "inputs/states/index.json的logs为空；按P8，唯一节点when、全部约束never和success均为null，无法执行状态绑定或验证success与评分器逐日志一致。25帧/秒远景录像也不能独立分辨所有1毫秒接触与0.05秒宽限。", + "文本验收未受评：instruction要求来球先在己方右发球区反弹，评分器inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:1200-1203只检查首次弹地球心x<=12.385米，不检查服务线及横向界线；固定发球的正常落点符合文字，因此不另造反弹节点，也不将发球区限制加严为额外评分条件（N1）。", + "三模式共用完全相同的任务模块、参数和冻结初态；privileged由两个新进程重放均成功且终态哈希相同才得分（inputs/grader/efe_embodied-olympics-fun-files/grade.py:106-115），standard与standard-realtime取最后一次现场尝试的原生裁判结果（inputs/grader/efe_embodied-olympics-fun-files/standard/server.py:6-11、157-178）。重放一致性、文件有效性和服务交付不属于语义世界状态，未伪造success绑定；这两次重放也不是两个独立任务场景。", + "文本称每个1毫秒物理步检查跌倒，实际通用跌倒检查位于10毫秒控制步末（inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:266-282），且已锁存成功后跳过该项；约束以实际检查时点为准，不自行增加落地后的站稳窗口。", + "球拍是所有合法路线必用的工具：击球阶段只由球与球拍接触触发，身体触球失败；但reset已使右手闭合握拍（inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/tennis_return.py:1117-1130、1181-1188），因此无需Y17拾取工具节点。" + ], + "conf": 6, + "conf_reason_zh": "文字、同源裁判和初态明确支持单次完整回球，oracle录像及评分记录确认四阶段全通过；但没有状态日志，短时接触只能保留精确文字判据，发球区与跌倒采样存在已列明差异。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "任务文本只描述一次完整回球（弹跳后持拍击球越网落入对场），单个goal事件节点合理；bounce/racket/over net/in是裁判阶段，折入criterion。约束覆盖评分器全部失败条件且均为立即终止，无状态日志故when/never为null并已列入待决问题，文字与评分器的发球区差异已列出。", + "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：['return_landed_in']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:45:35Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__tennis_return/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__tennis_return/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "86b8be745bf8709df895dbe0aecc366464cbb1c74f4a8ee07e3ca2586dd93883", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/embodied-olympics-fun/tennis_return/trajectories/embodied-olympics-fun-tennis-return-i00-privileged--mupx4ue.json b/phase2/embodied-olympics-fun/tennis_return/trajectories/embodied-olympics-fun-tennis-return-i00-privileged--mupx4ue.json new file mode 100644 index 0000000000000000000000000000000000000000..a685a4cf851899a80ff704c434917744cfbc4abf --- /dev/null +++ b/phase2/embodied-olympics-fun/tennis_return/trajectories/embodied-olympics-fun-tennis-return-i00-privileged--mupx4ue.json @@ -0,0 +1,195 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-tennis-return-i00-privileged--mupx4ue", + "taskKey": "embodied-olympics-fun/tennis_return", + "dag_sha256": "ccffc71d0aa2a0e2f5f39987360ec83305903d4551c5e4b8aa3a2ecae36782d8", + "outcome": { + "success": 1, + "reward": 1.0, + "source": "catalog", + "raw": { + "success": 1, + "replay_success": 1, + "deterministic": 1, + "n_actions": 581, + "sim_steps": 581, + "horizon": 1000, + "episode_over": 1, + "progress": 1.0, + "stages_reached": 4, + "stages_total": 4, + "fell": 0, + "replay_s": 7.85, + "bad_trajectory": 0, + "metric_grip_break_max_s": 0.004, + "metric_landing_x_m": 15.772, + "metric_landing_y_m": 3.598, + "grader_error": 0, + "video_rendered": 1, + "agent_sessions": 27, + "agent_resets": 13, + "agent_replays": 3, + "agent_cost_self_reported": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.773, + "peak_processes": 9, + "peak_disk_gb": 0.249, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0, + "session_log_tampered": 0, + "session_log_from_monitor": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-tennis-return-i00-privileged--mupx4ue/replay.mp4", + "duration_s": 5.84, + "fps": 25.0, + "frames": 146, + "sha256": "c6c442b895296c0a06b9aa55cabdaaa5e16f2043a8e52eec3989e07bbb8a9061" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "return_landed_in": { + "method": "visual", + "intervals": [ + [ + 5.8, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 3.8, + "log_step": null, + "text_zh": "观察：视频 0.00–3.80 s 的固定场地视角中，机器人握着黄绿色球拍从近侧底线附近侧移，随后右臂将球拍移到身体侧前方；双脚交替接触场地，身体保持直立。" + }, + { + "t": 4.08, + "log_step": null, + "text_zh": "观察：视频 3.84–4.04 s 来球接近机器人持拍区域，视频 4.08–4.16 s 球离开该区域并向球网方向上升，右手仍持拍；之后视频 4.40 s 可见球位于球网上方的画面区域。" + }, + { + "t": 5.8, + "log_step": 47, + "text_zh": "观察：log step 47 的官方交付文件重放器报告 success=true、failed=null、依次达到 bounce/racket/over the net/in，执行581个控制步，仿真 5.81 s 成功；落点为(15.772,3.598)米。inputs/run.json 的最终评分器字段同样为 success=1、replay_success=1、deterministic=1、stages_reached=4/4，且落点指标相同。" + }, + { + "t": 5.8, + "log_step": 47, + "text_zh": "推断：log step 47 是交付轨迹文件的正式重放，581个控制步×0.01秒=仿真 5.81 s，与视频总长5.84秒一致；其仿真 4.045 s 击球事件与视频 4.04–4.08 s 球在拍旁转向飞出的画面相符。据此将终末落入对场事件近似锚定在已检查的视频 5.80 s，误差约一帧；远侧落地瞬间在远景中难以独立分辨，节点终值以最终评分器成功及合法落点确认。" + } + ], + "contact_actual": [ + { + "who": "手持工具", + "how": "其他", + "note_zh": "推断：候选“击打”；现有怎么碰取值没有击打，按S16使用其他。机器人右手持黄绿色球拍击打网球；视频 4.04–4.16 s 拍旁来球转向并飞出，log step 47 的交付文件官方重放确认racket阶段。手握拍柄是同一工具作用链，不重复计抓握。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：视频 0.00–3.80 s 机器人双脚交替接触球场地面完成侧移，视频 4.04–5.80 s 双腿与躯干支撑挥拍后的直立身体。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "grip_break_max_s", + "unit": "s", + "threshold": 0.05, + "direction": "lower_better", + "value": 0.004 + }, + { + "name": "landing_x_m", + "unit": "m", + "threshold": [ + 12.385, + 24.27 + ], + "direction": "target", + "value": 15.772 + }, + { + "name": "landing_y_m", + "unit": "m", + "threshold": [ + -1.365, + 6.865 + ], + "direction": "target", + "value": 3.598 + } + ], + "failure": null, + "summary_zh": "观察：视频中机器人侧移就位后右手持拍回击来球并保持直立，最终评分器确认bounce、racket、over the net、in四阶段全部通过，唯一节点“回球落入对场”完成（1/1）。推断：结合log step 47交付文件正式重放及视频击球时刻的对应，节点约在视频 5.80 s 达成，落点(15.772,3.598)米在对场单打界内，无失败断点。观察：本次没有状态日志；25帧/秒远景难以分辨毫秒级握拍接触和远侧落地瞬间，精确接触时长及合法落地由评分器字段补足。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "成功运行，唯一目标节点由评分器 success=1、stages 4/4、落点(15.772,3.598)界内确认；视频帧中机器人持拍直立，未见失败迹象，时间锚点与接触标注（击打按S16记其他、腿/躯干支撑）合理，无失败断点。", + "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": "2026-10-29T10:06:48Z", + "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/embodied-olympics-fun-tennis-return-i00-privileged--mupx4ue/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-tennis-return-i00-privileged--mupx4ue/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "a594c0fca6d27e3d9075b27b8c45fecbebbd898d2a8815257a3483b3501f2e0a", + "video_sha256": "c6c442b895296c0a06b9aa55cabdaaa5e16f2043a8e52eec3989e07bbb8a9061", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/tennis_return/trajectories/embodied-olympics-fun-tennis-return-i00-standard-realtime--odjqwpy.json b/phase2/embodied-olympics-fun/tennis_return/trajectories/embodied-olympics-fun-tennis-return-i00-standard-realtime--odjqwpy.json new file mode 100644 index 0000000000000000000000000000000000000000..8b73700f8fe241386ffded0ff49993f7030fbace --- /dev/null +++ b/phase2/embodied-olympics-fun/tennis_return/trajectories/embodied-olympics-fun-tennis-return-i00-standard-realtime--odjqwpy.json @@ -0,0 +1,220 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-tennis-return-i00-standard-realtime--odjqwpy", + "taskKey": "embodied-olympics-fun/tennis_return", + "dag_sha256": "ccffc71d0aa2a0e2f5f39987360ec83305903d4551c5e4b8aa3a2ecae36782d8", + "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": 491, + "sim_steps": 1880, + "calls": 19, + "rows": 890, + "resets": 5, + "reset_budget": 5, + "sim_startups": 1, + "elapsed_s": 1787.9, + "video_rendered": 1, + "first_success": 0, + "first_reason_code": 2, + "realtime": 1, + "slowdown": 1.0, + "clock_s": 18.796, + "max_lag_s": 0.0968, + "stalls": 0, + "stalled_s": 0.0, + "stalled_share": 0.0, + "held_steps": 1028, + "late_steps": 0, + "requests_refused": 5, + "request_busy_s": 1.741, + "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/embodied-olympics-fun-tennis-return-i00-standard-realtime--odjqwpy/replay.mp4", + "duration_s": 4.92, + "fps": 25.0, + "frames": 123, + "sha256": "b416c7ef837278f9863f1eaa2e499207c201789c137055ccb8b17a2b19c89fed" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "return_landed_in": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 3.6, + "log_step": null, + "text_zh": "观察：视频 0.00–3.60 s 中机器人在近侧底线附近持拍侧移并开始摆拍；视频 3.60–4.36 s 逐帧可见球拍转动，但来球继续朝近侧底线外运动，没有被回击向球网。" + }, + { + "t": 4.16, + "log_step": 159, + "text_zh": "观察：视频 4.00–4.32 s 的小黄球从机器人画面左侧经过并越过近侧底线，视频 4.36–4.88 s 没有回球越网或落入对场；log step 159 最后一次现场尝试的服务端 outcome 为“the ball bounced twice before the racket hit it”，control_steps=491、done=True；log step 163 再次读取该尝试的 outcome，原生阶段仅为1。" + }, + { + "t": 4.88, + "log_step": 160, + "text_zh": "推断：本次视频的挥拍未使球转向对场，与 log step 159、163 的服务端未击球判负相符，因此完整的回球落入对场事件从未达到；log step 160 确认 episode=6 已结束，后续分析不是新的评分尝试。成功参考视频 4.12–4.36 s 可见球从持拍位置向球网方向离开，本次未出现这一回球运动。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：视频 0.00–4.88 s，机器人双脚在球场及近侧底线外地面交替承重，支撑侧移和右臂持拍摆动。球拍未击中网球，故不登记手持工具与球的击打接触；持拍抓握不重复计入工具接触链。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "grip_break_max_s", + "unit": "s", + "threshold": 0.05, + "direction": "lower_better", + "value": null + }, + { + "name": "landing_x_m", + "unit": "m", + "threshold": [ + 12.385, + 24.27 + ], + "direction": "target", + "value": null + }, + { + "name": "landing_y_m", + "unit": "m", + "threshold": [ + -1.365, + 6.865 + ], + "direction": "target", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "return_landed_in", + "t": 4.08, + "log_step": 159, + "where": "操作中", + "evidence_zh": "观察：视频 3.60–4.32 s 机器人正在持拍迎球，来球却从其画面左侧向底线外飞过，没有被球拍回击；log step 159、163 的最后一次现场尝试均返回“the ball bounced twice before the racket hit it”。断点取视频约 4.08 s 的错过来球处，失败步骤为完整回球。" + }, + "unmet": { + "return_landed_in": false + }, + "phenomena": [ + { + "cat": "运动", + "item": "动态时机", + "t": 4.08, + "log_step": 159, + "evidence_zh": "观察：视频 3.40–3.84 s 机器人已明显摆动球拍，视频 4.00–4.32 s 来球抵达身旁时未被迎上，继续飞向近侧底线外；log step 159 的服务端报告球在球拍击中前二次弹地。推断：机器人持拍动作没有与运动中的来球相遇，属于动态目标的接触时机未配合；不将同一次漏球再记为没建立接触。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志", + "evidence_zh": "推断：log step 146 的控制脚本预先生成侧移、摆拍和保持姿态，log step 156、159 修改后仍按固定行列与控制步提交，没有基于当前来球位置修正挥拍；视频 3.60–4.32 s 显示持拍动作已执行，却未使拍面与来球相遇，log step 159 的现场服务结果确认未击球。证据支持回球控制未实现，尚不足以细分为轨迹、反馈或参数问题。" + } + ], + "gap_zh": "本次没有状态日志，远景视频无法分辨毫秒级接触、手内握持细节及精确拍面与球的三维误差；三个原生指标的数值均未报告。log step 166、175 中 agent 自述最终分段延迟，但 log step 159 已将第三段提交起点改为334，服务端 plan_end=445 与该起点加111行一致，而“seg3 at 394”是脚本固定打印文本，故不能据此断言迟了0.6秒；该时序解释不作为根因。此前 log step 155 的掉拍属于另一次尝试，不属于本次评分视频。" + }, + "summary_zh": "评分轨迹中机器人持拍侧移并摆拍，但来球从身旁飞过，未回击越网，完整目标节点完成0/1；log step 159、163 的现场服务端记录只通过首次反弹阶段，并因击球前二次弹地判负。断点在回球操作中，现象为动态时机未配合，固定时序回球控制没有实现有效击球。无状态日志及精确接触数据，无法进一步确定控制误差的子类，且不采用 agent 对分段延迟的未证实解释。", + "conf": 7, + "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、log 一致：评分视频（最后一次现场尝试，control_steps=491）中球拍摆动但来球从身旁飞过，服务端判“击球前二次弹地”，return_landed_in 为 false，断点和动态时机/控制归因有证据支持，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": "2026-10-29T10:07:34Z", + "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/embodied-olympics-fun-tennis-return-i00-standard-realtime--odjqwpy/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-tennis-return-i00-standard-realtime--odjqwpy/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "9b819a9eae7edba674b44475b92d5a9dc86f361beb4aa7aa7ad97e4f46e1da15", + "video_sha256": "b416c7ef837278f9863f1eaa2e499207c201789c137055ccb8b17a2b19c89fed", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/tennis_serve/task_dag.json b/phase2/embodied-olympics-fun/tennis_serve/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..1eb3137cd27b854c2754d34e9520e4bb4866e9d7 --- /dev/null +++ b/phase2/embodied-olympics-fun/tennis_serve/task_dag.json @@ -0,0 +1,365 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "embodied-olympics-fun/tennis_serve", + "suite": "embodied-olympics-fun", + "base": "embodied-olympics-fun-tennis-serve-i00", + "title_zh": "网球发球", + "summary_zh": "从己方底线后合法抛球，用右手所持球拍击球过网，使球首次落入对角发球区且无犯规。", + "embodiment_zh": "两种模式均为Unitree H1-2双足人形机器人，27个关节，双臂各配Inspire五指灵巧手；左手开局持网球，右手以接触和摩擦握住自由球拍，双脚支撑地面，无轮式底盘。", + "task_class": "atomic", + "analysis_level": "step", + "spec": { + "version": "1.2", + "task": "embodied-olympics-fun-tennis-serve-i00", + "bench": "embodied-olympics-fun", + "progress": "live", + "entities": {}, + "milestones": [ + { + "id": "serve_in_box", + "zh": "发球入区", + "mode": "event", + "when": null, + "after": [], + "hold_s": 0, + "goal": true, + "criterion_zh": "场馆固定远景中，近侧底线后的机器人从左手放出小球，以右手持拍击出，球从黑色球网上方飞入远侧黄色标出的对角发球区并首次触地，才算发球入区；仍在准备、只抛球或击球、球未落地、落在别处或擦网后首次入区均未完成。验收为同一次合法发球依次满足离手全无接触0.1秒（确认时双脚在合法发球位置且过去0.5秒均落地、位移≤0.02米），击球前距上次碰本体≥0.1秒且来球速度≤5.0米每秒、击后最后碰拍0.05秒时球速严格增加，球心向前通过x=12.385米平面且|y-1|≤5.029米、z>0.914+0.156|y-1|/5.029米，首次触地球心投影满足12.38518.785米，或y不在[0.975,5.115]米内，即失败；相距不足0.02秒的地面接触算一次。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:57-84", + "inputs/task/privileged/environment/rfo_task.py:444-447", + "inputs/task/privileged/environment/rfo_task.py:636-657", + "inputs/task/privileged/environment/rfo_task.py:498-501" + ] + }, + { + "id": "second_let", + "zh": "再次擦网重发", + "never": null, + "criterion_zh": "第一次球擦网且仍合法落入目标发球区，只判重发并转用备用球，不算成功；第二次仍出现这样的擦网入区则失败。重发须先由左手持住备用球且球已离开杯子，之后遵守相同发球规则。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:572-575", + "inputs/task/privileged/environment/rfo_task.py:640-668" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "legal_serve_in", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:174-178; log per-frame flags.success from Referee.success, with ordered stage events and current serve id" + }, + { + "atom": "fall", + "level": "new_code", + "source": "log per-frame flags.failure and task referee event payloads / contact trackers; inputs/task/privileged/environment/rfo_task.py:363-374; inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:272-289" + }, + { + "atom": "racket_unheld", + "level": "new_code", + "source": "log per-frame flags.failure and task referee event payloads / contact trackers; inputs/task/privileged/environment/rfo_task.py:516-525; inputs/task/privileged/environment/rfo_task.py:563-567" + }, + { + "atom": "robot_or_racket_touches_net", + "level": "new_code", + "source": "log per-frame flags.failure and task referee event payloads / contact trackers; inputs/task/privileged/environment/rfo_task.py:508-510; inputs/task/privileged/environment/rfo_task.py:558-562" + }, + { + "atom": "spare_touched_before_let", + "level": "new_code", + "source": "log per-frame flags.failure and task referee event payloads / contact trackers; inputs/task/privileged/environment/rfo_task.py:570-575" + }, + { + "atom": "illegal_release_or_foot_fault", + "level": "new_code", + "source": "log per-frame flags.failure and task referee event payloads / contact trackers; inputs/task/privileged/environment/rfo_task.py:530-556; inputs/task/privileged/environment/rfo_task.py:582-602; inputs/task/privileged/environment/rfo_task.py:670-681" + }, + { + "atom": "toss_fault", + "level": "new_code", + "source": "log per-frame flags.failure and task referee event payloads / contact trackers; inputs/task/privileged/environment/rfo_task.py:576-585; inputs/task/privileged/environment/rfo_task.py:599-625; inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:304-347" + }, + { + "atom": "strike_fault", + "level": "new_code", + "source": "log per-frame flags.failure and task referee event payloads / contact trackers; inputs/task/privileged/environment/rfo_task.py:595-618; inputs/task/privileged/environment/rfo_task.py:626-635; inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:304-347" + }, + { + "atom": "served_ball_fault", + "level": "new_code", + "source": "log per-frame flags.failure and task referee event payloads / contact trackers; inputs/task/privileged/environment/rfo_task.py:57-84; inputs/task/privileged/environment/rfo_task.py:444-447; inputs/task/privileged/environment/rfo_task.py:636-657; inputs/task/privileged/environment/rfo_task.py:498-501" + }, + { + "atom": "second_let", + "level": "new_code", + "source": "log per-frame flags.failure and task referee event payloads / contact trackers; inputs/task/privileged/environment/rfo_task.py:572-575; inputs/task/privileged/environment/rfo_task.py:640-668" + } + ] + }, + "native_metrics": [ + { + "name": "speed_before_m_s", + "unit": "m/s", + "threshold": 5.0, + "direction": "lower_better", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:595-618" + ] + }, + { + "name": "strike_speed_m_s", + "unit": "m/s", + "threshold": null, + "direction": "higher_better", + "note_zh": "无固定绝对阈值：须严格大于同次碰拍前的speed_before_m_s；在最后碰拍满0.05秒后取值。", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:631-635" + ] + }, + { + "name": "clearance_m", + "unit": "m", + "threshold": 0, + "direction": "higher_better", + "note_zh": "球心过网时须严格为正，并满足两网柱之间的横向范围。", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:642-648" + ] + }, + { + "name": "landing_x_m", + "unit": "m", + "threshold": 12.385, + "direction": "higher_better", + "note_zh": "下界严格；同一指标另受18.785米上界约束。", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:651-657" + ] + }, + { + "name": "landing_x_m", + "unit": "m", + "threshold": 18.785, + "direction": "lower_better", + "note_zh": "上界包含发球线。", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:444-447", + "inputs/task/privileged/environment/rfo_task.py:651-657" + ] + }, + { + "name": "landing_y_m", + "unit": "m", + "threshold": 0.975, + "direction": "higher_better", + "note_zh": "下界含中央发球线的一半宽度。", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:444-447", + "inputs/task/privileged/environment/rfo_task.py:651-657" + ] + }, + { + "name": "landing_y_m", + "unit": "m", + "threshold": 5.115, + "direction": "lower_better", + "note_zh": "上界包含左单打边线。", + "sources": [ + "inputs/task/privileged/environment/rfo_task.py:57-60", + "inputs/task/privileged/environment/rfo_task.py:444-447", + "inputs/task/privileged/environment/rfo_task.py:651-657" + ] + } + ], + "open_questions_zh": [ + "无状态日志（inputs/states/index.json的logs为空）；依暂行P8，所有when、never和success为null，atoms列出需要新增记录的裁判世界状态，不能进行状态重放的全达成或成功一致性检验。成功录像为远景，0.05秒和0.1秒接触窗口、来球球速及压线无法逐项从25帧每秒画面精确复核，须保留该证据缺口。", + "粒度按S10：抛球→击球→过网→首次落入发球区是一段不可暂停的连续动作，合为一个发球入区事件；不得仅因裁判有四阶段而拆成四节点。擦网后取备用球只在条件分支出现，直接合法发球会跳过，故未加节点。", + "工具必需性依D2核对：评分器必须记录球拍击球且始终由右手握拍，所有合法路线都需球拍；reset在开局已经把拍柄放入右手并闭合手指，因此没有需要另计的拿拍步骤。", + "模式差异：两模式的任务模块和实例阈值一致，共用发球规则；特权模式评分另要求两次新进程重放成功且状态哈希相同（grade.py:106-115），标准模式只评分最后一次在线尝试（standard/server.py:157-178及efe_standard-files/v2_1/grade.py:90-100）。交付与复现有效性不是世界状态步骤，也未伪造为节点条件。", + "文本与代码差异：文字称“离手至击球”不许脚误，实际脚误检查从连续无接触满0.1秒、进入toss阶段才开始，feet0也在此刻采样（rfo_task.py:582-602、677-681）；此前只在离手确认时检查站位和过去0.5秒站定。按代码保留验收时刻，未增步骤。", + "文本与代码差异：文字称击球后可在己方任意处踏步，但代码在served阶段没有检查机器人是否跨到对方，只有人拍触网、跌倒等约束（rfo_task.py:558-567、626-668、670-681）；不增加未实现的击后半场禁则。", + "击后球速阈值是与本次碰拍前球速的严格比较，非固定速度；native_metrics的strike_speed_m_s.threshold为null，绝不使用成功录像的12.177米每秒调阈值。20秒为回合时限，依P21不另建约束。" + ], + "conf": 6, + "conf_reason_zh": "任务文字、初始握持、两模式代码和裁判阈值均已核实，连续发球合为一步符合S10；成功录像与目录成绩相符，但无逐帧状态日志且远景无法精确检查接触窗口，因此绑定及逐帧验收仍不可验证。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "单节点“发球入区”符合S10（抛球→击球→过网→落区为不可暂停的连续动作）；无状态日志，依P8 when/never/success为null；约束与裁判失败条件一致且均为立即终止；任务类别atomic合理。开放问题已列出文字/代码差异。无需修改。", + "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：['serve_in_box']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:47:55Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__tennis_serve/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__tennis_serve/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "742a7fbd17d873f3f73f1a6b3ecd6a8290fe7fcffa46787a7789049cb07497f6", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/embodied-olympics-fun/tennis_serve/trajectories/embodied-olympics-fun-tennis-serve-i00-privileged--egwog3n.json b/phase2/embodied-olympics-fun/tennis_serve/trajectories/embodied-olympics-fun-tennis-serve-i00-privileged--egwog3n.json new file mode 100644 index 0000000000000000000000000000000000000000..e7d06ae4253fca96f3f8cae01ec941589694dae0 --- /dev/null +++ b/phase2/embodied-olympics-fun/tennis_serve/trajectories/embodied-olympics-fun-tennis-serve-i00-privileged--egwog3n.json @@ -0,0 +1,235 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-tennis-serve-i00-privileged--egwog3n", + "taskKey": "embodied-olympics-fun/tennis_serve", + "dag_sha256": "75f52b1653f7c3603e4ad1e686c78bfabba387976b8d85a1aaa2f5742b889fd4", + "outcome": { + "success": 0, + "reward": 0.5, + "source": "catalog", + "raw": { + "success": 0, + "replay_success": 0, + "deterministic": 1, + "n_actions": 199, + "sim_steps": 199, + "horizon": 2000, + "episode_over": 1, + "progress": 0.5, + "stages_reached": 2, + "stages_total": 4, + "fell": 0, + "replay_s": 6.18, + "bad_trajectory": 0, + "metric_strike_speed_m_s": 11.671, + "grader_error": 0, + "video_rendered": 1, + "agent_sessions": 153, + "agent_resets": 4943, + "agent_replays": 4, + "agent_cost_self_reported": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 1.058, + "peak_processes": 31, + "peak_disk_gb": 1.002, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0, + "session_log_tampered": 0, + "session_log_from_monitor": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-tennis-serve-i00-privileged--egwog3n/replay.mp4", + "duration_s": 2.0, + "fps": 25.0, + "frames": 50, + "sha256": "bcb992e59e2e42e8f6d11f1be5cdd4e34bf78a6bc245df622a19527f7b603f91" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "serve_in_box": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 1.24, + "log_step": null, + "text_zh": "观察：视频 0.00–1.16 s，机器人站在近侧底线后，左臂抬球并放球，右手持续持拍准备；视频 1.20–1.32 s 右臂挥拍，小球离开拍旁向画面左前方飞出，机器人保持站立。" + }, + { + "t": 1.48, + "log_step": 298, + "text_zh": "观察：视频 1.32–1.52 s，小球向近侧左边线方向飞行并下降，未到球网便落向近侧边线附近；视频 1.52–1.96 s 画面保持结束姿态，没有过网或远侧黄色发球区落地事件。log step 298 对交付文件的独立回放返回 stages=[toss,strike]、failed=the serve landed on your own side、executed_rows=199；log step 316 两次新进程回放得到相同终态哈希，log step 358 再生交付文件仍得到同样失败。" + }, + { + "t": 1.96, + "log_step": null, + "text_zh": "推断：本次只完成抛球和击球，未完成过网并首次落入对角发球区，因此“发球入区”始终为假；评分器字段 success=0、stages_reached=2/4 与可见过程一致。成功参考视频 9.50–11.00 s 中球沿球场纵向飞往球网和远侧区域，供辨认验收位置，不作为本次达成证据。" + } + ], + "contact_actual": [ + { + "who": "手持工具", + "how": "其他", + "note_zh": "观察：视频 1.20–1.32 s，机器人右手持球拍挥动，小球从拍旁飞出；交付回放在 log step 298 确认 strike。候选接触名“击打”，现有方式没有瞬时打击，故用其他；同一握拍链路不重复标注。此前左手直接握持网球，抬起后放球，随后退让，抛球和击球依次发生。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：视频 0.00–1.96 s，机器人双脚支撑地面，腿和躯干维持站立以完成抬臂、放球及挥拍，没有可见倒地。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "speed_before_m_s", + "unit": "m/s", + "threshold": 5.0, + "direction": "lower_better", + "value": null + }, + { + "name": "strike_speed_m_s", + "unit": "m/s", + "threshold": null, + "direction": "higher_better", + "value": 11.671 + }, + { + "name": "clearance_m", + "unit": "m", + "threshold": 0, + "direction": "higher_better", + "value": null + }, + { + "name": "landing_x_m", + "unit": "m", + "threshold": 12.385, + "direction": "higher_better", + "value": null + }, + { + "name": "landing_x_m", + "unit": "m", + "threshold": 18.785, + "direction": "lower_better", + "value": null + }, + { + "name": "landing_y_m", + "unit": "m", + "threshold": 0.975, + "direction": "higher_better", + "value": null + }, + { + "name": "landing_y_m", + "unit": "m", + "threshold": 5.115, + "direction": "lower_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "serve_in_box", + "t": 1.24, + "log_step": 298, + "where": "操作中", + "evidence_zh": "观察：视频 1.20–1.32 s，机器人挥拍后球明显向画面左侧飞出；视频 1.44–1.52 s 球已降到己方半场左边线附近，尚未接近球网。log step 298 对交付轨迹的回放判为 the serve landed on your own side，只有 toss、strike 两阶段；该发球未被修复。" + }, + "phenomena": [ + { + "cat": "运动", + "item": "位姿不准", + "t": 1.24, + "log_step": 298, + "evidence_zh": "观察：视频 1.24–1.52 s，小球虽被球拍击出，却朝近侧左边线飞行并过早下降，未越过黑色球网，也未落入远侧黄色目标区；log step 298 的交付回放明确判球落在己方。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "log_step": 302, + "source_zh": "日志 / 视频", + "evidence_zh": "推断：机器人完成了左手放球和右手持拍击球，但挥拍所产生的出球方向与仰角未能将球送过网。视频 1.24–1.52 s 显示出球向侧方、飞行短；log step 296 记录交付脚本的右臂预摆、挥拍及随挥参数，log step 302 中相同挥拍时序参数 cs=0.6、nsw=8、S=116 的会话测量给出约 11.7 米每秒和方向 [0.76,0.64,0.12]，支持出球横向分量过大、向上分量偏小。该会话测量仅用作控制根因的辅助证据，不据此给视频定时；缺少实际碰拍姿态及跟踪误差，故不细分控制子类。" + } + ], + "gap_zh": "观察：inputs/ 中有运行日志、评分录像和成功参考，但没有逐帧状态日志。远景画面的小球只有数个像素，近侧落地的精确位置和瞬时接触不能单靠画面确定；己方落地由 log step 298 的交付文件回放结果补充确认。评分字段只报告击后球速 11.671 米每秒，未报告碰拍前球速、净过网高度或落点坐标，对应原生指标保持 null；会话试验中的数值不代替交付评分指标。缺少碰拍时拍面姿态、接触力和关节跟踪数据，无法进一步区分接触几何设计与轨迹跟踪问题。", + "unmet": { + "serve_in_box": false + } + }, + "summary_zh": "机器人完成抛球和持拍击球，评分器记原生阶段 2/4、击后球速 11.671 米每秒，但 DAG 的“发球入区”节点未完成，终态为 0/1。视频 1.24–1.52 s 显示击出球向侧方短距离飞行，交付回放在 log step 298 确认球落己方，断点为发球操作中的位姿不准，根因归控制。远景和缺少逐帧状态限制了精确接触、落点及控制子类的判断。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "标签与证据一致：发球入区为假，final=0，断点在发球操作中，现象位姿不准，根因控制（不细分）。视频时间都在 2.0 s 内，log step 引用的是 steps[].id，且与 step 298/302 的输出相符。", + "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": "2026-10-29T10:05:06Z", + "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/embodied-olympics-fun-tennis-serve-i00-privileged--egwog3n/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-tennis-serve-i00-privileged--egwog3n/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "6728156fe77511ba9703bc31717c64bacaab994dabfeb6d5a874e40f11de0b36", + "video_sha256": "bcb992e59e2e42e8f6d11f1be5cdd4e34bf78a6bc245df622a19527f7b603f91", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/tennis_serve/trajectories/embodied-olympics-fun-tennis-serve-i00-standard--cxryqce.json b/phase2/embodied-olympics-fun/tennis_serve/trajectories/embodied-olympics-fun-tennis-serve-i00-standard--cxryqce.json new file mode 100644 index 0000000000000000000000000000000000000000..154c59c62c8144f22c9496a3c5f9450f7fe76710 --- /dev/null +++ b/phase2/embodied-olympics-fun/tennis_serve/trajectories/embodied-olympics-fun-tennis-serve-i00-standard--cxryqce.json @@ -0,0 +1,217 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-tennis-serve-i00-standard--cxryqce", + "taskKey": "embodied-olympics-fun/tennis_serve", + "dag_sha256": "75f52b1653f7c3603e4ad1e686c78bfabba387976b8d85a1aaa2f5742b889fd4", + "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": 820, + "sim_steps": 2998, + "calls": 56, + "rows": 71, + "resets": 4, + "reset_budget": 5, + "sim_startups": 1, + "elapsed_s": 1661.0, + "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": 0.0, + "peak_memory_gb": 0.306, + "peak_processes": 10, + "peak_disk_gb": 1.14, + "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/embodied-olympics-fun-tennis-serve-i00-standard--cxryqce/replay.mp4", + "duration_s": 8.24, + "fps": 25.0, + "frames": 206, + "sha256": "d2ea606b523bc33a21f50a751462e11ab605cf34f158927d11e309ec4efa0ce2" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "serve_in_box": { + "method": "visual", + "intervals": [ + [ + 8.2, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.0, + "log_step": 93, + "text_zh": "观察：视频 0.00–6.24 s，机器人在近侧底线后站立，双臂调整到准备姿态；视频 6.24–6.48 s，左手附近仍有小球，右手持拍，未见备用球离开左侧支架。log step 93 的服务记录显示已重置为 episode 5；评分器字段 graded_by_live=1、resets=4，故本视频对应最后一次在线尝试，先前各次失败不计入本次节点进度。" + }, + { + "t": 6.6, + "log_step": 95, + "text_zh": "观察：视频 6.48–6.64 s，机器人右手球拍从准备位置向上挥过，左手向外移开；视频 6.68–6.72 s，小球亮点已与手和球拍分离并向球网方向运动。机器人双脚着地、身体保持站立，球拍仍随右手运动。log step 95 的服务记录给出本次释放结束 control_steps=645、挥拍结束 control_steps=660，均属 episode 5。" + }, + { + "t": 8.2, + "log_step": 100, + "text_zh": "观察：log step 100 中，机器人服务的裁判结果为 episode=5、control_steps=820、sim_time_s=8.2、done=true、outcome=success、stages_reached=4、stages_total=4；run.json 的评分器字段 success=1、live_success=1。视频 7.20–8.20 s 已逐帧检查，机器人始终站立且球拍留在右手；小球在远景中过小，不能单凭这些画面精确辨认过网及首次落地。" + }, + { + "t": 8.2, + "log_step": 100, + "text_zh": "推断：本次控制步长为 0.01 s，log step 95 的 control_steps=645–660 对应视频 6.45–6.60 s，已用视频 6.48–6.64 s 的挥拍核对；log step 100 的 control_steps=820 因而对应视频约 8.20 s，也与视频末帧一致。结合本次裁判四阶段成功记录，发球入区事件在视频约 8.2 s 达成并锁存；该时刻由控制步时钟定位，首次落地点及毫秒级合法性依赖裁判结果，未从 agent 工作时间换算。" + } + ], + "contact_actual": [ + { + "who": "手持工具", + "how": "其他", + "note_zh": "观察：视频 6.48–6.72 s，机器人右手持球拍向上挥动，小球随后向前飞离；实际工具接触为球拍击打网球。候选接触名“击打”，现有接触方式没有瞬时打击条目，故填其他。此前左手持球并释放，随后外移避让，属于依次发生的直接接触；右手握拍链路不重复标注。" + }, + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：视频 0.00–8.20 s，机器人左右脚支撑地面，腿和躯干在准备、挥拍及随挥后维持站立。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "speed_before_m_s", + "unit": "m/s", + "threshold": 5.0, + "direction": "lower_better", + "value": null + }, + { + "name": "strike_speed_m_s", + "unit": "m/s", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "clearance_m", + "unit": "m", + "threshold": 0, + "direction": "higher_better", + "value": null + }, + { + "name": "landing_x_m", + "unit": "m", + "threshold": 12.385, + "direction": "higher_better", + "value": null + }, + { + "name": "landing_x_m", + "unit": "m", + "threshold": 18.785, + "direction": "lower_better", + "value": null + }, + { + "name": "landing_y_m", + "unit": "m", + "threshold": 0.975, + "direction": "higher_better", + "value": null + }, + { + "name": "landing_y_m", + "unit": "m", + "threshold": 5.115, + "direction": "lower_better", + "value": null + } + ], + "failure": null, + "summary_zh": "本次评分采用第五次在线尝试：机器人站在底线后释放左手小球，用右手球拍击出，裁判在对应视频约 8.2 s 时确认四阶段成功，“发球入区”最终达成 1/1，failure 为 null。视频约 6.5 s 的挥拍与本次控制步时钟吻合，末帧仍见机器人站立持拍。无状态日志；远景小球过小，首次落地点、过网余量及 0.05/0.1 秒接触窗口无法独立精确复核，节点最终值与达成时间结合本次服务裁判记录判定，原生指标未报告，均填 null。", + "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 serve_in_box 达成（评分器 success、stages 4/4、control_steps=820 对应视频约 8.2 s）与视频末帧站立一致；控制步时钟换算有依据，failure 为 null 合理，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": "2026-10-29T10:05:07Z", + "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/embodied-olympics-fun-tennis-serve-i00-standard--cxryqce/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-tennis-serve-i00-standard--cxryqce/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "0629512d393c246a4ad44b0ff5ecec00ef8c0a9c5c0a237c81dc52bd18bc1dd2", + "video_sha256": "d2ea606b523bc33a21f50a751462e11ab605cf34f158927d11e309ec4efa0ce2", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/toes_to_bar/task_dag.json b/phase2/embodied-olympics-fun/toes_to_bar/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..593a91e0de9d06238c85efc452c17e185f2128e6 --- /dev/null +++ b/phase2/embodied-olympics-fun/toes_to_bar/task_dag.json @@ -0,0 +1,418 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "embodied-olympics-fun/toes_to_bar", + "suite": "embodied-olympics-fun", + "base": "embodied-olympics-fun-toes-to-bar-i00", + "title_zh": "悬垂举足触杠五次", + "summary_zh": "双手不脱杠且无禁碰违规，依次完成五次双脚同时在双手之间触杠并回到杠后、脚低于杠至少1.40米的悬垂。", + "embodiment_zh": "两种模式均为Unitree G1双足人形机器人（29个本体关节，约1.32米、35千克），左、右臂各装Dex3-1三指灵巧手，每手七个驱动关节，手部驱动力为标称的1.5倍；以双手接触摩擦悬挂在固定横杠上，双腿和躯干完成举足回垂，无轮式底盘。", + "task_class": "compositional", + "analysis_level": "step", + "spec": { + "version": "1.2", + "task": "embodied-olympics-fun-toes-to-bar-i00", + "bench": "embodied-olympics-fun", + "progress": "live", + "entities": {}, + "milestones": [ + { + "id": "toes_to_bar_1", + "zh": "第1次触杠回垂", + "mode": "event", + "when": null, + "after": [], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "累计完成至少1次完整举足触杠：场地固定视角中，银灰机器人双手留在细横杠上，双脚向上抬至两手之间并同时实际接触横杠（每只脚至少一个接触点的横向位置严格位于两手抓握中心之间），随后双脚一起降到杠下至少1.40米且脚中心都在杠后（世界坐标x<0），即本次回悬垂完成，不需额外停留。仅抬腿接近杠、仅单脚触杠、腿胫触杠、双脚未同时在两手之间触杠，或触杠后尚未降到规定的杠后悬垂位置，都不计本次；全程须满足下列握杠和禁碰约束。", + "step": { + "cat": "本体", + "item": "悬垂摆动" + }, + "contact": [ + { + "who": "双臂协作", + "how": "抓握", + "plan": 1, + "note_zh": "左、右灵巧手分别握住横杠的两处并共同承重；规则也允许手指勾住杠，保持任一手部接触即可，不强制持续拇指对握。" + }, + { + "who": "腿/躯干", + "how": "其他", + "plan": 1, + "note_zh": "候选：触碰；双脚在两手之间同时碰到横杠，此时双手仍握杠。脚部只需实际接触，无须击打、擦拭或承重，现有怎么碰条目无准确对应。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:4 任务要求五次举足触杠并各回悬垂", + "inputs/task/standard/instruction.md:5 两种模式的任务要求一致", + "inputs/task/privileged/tests/rfo_task.py:218-222 五个顺序阶段，回垂高度1.40米", + "inputs/task/privileged/tests/rfo_task.py:276-284 双脚同时与杠接触，接触点严格在两手中心之间", + "inputs/task/privileged/tests/rfo_task.py:286-312 触杠后回垂才增计数，第五次成功", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178 阶段按顺序锁存，全部阶段完成且无失败才成功", + "inputs/oracle/replay.mp4:0.00-28.68 五次举足与回垂的视觉参考" + ], + "conf": 6, + "note_zh": "每个计数包含触杠及其后的回悬垂，按重复次数建一个完整动作节点；中间触杠不单列，开局悬挂不计一次。允许借摆、屈膝及不同关节控制；不要求回垂时身体静止或另外检查肘角。该节点为已完成次数的锁存事件，后续抬腿不撤销之前的次数；其他接触用于脚的轻触，见该接触说明。" + }, + { + "id": "toes_to_bar_2", + "zh": "第2次触杠回垂", + "mode": "event", + "when": null, + "after": [ + "toes_to_bar_1" + ], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "累计完成至少2次完整举足触杠：场地固定视角中，银灰机器人双手留在细横杠上，双脚向上抬至两手之间并同时实际接触横杠（每只脚至少一个接触点的横向位置严格位于两手抓握中心之间），随后双脚一起降到杠下至少1.40米且脚中心都在杠后（世界坐标x<0），即本次回悬垂完成，不需额外停留。仅抬腿接近杠、仅单脚触杠、腿胫触杠、双脚未同时在两手之间触杠，或触杠后尚未降到规定的杠后悬垂位置，都不计本次；全程须满足下列握杠和禁碰约束。", + "step": { + "cat": "本体", + "item": "悬垂摆动" + }, + "contact": [ + { + "who": "双臂协作", + "how": "抓握", + "plan": 1, + "note_zh": "左、右灵巧手分别握住横杠的两处并共同承重；规则也允许手指勾住杠，保持任一手部接触即可，不强制持续拇指对握。" + }, + { + "who": "腿/躯干", + "how": "其他", + "plan": 1, + "note_zh": "候选：触碰；双脚在两手之间同时碰到横杠，此时双手仍握杠。脚部只需实际接触，无须击打、擦拭或承重，现有怎么碰条目无准确对应。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:4 任务要求五次举足触杠并各回悬垂", + "inputs/task/standard/instruction.md:5 两种模式的任务要求一致", + "inputs/task/privileged/tests/rfo_task.py:218-222 五个顺序阶段，回垂高度1.40米", + "inputs/task/privileged/tests/rfo_task.py:276-284 双脚同时与杠接触，接触点严格在两手中心之间", + "inputs/task/privileged/tests/rfo_task.py:286-312 触杠后回垂才增计数，第五次成功", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178 阶段按顺序锁存，全部阶段完成且无失败才成功", + "inputs/oracle/replay.mp4:0.00-28.68 五次举足与回垂的视觉参考" + ], + "conf": 6, + "note_zh": "每个计数包含触杠及其后的回悬垂，按重复次数建一个完整动作节点；中间触杠不单列，开局悬挂不计一次。允许借摆、屈膝及不同关节控制；不要求回垂时身体静止或另外检查肘角。该节点为已完成次数的锁存事件，后续抬腿不撤销之前的次数；其他接触用于脚的轻触，见该接触说明。" + }, + { + "id": "toes_to_bar_3", + "zh": "第3次触杠回垂", + "mode": "event", + "when": null, + "after": [ + "toes_to_bar_2" + ], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "累计完成至少3次完整举足触杠：场地固定视角中，银灰机器人双手留在细横杠上，双脚向上抬至两手之间并同时实际接触横杠（每只脚至少一个接触点的横向位置严格位于两手抓握中心之间），随后双脚一起降到杠下至少1.40米且脚中心都在杠后（世界坐标x<0），即本次回悬垂完成，不需额外停留。仅抬腿接近杠、仅单脚触杠、腿胫触杠、双脚未同时在两手之间触杠，或触杠后尚未降到规定的杠后悬垂位置，都不计本次；全程须满足下列握杠和禁碰约束。", + "step": { + "cat": "本体", + "item": "悬垂摆动" + }, + "contact": [ + { + "who": "双臂协作", + "how": "抓握", + "plan": 1, + "note_zh": "左、右灵巧手分别握住横杠的两处并共同承重；规则也允许手指勾住杠，保持任一手部接触即可，不强制持续拇指对握。" + }, + { + "who": "腿/躯干", + "how": "其他", + "plan": 1, + "note_zh": "候选：触碰；双脚在两手之间同时碰到横杠，此时双手仍握杠。脚部只需实际接触，无须击打、擦拭或承重，现有怎么碰条目无准确对应。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:4 任务要求五次举足触杠并各回悬垂", + "inputs/task/standard/instruction.md:5 两种模式的任务要求一致", + "inputs/task/privileged/tests/rfo_task.py:218-222 五个顺序阶段，回垂高度1.40米", + "inputs/task/privileged/tests/rfo_task.py:276-284 双脚同时与杠接触，接触点严格在两手中心之间", + "inputs/task/privileged/tests/rfo_task.py:286-312 触杠后回垂才增计数，第五次成功", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178 阶段按顺序锁存，全部阶段完成且无失败才成功", + "inputs/oracle/replay.mp4:0.00-28.68 五次举足与回垂的视觉参考" + ], + "conf": 6, + "note_zh": "每个计数包含触杠及其后的回悬垂，按重复次数建一个完整动作节点；中间触杠不单列，开局悬挂不计一次。允许借摆、屈膝及不同关节控制；不要求回垂时身体静止或另外检查肘角。该节点为已完成次数的锁存事件，后续抬腿不撤销之前的次数；其他接触用于脚的轻触，见该接触说明。" + }, + { + "id": "toes_to_bar_4", + "zh": "第4次触杠回垂", + "mode": "event", + "when": null, + "after": [ + "toes_to_bar_3" + ], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "累计完成至少4次完整举足触杠：场地固定视角中，银灰机器人双手留在细横杠上，双脚向上抬至两手之间并同时实际接触横杠（每只脚至少一个接触点的横向位置严格位于两手抓握中心之间），随后双脚一起降到杠下至少1.40米且脚中心都在杠后（世界坐标x<0），即本次回悬垂完成，不需额外停留。仅抬腿接近杠、仅单脚触杠、腿胫触杠、双脚未同时在两手之间触杠，或触杠后尚未降到规定的杠后悬垂位置，都不计本次；全程须满足下列握杠和禁碰约束。", + "step": { + "cat": "本体", + "item": "悬垂摆动" + }, + "contact": [ + { + "who": "双臂协作", + "how": "抓握", + "plan": 1, + "note_zh": "左、右灵巧手分别握住横杠的两处并共同承重；规则也允许手指勾住杠，保持任一手部接触即可，不强制持续拇指对握。" + }, + { + "who": "腿/躯干", + "how": "其他", + "plan": 1, + "note_zh": "候选：触碰；双脚在两手之间同时碰到横杠，此时双手仍握杠。脚部只需实际接触，无须击打、擦拭或承重，现有怎么碰条目无准确对应。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:4 任务要求五次举足触杠并各回悬垂", + "inputs/task/standard/instruction.md:5 两种模式的任务要求一致", + "inputs/task/privileged/tests/rfo_task.py:218-222 五个顺序阶段，回垂高度1.40米", + "inputs/task/privileged/tests/rfo_task.py:276-284 双脚同时与杠接触，接触点严格在两手中心之间", + "inputs/task/privileged/tests/rfo_task.py:286-312 触杠后回垂才增计数，第五次成功", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178 阶段按顺序锁存，全部阶段完成且无失败才成功", + "inputs/oracle/replay.mp4:0.00-28.68 五次举足与回垂的视觉参考" + ], + "conf": 6, + "note_zh": "每个计数包含触杠及其后的回悬垂，按重复次数建一个完整动作节点；中间触杠不单列，开局悬挂不计一次。允许借摆、屈膝及不同关节控制；不要求回垂时身体静止或另外检查肘角。该节点为已完成次数的锁存事件，后续抬腿不撤销之前的次数；其他接触用于脚的轻触，见该接触说明。" + }, + { + "id": "toes_to_bar_5", + "zh": "第5次触杠回垂", + "mode": "event", + "when": null, + "after": [ + "toes_to_bar_4" + ], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "累计完成至少5次完整举足触杠：场地固定视角中，银灰机器人双手留在细横杠上，双脚向上抬至两手之间并同时实际接触横杠（每只脚至少一个接触点的横向位置严格位于两手抓握中心之间），随后双脚一起降到杠下至少1.40米且脚中心都在杠后（世界坐标x<0），即本次回悬垂完成，不需额外停留。仅抬腿接近杠、仅单脚触杠、腿胫触杠、双脚未同时在两手之间触杠，或触杠后尚未降到规定的杠后悬垂位置，都不计本次；全程须满足下列握杠和禁碰约束。", + "step": { + "cat": "本体", + "item": "悬垂摆动" + }, + "contact": [ + { + "who": "双臂协作", + "how": "抓握", + "plan": 1, + "note_zh": "左、右灵巧手分别握住横杠的两处并共同承重；规则也允许手指勾住杠，保持任一手部接触即可，不强制持续拇指对握。" + }, + { + "who": "腿/躯干", + "how": "其他", + "plan": 1, + "note_zh": "候选：触碰；双脚在两手之间同时碰到横杠，此时双手仍握杠。脚部只需实际接触，无须击打、擦拭或承重，现有怎么碰条目无准确对应。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:4 任务要求五次举足触杠并各回悬垂", + "inputs/task/standard/instruction.md:5 两种模式的任务要求一致", + "inputs/task/privileged/tests/rfo_task.py:218-222 五个顺序阶段，回垂高度1.40米", + "inputs/task/privileged/tests/rfo_task.py:276-284 双脚同时与杠接触，接触点严格在两手中心之间", + "inputs/task/privileged/tests/rfo_task.py:286-312 触杠后回垂才增计数，第五次成功", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178 阶段按顺序锁存，全部阶段完成且无失败才成功", + "inputs/oracle/replay.mp4:0.00-28.68 五次举足与回垂的视觉参考" + ], + "conf": 6, + "note_zh": "每个计数包含触杠及其后的回悬垂，按重复次数建一个完整动作节点；中间触杠不单列，开局悬挂不计一次。允许借摆、屈膝及不同关节控制；不要求回垂时身体静止或另外检查肘角。该节点为已完成次数的锁存事件，后续抬腿不撤销之前的次数；其他接触用于脚的轻触，见该接触说明。" + } + ], + "constraints": [ + { + "id": "grip_lost", + "zh": "任一手脱杠", + "never": null, + "criterion_zh": "任一只手的手指、拇指和掌部均离开横杠，连续无接触超过0.05秒，即脱杠失败；不超过0.05秒的接触闪断允许，手指勾杠也算在杠上，不能用自由物体的held标志判断。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:137-179", + "inputs/task/privileged/tests/rfo_task.py:221-222" + ] + }, + { + "id": "forbidden_bar_contact", + "zh": "禁用部位碰杠", + "never": null, + "criterion_zh": "手和腿以外的机器人部位（上臂、前臂、躯干、骨盆或头）与横杠发生实际接触即失败；脚、胫、膝和大腿碰杠允许，但只有双脚满足指定位置的同时触杠才可计次。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:144-154", + "inputs/task/privileged/tests/rfo_task.py:178-180", + "inputs/task/privileged/tests/rfo_task.py:220", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:25-43" + ] + }, + { + "id": "upright_contact", + "zh": "碰到立柱", + "never": null, + "criterion_zh": "机器人的任何部位与两侧立柱或柱脚发生实际接触即失败；立柱和柱脚都由名称前缀post选入。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:46-49", + "inputs/task/privileged/tests/rfo_task.py:264-266", + "inputs/task/privileged/tests/rfo_task.py:290-291" + ] + }, + { + "id": "ground_contact", + "zh": "碰垫或地面", + "never": null, + "criterion_zh": "机器人的任何部位实际碰到蓝色垫面或地面即失败；不另外施加常规站立任务的骨盆高度或倾角跌倒阈值。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:50", + "inputs/task/privileged/tests/rfo_task.py:219", + "inputs/task/privileged/tests/rfo_task.py:292-293" + ] + }, + { + "id": "simulation_unstable", + "zh": "仿真失稳", + "never": null, + "criterion_zh": "仿真出现数值发散而被平台判为失稳（加速度异常警告、非有限位姿或线性自由度速度绝对值超过150米/秒），当次尝试立即失败；需由平台诊断确认，普通的大幅摆动不等于数值失稳。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:108-109", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:872-885", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:166-172" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "completed_reps_ge_1", + "level": "new_code", + "source": "需逐帧记录inputs/task/privileged/tests/rfo_task.py:308-310的Task.trackers[\"reps\"]或原生已锁存rep阶段列表；达到1次。现无状态日志，未绑定假定路径。" + }, + { + "atom": "completed_reps_ge_2", + "level": "new_code", + "source": "需逐帧记录inputs/task/privileged/tests/rfo_task.py:308-310的Task.trackers[\"reps\"]或原生已锁存rep阶段列表；达到2次。现无状态日志，未绑定假定路径。" + }, + { + "atom": "completed_reps_ge_3", + "level": "new_code", + "source": "需逐帧记录inputs/task/privileged/tests/rfo_task.py:308-310的Task.trackers[\"reps\"]或原生已锁存rep阶段列表；达到3次。现无状态日志，未绑定假定路径。" + }, + { + "atom": "completed_reps_ge_4", + "level": "new_code", + "source": "需逐帧记录inputs/task/privileged/tests/rfo_task.py:308-310的Task.trackers[\"reps\"]或原生已锁存rep阶段列表；达到4次。现无状态日志，未绑定假定路径。" + }, + { + "atom": "completed_reps_ge_5", + "level": "new_code", + "source": "需逐帧记录inputs/task/privileged/tests/rfo_task.py:308-310的Task.trackers[\"reps\"]或原生已锁存rep阶段列表；达到5次。现无状态日志，未绑定假定路径。" + }, + { + "atom": "referee_failure", + "level": "new_code", + "source": "需逐帧记录inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:161-172的Referee.failed及失手计时；用于禁碰、脱杠和数值失稳约束。" + }, + { + "atom": "referee_success", + "level": "new_code", + "source": "需逐帧记录inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:174-178的Referee.success；为模拟器原生成功状态，不是外部交付验证结果。" + } + ] + }, + "native_metrics": [ + { + "name": "reps", + "unit": "count", + "threshold": 5, + "direction": "higher_better" + }, + { + "name": "grip_break_max_s", + "unit": "s", + "threshold": 0.05, + "direction": "lower_better" + } + ], + "open_questions_zh": [ + "inputs/states/index.json的logs为空；按DECISIONS.md P8，五个节点、全部约束及success均为null条件。视频可见五次举足和回落，但精确同时接触、接触点边界及0.05秒脱杠需逐物理步日志核验，未虚构字段或估计阈值。", + "模式差异：运动规则、任务模块及冻结实例相同；privileged评分还要求有限且合法的轨迹在两个新进程中均成功且末状态哈希相同（inputs/grader/efe_embodied-olympics-fun-files/grade.py:87-115），standard按最后一次实时尝试的原生裁判判定评分（inputs/grader/efe_embodied-olympics-fun-files/standard/server.py:157-178；inputs/grader/efe_standard-files/v2_1/grade.py:89-99）。DAG只描述共同世界状态，交付复现条件无逐帧状态依据，success保持null。", + "脚部轻触横杠没有准确的接触方式条目；依DECISIONS.md S16、D16保留腿/躯干×其他，候选名为触碰，不推定脚需承重或击打。", + "规则段概述裁判每1毫秒判定；实现中双脚触杠和违规检查在物理子步发生，回悬垂与计次在10毫秒控制步末检查（inputs/task/privileged/tests/rfo_task.py:286-312；inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:867-885），此处以实现采样作为验收，不增设保持窗口。" + ], + "conf": 6, + "conf_reason_zh": "五次完整重复及顺序直接来自文本，阈值、禁碰与计数均有评分源码，两个模式任务模块和冻结实例一致，成功录像与原生五阶段记录相符；无状态日志，无法自动验证接触细节和success等价性，脚部接触方式需用其他。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "五个按次计数的节点（每次=双脚在两手之间同时触杠+回到杠后≥1.40米的悬垂）与指令及rfo_task.py的rep 1..5阶段一致，无多余或缺失节点；握杠、禁碰杠、立柱、地面、仿真失稳均为立即终止约束；无状态日志故when/success为null并已在open_questions_zh说明；60秒仅为视界，按P21不作约束。标签与conf合理。", + "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：['toes_to_bar_1', 'toes_to_bar_2', 'toes_to_bar_3', 'toes_to_bar_4', 'toes_to_bar_5']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:47:46Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__toes_to_bar/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__toes_to_bar/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "99d07409a975cf9bce51502e61a84da4395659fc5563604798230fdc1d91bb6c", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/embodied-olympics-fun/toes_to_bar/trajectories/embodied-olympics-fun-toes-to-bar-i00-privileged--bquh9w6.json b/phase2/embodied-olympics-fun/toes_to_bar/trajectories/embodied-olympics-fun-toes-to-bar-i00-privileged--bquh9w6.json new file mode 100644 index 0000000000000000000000000000000000000000..d4f1a209d3c981a519be2495a7a1eab0e4592dd2 --- /dev/null +++ b/phase2/embodied-olympics-fun/toes_to_bar/trajectories/embodied-olympics-fun-toes-to-bar-i00-privileged--bquh9w6.json @@ -0,0 +1,333 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-toes-to-bar-i00-privileged--bquh9w6", + "taskKey": "embodied-olympics-fun/toes_to_bar", + "dag_sha256": "ac964ab85ff6f7f76f4e7777abc98665ca1819d4122120d46d6b9b1067731cff", + "outcome": { + "success": 1, + "reward": 1.0, + "source": "catalog", + "raw": { + "success": 1, + "replay_success": 1, + "deterministic": 1, + "n_actions": 1462, + "sim_steps": 1462, + "horizon": 6000, + "episode_over": 1, + "progress": 1.0, + "stages_reached": 5, + "stages_total": 5, + "fell": 0, + "replay_s": 16.56, + "bad_trajectory": 0, + "metric_reps": 5.0, + "metric_time_s": 14.62, + "metric_touches": 5.0, + "grader_error": 0, + "video_rendered": 1, + "agent_sessions": 104, + "agent_resets": 473, + "agent_replays": 4, + "agent_cost_self_reported": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.543, + "peak_processes": 10, + "peak_disk_gb": 0.256, + "resource_emergency_kills": 0, + "resource_oom_kills": 0, + "resource_host_oom_kills": 0, + "resource_service_oom_kills": 0, + "session_log_tampered": 0, + "session_log_from_monitor": 0 + } + }, + "video": { + "kind": "hf", + "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-toes-to-bar-i00-privileged--bquh9w6/replay.mp4", + "duration_s": 14.64, + "fps": 25.0, + "frames": 366, + "sha256": "35792baae7121783245e745d4332d1af2387b1c5dfb29a35ff08393a1a091f31" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 5, + "base": 0, + "peak": 5, + "final": 5, + "nodes": { + "toes_to_bar_1": { + "method": "visual", + "intervals": [ + [ + 2.56, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 1.04, + "log_step": null, + "text_zh": "观察：视频 1.04–1.20 s，银灰机器人第1轮屈髋举起双腿，双脚到达双手之间的横杠处，左右手仍留在杠上；随后双脚离开杠向下回落。" + }, + { + "t": 2.56, + "log_step": null, + "text_zh": "观察：视频 2.40–2.64 s，机器人双腿降回近竖直悬垂，双脚远低于横杠，双手继续悬挂承重；以视频 2.56 s 为第1次回垂完成的近似锚点，视觉时间误差约0.2秒。" + }, + { + "t": 2.56, + "log_step": null, + "text_zh": "推断：本次视频五轮举足、回垂顺序清楚，且 inputs/run.json 的评分器记录 metric_reps=5、metric_touches=5、stages_reached=5、replay_success=1、deterministic=1，支持第1次完整重复已被原生裁判接受；毫米级双脚接触位置、脚中心在杠后及1.40米回垂阈值不能单靠远景量测。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "观察：第1轮视频 1.04–2.64 s，左右手各自环在横杠的两处，双臂共同承重维持悬挂，举足和回垂过程中手部均留在杠上；远景不能分辨每根手指的接触。" + }, + { + "who": "腿/躯干", + "how": "其他", + "note_zh": "候选：触碰。观察：视频 1.04–1.20 s，双脚并举到两手之间的横杠处；推断：结合本次评分器 metric_touches=5 和五阶段完成，双脚实际触杠被接受。脚部短暂触碰无需抓握、推扫或承重，现有怎么碰条目无准确对应。" + } + ] + }, + "toes_to_bar_2": { + "method": "visual", + "intervals": [ + [ + 5.52, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 4.12, + "log_step": null, + "text_zh": "观察：视频 4.12–4.28 s，银灰机器人第2轮屈髋举起双腿，双脚到达双手之间的横杠处，左右手仍留在杠上；随后双脚离开杠向下回落。" + }, + { + "t": 5.52, + "log_step": null, + "text_zh": "观察：视频 5.36–5.68 s，机器人双腿降回近竖直悬垂，双脚远低于横杠，双手继续悬挂承重；以视频 5.52 s 为第2次回垂完成的近似锚点，视觉时间误差约0.2秒。" + }, + { + "t": 5.52, + "log_step": null, + "text_zh": "推断：本次视频五轮举足、回垂顺序清楚，且 inputs/run.json 的评分器记录 metric_reps=5、metric_touches=5、stages_reached=5、replay_success=1、deterministic=1，支持第2次完整重复已被原生裁判接受；毫米级双脚接触位置、脚中心在杠后及1.40米回垂阈值不能单靠远景量测。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "观察：第2轮视频 4.12–5.68 s，左右手各自环在横杠的两处，双臂共同承重维持悬挂，举足和回垂过程中手部均留在杠上；远景不能分辨每根手指的接触。" + }, + { + "who": "腿/躯干", + "how": "其他", + "note_zh": "候选：触碰。观察：视频 4.12–4.28 s，双脚并举到两手之间的横杠处；推断：结合本次评分器 metric_touches=5 和五阶段完成，双脚实际触杠被接受。脚部短暂触碰无需抓握、推扫或承重，现有怎么碰条目无准确对应。" + } + ] + }, + "toes_to_bar_3": { + "method": "visual", + "intervals": [ + [ + 8.64, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 7.16, + "log_step": null, + "text_zh": "观察：视频 7.16–7.32 s，银灰机器人第3轮屈髋举起双腿，双脚到达双手之间的横杠处，左右手仍留在杠上；随后双脚离开杠向下回落。" + }, + { + "t": 8.64, + "log_step": null, + "text_zh": "观察：视频 8.48–8.76 s，机器人双腿降回近竖直悬垂，双脚远低于横杠，双手继续悬挂承重；以视频 8.64 s 为第3次回垂完成的近似锚点，视觉时间误差约0.2秒。" + }, + { + "t": 8.64, + "log_step": null, + "text_zh": "推断：本次视频五轮举足、回垂顺序清楚，且 inputs/run.json 的评分器记录 metric_reps=5、metric_touches=5、stages_reached=5、replay_success=1、deterministic=1，支持第3次完整重复已被原生裁判接受；毫米级双脚接触位置、脚中心在杠后及1.40米回垂阈值不能单靠远景量测。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "观察：第3轮视频 7.16–8.76 s，左右手各自环在横杠的两处，双臂共同承重维持悬挂，举足和回垂过程中手部均留在杠上；远景不能分辨每根手指的接触。" + }, + { + "who": "腿/躯干", + "how": "其他", + "note_zh": "候选：触碰。观察：视频 7.16–7.32 s，双脚并举到两手之间的横杠处；推断：结合本次评分器 metric_touches=5 和五阶段完成，双脚实际触杠被接受。脚部短暂触碰无需抓握、推扫或承重，现有怎么碰条目无准确对应。" + } + ] + }, + "toes_to_bar_4": { + "method": "visual", + "intervals": [ + [ + 11.68, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 10.2, + "log_step": null, + "text_zh": "观察：视频 10.20–10.32 s，银灰机器人第4轮屈髋举起双腿，双脚到达双手之间的横杠处，左右手仍留在杠上；随后双脚离开杠向下回落。" + }, + { + "t": 11.68, + "log_step": null, + "text_zh": "观察：视频 11.52–11.80 s，机器人双腿降回近竖直悬垂，双脚远低于横杠，双手继续悬挂承重；以视频 11.68 s 为第4次回垂完成的近似锚点，视觉时间误差约0.2秒。" + }, + { + "t": 11.68, + "log_step": null, + "text_zh": "推断：本次视频五轮举足、回垂顺序清楚，且 inputs/run.json 的评分器记录 metric_reps=5、metric_touches=5、stages_reached=5、replay_success=1、deterministic=1，支持第4次完整重复已被原生裁判接受；毫米级双脚接触位置、脚中心在杠后及1.40米回垂阈值不能单靠远景量测。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "观察：第4轮视频 10.20–11.80 s，左右手各自环在横杠的两处，双臂共同承重维持悬挂，举足和回垂过程中手部均留在杠上；远景不能分辨每根手指的接触。" + }, + { + "who": "腿/躯干", + "how": "其他", + "note_zh": "候选：触碰。观察：视频 10.20–10.32 s，双脚并举到两手之间的横杠处；推断：结合本次评分器 metric_touches=5 和五阶段完成，双脚实际触杠被接受。脚部短暂触碰无需抓握、推扫或承重，现有怎么碰条目无准确对应。" + } + ] + }, + "toes_to_bar_5": { + "method": "visual", + "intervals": [ + [ + 14.6, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 13.2, + "log_step": null, + "text_zh": "观察：视频 13.20–13.36 s，银灰机器人第5轮屈髋举起双腿，双脚到达双手之间的横杠处，左右手仍留在杠上；随后双脚离开杠向下回落。" + }, + { + "t": 14.6, + "log_step": null, + "text_zh": "观察：视频 14.44–14.60 s，机器人双腿降回近竖直悬垂，双脚远低于横杠，双手继续悬挂承重；以视频 14.60 s 为第5次回垂完成的近似锚点，视觉时间误差约0.2秒。" + }, + { + "t": 14.6, + "log_step": null, + "text_zh": "推断：本次视频五轮举足、回垂顺序清楚，且 inputs/run.json 的评分器记录 metric_reps=5、metric_touches=5、stages_reached=5、replay_success=1、deterministic=1，支持第5次完整重复已被原生裁判接受；毫米级双脚接触位置、脚中心在杠后及1.40米回垂阈值不能单靠远景量测。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "观察：第5轮视频 13.20–14.60 s，左右手各自环在横杠的两处，双臂共同承重维持悬挂，举足和回垂过程中手部均留在杠上；远景不能分辨每根手指的接触。" + }, + { + "who": "腿/躯干", + "how": "其他", + "note_zh": "候选：触碰。观察：视频 13.20–13.36 s，双脚并举到两手之间的横杠处；推断：结合本次评分器 metric_touches=5 和五阶段完成，双脚实际触杠被接受。脚部短暂触碰无需抓握、推扫或承重，现有怎么碰条目无准确对应。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "reps", + "value": 5.0, + "unit": "count", + "threshold": 5, + "direction": "higher_better" + }, + { + "name": "grip_break_max_s", + "value": null, + "unit": "s", + "threshold": 0.05, + "direction": "lower_better" + } + ], + "failure": null, + "summary_zh": "本次评分录像显示机器人双手悬挂，连续完成五轮举足至杠和回垂；五个事件节点分别约在视频 2.56、5.52、8.64、11.68、14.60 s 达成并锁存，评分器确认5次重复、5次触杠及确定性回放成功。录像末段仍保持悬挂，未见中断，故无失败断点；日志中的早期调试失败不属于这段交付轨迹。无状态日志，远景无法独立核验双脚接触点边界、精确回垂阈值或0.05秒脱杠窗口，节点验收由可见动作与本次评分器成功记录共同支持；评分器未报告 grip_break_max_s，保留为null。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "五个事件节点的时间（约2.56/5.52/8.64/11.68/14.60 s）与视频中五轮举足回垂一致，评分器 metric_reps=5、stages 5/5 支持全部达成；成功无断点，接触标注合理。远景较小，只能核对节奏。", + "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": "2026-10-29T10:05:32Z", + "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/embodied-olympics-fun-toes-to-bar-i00-privileged--bquh9w6/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-toes-to-bar-i00-privileged--bquh9w6/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "889167ceabbc83a3c0c7e4e6b5cb15d2ea535308eb766c06eb6864bf2386c1c9", + "video_sha256": "35792baae7121783245e745d4332d1af2387b1c5dfb29a35ff08393a1a091f31", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/toes_to_bar/trajectories/embodied-olympics-fun-toes-to-bar-i00-standard--l45qynp.json b/phase2/embodied-olympics-fun/toes_to_bar/trajectories/embodied-olympics-fun-toes-to-bar-i00-standard--l45qynp.json new file mode 100644 index 0000000000000000000000000000000000000000..207f4d3130623d85bb6ef43fa974cd25c6259e88 --- /dev/null +++ b/phase2/embodied-olympics-fun/toes_to_bar/trajectories/embodied-olympics-fun-toes-to-bar-i00-standard--l45qynp.json @@ -0,0 +1,248 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-toes-to-bar-i00-standard--l45qynp", + "taskKey": "embodied-olympics-fun/toes_to_bar", + "dag_sha256": "ac964ab85ff6f7f76f4e7777abc98665ca1819d4122120d46d6b9b1067731cff", + "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": 459, + "sim_steps": 3218, + "calls": 36, + "rows": 3218, + "resets": 5, + "reset_budget": 5, + "sim_startups": 1, + "elapsed_s": 3322.8, + "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": 3.997, + "peak_processes": 11, + "peak_disk_gb": 0.329, + "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/embodied-olympics-fun-toes-to-bar-i00-standard--l45qynp/replay.mp4", + "duration_s": 4.6, + "fps": 25.0, + "frames": 115, + "sha256": "8fbf04ce5fc3690841202cf16c8fcd0eff53c2df8c6431036342d70598105165" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 5, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "toes_to_bar_1": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0, + "log_step": null, + "text_zh": "观察：本次视频 0.00 s 机器人双手握住横杠悬挂，双脚在杠下；初始悬挂不是一次触杠回垂。" + }, + { + "t": 4.56, + "log_step": 192, + "text_zh": "观察：本次视频约 2.80–4.56 s 机器人摆动并抬起双腿，末段双脚仍高于完整回垂位置，未出现抬足后回到低位悬垂的完整循环。log step 192 在最后一次 reset 后的服务返回 rep1 outcome=\"let go of the bar\"、stages=0、done=true、resets_left=0；这里的 outcome 和 stages 是机器人服务所接入原生裁判的结果，不是 agent 自述。" + }, + { + "t": 4.16, + "log_step": null, + "text_zh": "观察：已逐帧查看本次视频 3.60–4.56 s 的机器人裁剪图，双脚上举接近横杠，但未能确认双脚同时实际触杠；成功参考视频约 1.20–3.80 s 可见举足至杠附近后再降回低位的完整循环。本次没有该回垂结果，不能把接近横杠计作完成。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "观察：本次视频 0.00–3.60 s，机器人的左右手分别环握横杠两处，共同承载悬挂和举腿摆动；末段服务报告脱杠（log step 192）。双脚虽上举，但未确认实际碰杠，因此不添加脚杠触碰。" + } + ] + }, + "toes_to_bar_2": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 4.56, + "log_step": 192, + "text_zh": "观察：本次视频 0.00–4.56 s 只有第一轮举腿摆动，片尾仍处于抬腿姿态，没有第2次触杠后回悬垂的动作；最后一次实时尝试的原生裁判在 log step 192 返回 stages=0 并因脱杠结束。推断：第1次完整动作未完成，按顺序锁存的第2次节点也未达到，且机器人尚未开始该次动作。" + } + ] + }, + "toes_to_bar_3": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 4.56, + "log_step": 192, + "text_zh": "观察：本次视频 0.00–4.56 s 只有第一轮举腿摆动，片尾仍处于抬腿姿态，没有第3次触杠后回悬垂的动作；最后一次实时尝试的原生裁判在 log step 192 返回 stages=0 并因脱杠结束。推断：第1次完整动作未完成，按顺序锁存的第3次节点也未达到，且机器人尚未开始该次动作。" + } + ] + }, + "toes_to_bar_4": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 4.56, + "log_step": 192, + "text_zh": "观察：本次视频 0.00–4.56 s 只有第一轮举腿摆动，片尾仍处于抬腿姿态，没有第4次触杠后回悬垂的动作；最后一次实时尝试的原生裁判在 log step 192 返回 stages=0 并因脱杠结束。推断：第1次完整动作未完成，按顺序锁存的第4次节点也未达到，且机器人尚未开始该次动作。" + } + ] + }, + "toes_to_bar_5": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 4.56, + "log_step": 192, + "text_zh": "观察：本次视频 0.00–4.56 s 只有第一轮举腿摆动，片尾仍处于抬腿姿态，没有第5次触杠后回悬垂的动作；最后一次实时尝试的原生裁判在 log step 192 返回 stages=0 并因脱杠结束。推断：第1次完整动作未完成，按顺序锁存的第5次节点也未达到，且机器人尚未开始该次动作。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "reps", + "value": 0, + "unit": "count", + "threshold": 5, + "direction": "higher_better" + }, + { + "name": "grip_break_max_s", + "value": null, + "unit": "s", + "threshold": 0.05, + "direction": "lower_better" + } + ], + "failure": { + "breakpoint": { + "ref": "toes_to_bar_1", + "t": 4.5, + "log_step": 192, + "where": "操作中", + "evidence_zh": "观察：本次视频约 3.60–4.56 s 机器人正在第一次举腿，双脚抬高后尚未降回完整悬垂；log step 192 的最后一次 reset 后，原生裁判返回 rep1 outcome=\"let go of the bar\"、stages=0，并结束该次尝试。推断：第一次完整触杠回垂因手部接触未保持而中断；视频约 4.5 s 仅作末段导航锚点，不是由 agent 工作时间换算的精确脱杠时刻。" + }, + "unmet": { + "toes_to_bar_1": false, + "toes_to_bar_2": false, + "toes_to_bar_3": false, + "toes_to_bar_4": false, + "toes_to_bar_5": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "没保持住", + "t": 4.5, + "log_step": 192, + "evidence_zh": "观察：本次视频 0.00–3.60 s 机器人已由双手握杠悬挂并开始摆腿；视频 4.00–4.56 s 仍在第一次举足过程中。最后一次尝试的机器人服务在 log step 192 明确报告原生裁判结果 \"let go of the bar\"，完成阶段数为0，随即 done=true。推断：已有的手杠承载接触未保持到完成回垂；不把同一脱杠事件再重复标为平衡或违反约束。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志", + "evidence_zh": "推断：agent 用焊接手腕的本地模型优化摆动（log step 61 的说明及 log step 62 的 weld 代码），而真实任务的双手依靠接触摩擦承重，模型不能检验手从杠上滑脱。最终控制器在 log step 192 为左右手持续发送闭合值1，执行 log step 191 构造的摆腿轨迹，却仍被原生裁判判为脱杠；结合本次视频 2.80–4.56 s 的举腿运动，支持动态握杠控制未能维持承载接触。没有足够证据区分手部接触几何、关节跟踪或负载反馈问题，故停在控制总类。" + } + ], + "gap_zh": "本次没有状态日志。固定远景虽能判断未完成回垂，但无法分辨哪只手先失去接触、接触中断的精确时长或腕部力矩，grip_break_max_s 未报告；reps=0 取自 log step 192 原生裁判 stages=0（五个阶段分别对应一次完整重复）。agent 在 log step 197 自述各次尝试均在脚触杠后脱杠，这属于跨尝试总结；本次视频未确认双脚实际接触，因此未据此添加脚杠接触，也不采信其腕关节屈服假说为已证实根因。" + }, + "summary_zh": "机器人在本次视频中保持初始双手悬挂，随后开始第一次摆动举足，但直到视频 4.56 s 仍未完成触杠后的回悬垂，五个节点均未达到。最后一次实时尝试的原生裁判返回脱杠、0阶段（log step 192），断点为第1次触杠回垂，现象为没保持住，根因保守归为控制。没有逐步接触或力矩日志，无法确认双脚是否同时触杠、先脱杠的手及具体失握机制。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "标签与视频和日志一致：视频 0–4.56 s 只有一次举腿摆动，未见回垂；log step 192 原生裁判返回 let go of the bar、stages=0；五个节点未达到，断点 toes_to_bar_1、接触/没保持住、控制根因保守且证据已标注为推断。", + "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": "2026-10-29T10:05:50Z", + "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/embodied-olympics-fun-toes-to-bar-i00-standard--l45qynp/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-toes-to-bar-i00-standard--l45qynp/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "c2d80892e05765e0fd0733b9315bbda3d099ad7f74928c7ed4905191c4ec6114", + "video_sha256": "8fbf04ce5fc3690841202cf16c8fcd0eff53c2df8c6431036342d70598105165", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/triple_jump/task_dag.json b/phase2/embodied-olympics-fun/triple_jump/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..41dc4bbcdd4bcdf7c41a46be67f5da2277e988f1 --- /dev/null +++ b/phase2/embodied-olympics-fun/triple_jump/task_dag.json @@ -0,0 +1,407 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "embodied-olympics-fun/triple_jump", + "suite": "embodied-olympics-fun", + "base": "embodied-olympics-fun-triple-jump-i00", + "title_zh": "三级跳远", + "summary_zh": "机器人依次完成起跳线后单脚起跳同脚落地、跨步跳异脚落地及跃入沙坑，并通过首次触沙后两秒的合法落地验收，最近触沙成绩至少为七米。", + "embodiment_zh": "两种模式均为 Unitree G1 双足人形机器人，29 个关节，身高 1.32 米、质量约 33 千克；双臂末端为无可动手指的固定拳垫，无可抓物体、无轮式底盘，跳跃由腿足和躯干完成。", + "task_class": "compositional", + "analysis_level": "step", + "spec": { + "version": "1.2", + "task": "embodied-olympics-fun-triple-jump-i00", + "bench": "embodied-olympics-fun", + "progress": "live", + "entities": {}, + "milestones": [ + { + "id": "hop_landed", + "zh": "单足跳同脚落地", + "level": "semantic", + "mode": "event", + "when": null, + "after": [], + "hold_s": 0, + "goal": true, + "freeze_held": false, + "criterion_zh": "在跑道或场地相机中，机器人从白色起跳线后单脚蹬地，完整离地至少 0.05 秒，骨盆比离地时升高至少 0.05 米；首次越线着地只由该起跳脚接触红色跑道，所有落地点均在宽 1.22 米的跑道内且尚未到沙坑。另一脚须比起跳脚提前至少 0.05 秒离地，起跳支撑点须在跑道宽度内；未形成合格腾空、双脚起跳、异脚或其他部位先着地均未完成，裁判对应 hop 阶段。", + "step": { + "cat": "本体", + "item": "跳跃" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "左脚或右脚在红色跑道上支撑、蹬地，同一只脚承接这次落地；左右脚均可作为起跳脚，腾空时无外部接触。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:104-110 初态、参数和阶段", + "inputs/task/privileged/tests/rfo_task.py:168-218 单足跳验收", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:353-394 腾空检测", + "inputs/task/privileged/tests/rfo_task.py:153-154 场地与跑道相机" + ], + "conf": 6, + "note_zh": "助跑并入本步，不设助跑或起跳节点。一次跳跃的起跳、腾空和落地不可分开；落地后可原地单足支撑停顿，下一次跳跃才是下一步。事件达成后保留，不要求起跳脚一直接地。" + }, + { + "id": "step_landed", + "zh": "跨步跳异脚落地", + "level": "semantic", + "mode": "event", + "when": null, + "after": [ + "hop_landed" + ], + "hold_s": 0, + "goal": true, + "freeze_held": false, + "criterion_zh": "同脚落地后，机器人由这只脚再次蹬地，完整离地至少 0.05 秒，首次着地只由另一只脚接触红色跑道，所有落地点仍在跑道两条白边线之间且位于沙坑前，裁判对应 step 阶段。从上一次落地到本次腾空，只允许上次落地脚接触地面，其接地时足部几何中心相对裁判锚点的水平位移不得超过 0.30 米；原脚落地、双脚或其他部位先着地、仅挪步而未合格腾空均未完成。", + "step": { + "cat": "本体", + "item": "跳跃" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "前一步落地脚支撑红色跑道并蹬地，另一只脚支撑红色跑道承接落地；两脚依次接地，腾空时无外部接触。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:108-110 腾空和足部位移阈值", + "inputs/task/privileged/tests/rfo_task.py:168-171 跑道接触范围", + "inputs/task/privileged/tests/rfo_task.py:219-243 跨步跳和中间支撑规则", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:353-394 腾空检测" + ], + "conf": 6, + "note_zh": "次序由文本和裁判共同要求。裁判锚点取落地后下一次检测到该脚接地时的足部几何中心，而非精确的首次落地接触点；本步完成后可原地单足支撑停顿。" + }, + { + "id": "jump_in_sand", + "zh": "跃入沙坑", + "level": "semantic", + "mode": "event", + "when": null, + "after": [ + "step_landed" + ], + "hold_s": 0, + "goal": true, + "freeze_held": false, + "criterion_zh": "机器人由跨步跳落地脚蹬地，完整离地至少 0.05 秒，首次着地只接触黄色沙坑，不能同时接触坑外硬地；随后从首次触沙起完整经过 2.0 秒，期间任一部位的触沙点中最靠近起跳线者为成绩，位于线后至少 7.0 米，且不得在该成绩点更靠近起跳线的一侧接触坑外地面。场地相机应可见身体首次落入黄色区域，裁判须依次记录 jump 和 landing；落在坑前、落地观察未满或期间出现上述坑外接触均未完成，允许用任意部位触沙及在坑内摔倒，不要求站稳。", + "step": { + "cat": "本体", + "item": "跳跃" + }, + "contact": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "跨步跳落地脚在红色跑道上支撑并蹬地，通常由脚、腿或躯干接触黄色沙坑承重；起跳、触沙依次发生，腾空时无外部接触。", + "plan": 1 + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:48-57 沙坑为 x=7 至 16 米、宽 2.75 米", + "inputs/task/privileged/tests/rfo_task.py:108-110 腾空和落地窗口参数", + "inputs/task/privileged/tests/rfo_task.py:219-263 最后一次跳跃和落地验收", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-178 阶段按序锁存后成功" + ], + "conf": 6, + "note_zh": "将裁判 jump 与 landing 两个标志合为一次跃入沙坑的验收；hold_s=0，2.0 秒由首次触沙时刻起算并在判据中保留，不能对已完成的 landing 再加一次等待。最终成绩不另设节点，代码以沙坑几何保证至少 7.0 米。" + } + ], + "constraints": [ + { + "id": "takeoff_line_fault", + "zh": "起跳线犯规", + "never": null, + "criterion_zh": "在有效单足跳落地前，接地脚的整个几何范围（包括脚上悬空的部分）达到或越过起跳线，或未经合格腾空就有身体部位在起跳线及其前方接地，裁判立即判负。合格单足跳本身的越线落地不属于此犯规。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:190-200" + ] + }, + { + "id": "hop_fault", + "zh": "单足跳犯规", + "never": null, + "criterion_zh": "首次越线着地所结束的腾空不满足单脚先后离地至少相差 0.05 秒、起跳支撑点位于跑道宽度内、骨盆上升至少 0.05 米、起跳脚单独落在跑道且不触沙中的任一项，裁判立即判负。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:201-218" + ] + }, + { + "id": "interphase_contact_fault", + "zh": "跳间接触犯规", + "never": null, + "criterion_zh": "单足跳落地至跨步跳、跨步跳落地至最后一跳之间，除上一阶段落地脚外的任何身体部位触地（包括悬空腿擦地），裁判立即判负。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:219-225" + ] + }, + { + "id": "landing_foot_moved", + "zh": "落地脚移位", + "never": null, + "criterion_zh": "两次中间落地后至下一次合格腾空前，落地脚接地时足部几何中心距裁判保存的水平锚点超过 0.30 米，裁判立即判负；锚点在该阶段首次进入接地保持检查时设置。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:226-233" + ] + }, + { + "id": "step_fault", + "zh": "跨步跳犯规", + "never": null, + "criterion_zh": "跨步跳的首次着地并非另一只脚单独接触跑道，或触沙，或任一硬地接触点不在跑道宽度内及沙坑之前，裁判立即判负。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:235-243" + ] + }, + { + "id": "jump_fault", + "zh": "入坑犯规", + "never": null, + "criterion_zh": "最后一跳首次着地未接触沙坑，或同时有任何硬地接触，裁判立即判负；首次触沙可以使用任一身体部位。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:244-249" + ] + }, + { + "id": "landing_outside_nearer", + "zh": "落地坑外回触", + "never": null, + "criterion_zh": "首次触沙起 2.0 秒内，在比当前所有触沙点的最近成绩更靠近起跳线的位置触及坑外硬地，裁判立即判负。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:250-263" + ] + }, + { + "id": "left_venue", + "zh": "离开场地", + "never": null, + "criterion_zh": "机器人触及场地地块下方的降低地面即判离场：场地地块横向范围为轴线左右各 8 米，纵向从起跳线后 45 米至沙坑远端再前 6 米（x=22 米）；仅在空中越过边界尚不是代码的触发条件。", + "consequence": "立即终止", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:41-58", + "inputs/task/privileged/tests/rfo_task.py:156-177" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "hop_landed", + "level": "new_code", + "source": "inputs/task/privileged/tests/rfo_task.py:216-218；记录 ref.stages 中的 hop 及其足侧与腾空数据" + }, + { + "atom": "step_landed", + "level": "new_code", + "source": "inputs/task/privileged/tests/rfo_task.py:241-243；记录 ref.stages 中的 step 及其足侧与腾空数据" + }, + { + "atom": "jump_in_sand", + "level": "new_code", + "source": "inputs/task/privileged/tests/rfo_task.py:244-263；记录 ref.stages 中的 landing、首次触沙时刻和 mark" + }, + { + "atom": "episode_success", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:174-178；记录每帧原生 ref.success，不能使用最终交付或重放有效性判决" + }, + { + "atom": "takeoff_line_fault", + "level": "new_code", + "source": "inputs/task/privileged/tests/rfo_task.py:190-200；需逐帧记录相关接触/足部状态及 ref.failed 对应原生犯规原因" + }, + { + "atom": "hop_fault", + "level": "new_code", + "source": "inputs/task/privileged/tests/rfo_task.py:201-218；需逐帧记录相关接触/足部状态及 ref.failed 对应原生犯规原因" + }, + { + "atom": "interphase_contact_fault", + "level": "new_code", + "source": "inputs/task/privileged/tests/rfo_task.py:219-225；需逐帧记录相关接触/足部状态及 ref.failed 对应原生犯规原因" + }, + { + "atom": "landing_foot_moved", + "level": "new_code", + "source": "inputs/task/privileged/tests/rfo_task.py:226-233；需逐帧记录相关接触/足部状态及 ref.failed 对应原生犯规原因" + }, + { + "atom": "step_fault", + "level": "new_code", + "source": "inputs/task/privileged/tests/rfo_task.py:235-243；需逐帧记录相关接触/足部状态及 ref.failed 对应原生犯规原因" + }, + { + "atom": "jump_fault", + "level": "new_code", + "source": "inputs/task/privileged/tests/rfo_task.py:244-249；需逐帧记录相关接触/足部状态及 ref.failed 对应原生犯规原因" + }, + { + "atom": "landing_outside_nearer", + "level": "new_code", + "source": "inputs/task/privileged/tests/rfo_task.py:250-263；需逐帧记录相关接触/足部状态及 ref.failed 对应原生犯规原因" + }, + { + "atom": "left_venue", + "level": "new_code", + "source": "inputs/task/privileged/tests/rfo_task.py:41-58；inputs/task/privileged/tests/rfo_task.py:156-177；需逐帧记录相关接触/足部状态及 ref.failed 对应原生犯规原因" + } + ] + }, + "native_metrics": [ + { + "name": "hop_air_s", + "unit": "s", + "threshold": 0.05, + "direction": "higher_better", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:109,183,206" + ] + }, + { + "name": "step_air_s", + "unit": "s", + "threshold": 0.05, + "direction": "higher_better", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:109,235" + ] + }, + { + "name": "jump_air_s", + "unit": "s", + "threshold": 0.05, + "direction": "higher_better", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:109,235" + ] + }, + { + "name": "hop_rise_m", + "unit": "m", + "threshold": 0.05, + "direction": "higher_better", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:110,205-212" + ] + }, + { + "name": "hop_m", + "unit": "m", + "threshold": null, + "direction": "higher_better", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:217 记录相对起跳线的落地点，不是该次跳跃的长度" + ] + }, + { + "name": "step_m", + "unit": "m", + "threshold": null, + "direction": "higher_better", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:242 记录相对起跳线的落地点，不是该次跳跃的长度" + ] + }, + { + "name": "mark_m", + "unit": "m", + "threshold": 7.0, + "direction": "higher_better", + "sources": [ + "inputs/task/privileged/tests/rfo_task.py:48-57,252-263 沙坑接触几何隐含至少 7 米，无额外显式成绩比较" + ] + } + ], + "open_questions_zh": [ + "没有成功示例：inputs/oracle_none.txt 说明 16 条目录运行全部失败；没有可观看的成功视频，也没有 solution 文件。步骤仅依据文本、冻结初态和裁判代码规划（N2），相机可见线索依据场地颜色和相机设置，尚未由成功视频核实。", + "inputs/states/index.json 的 logs 为空，所有节点 when、约束 never 和 success 均为 null（P8）；atoms 的 new_code 只说明将来需记录的原生阶段与接触信息，未假定任何状态日志字段已经存在。", + "S10 粒度解释：文本区分三次跳跃，代码允许在两次中间落地后原地单足支撑任意时长（除总时限），因此保留三个可停顿续做的步骤；每一次的起跳、腾空、落地合为一步，最后两秒并入入坑验收，不设第四个落地节点。", + "模式差异：运动裁判和冻结参数相同，三个 rfo_task.py 副本哈希相同；privileged 的 grade.py:87-115 要求合法轨迹文件、两次新进程重放都成功且终态哈希相同，standard 的 standard/server.py:157-178 与 efe_standard-files/v2_1/grade.py:93-99 判定实时回合成功及有效终结。交付和执行有效性不属于世界状态节点，因无日志 success 保持 null，不能声称已验证与目录判决相等。", + "任务设计问题：privileged/instruction.md:12 称坐标原点在初始骨盆正下方，但该文本的规则及 rfo_task.py:105,115-121 设起跳线 x=0、初始骨盆 x=-20，冻结 instance.json 的 t0.pelvis 也为 [-20,0,0.793,0]。这里所有数值位置采用裁判的起跳线坐标系，避免将线和沙坑错移二十米。", + "任务说明说落地脚不能偏离首次落地点超过 0.30 米；代码 rfo_task.py:226-233 的 anchor 却在落地后下一次检测到该脚接地时才设置。采用代码锚点定义验收和约束，视频只能近似判断此差异。", + "文本要求成绩至少 7.0 米，代码 rfo_task.py:260-263 不再显式比较 mark 阈值；因沙坑从 x=7 起，合格触沙几何已隐含该要求。hop_m 和 step_m 是相对起跳线的落地位置，只记录，无独立及格距离阈值。", + "60 秒是回合时限（rfo_task.py:104、referee.py:284-289），按 P21 不另建约束；两秒落地窗口必须在回合结束前完成。fall_rule=False，摔倒本身不判负，只有具体阶段的非法接触才判负；未加入站稳要求。" + ], + "conf": 6, + "conf_reason_zh": "文本、两模式一致的场景和裁判代码支持三步及其先后，阈值均有来源；没有成功视频或状态日志，无法实测绑定、可停顿策略及视频判据，且文本存在坐标和足部锚点表述差异。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "三步（单足跳、跨步跳、入坑跳）与文本一致，助跑和最后两秒落地窗口并入各步；链式边是规则要求的阶段顺序；各犯规约束与裁判失败条件对应且为立即终止；无状态日志故 when/never/success 为 null 且已在 open_questions_zh 说明；文本与代码的不一致（坐标原点、锚点、7.0 米阈值）均已列出。未发现需要修改的内容。", + "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：['hop_landed', 'step_landed', 'jump_in_sand']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:45:48Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__triple_jump/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__triple_jump/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "5b0bca9569bbe0ffece10b50bce92920a5bf1029f005f252c9c845f405ea60ce", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/embodied-olympics-fun/triple_jump/trajectories/embodied-olympics-fun-triple-jump-i00-standard--dgs3xjw.json b/phase2/embodied-olympics-fun/triple_jump/trajectories/embodied-olympics-fun-triple-jump-i00-standard--dgs3xjw.json new file mode 100644 index 0000000000000000000000000000000000000000..5fe578240993b24b29ab0d62c08d3914e274e3f3 --- /dev/null +++ b/phase2/embodied-olympics-fun/triple_jump/trajectories/embodied-olympics-fun-triple-jump-i00-standard--dgs3xjw.json @@ -0,0 +1,254 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-triple-jump-i00-standard--dgs3xjw", + "taskKey": "embodied-olympics-fun/triple_jump", + "dag_sha256": "0baf08d04fbf745076feddee38258dc6f6ef2b10629220cbb9e20880d00d44ba", + "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": 681, + "sim_steps": 7052, + "calls": 222, + "rows": 3552, + "resets": 5, + "reset_budget": 5, + "sim_startups": 1, + "elapsed_s": 3157.8, + "video_rendered": 1, + "first_success": 0, + "first_reason_code": 3, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.0, + "peak_memory_gb": 0.556, + "peak_processes": 10, + "peak_disk_gb": 0.329, + "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/embodied-olympics-fun-triple-jump-i00-standard--dgs3xjw/replay.mp4", + "duration_s": 6.84, + "fps": 25.0, + "frames": 171, + "sha256": "4da02b67bfceede224b5c4d215c076fb2138e25e64318fc996cf9f855a0e6614" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 3, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "hop_landed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 6.24, + "log_step": 78, + "text_zh": "观察：视频 5.20–6.24 s，机器人沿红色跑道向白色起跳线助跑，仍位于线后；log step 78 明确重置为 episode 6，并输出左腿支撑后执行 hop 推蹬。" + }, + { + "t": 6.72, + "log_step": 78, + "text_zh": "观察：视频 6.32–6.56 s，机器人在线前离地并向前俯转；视频 6.68–6.80 s，身体降至跑道、在起跳线附近扑倒，没有出现越线后由起跳脚单独承接的合格落地。log step 78 的服务终态为 episode 6、done=true、reason=failed。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：视频 5.20–6.32 s，机器人双脚交替接触红色跑道承重并尝试蹬地；视频 6.72–6.80 s，倒地后的腿及躯干接触跑道。不能由远景可靠区分最后一次支撑的足侧。" + } + ] + }, + "step_landed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 6.8, + "log_step": 78, + "text_zh": "观察：视频 6.32–6.80 s 仅见首次跳跃尝试及扑倒，没有合格的同脚落地，也没有随后由该脚再次蹬地、另一脚单独落地的跨步跳；log step 78 在 hop 阶段后即报告回合 failed，未输出 step 推蹬。" + } + ] + }, + "jump_in_sand": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 6.8, + "log_step": 78, + "text_zh": "观察：视频全程 0–6.80 s 黄色沙坑内没有机器人；末帧机器人倒在白色起跳线附近的红色跑道上，未完成跨步跳或进入沙坑，亦无触沙后的两秒验收。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "hop_air_s", + "unit": "s", + "threshold": 0.05, + "direction": "higher_better", + "value": null + }, + { + "name": "step_air_s", + "unit": "s", + "threshold": 0.05, + "direction": "higher_better", + "value": null + }, + { + "name": "jump_air_s", + "unit": "s", + "threshold": 0.05, + "direction": "higher_better", + "value": null + }, + { + "name": "hop_rise_m", + "unit": "m", + "threshold": 0.05, + "direction": "higher_better", + "value": null + }, + { + "name": "hop_m", + "unit": "m", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "step_m", + "unit": "m", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "mark_m", + "unit": "m", + "threshold": 7.0, + "direction": "higher_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "hop_landed", + "t": 6.5, + "log_step": 78, + "where": "操作中", + "evidence_zh": "观察：视频 6.32–6.56 s，机器人第一次跳跃时身体持续向前俯转；视频 6.68–6.80 s 在起跳线附近扑倒，未完成规定的同脚落地，随后回合结束。log step 78 的服务状态为 episode 6、done=true、reason=failed；log step 79 重置被拒绝，之后未修复该步。" + }, + "unmet": { + "hop_landed": false, + "step_landed": false, + "jump_in_sand": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "平衡", + "t": 6.5, + "log_step": 78, + "evidence_zh": "观察：视频 6.32–6.64 s，机器人腾空后上身不断向前翻俯，未能把足部放回身体下方承接；视频 6.72–6.80 s 身体横倒在起跳线附近，失去直立支撑。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "t": 6.5, + "log_step": 78, + "source_zh": "视频", + "evidence_zh": "推断：视频 5.20–6.12 s 助跑时机器人保持直立，转入首次跳跃后（视频 6.24–6.80 s）连续前倾并扑倒，说明此次推蹬至腾空、落地的全身运动控制未维持可用的落地姿态。log step 78 同时记录脚本执行 hop 推蹬后未检测到落足，回合已失败；证据不足以细分为特定增益、接触几何或轨迹求解错误。" + } + ], + "gap_zh": "观察：inputs/ 没有状态日志及成功参考视频，评分字段也未报告阶段标志、原生腾空时长、骨盆升高或落点距离；远景不能精确核查足侧、离地先后及最先触地的身体部位，故原生指标全部留 null。agent 在 log step 89 自述视觉锚点错误导致助跑越线犯规，但视频 6.24–6.56 s 仍显示机器人在线前起跳、前俯，未见其所说的起跳前接地脚越线；不据此增标感知或任务设计根因，具体原生判负条款无法确认。log step 78 的 EpisodeOver 出现在服务已判 failed 之后，不作为先行的脚本崩溃原因。" + }, + "summary_zh": "评分对象为 log step 78 的最后一次实时回合 episode 6；机器人完成助跑并尝试首次跳跃，但未完成合格的同脚落地，最终三个节点均未达成。视频 6.32–6.80 s 显示机器人腾空后持续前俯并在起跳线附近扑倒，断点为单足跳同脚落地，现象为平衡失稳，根因保守归为控制。缺少状态及原生阶段数据，不能验证 agent 在 log step 89 所称的视觉锚点误判、起跳前越线原因；后续脚本修改没有重跑，不改变本回合标签。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "评分对象为 episode 6（log id 78，control_steps 681 ≈ 视频 6.84 s，对应得上）。视频中机器人助跑后在线前腾空、前俯，于起跳线附近扑倒，三个节点均未达成，断点 hop_landed、平衡、控制均有帧证据支持；labeller 未采纳 agent 自述的『起跳前越线犯规』也合理（视频 6.0–6.4 s 机器人仍在线后约 1–3 米处起跳）。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": "2026-10-29T10:02:58Z", + "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/embodied-olympics-fun-triple-jump-i00-standard--dgs3xjw/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-triple-jump-i00-standard--dgs3xjw/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "e9bcca7155dfa9ba06fc5e2ccbae6a6f11b56086e07ab15b76cf14763b84c75f", + "video_sha256": "4da02b67bfceede224b5c4d215c076fb2138e25e64318fc996cf9f855a0e6614", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/triple_jump/trajectories/embodied-olympics-fun-triple-jump-i00-standard--nzkqtsz.json b/phase2/embodied-olympics-fun/triple_jump/trajectories/embodied-olympics-fun-triple-jump-i00-standard--nzkqtsz.json new file mode 100644 index 0000000000000000000000000000000000000000..f978126da6ecdab0bfd6492221999b5618a2f30c --- /dev/null +++ b/phase2/embodied-olympics-fun/triple_jump/trajectories/embodied-olympics-fun-triple-jump-i00-standard--nzkqtsz.json @@ -0,0 +1,256 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-triple-jump-i00-standard--nzkqtsz", + "taskKey": "embodied-olympics-fun/triple_jump", + "dag_sha256": "0baf08d04fbf745076feddee38258dc6f6ef2b10629220cbb9e20880d00d44ba", + "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": 701, + "sim_steps": 701, + "calls": 7, + "rows": 701, + "resets": 0, + "reset_budget": 50, + "sim_startups": 1, + "elapsed_s": 667.5, + "video_rendered": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.687, + "peak_processes": 10, + "peak_disk_gb": 2.27, + "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/embodied-olympics-fun-triple-jump-i00-standard--nzkqtsz/replay.mp4", + "duration_s": 7.04, + "fps": 25.0, + "frames": 176, + "sha256": "23b251e6d9cc6e9cc96e884913c06eb01fdde865c854f59a1c3a34c7b40fb44c" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 3, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "hop_landed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 6.8, + "log_step": 28, + "text_zh": "观察：视频 5.4–7.0 s 中，机器人从画面左侧跑向白色起跳线，身体保持直立，足部逐渐偏到红色跑道远侧白边线外；未出现跑道内的合格越线同脚落地。log step 28 的实时服务返回 done=True、reason=failed、outcome=\"took off outside the board's ends\"、stages_reached=0/4。该 outcome 是服务内运动裁判的判决，不是 agent 的自述。" + }, + { + "t": 7.0, + "log_step": 28, + "text_zh": "推断：log step 3 给出 control_dt=0.01 s，log step 28 的 control_steps=701 对应仿真 7.01 s；与视频 7.0 s 机器人抵达起跳线附近、随后视频结束的事件一致，因此将该裁判终止结果对应到本视频末端。起跳支撑点超出板的横向范围，单足跳未获接受。" + } + ], + "contact_actual": [ + { + "who": "腿/躯干", + "how": "支撑", + "note_zh": "观察：视频 5.4–7.0 s，机器人双脚交替接地承重、蹬地助跑，沿红色跑道远侧边缘前进，末段偏至邻接绿色地面；未见躯干倒地或双手触地。" + } + ] + }, + "step_landed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 7.0, + "log_step": 28, + "text_zh": "推断：视频 5.4–7.0 s 仅见机器人助跑至起跳线附近，未形成已验收单足跳之后的跨步跳异脚落地；log step 28 的实时裁判阶段计数仍为 0/4，且因起跳点超出板端立即终止。必需前置 hop_landed 未达成，故本事件未达成，未实际开展本阶段。" + } + ] + }, + "jump_in_sand": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 7.0, + "log_step": 28, + "text_zh": "观察：已检查视频全程及视频 6.0–7.0 s 的逐帧画面，黄色沙坑始终空着；末帧机器人仍在白色起跳线附近，与沙坑之间尚有红色跑道，没有跃入沙坑或随后两秒的触沙验收。log step 28 的实时裁判已终止回合且阶段计数为 0/4。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "hop_air_s", + "unit": "s", + "threshold": 0.05, + "direction": "higher_better", + "value": null + }, + { + "name": "step_air_s", + "unit": "s", + "threshold": 0.05, + "direction": "higher_better", + "value": null + }, + { + "name": "jump_air_s", + "unit": "s", + "threshold": 0.05, + "direction": "higher_better", + "value": null + }, + { + "name": "hop_rise_m", + "unit": "m", + "threshold": 0.05, + "direction": "higher_better", + "value": null + }, + { + "name": "hop_m", + "unit": "m", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "step_m", + "unit": "m", + "threshold": null, + "direction": "higher_better", + "value": null + }, + { + "name": "mark_m", + "unit": "m", + "threshold": 7.0, + "direction": "higher_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "hop_landed", + "t": 6.9, + "log_step": 28, + "where": "操作中", + "evidence_zh": "观察：视频 6.6–7.0 s，机器人在起跳线附近继续跑动，足部位于跑道远侧边线外，没有完成合法的越线同脚落地；log step 28 的实时运动裁判以“took off outside the board's ends”判负，阶段 0/4。助跑属于单足跳这一步，后续无重置或修复；log step 44 的模型过滤提示出现在本回合已经判负之后。" + }, + "unmet": { + "hop_landed": false, + "step_landed": false, + "jump_in_sand": false + }, + "phenomena": [ + { + "cat": "全身", + "item": "步态", + "t": 6.9, + "log_step": 28, + "evidence_zh": "观察：视频 6.0–7.0 s 逐帧可见机器人保持直立、交替迈腿，沿跑道远侧偏出的路线跑向起跳线，未摔倒；起跳支撑位置没有保持在允许的跑道宽度内，log step 28 的裁判对应判为起跳超出板端。" + } + ], + "root_causes": [ + { + "cat": "认知", + "item": "感知", + "source_zh": "日志", + "evidence_zh": "推断：agent 的视觉定位给助跑纠偏提供了错误位置。log step 21 的脚本优先采用场地相机的背景差分定位，只要结果非空就舍弃跑道相机结果；在仿真 3.0、4.0、5.0 s 三次均报告同一位置 [-1.609,-0.820] 米，并据此持续发出正向横移 vy=0.12 m/s、正向转向命令。视频 3.0–5.0 s 机器人尚未进入该场地相机视野，视频 5.4–7.0 s 才从左侧进入并向远侧边线外偏移；log step 24 改用机器人足部像素重新反投影，在仿真 6.0 s 得到 y=+0.472 米，与此前被持续使用的负侧偏差符号相反。错误的目标定位使脚本未及时停止向远侧纠偏，最终板端外起跳由 log step 28 的裁判结果证实。" + }, + { + "cat": "认知", + "item": "推理/规划", + "source_zh": "日志", + "evidence_zh": "推断：agent 在唯一受评实时回合中采用了先跑过起跳线查看裁判的诊断策略，没有在越线前切换到所需的单足跳。log step 10 的笔记明确知道单脚起跳、同脚落地要求；但 log step 28 的调用标签为“Running across take-off line to inspect referee output”，实际脚本仅提交 run 模式、前进速度 3.8 m/s 和转向修正，服务随后立即判负。log step 15、21、28 的服务记录始终为 episode=1，resets=0，诊断性跑过线直接成为受评轨迹；这一策略即使横向定位正确，也没有安排同脚落地的跳跃序列。" + } + ], + "gap_zh": "场地相机在视频约 0–5.2 s 没有拍到机器人，后段机器人尺寸很小；没有状态日志，不能精确重建每只脚的接触时刻、腾空时长、骨盆升高或横向偏移量，未达成阶段由后段画面与同一实时回合的原生裁判结果共同判定。DAG 的七项原生指标在 reward 和服务终止结果中均未报告，全部填 null。log step 29–43 是判负之后的本地分析，未作为新的受评运动；log step 44 的过滤中断也未改变此前已固定的物理失败结果。" + }, + "summary_zh": "机器人完成了助跑，但在视频约 6.9 s 偏到跑道远侧边线外接近起跳线，裁判以起跳超出板端判负，三个节点均未达成（0/3）。日志支持视觉误定位导致错误纠偏，以及 agent 在同一受评回合中直接跑过线查看裁判、未切换到合格单足跳的策略问题；模型过滤发生在回合已经判负之后。视频前段机器人不在场地相机视野内，且缺少状态日志，精确足部接触与跳跃数值无法核实。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "三个节点均未达成、断点 hop_landed 与裁判 outcome（log step 28，服务读数）一致；视频 5.3–7.0 s 机器人确实偏到跑道远侧边线外的草地上。感知（log step 21 场地相机定位恒定不变、step 24 反投影 y=+0.47）和推理/规划（step 28 诊断性跑过线直接成为受评轨迹）均有日志支撑，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": "2026-10-29T10:04:07Z", + "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/embodied-olympics-fun-triple-jump-i00-standard--nzkqtsz/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-triple-jump-i00-standard--nzkqtsz/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "6b05055b725e109e9c022d67ed4cfc642ca8dce62882238dbf8710ba3c8d9631", + "video_sha256": "23b251e6d9cc6e9cc96e884913c06eb01fdde865c854f59a1c3a34c7b40fb44c", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/volleyball_spike/task_dag.json b/phase2/embodied-olympics-fun/volleyball_spike/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..9f13ec8dbe8e0cb91688e9b4ee01c2696ae45af2 --- /dev/null +++ b/phase2/embodied-olympics-fun/volleyball_spike/task_dag.json @@ -0,0 +1,388 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "embodied-olympics-fun/volleyball_spike", + "suite": "embodied-olympics-fun", + "base": "embodied-olympics-fun-volleyball-spike-i00", + "title_zh": "排球扣杀", + "summary_zh": "机器人合法腾空手扣下落排球，使球越网首次落入对方场内，并无犯规地落地站稳一秒。", + "embodiment_zh": "三个模式均为双足 Unitree H1-2 人形机器人，27 关节、身高 1.78 米、质量 67 千克；双臂腕端各有固定拳垫，无工作手指，不能抓握排球，无轮式底盘，双腿负责起跳及落地支撑。", + "task_class": "atomic", + "analysis_level": "step", + "spec": { + "version": "1.2", + "task": "embodied-olympics-fun-volleyball-spike-i00", + "bench": "embodied-olympics-fun", + "progress": "live", + "entities": {}, + "milestones": [ + { + "id": "spike_in_court_landed", + "zh": "扣球并落地站稳", + "mode": "event", + "when": null, + "after": [], + "hold_s": 0, + "goal": true, + "criterion_zh": "场馆斜俯视中，机器人在网前起跳，手部短促击中已过最高点而下落的球，全部触点高于网顶 2.43 米，首次触球时身体腾空且该次飞行骨盆较离地时升高至少 0.10 米；触球跨度不超过 0.10 秒、脱离身体连续 0.03 秒后确认击球结束，离手球速不低于来球速且竖直速度向下，球随后从两标志杆之间越过球网平面（球心高度至少 2.43 米，可碰网但不可碰标志杆），首次触地球心在对方橙色场区含边线内（起点坐标 x 为 3 至 12 米、y 为 -8 至 1 米）。机器人落地后，在球已落入场内且该次起跳合格的前提下，连续 100 个控制步（名义 1 秒）至少一脚触地、触地部位仅为脚、躯干竖直余弦至少 0.94、骨盆速度不超过 0.35 米每秒，并全程没有下列裁判犯规，才算完成；仅挥臂、原地垫球、球未入对方场区、仍在空中或落地后未站稳均未完成。", + "step": { + "cat": "动态", + "item": "击打" + }, + "contact": [ + { + "who": "手指/夹爪", + "how": "其他", + "plan": 1, + "note_zh": "候选名称：击打；左手或右手腕端拳垫直接短促击打排球，现有接触方式没有击打，不能归为抓握或推/扫。" + }, + { + "who": "双臂协作", + "how": "其他", + "plan": 2, + "note_zh": "候选名称：击打；左右手腕端拳垫同时短促击打同一排球，文本允许一手或双手，现有接触方式没有击打。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/privileged/instruction.md:3-29", + "inputs/task/standard/instruction.md:3-29", + "inputs/task/standard-realtime/instruction.md:3-29", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:742-784 击球起止、速度与触球跨度", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:872-880 阶段与实例阈值", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:927-962 场地坐标、起点与飞行追踪", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:981-1052 合法扣球、越网与首次落点", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:1054-1062 起跳确认及站立验收", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:149-184 阶段锁存、成功及控制步保持计数", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:292-301 站立接触、竖直余弦与速度", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:353-394 飞行边界和骨盆高度", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:1103-1105 躯干竖直余弦", + "inputs/task/privileged/environment/instance.json:1 参数与冻结初态", + "inputs/oracle/replay.mp4 斜俯视可见网前起跳、挥手击球、落地及末段站立" + ], + "conf": 6, + "note_zh": "按 S10，一次连续扣球连同球的飞行、落点及机器人落地站稳合为一步，四个裁判阶段只作内部顺序验收。类型按送出排球的击打动作确定；先后发生的双脚起跳和落地支撑写于此处，不与空中手球接触列成同时接触；允许任意单手或双手，不限定成功录像的右手。使用已包含站立计数的完成事件，hold_s=0 避免再加一次保持窗口；100 个 0.01 秒控制步的首末采样跨度为 0.99 秒。球落地后为死球，不要求它最后留在场内；完成事件锁存。" + } + ], + "constraints": [ + { + "id": "machine_interference", + "zh": "干扰发球", + "never": null, + "criterion_zh": "在排球尚未离开发球机轮组时，机器人任意部位接触发球机或待发球；录像中表现为机器人触及双轮机器或从机器内碰球。发球后的机器接触未受源码此分支检查，文本差异见待议项。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:967-979", + "inputs/task/privileged/instruction.md:25" + ] + }, + { + "id": "invalid_spike", + "zh": "扣球犯规", + "never": null, + "criterion_zh": "球成为死球前，发生非手部触球、在来球下降标志建立前首次触球、首次触球时未腾空、任一手球触点不高于 2.43 米、触球跨度超过 0.10 秒，或最后离手球竖直速度非负或速率小于来球速；击球所在飞行结束时，骨盆最大升高四舍五入到四位小数后不足 0.10 米也立即失败。录像可见以臂或身体碰球、脚未离地击球、低于网顶击球等；精确速度及毫秒边界需裁判事件核实。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:981-991", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:998-1032", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:755-784" + ] + }, + { + "id": "second_ball_touch", + "zh": "二次触球", + "never": null, + "criterion_zh": "第一次击球已被连续离体 0.03 秒确认结束后、球首次合法落入场内之前，任一手再次触球即失败；同一次至多 0.10 秒击球跨度内的接触由同一击球验收。合法落地后的死球接触不再受此规则判断。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:994-1006", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:755-781" + ] + }, + { + "id": "invalid_net_crossing", + "zh": "越网犯规", + "never": null, + "criterion_zh": "球成为死球前碰到任一标志杆，或未经击球便越网，或第一次从本方穿过 x=3 米网平面的插值球心高度小于 2.43 米、y 在 [-8,1] 米之外即失败；录像中可见从网下或标志杆外穿过。排球接触网面本身允许。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:1003-1004", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:1033-1043", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:400-423" + ] + }, + { + "id": "invalid_first_landing", + "zh": "首次落点犯规", + "never": null, + "criterion_zh": "排球第一次触地在合法扣球完成之前，或尚未合法越网，或该时刻球心不在对方场区 x=[3,12] 米、y=[-8,1] 米内即失败；落在己方橙色半场或蓝色界外区不合格，压边线按场内接受。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:927-930", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:1044-1052" + ] + }, + { + "id": "robot_net_contact", + "zh": "机器人触网", + "never": null, + "criterion_zh": "整次尝试直到完成站稳，机器人任意部位接触球网、任一标志杆或网柱即失败；近网挥手和落地时仍须与这些部件保持分离。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:811-819", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:952-957", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:992-994" + ] + }, + { + "id": "centre_line_fault", + "zh": "越过中线", + "never": null, + "criterion_zh": "任一脚触地时，该脚全部地面接触点均在对方一侧且超出中线轴 0.025 米（起点坐标 x>3.025 米），判脚完全越过中线并终止；仅身体腾空伸过中线或脚仍压住中线不满足此判据。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:832-839", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:1054-1058" + ] + }, + { + "id": "robot_fallen", + "zh": "机器人跌倒", + "never": null, + "criterion_zh": "控制步检查时骨盆高度低于地面上方 0.35 米，或除左脚、右脚外任何机器人部位接触地面即失败；录像中可见跪地、手撑地、身体倒地等，起跳腾空本身不是跌倒。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:873-876", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:272-289", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/task.py:23-25" + ] + }, + { + "id": "simulation_unstable", + "zh": "仿真数值发散", + "never": null, + "criterion_zh": "仿真出现加速度数值警告、非有限位姿，或源码所检查的线速度自由度绝对值超过 150 米每秒，服务以仿真不稳定结束尝试；仅凭画面轻微晃动不能认定，需服务的对应终止原因或明确数值爆炸证据。", + "consequence": "立即终止", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:108-109", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:867-885", + "inputs/task/privileged/instruction.md:145" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "dyn.spike_in_court_landed", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:981-1062；需新增逐帧记录裁判原生有序 stages、trackers.jump_ok、_hold[stand] 与 failed，组成全部四阶段完成且无犯规的世界事件；当前无状态日志。" + }, + { + "atom": "dyn.machine_interference", + "level": "new_code", + "source": "需新增记录此物理步或控制步的犯规检查结果及其所用接触、速度或数值稳定状态；依据 inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:967-979; inputs/task/privileged/instruction.md:25" + }, + { + "atom": "dyn.invalid_spike", + "level": "new_code", + "source": "需新增记录此物理步或控制步的犯规检查结果及其所用接触、速度或数值稳定状态；依据 inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:981-991; inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:998-1032; inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:755-784" + }, + { + "atom": "dyn.second_ball_touch", + "level": "new_code", + "source": "需新增记录此物理步或控制步的犯规检查结果及其所用接触、速度或数值稳定状态；依据 inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:994-1006; inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:755-781" + }, + { + "atom": "dyn.invalid_net_crossing", + "level": "new_code", + "source": "需新增记录此物理步或控制步的犯规检查结果及其所用接触、速度或数值稳定状态；依据 inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:1003-1004; inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:1033-1043; inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:400-423" + }, + { + "atom": "dyn.invalid_first_landing", + "level": "new_code", + "source": "需新增记录此物理步或控制步的犯规检查结果及其所用接触、速度或数值稳定状态；依据 inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:927-930; inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:1044-1052" + }, + { + "atom": "dyn.robot_net_contact", + "level": "new_code", + "source": "需新增记录此物理步或控制步的犯规检查结果及其所用接触、速度或数值稳定状态；依据 inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:811-819; inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:952-957; inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:992-994" + }, + { + "atom": "dyn.centre_line_fault", + "level": "new_code", + "source": "需新增记录此物理步或控制步的犯规检查结果及其所用接触、速度或数值稳定状态；依据 inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:832-839; inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:1054-1058" + }, + { + "atom": "dyn.robot_fallen", + "level": "new_code", + "source": "需新增记录此物理步或控制步的犯规检查结果及其所用接触、速度或数值稳定状态；依据 inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:873-876; inputs/grader/efe_embodied-olympics-fun-files/rfo/referee.py:272-289; inputs/grader/efe_embodied-olympics-fun-files/rfo/task.py:23-25" + }, + { + "atom": "dyn.simulation_unstable", + "level": "new_code", + "source": "需新增记录此物理步或控制步的犯规检查结果及其所用接触、速度或数值稳定状态；依据 inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:108-109; inputs/grader/efe_embodied-olympics-fun-files/rfo/world.py:867-885; inputs/task/privileged/instruction.md:145" + } + ] + }, + "native_metrics": [ + { + "name": "contact_height_m", + "unit": "m", + "threshold": 2.43, + "direction": "higher_better", + "note_zh": "最低手球触点必须严格高于阈值；记录值仅保留三位小数，判定使用未舍入值。", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:1007-1027" + ] + }, + { + "name": "jump_rise_m", + "unit": "m", + "threshold": 0.1, + "direction": "higher_better", + "note_zh": "击球所在整段飞行的骨盆最高点相对离地高度；按源码四位小数舍入后比较。", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:984-991" + ] + }, + { + "name": "speed_in_m_s", + "unit": "m/s", + "threshold": null, + "direction": null, + "note_zh": "首次手球接触之前的球速，是离手速度的比较基准。", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:755-784", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:1024-1032" + ] + }, + { + "name": "speed_out_m_s", + "unit": "m/s", + "threshold": "speed_in_m_s", + "direction": "higher_better", + "note_zh": "离手速率至少为来球速率的 1.0 倍，并另要求离手竖直速度严格为负；不是固定速度阈值。", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:1024-1032" + ] + }, + { + "name": "air_s", + "unit": "s", + "threshold": null, + "direction": null, + "note_zh": "击球所在飞行时长，记录但没有正的最小时长门槛。", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:958-959", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:984-991" + ] + }, + { + "name": "landing_x_m", + "unit": "m", + "threshold": [ + 3.0, + 12.0 + ], + "direction": "target", + "note_zh": "首次触地时球心的起点坐标，闭区间；数值记录保留三位小数。", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:927-930", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:1044-1052" + ] + }, + { + "name": "landing_y_m", + "unit": "m", + "threshold": [ + -8.0, + 1.0 + ], + "direction": "target", + "note_zh": "首次触地时球心的起点坐标，闭区间；数值记录保留三位小数。", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:927-930", + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:1044-1052" + ] + }, + { + "name": "ball_height_m", + "unit": "m", + "threshold": null, + "direction": null, + "note_zh": "当前球心高度，仅为过程诊断，不要求成功时仍高于球网。", + "sources": [ + "inputs/grader/efe_embodied-olympics-fun-files/rfo/tasks/volleyball_spike.py:1059" + ] + } + ], + "open_questions_zh": [ + "inputs/states/index.json 的 logs 为空；按 P8 所有 when、never 和 success 均为 null，不能从源码假定存在日志字段。录像能显示起跳挥手、球的飞行与落地站立，但最低触点、速度比、毫秒触球跨度及精确落点边界仍需原生裁判数据核实；原语按 new_code 标记待补日志。", + "粒度按 DECISIONS.md S10：起跳扣球、越网、球入场与落地站稳不能暂停后另行完成，合并为一个原子步骤；文本列出的四个阶段仅规定验收顺序，不按裁判阶段拆节点。", + "任务设计问题：文本笼统禁止机器人碰发球机，但源码 launched() 仅在球尚未出轮时检查机器与待发球接触，发球后不再检查机器（instruction.md:25；volleyball_spike.py:967-979）。本文件机器干扰约束明确标出评分器实际覆盖的发球前范围，发球后的文本禁令未被实现。", + "任务设计问题：文本要求所有手球接触都发生在腾空中，源码只在首次触球事件检查 fl.air，后续接触只检查高度、时长与球速；一次击球可在 0.03 秒离体确认前继续接触，并未逐次重验腾空（volleyball_spike.py:755-781,1011-1023）。验收按实际裁判边界描述，不能凭一次成功录像确定边界例是否会出现。", + "文本的站立速度描述为低于 0.35 米每秒，源码为不超过 0.35；约 20 度实际使用躯干竖直余弦至少 0.94，且只要求至少一脚接地、没有其他部位接地，不要求双脚同时落地（referee.py:292-301）。越网按球心高度至少网高判断，首次落点按触地时球心在闭区间场地内判断；击球触点则严格高于网高。", + "三模式的冻结起点、机器人和参数相同，使用同一任务模块与裁判；privileged 还要求交付轨迹在两个新进程重放均成功且末态哈希相同，standard 与 standard-realtime 评服务内最后一次尝试的实时裁判结果（grade.py:106-115；standard/server.py:1-12,166-178；standard-files/v2_1/grade.py:93-99）。这些交付和有效性门槛不增设语义节点；当前没有状态日志可表达完整 success。", + "成功参照为 nhymezu 的评分录像和 run.json 的 reward；已浏览起跳击球与末段站立。录像分辨率有限，不能独立核准全部毫秒级条件，未使用其右手、助跑时长、击球角度或落点数值作为通用要求；球可碰网，离手向下不意味着整个后续球路始终向下。" + ], + "conf": 6, + "conf_reason_zh": "文本、同源裁判和一致初态支持一个连续扣球步骤及全部阈值，成功录像显示相应动作且原生评分成功；但没有状态日志，精细接触和速度只能依赖裁判语义，且机器接触和腾空检查存在文本与实现边界差异。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "单个原子节点（扣球→越网→入场→落地站稳）按 S10 合并合理，成功条件各合取项均在 criterion_zh 内，约束对应裁判失败条件，无状态日志故 when/never/success 全为 null 符合 P8，待议项已列出文本与裁判差异。小备注：simulation_unstable 属平台终止而非文本规则，保留无害。", + "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：['spike_in_court_landed']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:46:36Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__volleyball_spike/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/embodied-olympics-fun__volleyball_spike/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "1ff103294e89c7dd9518b238cfc14d4c576be306f9ab8a7e72a3eea1601643f0", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/embodied-olympics-fun/volleyball_spike/trajectories/embodied-olympics-fun-volleyball-spike-i00-privileged--gat4ufp.json b/phase2/embodied-olympics-fun/volleyball_spike/trajectories/embodied-olympics-fun-volleyball-spike-i00-privileged--gat4ufp.json new file mode 100644 index 0000000000000000000000000000000000000000..2a397ad6f4cd21ee2d7628047ea0fc0fda0553ae --- /dev/null +++ b/phase2/embodied-olympics-fun/volleyball_spike/trajectories/embodied-olympics-fun-volleyball-spike-i00-privileged--gat4ufp.json @@ -0,0 +1,233 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-volleyball-spike-i00-privileged--gat4ufp", + "taskKey": "embodied-olympics-fun/volleyball_spike", + "dag_sha256": "f4bddfc185de81e8d53ee7f0cfb80e88756f3bb0aa176b76645d88cc0c11176b", + "outcome": { + "success": 1, + "reward": 1.0, + "source": "catalog", + "raw": { + "success": 1, + "replay_success": 1, + "deterministic": 1, + "n_actions": 666, + "sim_steps": 666, + "horizon": 1200, + "episode_over": 1, + "progress": 1.0, + "stages_reached": 4, + "stages_total": 4, + "fell": 0, + "replay_s": 7.31, + "bad_trajectory": 0, + "metric_air_s": 0.575, + "metric_ball_height_m": 1.096, + "metric_contact_height_m": 2.479, + "metric_jump_rise_m": 0.3761, + "metric_landing_x_m": 7.58, + "metric_landing_y_m": -0.168, + "metric_speed_in_m_s": 7.01, + "metric_speed_out_m_s": 7.62, + "grader_error": 0, + "video_rendered": 1, + "agent_sessions": 1956, + "agent_resets": 0, + "agent_replays": 11, + "agent_cost_self_reported": 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/embodied-olympics-fun-volleyball-spike-i00-privileged--gat4ufp/replay.mp4", + "duration_s": 6.68, + "fps": 25.0, + "frames": 167, + "sha256": "4ad27939f479bd1f811df7037ff8df6532af5045a328b5968b13de22ac7fd48a" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "spike_in_court_landed": { + "method": "visual", + "intervals": [ + [ + 6.64, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 4.32, + "log_step": null, + "text_zh": "观察：视频 4.20–4.40 s 的逐帧放大图中，机器人在网前腾空，一侧手部向下落的排球挥动；视频 4.28–4.36 s 球经过腕端拳垫附近，随后向球网另一侧飞离。机器人没有包握或持住球。" + }, + { + "t": 4.88, + "log_step": null, + "text_zh": "观察：视频 4.40–4.96 s 中，球向对方半场运动，在远侧橙色场地近地后反弹；机器人约在视频 4.6 s 已以双脚落地，屈膝前倾后逐步直起身体。首次触地约在视频 4.8–4.9 s，画面中球很小，落点精确坐标采用评分器记录。" + }, + { + "t": 6.64, + "log_step": null, + "text_zh": "观察：视频 5.6–6.64 s 中机器人脚部留在地面，躯干直立且位置变化很小；已检查视频末段，包括视频 6.32–6.64 s 每帧，未见倒地或其他部位撑地。" + }, + { + "t": 6.64, + "log_step": 163, + "text_zh": "观察：inputs/run.json 的评分器 reward 为 success=1、replay_success=1、deterministic=1、stages_reached=4/4、fell=0；触球高度 2.479 m、跳跃升高 0.3761 m、来球速率 7.01 m/s、离手速率 7.62 m/s、首次落点 (7.580, -0.168) m。log step 163 中交付轨迹的渲染重放也返回全部四阶段完成、执行 666 行和 success=true；log step 179 核对最终生成文件字节一致。" + }, + { + "t": 6.64, + "log_step": null, + "text_zh": "推断：结合视频末段站立及评分器四阶段成功结果，将包含落地站稳验收的整个事件记为视频 6.64 s 末帧附近完成，时间为约数；视频 4.3 s 的击球本身尚不足以完成此节点。无状态日志，25 fps 远景不能独立核实毫秒级接触跨度及精确速度、站立速度阈值，这些细节依据本次裁判成功结果判定。" + } + ], + "contact_actual": [ + { + "who": "手指/夹爪", + "how": "其他", + "note_zh": "观察：视频 4.28–4.36 s，机器人单侧腕端拳垫短促击打排球；候选名称：击打，现有接触方式没有对应条目，非抓握。此前机器人双脚接触场地完成助跑与起跳，之后约在视频 4.6 s 双脚落地并支撑站立；这些脚地支撑与腾空击球先后发生。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "contact_height_m", + "unit": "m", + "threshold": 2.43, + "direction": "higher_better", + "value": 2.479 + }, + { + "name": "jump_rise_m", + "unit": "m", + "threshold": 0.1, + "direction": "higher_better", + "value": 0.3761 + }, + { + "name": "speed_in_m_s", + "unit": "m/s", + "threshold": null, + "direction": null, + "value": 7.01 + }, + { + "name": "speed_out_m_s", + "unit": "m/s", + "threshold": "speed_in_m_s", + "direction": "higher_better", + "value": 7.62 + }, + { + "name": "air_s", + "unit": "s", + "threshold": null, + "direction": null, + "value": 0.575 + }, + { + "name": "landing_x_m", + "unit": "m", + "threshold": [ + 3.0, + 12.0 + ], + "direction": "target", + "value": 7.58 + }, + { + "name": "landing_y_m", + "unit": "m", + "threshold": [ + -8.0, + 1.0 + ], + "direction": "target", + "value": -0.168 + }, + { + "name": "ball_height_m", + "unit": "m", + "threshold": null, + "direction": null, + "value": 1.096 + } + ], + "failure": null, + "summary_zh": "机器人在网前起跳，以单侧腕端拳垫击中下落排球，球越网后首次落入对方场区，机器人落地调整后直立站稳；唯一节点“扣球并落地站稳”在视频末段约 6.64 s 完成，评分器确认四阶段全部通过。视频与成功参考呈现的起跳、击球、落地站立过程一致；本次没有状态日志，精确触点高度、速度及毫秒级验收依据本次评分器结果，节点完成时刻仅为视频近似锚点。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "单节点成功运行，评分器 success=1、stages 4/4、fell=0；视频中网前起跳、单侧腕端击球、球越网、落地站立与标注一致，节点完成于视频末段约 6.64 s（666 步≈6.66 s，与视频时长吻合）。contact_actual 为 手指/夹爪+其他（击打），符合 S16。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": "2026-10-29T10:04:27Z", + "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/embodied-olympics-fun-volleyball-spike-i00-privileged--gat4ufp/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-volleyball-spike-i00-privileged--gat4ufp/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "01b542723c3f5492c0ba9a34dc74caed11b2c7e548b1296a7b051b18cf23eaa0", + "video_sha256": "4ad27939f479bd1f811df7037ff8df6532af5045a328b5968b13de22ac7fd48a", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-fun/volleyball_spike/trajectories/embodied-olympics-fun-volleyball-spike-i00-standard-realtime--j35rzt9.json b/phase2/embodied-olympics-fun/volleyball_spike/trajectories/embodied-olympics-fun-volleyball-spike-i00-standard-realtime--j35rzt9.json new file mode 100644 index 0000000000000000000000000000000000000000..219cbda2263feaac967c839fc670abed4f9b1d1a --- /dev/null +++ b/phase2/embodied-olympics-fun/volleyball_spike/trajectories/embodied-olympics-fun-volleyball-spike-i00-standard-realtime--j35rzt9.json @@ -0,0 +1,240 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-fun-volleyball-spike-i00-standard-realtime--j35rzt9", + "taskKey": "embodied-olympics-fun/volleyball_spike", + "dag_sha256": "f4bddfc185de81e8d53ee7f0cfb80e88756f3bb0aa176b76645d88cc0c11176b", + "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": 0, + "sim_steps": 0, + "calls": 1, + "rows": 0, + "resets": 0, + "reset_budget": 5, + "sim_startups": 1, + "elapsed_s": 1694.9, + "video_rendered": 1, + "first_success": 0, + "first_reason_code": 4, + "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": 0, + "request_busy_s": 0.227, + "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/embodied-olympics-fun-volleyball-spike-i00-standard-realtime--j35rzt9/replay.mp4", + "duration_s": 0.04, + "fps": 25.0, + "frames": 1, + "sha256": "d70e84dc934beef48c440caa742c5edf0a0c7436521d5ea1ad92c79fb425ac8c" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "spike_in_court_landed": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察：已检查本次评分视频的全部一帧（视频 0.00 s，总长 0.04 s）；场馆斜俯视中机器人站在己方场内，双臂未举起扣球，画面未显示起跳、手球接触或球越网落入对方场区，扣球并落地站稳事件未发生。" + }, + { + "t": null, + "log_step": 6, + "text_zh": "观察：log step 6 中 agent 仅向机器人服务请求 seconds=0 的初始观测，服务返回 control_steps=0、rows=0、running=false、planned_steps=0；log step 45 的服务状态仍为 running=False。评分器 reward 同时报 n_actions=0、sim_steps=0、rows=0、clock_s=0.0，与评分视频停在初态一致。" + }, + { + "t": null, + "log_step": 53, + "text_zh": "观察：log step 53 的最终消息为“This request was blocked by Gemini's filters”；此前 log step 25–52 的动作测试在 agent 自建的本地 MuJoCo 场景内进行，未向评分服务提交运动。推断：外部过滤在评分回合启动之前中断会话，离线测试不构成本节点的实际完成。" + } + ], + "contact_actual": [ + { + "who": "不适用", + "how": "不适用", + "note_zh": "观察：评分器记录 n_actions=0、sim_steps=0、rows=0，视频 0.00 s 仅有初态；没有向评分服务交付运动轨迹，故扣球步骤的轨迹级接触不适用。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "contact_height_m", + "unit": "m", + "threshold": 2.43, + "direction": "higher_better", + "value": null + }, + { + "name": "jump_rise_m", + "unit": "m", + "threshold": 0.1, + "direction": "higher_better", + "value": null + }, + { + "name": "speed_in_m_s", + "unit": "m/s", + "threshold": null, + "direction": null, + "value": null + }, + { + "name": "speed_out_m_s", + "unit": "m/s", + "threshold": "speed_in_m_s", + "direction": "higher_better", + "value": null + }, + { + "name": "air_s", + "unit": "s", + "threshold": null, + "direction": null, + "value": null + }, + { + "name": "landing_x_m", + "unit": "m", + "threshold": [ + 3.0, + 12.0 + ], + "direction": "target", + "value": null + }, + { + "name": "landing_y_m", + "unit": "m", + "threshold": [ + -8.0, + 1.0 + ], + "direction": "target", + "value": null + }, + { + "name": "ball_height_m", + "unit": "m", + "threshold": null, + "direction": null, + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "unknown", + "t": null, + "log_step": 53, + "where": "首步前", + "evidence_zh": "观察：log step 53 返回 Gemini 过滤器阻断提示；log step 6 的机器人服务观测显示 control_steps=0、running=false，log step 45 状态仍为 running=False；评分器记录 n_actions=0、sim_steps=0、rows=0，本次视频 0.00 s 为唯一初始帧。推断：评分运动尚未启动或交付，结果固定前即被模型供应商过滤中断，无法指定物理失败步骤，按外因中断记 unknown／首步前。" + }, + "unmet": { + "spike_in_court_landed": false + }, + "phenomena": [], + "root_causes": [], + "gap_zh": "观察：inputs/ 提供的评分视频只有视频 0.00 s 一帧，没有状态日志；评分器未报告 DAG 所列的触球高度、起跳升高、入球与出球速度、腾空时长、落点坐标及球高，各原生指标值均留 null。log step 53 是供应商过滤阻断，评分服务在 log step 6、45 尚未运行且最终动作数为零；不存在可检查的评分运动或物理失误。按外因中断规则，现象和根因均留空，不将离线 MuJoCo 测试中的跌倒、击球动作或调参结果移作本次评分证据，也不标没尝试。" + }, + "summary_zh": "本次评分视频只有视频 0.00 s 的初始站立画面，扣球并落地站稳节点未达到，最终为 0/1；服务及评分器记录均显示未提交动作、未推进仿真。log step 53 返回 Gemini 过滤器阻断提示，属于结果固定前的外因中断，断点记 unknown／首步前，现象及根因留空，不作机器人能力失败归因。会话中的本地 MuJoCo 测试未进入评分回合；缺少评分运动与状态日志，原生指标亦未报告。", + "conf": 10, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 9, + "summary_zh": "标注与证据一致：评分视频仅一帧初态，评分器 n_actions=0、sim_steps=0，log step 53 为 Gemini 过滤器阻断（交付前外因中断），按 FAIL-5 记 unknown／首步前、现象根因为空且 gap_zh 说明，节点 false 正确。", + "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": "2026-10-29T10:03:06Z", + "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/embodied-olympics-fun-volleyball-spike-i00-standard-realtime--j35rzt9/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-fun-volleyball-spike-i00-standard-realtime--j35rzt9/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "2df6be7707fc4e40571034102aed8a94b4d910ad9a7ffe97a92607ba708e15c8", + "video_sha256": "d70e84dc934beef48c440caa742c5edf0a0c7436521d5ea1ad92c79fb425ac8c", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-match/tabletennis_match_i04/task_dag.json b/phase2/embodied-olympics-match/tabletennis_match_i04/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..a7f83c99a77844bb3dec05a87131d343f188dd29 --- /dev/null +++ b/phase2/embodied-olympics-match/tabletennis_match_i04/task_dag.json @@ -0,0 +1,226 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "embodied-olympics-match/tabletennis_match_i04", + "suite": "embodied-olympics-match", + "base": "embodied-olympics-match-tabletennis-match-i04", + "title_zh": "战胜四级乒乓球对手", + "summary_zh": "A 方在与四级脚本对手的有效乒乓球比赛结束时被裁判判为整场胜方。", + "embodiment_zh": "标准实时模式：一条七关节 Franka Panda 机械臂，末端为法兰固定安装的球拍，无灵巧手或二指夹爪；机械臂安装在端线后的横纵双轴导轨底座上，机器人抛球装置完成发球抛球。观察来自机器人后方与侧方两台相机。", + "task_class": "atomic", + "analysis_level": "outcome_only", + "spec": { + "version": "1.2", + "task": "embodied-olympics-match-tabletennis-match-i04", + "bench": "embodied-olympics-match", + "progress": "live", + "entities": {}, + "milestones": [ + { + "id": "match_won", + "zh": "赢得比赛", + "level": "semantic", + "mode": "state", + "when": null, + "after": [], + "hold_s": 0, + "freeze_held": false, + "goal": true, + "criterion_zh": "比赛已经结束，裁判最终胜方为 A（我方）：通常我方至少 5 分且领先至少 2 分，或先到 7 分；若含重发共进行 40 分则按比分领先方、平分时最近一个得分的胜方裁决。视频或比赛记录须能确认最后一分后的终局比分与裁判胜方；仅击中球、赢得一分、暂时领先或尚在等待回顾均未完成，A 方被判弃权或取消资格也未完成。", + "step": { + "cat": "对抗", + "item": "获胜" + }, + "contact": [ + { + "who": "手持工具", + "how": "其他", + "plan": 1, + "note_zh": "候选方式：击打。Franka Panda 机械臂法兰固定安装的球拍拍面击打橙色乒乓球；现有接触方式没有击打，按 S16 标其他。球拍是安装工具，无手指抓握球拍的接触或拾取步骤。" + }, + { + "who": "底盘", + "how": "支撑", + "plan": 1, + "note_zh": "横纵双轴导轨底座承载 Franka Panda 机械臂，在本方端线后移动以把安装球拍送至击球位置；与挥拍击球属于同一典型方案。" + } + ], + "optional": false, + "verify_zh": null, + "sources": [ + "inputs/task/standard-realtime/instruction.md:3-4", + "inputs/task/standard-realtime/instruction.md:45-47", + "inputs/task/standard-realtime/instruction.md:59", + "inputs/grader/efe_embodied-olympics-match-files/sim/sport/sim/match.py:78-85", + "inputs/grader/efe_embodied-olympics-match-files/sim/sport/sim/match.py:157-163", + "inputs/grader/efe_embodied-olympics-match-files/sim/sport/sim/match.py:210-223", + "inputs/grader/efe_embodied-olympics-match-files/sim/sport/sim/match.py:736-759", + "inputs/grader/efe_embodied-olympics-match-files/sim/variants/98e08661a3/sport/plugin.py:194-212", + "inputs/grader/efe_embodied-olympics-match-files/privileged/environment/fullsim/core/outcome.py:26-53", + "inputs/grader/efe_embodied-olympics-match-files/shared/tests/grade.py:96-138", + "inputs/grader/efe_embodied-olympics-match-files/privileged/environment/fullsim/sports/tabletennis/sim/scene/tt_scene.xml:6-14" + ], + "conf": 7, + "note_zh": "按竞技比赛规则使用 outcome_only，仅分析整场获胜，不将发球、挥拍、回球、逐分回顾或到达某一比分拆成步骤。对局从 0:0、无胜方开始；两分轮换发球，4:4 起每分轮换；接发方第 13 次合法回球直接赢得该分，发球连续第 3 次合法擦网重发改判失分。普通回球或发球失误仅计入比分，可在之后追回，不是整场历史否决；触桌合法，触网判失分。球拍和轨道接触描述来自场景与典型击球方案，不限定成功轨迹每分都移动底座。" + } + ], + "constraints": [ + { + "id": "start_forfeit", + "zh": "未按时开赛", + "never": null, + "criterion_zh": "比赛服务启动后 1800 秒真实时间内，A 方仍未发送开始请求，裁判以 no_show 结束并判 A 方负；会话在开始请求前结束也按 no_show 判负。", + "consequence": "立即终止", + "sources": [ + "inputs/task/standard-realtime/instruction.md:11-12", + "inputs/task/standard-realtime/environment/docker-compose.yaml:47", + "inputs/grader/efe_embodied-olympics-match-files/privileged/environment/fullsim/core/service.py:300-308", + "inputs/grader/efe_embodied-olympics-match-files/privileged/environment/fullsim/core/service.py:323-362" + ] + }, + { + "id": "review_forfeit", + "zh": "回顾超时", + "never": null, + "criterion_zh": "非决胜分结束后世界冻结等待回顾，A 方在冻结后的 300 秒真实时间内未发送 ready；服务以 ready_timeout 立即结束比赛，即使 A 方比分领先也判负。", + "consequence": "立即终止", + "sources": [ + "inputs/task/standard-realtime/instruction.md:9-10", + "inputs/task/standard-realtime/environment/docker-compose.yaml:48", + "inputs/grader/efe_embodied-olympics-match-files/privileged/environment/fullsim/core/realtime.py:344-361", + "inputs/grader/efe_embodied-olympics-match-files/sim/variants/98e08661a3/sport/plugin.py:194-203" + ] + }, + { + "id": "serve_clock_forfeit", + "zh": "发球延误", + "never": null, + "criterion_zh": "A 方发球窗口开放后 5 秒比赛时间内仍未抛球，裁判记录 serve_fault:undue_delay，立即结束该分并给对手加一分；A 方累计第 3 次发球延误时，裁判直接取消其资格并结束整场比赛，不受当前领先比分影响。前两次的终止范围仅为该分，并不否决后来赢得整场比赛。", + "consequence": "立即终止", + "sources": [ + "inputs/task/standard-realtime/instruction.md:47", + "inputs/task/standard-realtime/instruction.md:59", + "inputs/grader/efe_embodied-olympics-match-files/privileged/environment/fullsim/sports/tabletennis/sim/serve.py:55-56", + "inputs/grader/efe_embodied-olympics-match-files/sim/sport/sim/match.py:251-254", + "inputs/grader/efe_embodied-olympics-match-files/sim/sport/sim/match.py:487-499", + "inputs/grader/efe_embodied-olympics-match-files/sim/sport/sim/match.py:751-759" + ] + }, + { + "id": "match_abandoned", + "zh": "提前结束会话", + "never": null, + "criterion_zh": "A 方已开始比赛，但会话结束并收集结果时比赛仍未结束，服务以 abandoned 判 A 方整场负；开赛前结束属于未按时开赛约束。7200 秒会话预算本身不另作约束。", + "consequence": "立即终止", + "sources": [ + "inputs/task/standard-realtime/instruction.md:12", + "inputs/grader/efe_embodied-olympics-match-files/privileged/environment/fullsim/core/service.py:371-381", + "inputs/grader/efe_embodied-olympics-match-files/privileged/environment/fullsim/core/realtime.py:459-468" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "match_won", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-match-files/sim/variants/98e08661a3/sport/plugin.py:194-212; inputs/grader/efe_embodied-olympics-match-files/privileged/environment/fullsim/core/outcome.py:26-53; log live service phase, reason, forfeiting and referee winner per frame; no external grader verdict binding" + }, + { + "atom": "start_forfeit", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-match-files/privileged/environment/fullsim/core/service.py:323-362; log prepare phase, joined, start deadline and no_show event" + }, + { + "atom": "review_forfeit", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-match-files/privileged/environment/fullsim/core/realtime.py:344-361; log review phase, ready deadline and ready_timeout event" + }, + { + "atom": "serve_clock_forfeit", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-match-files/sim/sport/sim/match.py:487-499; log serve open/deadline, toss state, delay_faults[0], disqualified and point-end event" + }, + { + "atom": "match_abandoned", + "level": "new_code", + "source": "inputs/grader/efe_embodied-olympics-match-files/privileged/environment/fullsim/core/service.py:371-381; log in_play at collect and abandoned event" + } + ] + }, + "native_metrics": [ + { + "name": "point_share", + "unit": "比例", + "threshold": null, + "direction": "higher_better" + } + ], + "open_questions_zh": [ + "没有成功示范：inputs/oracle_none.txt 说明本任务 8 条目录运行均失败；只能依据任务文本、场景、裁判及相机参考程序制定 DAG，未从成功视频核对终局视觉提示（暂行：DECISIONS.md N2、N8）。", + "inputs/states/index.json 的 logs 为空；全部节点 when、全部约束 never 与 success 均为 null，不能安全绑定任何逐帧字段，也不能进行回放一致性核验。atoms 列出需补录的原生比赛状态，不把未记录的最终评分当作世界状态条件（暂行：DECISIONS.md P8）。", + "评分器成功条件是有效记录且实际裁判胜方为 A，并排除首次动作后世界步错误导致的 agent_sim_error；有效性要求 finalized、live 或一致回放、complete、无服务或计量及代理错误。当前实例按 live 评分，不要求重放一致性；这些记录有效性条件属于 success 的验收范围，不增设语义节点，现因缺日志保留 success=null（inputs/grader/efe_embodied-olympics-match-files/shared/tests/grade.py:100-138）。", + "发球时钟的前两次延误只终止该分、第三次才终止整场。按 DECISIONS.md N7，判分负的裁判时钟也须列为约束；serve_clock_forfeit 的 consequence=立即终止 在前两次指该分，并非整场，当前后果枚举没有单独的判分负值，不能把前两次延误当成整场历史否决。", + "point_share 是独立记录的连续得分占比，无固定成功阈值，故 threshold=null；有得分时为我方得分除双方总分，零得分时为 0.5，服务层弃权时为 0。它不单独产生另一个得分节点，也不能代替裁判胜方（inputs/grader/efe_embodied-olympics-match-files/sim/sport/sport.toml:203-212；inputs/grader/efe_embodied-olympics-match-files/privileged/environment/fullsim/core/score.py:70-89）。", + "README 的 partial 参考解并非本任务已成功的目录运行：文本报告相机参考在四级对手上 0/40，真值参考在本种子曾以 5:3 获胜；未把真值策略或这个比分当作必须的路线或验收阈值。" + ], + "conf": 7, + "conf_reason_zh": "单一比赛结果、规则阈值、初始比分和四类处罚均有文本与实际装载源码支持，未引入轨迹顺序或固定比分；但无成功示范和状态日志，无法核验绑定与画面提示，且发球时钟约束的分级后果超出当前枚举的细分能力。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "outcome_only 单节点 match_won（对抗·获胜）、四个裁判时钟约束与 N7/S16 一致；when/never/success 为 null 因无状态日志，已在 open_questions_zh 说明。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：['match_won']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:49:45Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/embodied-olympics-match__tabletennis_match_i04/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/embodied-olympics-match__tabletennis_match_i04/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "430a238f22af1cecc3124f99488f599e6a76307ffb0835ed1e127d2a5bcab0de", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/embodied-olympics-match/tabletennis_match_i04/trajectories/embodied-olympics-match-tabletennis-match-i04-standard-realtime--ah6jnwf.json b/phase2/embodied-olympics-match/tabletennis_match_i04/trajectories/embodied-olympics-match-tabletennis-match-i04-standard-realtime--ah6jnwf.json new file mode 100644 index 0000000000000000000000000000000000000000..34d0f454464aee8ba7d028bf68e8ff7fdc2c06fe --- /dev/null +++ b/phase2/embodied-olympics-match/tabletennis_match_i04/trajectories/embodied-olympics-match-tabletennis-match-i04-standard-realtime--ah6jnwf.json @@ -0,0 +1,266 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-match-tabletennis-match-i04-standard-realtime--ah6jnwf", + "taskKey": "embodied-olympics-match/tabletennis_match_i04", + "dag_sha256": "a2b9c83262187d1f5d8ea2dae7c8fd0eadbfa1ed17c7e40e13e93d451ca941b7", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "valid_trial": 1, + "technical_failure": 0, + "grader_error": 0, + "missing_log": 0, + "deterministic": 1, + "match_over": 1, + "seed": 34, + "official_config": 0, + "video_rendered": 1, + "direct_play": 0, + "lowlevel_actions": 0, + "solo": 0, + "no_show": 0, + "abandoned": 0, + "hidden_seed": 0, + "hidden_seed_index": -1, + "A_continuous_score": 0.0, + "continuous_score": 0.0, + "A_won": 0, + "A_no_show": 0, + "A_transient_api_errors": 0, + "A_unrecovered_api_errors": 0, + "won": 0, + "sim_error": 0, + "replay_error": 0, + "meter_error": 0, + "proxy_error": 0, + "timing_invalid": 0, + "points_for": 0, + "points_against": 5, + "replay_points_for": 0, + "replay_points_against": 5, + "replay_joint_targets_equal": 1, + "finalized": 1, + "revision": 8, + "timing_lag_enforced": 0, + "rejected_acts": 0, + "nocode_tool_calls": 0, + "nocode_violations": 0, + "nocode_forfeit": 0, + "gate_blocks": 0, + "metered_requests": 0, + "metered_output_tokens": 0, + "metered_input_tokens": 0, + "rate_limit_retries": 0, + "transient_api_errors": 0, + "unrecovered_api_errors": 0, + "level": 4.0, + "slowdown": 10.0, + "our_shots": 2.0, + "balls_to_you": 2.0, + "good_shots": 0.0, + "n_decisions": 18.0, + "refused_not_model": 0.0, + "dropped_acts": 0.0, + "n_actions": 1372.0, + "sim_steps": 13720.0, + "steps": 1372, + "n_reviews": 4.0, + "n_calls": 501.0, + "n_observe": 481.0, + "n_act": 20.0, + "elapsed_s": 860.0, + "max_lag_s": 0.0, + "sim_lag_late_max_s": 2.0374201519880444, + "sim_lag_late_p99_s": 1.4392013920843596, + "sim_effective_slowdown": 10.048259698202747, + "sim_run_wall_s": 137.86212305934168, + "ladder_points": 5, + "ladder_balls": 2, + "ladder_points_touched": 2, + "ladder_points_active": 2, + "ladder_points_over_net": 0, + "ladder_points_landed_in": 0, + "ladder_points_returned": 0, + "ladder_points_won": 0, + "ladder_points_won_expedite": 0, + "ladder_points_won_rally": 0, + "ladder_balls_touched": 2, + "ladder_balls_active": 2, + "ladder_balls_over_net": 0, + "ladder_balls_landed_in": 0, + "ladder_balls_returned": 0, + "serves": 3, + "serves_good": 0, + "serve_faults": 3, + "serve_points_won": 0, + "delay_faults": 0, + "disqualified": 0, + "infra_error": 0, + "realtime": 1, + "opponent_level": 4, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.196, + "peak_processes": 8, + "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/embodied-olympics-match-tabletennis-match-i04-standard-realtime--ah6jnwf/replay.mp4", + "duration_s": 13.9, + "fps": 10.0, + "frames": 139, + "sha256": "e5d76f4b45b974f1e33b41d561b6b54115aaf7bcbf667f1916f7b473ee3257eb" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "match_won": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": 19, + "text_zh": "观察：视频 0.0 s 为本方机械臂在球台左端的开局画面；log step 19 的比赛服务 start 回复为 score=0:0、phase=running。本次日志记录的是这一场实时比赛，逐分冻结回顾后继续，无重置另开比赛。" + }, + { + "t": 13.7, + "log_step": 43, + "text_zh": "观察：视频 12.1–12.8 s 可见本方抛球上升后下落，视频约 12.9 s 球从球拍附近飞向台面，视频 13.2 s 左半台反弹，视频 13.7–13.8 s 停在中网附近。log step 43 的比赛服务报告 phase=over、reason=match_over、score=0:5，末分由对手获胜，原因为 serve_fault:short（接发方落点距网 0.15 m，要求至少 0.20 m）；评分器 points_for=0、points_against=5、A_won=0，故赢得比赛未成立。" + }, + { + "t": 12.9, + "log_step": 43, + "text_zh": "推断：本场服务控制步与比赛仿真时钟的关系为 step×0.01 s（log step 43：step=1372、t_sim=13.72）；该日志的末次发球 took_effect=1197，击拍事件相对其后 0.866 s，即约视频 12.84 s，与视频 12.8–12.9 s 球由拍边向台面飞出吻合。视频末尾仅多出约 0.2 s，以上仅用于核对服务记录属于本段视频，未将 agent 工作时间换算为视频时间。" + } + ], + "contact_actual": [ + { + "who": "手持工具", + "how": "其他", + "note_zh": "候选方式：击打；现有方式枚举无击打。视频 7.7–8.0 s、10.2–10.5 s 的接发尝试以及视频约 12.9 s 的末次发球中，Franka Panda 法兰固定球拍作用于乒乓球；末次击拍由 log step 43 服务 heard.paddle 佐证，不虚构抓握球拍。前两次抛球另碰到机器人本体，由 log step 22、29 裁判记录确认，画面不足以分辨具体连杆。" + }, + { + "who": "底盘", + "how": "支撑", + "note_zh": "观察：视频 6.6–7.5 s 轨道底座承载机械臂沿导轨移动靠近击球位置，视频 8.1–8.6 s 后退复位；比赛中实际发生底座移动。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "point_share", + "value": 0.0, + "unit": "比例", + "threshold": null, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "match_won", + "t": 13.7, + "log_step": 43, + "where": "操作中", + "evidence_zh": "观察：视频 13.2–13.7 s 本方末次发球先在本方台面反弹后过网，并在对方靠网处落台；log step 43 的比赛服务将该球判为 serve_fault:short，比分从 0:4 变为 0:5，phase=over、done=true、reason=match_over。评分器亦记 A_won=0、match_over=1；按 outcome_only，断点是整场获胜节点，视频 13.7 s 取决胜分落台附近为锚点。" + }, + "unmet": { + "match_won": false + }, + "phenomena": [ + { + "cat": "运动", + "item": "被挡/碰撞", + "t": 1.5, + "log_step": 29, + "evidence_zh": "观察：视频 1.4–1.5 s 与 5.0–5.1 s 抛出的球均紧邻本方机械臂末端后消失，没有形成发球过网。log step 22、29 的比赛服务裁判逐分记录前两分均为 serve_fault:toss_touched: the toss touched robot server before it was struck，比分依次 0:1、0:2；这是抛球先碰本体而失分的重复模式。具体碰撞部位不能仅从该远景辨认，腕部说法仅为 agent 在 log step 32 的判断。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "抓取和接触设计", + "alt_feasible": "已验证有", + "t": 5.0, + "log_step": 32, + "source_zh": "日志", + "evidence_zh": "推断：发球抛球通道与机器人末端占据的空间不适配。log step 22、28 的实际服务回复显示两次均在 [0.25,-1.55,0.15] 抛球，虽 log step 27 将高度从 0.35 m 增至 0.8 m 并延后挥拍，两分仍由裁判记录 toss_touched（log step 29）。log step 32 才把抛球横向位置改为 x=-0.3、底座 x=-0.1；log step 43 的末次发球在该位置得到 paddle、本方落台、对方落台三项服务事件，视频 12.1–13.7 s 也显示球完成抛起、击拍和两侧落台。这验证了可避开本体碰撞的替代接触布局；该替代发球仍因过短失分，不代表验证了获胜策略。" + } + ], + "gap_zh": "无状态日志，远景视频且仅 10 fps，无法独立辨认前两次抛球碰到哪个连杆及瞬时接触面，碰撞事实由同场实时服务的裁判记录补足。log step 33 将第三分记录为 volley，agent 在 log step 39、47 自述为两次落台后才击球，两者存在表述差异；不据此另加动态时机现象。第四分回球落本方、第五分发球过短各出现一次，按 outcome_only 的重复失分模式规则不另列现象；所列根因只解释前两分的重复碰撞。" + }, + "summary_zh": "本方机械臂移动底座并尝试三次发球、两次接发，最终以 0:5 输给四级对手，赢得比赛节点始终未达成，终态进度 0/1；得分占比为 0。断点取视频约 13.7 s 决胜分，裁判判末次发球过短；整场可确认的重复失分模式是前两次抛球碰本体，抛球位置与机械臂末端空间冲突，换侧后已避免该碰撞。无状态日志，远景画面不足以确定具体碰撞连杆，依同场实时裁判记录确认接触事实。", + "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 outcome_only 标注：match_won=false、终态 0:5、断点 match_won、反复抛球碰本体的现象与根因均有日志（裁判记录 toss_touched）与视频支持；时钟说明（step×0.01 s）有核对，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": "2026-10-29T10:07:41Z", + "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/embodied-olympics-match-tabletennis-match-i04-standard-realtime--ah6jnwf/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-match-tabletennis-match-i04-standard-realtime--ah6jnwf/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "f2e797741932d9f6011081fbf0461d923d0c4e0c967bbac4a3882185919fbcbb", + "video_sha256": "e5d76f4b45b974f1e33b41d561b6b54115aaf7bcbf667f1916f7b473ee3257eb", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/embodied-olympics-match/tabletennis_match_i04/trajectories/embodied-olympics-match-tabletennis-match-i04-standard-realtime--anp2xgt.json b/phase2/embodied-olympics-match/tabletennis_match_i04/trajectories/embodied-olympics-match-tabletennis-match-i04-standard-realtime--anp2xgt.json new file mode 100644 index 0000000000000000000000000000000000000000..5cf20ab8454f9ac46f273574f39b26a41d1d7b67 --- /dev/null +++ b/phase2/embodied-olympics-match/tabletennis_match_i04/trajectories/embodied-olympics-match-tabletennis-match-i04-standard-realtime--anp2xgt.json @@ -0,0 +1,274 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "embodied-olympics-match-tabletennis-match-i04-standard-realtime--anp2xgt", + "taskKey": "embodied-olympics-match/tabletennis_match_i04", + "dag_sha256": "a2b9c83262187d1f5d8ea2dae7c8fd0eadbfa1ed17c7e40e13e93d451ca941b7", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "valid_trial": 1, + "technical_failure": 0, + "grader_error": 0, + "missing_log": 0, + "deterministic": 1, + "match_over": 1, + "seed": 34, + "official_config": 0, + "video_rendered": 1, + "direct_play": 0, + "lowlevel_actions": 0, + "solo": 0, + "no_show": 0, + "abandoned": 0, + "hidden_seed": 0, + "hidden_seed_index": -1, + "A_continuous_score": 0.0, + "continuous_score": 0.0, + "A_won": 0, + "A_no_show": 0, + "A_transient_api_errors": 0, + "A_unrecovered_api_errors": 0, + "won": 0, + "sim_error": 0, + "replay_error": 0, + "meter_error": 0, + "proxy_error": 0, + "timing_invalid": 0, + "points_for": 0, + "points_against": 5, + "replay_points_for": 0, + "replay_points_against": 5, + "replay_joint_targets_equal": 1, + "finalized": 1, + "revision": 8, + "timing_lag_enforced": 0, + "rejected_acts": 0, + "nocode_tool_calls": 0, + "nocode_violations": 0, + "nocode_forfeit": 0, + "gate_blocks": 0, + "metered_requests": 0, + "metered_output_tokens": 0, + "metered_input_tokens": 0, + "rate_limit_retries": 0, + "transient_api_errors": 0, + "unrecovered_api_errors": 0, + "level": 4.0, + "slowdown": 10.0, + "our_shots": 1.0, + "balls_to_you": 2.0, + "good_shots": 0.0, + "n_decisions": 7.0, + "refused_not_model": 6.0, + "dropped_acts": 0.0, + "n_actions": 1453.0, + "sim_steps": 14530.0, + "steps": 1453, + "n_reviews": 4.0, + "n_calls": 276.0, + "n_observe": 269.0, + "n_act": 7.0, + "elapsed_s": 1631.8, + "max_lag_s": 0.0, + "sim_lag_late_max_s": 1.741841624956578, + "sim_lag_late_p99_s": 0.9204254467319699, + "sim_effective_slowdown": 10.002221549968173, + "sim_run_wall_s": 145.33227912103757, + "ladder_points": 5, + "ladder_balls": 2, + "ladder_points_touched": 1, + "ladder_points_active": 0, + "ladder_points_over_net": 0, + "ladder_points_landed_in": 0, + "ladder_points_returned": 0, + "ladder_points_won": 0, + "ladder_points_won_expedite": 0, + "ladder_points_won_rally": 0, + "ladder_balls_touched": 1, + "ladder_balls_active": 0, + "ladder_balls_over_net": 0, + "ladder_balls_landed_in": 0, + "ladder_balls_returned": 0, + "serves": 2, + "serves_good": 0, + "serve_faults": 2, + "serve_points_won": 0, + "delay_faults": 1, + "disqualified": 0, + "infra_error": 0, + "realtime": 1, + "opponent_level": 4, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0, + "peak_memory_gb": 0.383, + "peak_processes": 15, + "peak_disk_gb": 1.0, + "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/embodied-olympics-match-tabletennis-match-i04-standard-realtime--anp2xgt/replay.mp4", + "duration_s": 14.7, + "fps": 10.0, + "frames": 147, + "sha256": "f19aa36737d7b45c54540eefcabd7e0c93eaf1a948e798ebc9a1b7592cb7bdd1" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "match_won": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0, + "log_step": 105, + "text_zh": "观察：视频 0.0 s 机械臂在本方端线外持拍待命；log step 105 比赛服务接受开赛，比分为我方 0、对手 0，尚无胜方。" + }, + { + "t": 13.1, + "log_step": 148, + "text_zh": "观察：视频 10.4–11.5 s 机械臂移动导轨底座并伸拍，来球从拍旁经过；视频 12.8–13.2 s 再次挥拍。log step 148 的比赛服务记录两次接球分别为 missed: no paddle contact 与 fault: hit before the ball bounced on your half (volley)，后一次有出球速度 1.82。" + }, + { + "t": 14.5, + "log_step": 142, + "text_zh": "观察：视频 14.2–14.6 s 机械臂做最后一次发球动作后画面结束，没有获胜提示。log step 142、144 的比赛服务终局为 phase=over、done=true、reason=match_over、比分 0:5；最后一分裁判记为 serve_fault:not_own_court（发球未先触本方台面）。评分器字段 match_over=1、A_won=0、points_for=0、points_against=5，确认整场未获胜。" + } + ], + "contact_actual": [ + { + "who": "手持工具", + "how": "其他", + "note_zh": "观察：视频 12.8–13.2 s Franka Panda 机械臂用法兰固定球拍迎击来球，log step 148 的服务 shot 记录确认一次有速度的球拍触球，但判为截击。候选方式：击打；接触方式枚举没有击打，因此标其他。未把未触球的挥拍或球拍固定安装虚构为抓握。" + }, + { + "who": "底盘", + "how": "支撑", + "note_zh": "观察：视频 10.4–10.8 s 导轨底座承载机械臂横移，配合球拍伸向来球；视频 12.8–13.2 s 再次调整底座位置并挥拍。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "point_share", + "value": 0.0, + "unit": "比例", + "threshold": null, + "direction": "higher_better" + } + ], + "failure": { + "breakpoint": { + "ref": "match_won", + "t": 14.5, + "log_step": 142, + "where": "操作中", + "evidence_zh": "观察：视频 14.2–14.6 s 为最后一次发球动作；log step 142 的比赛服务将该分以 serve_fault:not_own_court 判给对手，比分从 0:4 变为 0:5 并报告 MATCH DONE。该任务为 outcome_only，断点取整场结果节点 match_won，锚定决胜分。推断：log step 8 的 control_dt=0.01，服务 control step 1453 对应视频约 14.53 s；已用视频 14.2 s 附近的抛球及最后动作核对 log step 142 的抛球 control step 1412、projected_at_step 1422，未使用 agent 工作时间换算视频时间。" + }, + "phenomena": [ + { + "cat": "接触", + "item": "没建立接触", + "t": 10.7, + "log_step": 148, + "evidence_zh": "推断：未能用球拍接触球在首分与第三分重复出现。首分 log step 111 比赛服务接受抛球，log step 113、115 的 agent 发球脚本报错且未发出击球动作；视频 1.4–2.5 s 机械臂维持待机姿态，没有随抛球挥拍，log step 117 随后记录比分 0:1。第三分视频 10.4–10.8 s 机械臂虽伸拍，橙色来球从拍旁继续向本方端线外运动，log step 148 的服务记录明确为 missed: no paddle contact。第四分后来发生的截击接触不改变前述两分已失分的事实。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志", + "t": 2.3, + "log_step": 130, + "note_zh": "候选细分：控制脚本异常（接口方法误用）。按 D38 保留在控制；不是外部网络故障。", + "evidence_zh": "观察：log step 66 的 agent HTTP 封装对所有接口使用 requests.post，status() 也走此封装；log step 93 的主循环遇 error 即继续等待，因此发球控制未运行。log step 113、115 手动发球脚本分别出现 KeyError 与 TypeError；log step 127 继续反复得到 HTTP 404 unknown endpoint。log step 130 对同一状态接口测试显示 POST 为 404、GET 为 200；log step 132 改用 GET 后成功读到比分 0:2，随后 log step 136 才恢复控制。推断：接口误用使前两分期间控制停滞，是首分未挥拍及第二分未抛球的直接控制原因，与视频 0.0–8.1 s 基本静止一致。" + }, + { + "cat": "控制", + "item": "轨迹或逆解", + "source_zh": "日志", + "t": 10.5, + "log_step": 140, + "evidence_zh": "观察：log step 140 的比赛服务对第三分回球指令报告 command not achievable accurately、IK residual 0.131、关节限速导致速度缩放 x0.05；第四分两次指令也报告 IK residual 0.047/0.049、底座无法及时到位及短挥拍。视频 10.4–10.8 s 可见机械臂伸拍仍漏球，视频 12.8–13.2 s 再次迎球，服务在 log step 148 判为截击。log step 142 最后一次发球指令仍报告 normal_error_deg=25.348 和关节限速。推断：agent 生成的击球轨迹不能按要求的位置、朝向与时间准确执行，接口修复后仍没有合法回球；现有证据不进一步把这些偏差归因到某个单独参数。" + } + ], + "gap_zh": "视频为单一远景、10 fps，没有比分叠字，也没有状态日志或成功示范；终局比分和截击判罚以该单局实时比赛服务及评分器记录补足。首分球在臂旁过小，未逐帧分清其全部飞行和碰撞；未触拍由未发出击球动作、静止机械臂及随后失分共同推断。末分发球失误和第四分截击各出现一次，未单独列为重复失分模式；无法仅凭画面分离击球几何、轨迹跟踪和感知误差各自的贡献。比赛结束后的离线修正与测试没有进入该评分回合。", + "unmet": { + "match_won": false + } + }, + "summary_zh": "我方完成开赛并尝试发球、两次接发球，但比赛服务与评分器确认最终以 0:5 落败，“赢得比赛”始终未达成，节点终态为 0/1。断点在视频约 14.5 s 的决胜分；首分和第三分重复未建立球拍接触，日志证实前段控制脚本误用状态接口，修复后击球指令仍多次报告逆解或执行偏差。视频球体较小且无比分叠字，终局与具体判罚依赖该局服务记录，赛后离线修正不计入进度。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "match_won 未达成、断点、两个根因与日志和视频一致（0:5，首分未发球击球、第三分漏球、第四分截击、末分发球失误）；时钟换算有明确的 control step 核对，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": "2026-10-29T10:08:11Z", + "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/embodied-olympics-match-tabletennis-match-i04-standard-realtime--anp2xgt/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/embodied-olympics-match-tabletennis-match-i04-standard-realtime--anp2xgt/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "e9240de875b6dc24e008d7f84c94d23e075221deb337e5a45c14a3c38f6f3051", + "video_sha256": "f19aa36737d7b45c54540eefcabd7e0c93eaf1a948e798ebc9a1b7592cb7bdd1", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/roboarc/blanket_flatten_04/task_dag.json b/phase2/roboarc/blanket_flatten_04/task_dag.json new file mode 100644 index 0000000000000000000000000000000000000000..24b117f02511180197c48d47eaf3048b874e2fd4 --- /dev/null +++ b/phase2/roboarc/blanket_flatten_04/task_dag.json @@ -0,0 +1,257 @@ +{ + "schema": "efe-task-dag/1", + "taxonomy": "v2", + "taskKey": "roboarc/blanket_flatten_04", + "suite": "roboarc", + "base": "roboarc-blanket-flatten-04-i01", + "title_zh": "桌面毯子展平", + "summary_zh": "毯子充分展开并平铺在桌上、无过度拉伸，机器人松开并撤离后，覆盖、贴桌、净空与稳定性条件同时通过持续确认。", + "embodiment_zh": "两种模式均为固定基座上的两条六自由度 ALOHA（ViperX）机械臂，左、右臂各配一个平行二指夹爪，可辅助抓取局部布片，无移动底盘；特权模式可读物体状态和重置，标准模式仅有相机与本体感知、单次实时回合。", + "task_class": "atomic", + "analysis_level": "step", + "spec": { + "version": "1.2", + "task": "roboarc-blanket-flatten-04-i01", + "bench": "roboarc", + "progress": "live", + "entities": {}, + "milestones": [ + { + "id": "blanket_flattened", + "zh": "毯子展平", + "mode": "state", + "when": null, + "after": [], + "hold_s": 0.5, + "freeze_held": false, + "goal": true, + "criterion_zh": "高位相机可见毯面展开成接近完整的单层片材，折回的角、卷边和重叠明显消除，低位相机可见毯子落在桌上而非悬空绷紧：严格验收要求投影并集面积至少为原材料面积的95%（0.2736平方米，重叠只计一次、孔洞不计），所有粒子中心相对桌面高度在负8至30毫米内，每条材料边长度不超过原长的1.15倍；两爪松开且辅助抓取解除，所有机器人碰撞盒距向外扩张8毫米的整张毯子包围盒至少5毫米，每个目标点速度不超过30毫米每秒、相对确认起点位移不超过3毫米，以上同时连续保持0.5秒才完成。仍有大片折叠、角悬起、过度拉伸、夹爪夹住或围住毯边、持续滑动时均未完成，已完成后再失去这些条件则失去该状态；特权模式回放的末尾补充保持段只把覆盖率与速度放宽为92%与80毫米每秒，其他条件不变，且两次回放都须成功、至少一次已在提交动作内按严格阈值完成确认。", + "step": { + "cat": "材料", + "item": "展平" + }, + "contact": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "左、右二指夹爪分别捏住毯子上互不重叠的布片或边角，一侧固定、另一侧展开，或同时展开后放回桌面；局部改抓依次发生，不另列同时接触。", + "plan": 1 + }, + { + "who": "手指/夹爪", + "how": "抓握", + "note_zh": "单侧二指夹爪逐次捏住毯子的松散边或角，利用其余毯面在桌上的支承依次翻开、铺展；另一只夹爪无需同时接触毯子。", + "plan": 2 + } + ], + "optional": false, + "verify_zh": "待核实：输入未包含冻结实例的粒子初态或初始图像；成功记录的步骤10仅描述毯子最初约对折。", + "sources": [ + "task.json:1 description", + "task/standard/instruction.md:3-13 任务、严格目标与连续确认", + "task/standard/instruction.md:24-27 原面积、包围盒净空与实时评分", + "task/privileged/instruction.md:38-87 公开的 evaluate_state 摘录：投影并集、粒子高度、边长比、速度", + "task/privileged/README.md:58-75 全部验收阈值与特权回放容差", + "grader/efe_roboarc-files/rcb_topogym.py:2340-2365 目标与松手、速度、漂移验收合取", + "grader/efe_roboarc-files/rcb_topogym.py:2974-2989 末尾补充保持段与容差", + "task/standard/tests/grade.py:55-88 实时成功及复核门控", + "task/privileged/tests/grade.py:123-134 两次成功且至少一次严格成功", + "oracle/run.json:1 steps[id=10,245,288,296,313,315] 初始折叠、单爪与双爪展开、最终服务确认" + ], + "conf": 6, + "note_zh": "展开、铺平是同一张毯子的一个材料状态目标；覆盖率、贴桌高度、拉伸、松手、撤离与静置均是验收，不建抓取、抬起、角点、撤臂或静置节点，也不规定展开次序或必须双手。方案1为典型双臂方案，方案2为单臂依次处理的替代方案；未将成功记录中的每次尝试转成步骤。关闭持物冻结，使重新抓取或抬起造成的验收失效能直接反映。" + } + ], + "constraints": [ + { + "id": "cloth_below_table", + "zh": "毯子跌落", + "never": null, + "criterion_zh": "任一毯子粒子中心低于桌面超过100毫米，即最小相对高度小于负0.100米，立即失败；只越过桌沿但尚未达到该高度不等于此项已触发。", + "consequence": "立即终止", + "sources": [ + "task/standard/instruction.md:15-17", + "task/privileged/instruction.md:43-45" + ] + }, + { + "id": "forced_breakdown", + "zh": "强迫导致崩溃", + "never": null, + "criterion_zh": "布料粒子速度超过20米每秒或出现非有限速度而触发模拟崩溃，且该时刻或此前0.25秒内至少一个物理步有手臂强迫施力：某手臂驱动达到力矩上限的95%，同时该臂夹爪或所持物对外总压力至少20牛。此组合归为机器人失败并终止；单独强迫施力仅记录，未伴随近期强迫施力的模拟崩溃属于基础设施错误。", + "consequence": "立即终止", + "sources": [ + "task/standard/instruction.md:22-23", + "grader/efe_roboarc-files/rcb_topogym.py:1253-1255", + "grader/efe_roboarc-files/rcb_topogym.py:1589-1628", + "grader/efe_roboarc-files/rcb_topogym.py:2917-2943" + ] + } + ], + "success": null, + "atoms": [ + { + "atom": "blanket_coverage_ratio", + "level": "new_code", + "source": "需记录 evaluate_state 的 coverage_ratio（真实投影并集，不是外包矩形）；task/privileged/instruction.md:67-83。" + }, + { + "atom": "blanket_height_range_m", + "level": "new_code", + "source": "需记录所有 blanket 粒子中心相对 table_height_m 的最小和最大高度；task/privileged/instruction.md:43-58。" + }, + { + "atom": "blanket_maximum_stretch_ratio", + "level": "new_code", + "source": "需记录粒子材料边长与 rest_lengths 的最大比值；task/privileged/instruction.md:47-59。" + }, + { + "atom": "blanket_maximum_speed_m_s", + "level": "new_code", + "source": "需记录 blanket 粒子最大线速度；task/privileged/instruction.md:55-60，grader/efe_roboarc-files/rcb_topogym.py:2977-2982。" + }, + { + "atom": "blanket_released", + "level": "new_code", + "source": "需记录全部辅助抓取已解除且目标已脱离机器人；task/standard/instruction.md:11。" + }, + { + "atom": "blanket_release_clearance_m", + "level": "new_code", + "source": "需记录每个机器人碰撞盒到扩张8毫米的整体布料包围盒的最小净空；task/standard/instruction.md:25，final_settling 辅助函数源码未提供。" + }, + { + "atom": "blanket_confirmation_drift_m", + "level": "new_code", + "source": "需记录相对当前确认窗口起点的最大目标点位移；task/standard/instruction.md:12，grader/efe_roboarc-files/rcb_topogym.py:2349-2365。" + }, + { + "atom": "blanket_confirmation_elapsed_s", + "level": "new_code", + "source": "需记录所有最终条件同时为真期间的连续确认时长，任一条件失效就重置；task/standard/instruction.md:13。" + }, + { + "atom": "cloth_breakdown_with_recent_forcing", + "level": "new_code", + "source": "需记录粒子速度或非有限状态崩溃，以及此前0.25秒逐物理步的驱动力矩占比和对外压力；grader/efe_roboarc-files/rcb_topogym.py:1553-1567,1589-1628,2917-2943。" + } + ] + }, + "native_metrics": [ + { + "name": "coverage_ratio", + "unit": "ratio", + "threshold": 0.95, + "direction": "higher_better" + }, + { + "name": "maximum_height_m", + "unit": "m", + "threshold": 0.03, + "direction": "lower_better" + }, + { + "name": "minimum_height_m", + "unit": "m", + "threshold": -0.008, + "direction": "higher_better" + }, + { + "name": "maximum_stretch_ratio", + "unit": "ratio", + "threshold": 1.15, + "direction": "lower_better" + }, + { + "name": "maximum_speed_m_s", + "unit": "m/s", + "threshold": 0.03, + "direction": "lower_better" + }, + { + "name": "maximum_drift_m", + "unit": "m", + "threshold": 0.003, + "direction": "lower_better" + }, + { + "name": "release_clearance_m", + "unit": "m", + "threshold": 0.005, + "direction": "higher_better" + }, + { + "name": "confirmation_elapsed_s", + "unit": "s", + "threshold": 0.5, + "direction": "higher_better" + } + ], + "open_questions_zh": [ + "输入 states/index.json 的 logs 为空，无法核对任何状态字段；依暂行规则 DECISIONS.md P8，唯一节点的 when、约束的 never 和 success 均为 null，列出的原子均需增加日志提取，不能据成功记录伪造动态绑定。", + "成功记录存在但无可用视频或状态日志（oracle/oracle.json）；oracle/run.json 的步骤10称初始毯子约对折，步骤315的服务读数确认在10223个控制步、模拟255.575秒成功，reward.live_success=1、resumes=0。未直接看到相机画面，视觉判据据任务相机定义与记录中的展开、落桌、撤爪描述制定，未用代理覆盖率估计替代评分器指标（DECISIONS.md N2、N8）。", + "冻结实例、场景粒子初态和 topogym 任务类完整源码不在输入中（task/privileged/README.md:79-80）；可核对的任务实现仅有 instruction.md:38-87 的部分摘录，松手、漂移、连续确认等委托函数未完整提供。没有证据支持增添额外必经预备节点，初态也不能量化。", + "模式差异：标准模式按实时严格条件成功（task/standard/instruction.md:27）；特权模式要求两次回放均成功，至少一次在提交动作内严格成功，末尾最多2秒补充保持段才放宽覆盖率到92%、速度到80毫米每秒，3毫米漂移及0.5秒窗口不变（task/privileged/README.md:75）。共用一个展平节点，判据注明此容差；native_metrics 给严格阈值，单条轨迹的末帧不能表达跨两次回放的判定，success 保留 null。", + "强迫施力本身只记录、不评分，不设约束；仅与随后0.25秒内的模拟崩溃组合时才是失败（task/standard/instruction.md:22-23，DECISIONS.md S13）。桌面限位、静态物体限位、力矩上限是控制保护，未另列失败约束。", + "评分器要求，文本未提：两个模式的通用 grade.py 均在成功带 review 记录时暂缓计为成功（标准:84-88，特权:131-134）；rcb_topogym.py:1525-1543 只对具有 rigid_measurements 的任务启用深接触复核，当前缺少毯子任务类源码，无法确认此通用分支是否适用于该柔性毯子，未将刚体专属深度阈值推定为布料失败约束。" + ], + "conf": 6, + "conf_reason_zh": "单节点粒度、材料展平类型和物理验收阈值有明确文本与公开实现摘录支持，成功记录的服务判定与之相容；但没有状态日志、可用视频或冻结初态，委托验收及复核适用性不能完整检查，也无法验证跨模式的回放判定绑定。", + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 7, + "summary_zh": "任务文本只描述一个目标「展平毯子并松手撤离」，单节点 atomic、松手撤离折入判据、success 为 null（无状态日志，符合 P8）、约束为评分器的失败条件，均合规；阈值与评分器一致。未发现需要修改的内容。", + "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：['blanket_flattened']" + ] + }, + "provenance": { + "prompt": "prompts/01_task_dag.md", + "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25", + "engine": "codex", + "model": "gpt-6.1-sol", + "effort": "high", + "created": "2026-10-29T09:39:30Z", + "review_prompt": "prompts/02_task_dag_review.md", + "review_prompt_sha256": "305c2c5feb2340e0c426ebfe883251f74824cbde2c3e230baa9c86fc05c31550", + "reconcile_prompt": "prompts/05_reconcile.md", + "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2", + "calls": [ + { + "stage": "dag", + "round": 0, + "ok": true, + "space": "work/tasks/roboarc__blanket_flatten_04/01_dag" + }, + { + "stage": "dag_review", + "round": 1, + "ok": true, + "space": "work/tasks/roboarc__blanket_flatten_04/02_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs_manifest_sha256": "e3eaaa739f85bf3edfdb7f1410812a19f69818f00755e34b43d58c89d67def8c", + "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53" + } +} diff --git a/phase2/roboarc/blanket_flatten_04/trajectories/roboarc-blanket-flatten-04-i01-privileged--44za5cn.json b/phase2/roboarc/blanket_flatten_04/trajectories/roboarc-blanket-flatten-04-i01-privileged--44za5cn.json new file mode 100644 index 0000000000000000000000000000000000000000..b1eee754e8c73b3ab2148cc0e1296f08cf2ca716 --- /dev/null +++ b/phase2/roboarc/blanket_flatten_04/trajectories/roboarc-blanket-flatten-04-i01-privileged--44za5cn.json @@ -0,0 +1,237 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "roboarc-blanket-flatten-04-i01-privileged--44za5cn", + "taskKey": "roboarc/blanket_flatten_04", + "dag_sha256": "a99674e5c44990e94d9aa8a3601a1e41c800bc3a09b75ed544b76e83470ebada", + "outcome": { + "success": 1, + "reward": 1.0, + "source": "catalog", + "raw": { + "success": 1, + "success_pending_review": 0, + "review_events": 0, + "review_max_depth_mm": 0.0, + "forcing_events": 0, + "all_success": 1, + "deterministic": 1, + "gpu_backend": 1, + "infra_error": 0, + "simulator_instability": 0, + "instance_ok": 1, + "instance_exact": 1, + "initial_state_generated": 0, + "instance_realized": 0, + "assets_exact": 1, + "runtime_binding_ok": 1, + "n_actions": 1049, + "submitted_actions": 1089, + "sim_steps": 1049, + "physics_steps": 10490, + "sim_time_s": 26.22499999999958, + "final_reward": 1.0, + "replay_wall_s": 710.2, + "replays": 2, + "grader_error": 0, + "video_rendered": 1, + "replay_hold_steps": 0, + "replay_tolerance_needed": 0, + "strict_goal_reached": 1, + "strict_goal_row": 1049, + "strict_replay_required": 0, + "agent_sessions": 4, + "agent_resets": 25, + "agent_sim_steps": 6294, + "agent_records": 13, + "agent_replays": 6, + "agent_cost_self_reported": 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/roboarc-blanket-flatten-04-i01-privileged--44za5cn/replay.mp4", + "duration_s": 26.333, + "fps": 15.0, + "frames": 395, + "sha256": "09c786432c6c4928357ebc4df7481bd84fbcde0298e19b599058beec19343af5" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 1, + "final": 1, + "nodes": { + "blanket_flattened": { + "method": "visual", + "intervals": [ + [ + 26.2, + null + ] + ], + "final": true, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察：视频 0.0 s，蓝色方格毯呈窄条状，左端及近侧可见折回、重叠，尚未展成完整单层毯面。" + }, + { + "t": 6.0, + "log_step": null, + "text_zh": "观察：视频 2.0–4.5 s，机器人左、右二指夹爪分别合拢在毯子两侧的局部布片上；视频 5.0–9.0 s，两爪同时抬起各自夹住的布片，毯子离桌悬垂，原先折叠的毯面随之展开。" + }, + { + "t": 16.5, + "log_step": null, + "text_zh": "观察：视频 10.0–16.5 s，机器人逐渐降低两爪，将大部分毯面铺到桌上并张开夹爪、向上撤离；毯子靠机器人一侧仍有一条折回的边。视频 18.5–23.0 s，两爪再次下降夹取该侧毯边并向外移动，修整这条折边。" + }, + { + "t": 26.2, + "log_step": 128, + "text_zh": "观察：视频 23.5–25.6 s，机器人先后张开夹爪并抬离毯边；视频 25.667–26.267 s，毯子呈完整近矩形单层片材平铺于桌面，无明显折回、悬起或继续滑移，两爪已离开毯面。inputs/run.json 的 reward 报告 success=1、all_success=1、strict_goal_reached=1、strict_goal_row=1049、replay_hold_steps=0、replay_tolerance_needed=0；log step 128 的最终交付文件复验也报告两次严格成功且结果确定。" + }, + { + "t": 26.2, + "log_step": null, + "text_zh": "推断：末段视频形态与评分器严格成功结果一致，唯一目标在视频约 26.2 s 完成并保持至结束，时间精度约半秒；将静置确认完成计入达成时刻，不把先前仅外观接近平铺或仍被夹爪接触的阶段计为完成。没有状态日志，无法从图像独立量测全部粒子的高度、拉伸、速度、漂移及毫米级净空，这些联合验收由严格成功字段支持。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "观察：视频 2.0–9.0 s，左、右平行二指夹爪分别夹住毯子两侧的局部布片，共同抬起并展开；视频 10.0–16.5 s 共同降低铺回桌面，张爪撤离。视频 18.5–23.0 s，两爪再次夹住靠机器人一侧的不同毯边位置并外移修整，随后先后松开、撤离。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "coverage_ratio", + "unit": "ratio", + "threshold": 0.95, + "direction": "higher_better", + "value": null + }, + { + "name": "maximum_height_m", + "unit": "m", + "threshold": 0.03, + "direction": "lower_better", + "value": null + }, + { + "name": "minimum_height_m", + "unit": "m", + "threshold": -0.008, + "direction": "higher_better", + "value": null + }, + { + "name": "maximum_stretch_ratio", + "unit": "ratio", + "threshold": 1.15, + "direction": "lower_better", + "value": null + }, + { + "name": "maximum_speed_m_s", + "unit": "m/s", + "threshold": 0.03, + "direction": "lower_better", + "value": null + }, + { + "name": "maximum_drift_m", + "unit": "m", + "threshold": 0.003, + "direction": "lower_better", + "value": null + }, + { + "name": "release_clearance_m", + "unit": "m", + "threshold": 0.005, + "direction": "higher_better", + "value": null + }, + { + "name": "confirmation_elapsed_s", + "unit": "s", + "threshold": 0.5, + "direction": "higher_better", + "value": null + } + ], + "failure": null, + "summary_zh": "观察：机器人双爪分别夹住毯子两侧，抬起展开后逐渐落桌，再次夹取并外移修整折边，最终松开撤离；视频 25.667–26.267 s 毯子平铺并保持稳定。推断：结合评分器两次回放均严格成功、无末尾补充保持或容差的记录，唯一目标“毯子展平”在视频约 26.2 s 达成，最终完成 1/1 节点，无失败断点。无状态日志，联合验收的细微物理阈值及首次确认时刻不能仅凭图像精确复核；reward 未逐项报告 DAG 的八个原生指标数值，均保留 null。", + "conf": 8, + "review": { + "status": "agreed", + "rounds": 1, + "reviewers": [ + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 1, + "verdict": "agree", + "conf": 8, + "summary_zh": "单节点 blanket_flattened 在视频约 26.2 s 达成与 strict_goal_row=1049（sim 26.225 s）一致；双爪抓握、铺平、再修整折边的帧所见与接触标注相符；成功运行无失败断点。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": "2026-10-29T09:54:01Z", + "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/roboarc-blanket-flatten-04-i01-privileged--44za5cn/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/roboarc-blanket-flatten-04-i01-privileged--44za5cn/04_review_r1" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "7d7037aed164bbda9156ca285998f86ef399389c24b676382769478173dce957", + "video_sha256": "09c786432c6c4928357ebc4df7481bd84fbcde0298e19b599058beec19343af5", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +} diff --git a/phase2/roboarc/blanket_flatten_04/trajectories/roboarc-blanket-flatten-04-i01-privileged--jbdxpbq.json b/phase2/roboarc/blanket_flatten_04/trajectories/roboarc-blanket-flatten-04-i01-privileged--jbdxpbq.json new file mode 100644 index 0000000000000000000000000000000000000000..09dee9d18dc324d54293fd6bd8a997d49fc4fad7 --- /dev/null +++ b/phase2/roboarc/blanket_flatten_04/trajectories/roboarc-blanket-flatten-04-i01-privileged--jbdxpbq.json @@ -0,0 +1,284 @@ +{ + "schema": "efe-trajectory-labels/1", + "taxonomy": "v2", + "runId": "roboarc-blanket-flatten-04-i01-privileged--jbdxpbq", + "taskKey": "roboarc/blanket_flatten_04", + "dag_sha256": "a99674e5c44990e94d9aa8a3601a1e41c800bc3a09b75ed544b76e83470ebada", + "outcome": { + "success": 0, + "reward": 0.0, + "source": "catalog", + "raw": { + "success": 0, + "success_pending_review": 0, + "review_events": 0, + "review_max_depth_mm": 0.0, + "forcing_events": 0, + "all_success": 0, + "deterministic": 1, + "gpu_backend": 1, + "infra_error": 0, + "simulator_instability": 0, + "instance_ok": 1, + "instance_exact": 1, + "initial_state_generated": 0, + "instance_realized": 0, + "assets_exact": 1, + "runtime_binding_ok": 1, + "n_actions": 1699, + "submitted_actions": 1699, + "sim_steps": 1779, + "physics_steps": 17790, + "sim_time_s": 44.474999999998545, + "final_reward": 0.0, + "replay_wall_s": 1101.3, + "replays": 2, + "grader_error": 0, + "video_rendered": 1, + "replay_hold_steps": 80, + "replay_tolerance_needed": 0, + "strict_goal_reached": 0, + "strict_goal_row": 0, + "strict_replay_required": 0, + "agent_sessions": 3, + "agent_resets": 36, + "agent_sim_steps": 23612, + "agent_records": 15, + "agent_replays": 2, + "agent_cost_self_reported": 1, + "resource_enforced": 0, + "limits_violation_observed": 0, + "resource_monitor_missing": 0, + "peak_gpu_memory_gb": 0.979, + "peak_memory_gb": 3.258, + "peak_processes": 14, + "peak_disk_gb": 1.386, + "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/roboarc-blanket-flatten-04-i01-privileged--jbdxpbq/replay.mp4", + "duration_s": 44.533, + "fps": 15.0, + "frames": 668, + "sha256": "5f0b7f4abc512694280e39135388b6dd72826621acff4f971f692342359e1e1a" + }, + "state": null, + "progress": { + "method": "visual", + "denominator": 1, + "base": 0, + "peak": 0, + "final": 0, + "nodes": { + "blanket_flattened": { + "method": "visual", + "intervals": [], + "final": false, + "evidence": [ + { + "t": 0.0, + "log_step": null, + "text_zh": "观察：视频 0.0 s，蓝色毯子呈窄长的折叠片状，左侧与近侧有鼓起的折皱，未展成完整单层毯面。" + }, + { + "t": 14.5, + "log_step": null, + "text_zh": "观察：视频 6.5–14.5 s，机器人左爪夹住布片并抬起、向远侧翻开后放回桌上；投影范围增大，但左侧仍有折回布片和连续皱脊，近侧边界仍向内收缩，尚未展平。" + }, + { + "t": 30.0, + "log_step": null, + "text_zh": "观察：视频 16.5–35.5 s，机器人主要用左爪在近侧边缘不同位置下降、闭爪、局部提拉和松开；毯边有小幅移动，左侧折回与近侧皱脊始终没有消除。" + }, + { + "t": 44.467, + "log_step": null, + "text_zh": "观察：视频 37.5–44.467 s，机器人停止进一步铺展；视频 43.4–44.467 s 的逐帧末段仍见左侧及近侧折皱、局部隆起，未形成充分展开且贴桌的单层片材。评分器 reward.success=0、strict_goal_reached=0，与该可见未完成状态一致。" + } + ], + "contact_actual": [ + { + "who": "双臂协作", + "how": "抓握", + "note_zh": "观察：视频约 4.5–14.5 s，机器人左二指夹爪夹住毯子的局部布片，抬起并翻开，右爪同时在邻近布面辅助压住后撤开；未把右爪也判成有效抓握。视频约 16.5–35.5 s，左爪单独依次改抓近侧局部布边并提拉，部分闭爪后空退。" + }, + { + "who": "双臂协作", + "how": "压/擦", + "note_zh": "观察：视频约 4.0–6.5 s，机器人右夹爪指端落在邻近毯面上压住布料，配合左爪的夹持；后续右爪抬离，主要由左爪继续操作。" + } + ] + } + }, + "constraints": {} + }, + "native_metrics": [ + { + "name": "coverage_ratio", + "unit": "ratio", + "threshold": 0.95, + "direction": "higher_better", + "value": null + }, + { + "name": "maximum_height_m", + "unit": "m", + "threshold": 0.03, + "direction": "lower_better", + "value": null + }, + { + "name": "minimum_height_m", + "unit": "m", + "threshold": -0.008, + "direction": "higher_better", + "value": null + }, + { + "name": "maximum_stretch_ratio", + "unit": "ratio", + "threshold": 1.15, + "direction": "lower_better", + "value": null + }, + { + "name": "maximum_speed_m_s", + "unit": "m/s", + "threshold": 0.03, + "direction": "lower_better", + "value": null + }, + { + "name": "maximum_drift_m", + "unit": "m", + "threshold": 0.003, + "direction": "lower_better", + "value": null + }, + { + "name": "release_clearance_m", + "unit": "m", + "threshold": 0.005, + "direction": "higher_better", + "value": null + }, + { + "name": "confirmation_elapsed_s", + "unit": "s", + "threshold": 0.5, + "direction": "higher_better", + "value": null + } + ], + "failure": { + "breakpoint": { + "ref": "blanket_flattened", + "t": 14.5, + "log_step": null, + "where": "操作中", + "evidence_zh": "观察：视频 14.5 s，机器人完成首次翻开并放回毯面后，左侧折回布片与皱脊仍在；视频 16.5–35.5 s 的后续局部牵拉未将其消除，视频 43.4–44.467 s 末段仍保持未展平形状。断点取这一次始终未完成的展平操作，视频 14.5 s 为首次翻开后残留折皱的导航锚点。" + }, + "unmet": { + "blanket_flattened": false + }, + "phenomena": [ + { + "cat": "接触", + "item": "没建立接触", + "t": 20.6, + "log_step": null, + "evidence_zh": "推断：视频 19.8–21.2 s 的连续帧中，机器人左爪在近侧偏左布边闭合后向外、向上回撤；视频 20.6 s、21.2 s 主视图中指间可见浅色桌面，没有夹起的蓝色布片，邻近布边仍留在桌面，没有随爪抬起。因此这次局部改抓未建立牵拉所需的夹持；不是由最终评分为零推得，也不表示此前翻开布片时从未抓住。", + "note_zh": "仅标注该次局部改抓；未观察到先夹住布片再滑脱的过程，不标“没保持住”。log step 248 的多次 pull False 仅为会话脚本旁证，不与该视频时刻逐次对应。" + } + ], + "root_causes": [ + { + "cat": "控制", + "item": "控制", + "source_zh": "日志、视频", + "evidence_zh": "推断：log step 224 的生成脚本及 log step 242 的最终修改采用局部闭爪、抬起并外拉的动作；log step 248 记录该最终生成过程多次 pull 返回 False（脚本在 object_in_hand 为空时放弃该次牵拉），其余局部牵拉也只小幅改变毯面。视频 16.5–35.5 s 与之相容：左爪反复在边缘闭合、上移，布料有时仅小幅变形，左侧和近侧折皱持续存在。所需的可靠夹持及有效铺展未实现，支持控制层面的失败；现有证据不足以可靠细分其控制子类。" + } + ], + "gap_zh": "没有状态日志，无法核实每次夹持的接触对、粒子高度、拉伸和净空等精确量，也不能进一步区分接触几何与运动跟踪问题。评分器 reward 未报告 DAG 所列八项连续指标，native_metrics.value 均保留 null；log step 250 的生成脚本自测数值属于会话证据，未当成评分回放末态测量。参考成功运行的 oracle.json 明确无视频，无法作成功画面对照。 视频局部改抓的指间与布边可辨认，支持所标一次未建立夹持，但不能从这些画面确定其余每次改抓是否夹住或滑脱。" + }, + "summary_zh": "机器人先以左爪夹持、右爪辅助压住毯面，翻开上层后再由左爪多次局部牵拉，毯面明显展开，但“毯子展平”始终未达成，最终为 0/1。视频 14.5 s 起留下的左侧及近侧折皱在后续牵拉和视频 44.467 s 末帧仍可见；视频 19.8–21.2 s 的一次局部改抓中，左爪闭合后空退，未建立牵拉所需的夹持，标为接触·没建立接触。最终脚本的多次夹持失败反馈与视频中的低效局部铺展支持控制根因。没有状态日志或评分器连续指标，精确验收量及控制子类不能确定。", + "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（blanket_flattened 未达、断点 14.5 s 操作中、末帧 0/1、左爪主导的双臂协作/抓握+压）与视频一致；问题在 phenomena：“折叠/褶皱残留”只是重述未达的展平指标（N10），不是机器人失败现象。", + "n_issues": 1, + "n_major": 1 + }, + { + "engine": "claude", + "model": "claude-sonnet-5-5", + "round": 2, + "verdict": "agree", + "conf": 6, + "summary_zh": "节点未达、断点、接触标注与视频一致：视频 6、14.5、20.6、30、44 s 帧显示毯子左侧及近侧折皱始终存在，末帧未展平；单一现象和控制根因有证据且未过度细分。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": "2026-10-29T10:08:17Z", + "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/roboarc-blanket-flatten-04-i01-privileged--jbdxpbq/03_label" + }, + { + "stage": "traj_review", + "round": 1, + "ok": true, + "space": "work/runs/roboarc-blanket-flatten-04-i01-privileged--jbdxpbq/04_review_r1" + }, + { + "stage": "reconcile", + "round": 1, + "ok": true, + "space": "work/runs/roboarc-blanket-flatten-04-i01-privileged--jbdxpbq/05_reconcile_r1" + }, + { + "stage": "traj_review", + "round": 2, + "ok": true, + "space": "work/runs/roboarc-blanket-flatten-04-i01-privileged--jbdxpbq/04_review_r2" + } + ], + "review_engine": "claude", + "review_model": "claude-sonnet-5-5", + "inputs": { + "run_json_sha256": "35f13e4cae9623aa57fae44a44ba5f9d71927fd9be3b3ac9222a6b0027aef6ec", + "video_sha256": "5f0b7f4abc512694280e39135388b6dd72826621acff4f971f692342359e1e1a", + "states_sha256": null, + "video_choice": "HF replay.mp4" + } + } +}