Blame for phase2/humanoidbench/manip_cube/task_dag.json

{
 "schema": "efe-task-dag/1",
 "taxonomy": "v2",
 "taskKey": "humanoidbench/manip_cube",
 "suite": "humanoidbench",
 "base": "humanoidbench-manip-cube-i00",
 "title_zh": "双手方块调姿",
 "summary_zh": "保持两个方块不掉落并将其转向目标朝向，使整段朝向、手块距离及站立控制奖励之和达到370。",
 "embodiment_zh": "三个模式均为Unitree H1双足人形机器人，左、右臂各装一只Shadow灵巧手，合计61个位置伺服执行器；双足与腿躯干支撑，无轮式底盘。特权模式可读完整物理状态，标准及实时模式通过头部双相机、房间相机和本体感觉控制，同一固定初态。",
 "task_class": "atomic",
 "analysis_level": "step",
 "spec": {
  "version": "1.2",
  "task": "humanoidbench-manip-cube-i00",
  "bench": "humanoidbench",
  "progress": "live",
  "entities": {},
  "milestones": [
   {
    "id": "cubes_reoriented",
    "zh": "转好两个方块",
    "mode": "event",
    "when": null,
    "after": [],
    "hold_s": 0,
    "goal": true,
    "freeze_held": false,
    "criterion_zh": "两只手附近的彩色方块被转向悬浮半透明目标方块所示的朝向；画面中可比较各色面的方向，初始平掌上方的方块尚未转向目标即未完成。验收以整段每步奖励之和达到370为准：朝向奖励占0.5、手块接近奖励占0.3、站立与低用力及低水平速度的乘积占0.2；瞬间看似对齐不代表已达标，视频也不能识别四元数符号或精确累计得分，须以原生累计奖励或该集成功记录确认。",
    "step": {
     "cat": "位姿",
     "item": "调整位姿"
    },
    "contact": [
     {
      "who": "双臂协作",
      "how": "抓握",
      "plan": 1,
      "note_zh": "左手灵巧手握左侧方块，右手灵巧手握右侧方块，各自随手腕和手臂转向目标；两手不必同时开始。"
     },
     {
      "who": "腿/躯干",
      "how": "支撑",
      "plan": 1,
      "note_zh": "人形机器人的双脚支撑地面，腿与躯干在双手转动方块期间承重。"
     },
     {
      "who": "双臂协作",
      "how": "手内调整",
      "plan": 2,
      "note_zh": "左、右灵巧手的手指分别在手内重定向各自方块，不要求两手遵循固定先后。"
     },
     {
      "who": "腿/躯干",
      "how": "支撑",
      "plan": 2,
      "note_zh": "人形机器人的双脚支撑地面，腿与躯干在双手手内调整方块期间承重。"
     }
    ],
    "optional": false,
    "verify_zh": null,
    "sources": [
     "inputs/task/privileged/instruction.md:3-5 两个方块转到目标朝向且不得掉落",
     "inputs/task/privileged/instruction.md:12-23 整段累计奖励阈值370及四元数符号敏感",
     "inputs/task/privileged/tests/rewards.json:7 Cube.get_reward的完整公式：朝向四元数差范数，margin=0.3；手块距离bounds=(0,0.1)m、margin=0.5m；权重0.5/0.3/0.2",
     "inputs/task/privileged/tests/rewards.json:11 站立奖励头部高度参数1.65m；并非跌倒终止阈值",
     "inputs/grader/efe_humanoidbench-files/privileged/environment/rcb_hb.py:371-379 每步累加奖励",
     "inputs/grader/efe_humanoidbench-files/privileged/environment/rcb_hb.py:425-429 原生成功条件episode_return >= success_bar",
     "inputs/oracle/replay.mp4 @ 0.00,0.10,0.30,0.60,5.00s 初始平掌托块，随后两手转动并持块；部分方块受手掌遮挡"
    ],
    "conf": 6,
    "note_zh": "按DECISIONS.md D6的单一累计量规则合为一个语义步骤，未另建积分、抓取、拿起、身体平衡或松手节点。mode=event：验收为已累积的奖励，达标后不因瞬时朝向变化撤销；不添加瞬时误差硬阈值，也不把margin=0.3当作通过阈值。典型方案先握住方块、再用腕臂旋转；也可在手内调整，两种接触方案分开列。允许持物达标，freeze_held=false。"
   }
  ],
  "constraints": [
   {
    "id": "cube_dropped",
    "zh": "方块掉落",
    "never": null,
    "criterion_zh": "任一方块跌到任务原生低高度终止界限以下即结束该集；视频中表现为方块从手边下坠至低处。短暂离手不等于已触发低高度终止，给定文件未包含该数值阈值。",
    "consequence": "立即终止",
    "sources": [
     "inputs/task/privileged/instruction.md:21-23",
     "inputs/task/standard/instruction.md:10-14",
     "inputs/task/standard/environment/docker-compose.yaml:27"
    ]
   },
   {
    "id": "pelvis_too_low",
    "zh": "骨盆过低",
    "never": null,
    "criterion_zh": "机器人骨盆降到任务原生终止高度以下即结束该集；画面可见身体明显塌低或倒下，但不能把任意倾斜、坐下或侧躺直接当作触发，给定文件未包含该数值阈值。",
    "consequence": "立即终止",
    "sources": [
     "inputs/task/privileged/instruction.md:21-23",
     "inputs/task/standard/instruction.md:10-14",
     "inputs/task/standard/environment/docker-compose.yaml:27"
    ]
   }
  ],
  "success": null,
  "atoms": [
   {
    "atom": "episode_return_ge_success_bar",
    "level": "new_code",
    "source": "需记录Session._ret与Session.success_bar：inputs/grader/efe_humanoidbench-files/privileged/environment/rcb_hb.py:373,425-429；世界过程奖励累计达到370，当前无状态日志。"
   },
   {
    "atom": "cube_below_termination_height",
    "level": "new_code",
    "source": "需记录Cube.get_terminated对左右方块高度的原生判据；inputs/task/standard/instruction.md:10-14确认低高度终止，原任务模块与数值阈值未提供。"
   },
   {
    "atom": "pelvis_below_termination_height",
    "level": "new_code",
    "source": "需记录Cube.get_terminated对骨盆高度的原生判据；inputs/task/standard/instruction.md:10-14确认低高度终止，原任务模块与数值阈值未提供。"
   }
  ]
 },
 "native_metrics": [
  {
   "name": "episode_return",
   "unit": "reward",
   "threshold": 370,
   "direction": "higher_better"
  }
 ],
 "open_questions_zh": [
  "inputs/states/index.json的logs为空；所有when、never及success均为null，无法进行状态日志绑定和回放一致性检验（暂行：DECISIONS.md P8）。",
  "粒度读取：文本一句话给出两个方块的同一调姿目标；按Step 0.3及Step 3.7核对后，因评分器只检查跨双手、跨整集的单一累计奖励阈值而不逐块计数，依DECISIONS.md D6的当前规则视为一个联合调姿步骤，并依P17将task_class定为atomic。保留单节点，不构造第1/2块的任意误差阈值；累计量的粒度规则仍为暂行规则。",
  "任务设计问题：任务文字要求两个方块达到目标朝向，实际成功是累计稠密奖励>=370；接近方块、站立等项可补偿部分朝向误差，也不要求结束时两个方块同时精确对齐。四元数直接相减还区分q与-q，物理朝向相同却可能得分不同；说明已告知此规则，视频不能区分该符号（inputs/task/privileged/tests/rewards.json:7；instruction.md:12-23）。",
  "给定输入没有Cube.get_terminated完整源码，也没有冻结instance.npz或状态日志；只能由说明与录像确认初态和低高度终止规则，不能核实具体骨盆、方块终止高度。不采用奖励的头部高度1.65m、代理通用fallen阈值或oracle试验高度替代。",
  "模式差异：共同物理验收均是累计奖励>=370；特权模式对动作文件进行两次独立重放并核对最终状态哈希，标准及实时模式判最后一集的现场结果且达标立即停止。确定性、交付格式、基础设施有效性不建世界状态节点；特权模式积分达标后仍可能继续至500步，之后低高度终止不构成额外的历史否决（grade.py:72-76；rcb_hb.py:425-443；server.py:267-270；rcb_service.py:695-705）。",
  "成功录像仅256像素且部分方块被手遮挡；可见平掌托块到握持转动的变化，无法精确判定朝向误差、四元数符号与370分达标时刻。节点需结合原生评分记录确认，不将oracle的极小终态误差或500步保持设成要求。"
 ],
 "conf": 6,
 "conf_reason_zh": "文本、恢复的奖励公式与会话累计条件一致；但没有状态日志、原生低高度终止源码及可由视频精确辨识的累计得分。",
 "review": {
  "status": "agreed",
  "rounds": 2,
  "reviewers": [
   {
    "engine": "claude",
    "model": "claude-sonnet-5-5",
    "round": 1,
    "verdict": "fix",
    "conf": 6,
    "summary_zh": "单节点DAG（D6单一累计量）基本合理，约束与成功条件与评分器一致；但task_class写成compositional而文本步骤清单只有一个步骤，与DAG自身及P17矛盾，应改为atomic（或按文本拆成两块的步骤，但评分器只有一个累计阈值，D6不支持）。无状态日志，无法做回放检验。",
    "n_issues": 1,
    "n_major": 1
   },
   {
    "engine": "claude",
    "model": "claude-sonnet-5-5",
    "round": 2,
    "verdict": "agree",
    "conf": 7,
    "summary_zh": "单节点（两个方块调姿）符合D6与P17，task_class=atomic已按上一轮修正；两个终止约束与说明一致，success为null合理（累计奖励无日志字段），open_questions_zh已列出任务设计问题与文本/评分差异。无状态日志，无法做回放检验。",
    "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：['cubes_reoriented']"
  ]
 },
 "provenance": {
  "prompt": "prompts/01_task_dag.md",
  "prompt_sha256": "b8e79c84b73950529a4fa6b988412b2d221d5df4e052adafcd9605afac64aa25",
  "engine": "codex",
  "model": "gpt-6.1-sol",
  "effort": "high",
  "created": "2026-10-29T09:52:26Z",
  "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/humanoidbench__manip_cube/01_dag"
   },
   {
    "stage": "dag_review",
    "round": 1,
    "ok": true,
    "space": "work/tasks/humanoidbench__manip_cube/02_review_r1"
   },
   {
    "stage": "reconcile",
    "round": 1,
    "ok": true,
    "space": "work/tasks/humanoidbench__manip_cube/05_reconcile_r1"
   },
   {
    "stage": "dag_review",
    "round": 2,
    "ok": true,
    "space": "work/tasks/humanoidbench__manip_cube/02_review_r2"
   }
  ],
  "review_engine": "claude",
  "review_model": "claude-sonnet-5-5",
  "inputs_manifest_sha256": "b28c38ff5b0ee635b183c6058cb4698e75b55910e9c4dd9dfaf6c5f9c06ee58d",
  "repo_commit": "80cac7a5a227db930c659c3307a1e79c2a84bc53"
 }
}