8×
8 倍速 · 原始 137 s，播放 17 s8x speed · original 137 s, plays in 17 s
SLR-01原版：钩杆三段短拉取物Original: hook retrieval in three short pulls
- 条件Conditions
- 名义配置。35 mm、60 g 黄色小盒子位于 0.60 m 高台面上的窄缝深处；窄缝宽 45 mm、净高约 160 mm、长 0.50 m，两侧与顶盖固定；台面上放有一根 0.46 m 自由钩杆；摩擦 0.28–0.40。超时 240 s；物理 200 Hz、控制 50 Hz；参考解读取特权仿真状态。Nominal configuration. A 35 mm, 60 g yellow box sits deep inside a slot on a 0.60 m work surface; the slot is 45 mm wide, about 160 mm high and 0.50 m long, with fixed sides and roof; a free 0.46 m hook tool lies on the table; friction 0.28–0.40. 240 s timeout; physics at 200 Hz, control at 50 Hz; the reference reads privileged simulator state.
- 任务设计Task design
- 盒子须完整放入台面目标区，由台面支撑、与机器人和工具均无接触并静置两秒；不得掉落，不得移动固定结构。G1 腕部内切球直径约 52.37 mm，大于 45 mm 缝宽，正面直接伸手不可达（只覆盖正面直伸，不是全局证明）。参考专家：抓起钩杆并拇指对握 → 在入口外校正位姿 → 越过目标后下落勾住 → 最多三段短拉，段间放开由台面支撑的目标并后退 0.12 m 换位 → 脱离并静置。The box must end fully inside the tabletop goal zone, supported by the table, out of contact with robot and tool and still for two seconds, without dropping it or moving the fixed structure. The G1 wrist's inscribed sphere (about 52.37 mm) exceeds the 45 mm slot, so a straight frontal reach is excluded (this covers only frontal reaching, not a global proof). Reference expert: grasp the hook with thumb opposition → correct its pose outside the entrance → pass beyond the target and drop the hook behind it → at most three short pulls, releasing the table-supported box and stepping back 0.12 m between pulls → disengage and settle.
- 工具Tools
- 0.46 m 自由钩杆（钩头高 50 mm，约 0.24 kg）Free 0.46 m hook tool (50 mm hook head, about 0.24 kg)
- Embodiment
- Unitree G1（自由基座）＋Dex3 手Unitree G1 (free base) with Dex3 hands
- 是否拒答Refusal
- 否No
- 见证结果Witness result
- 选定见证 136.66 s 完整成功，成功分 100，三层映射独立复核通过；固定结构最大接触约 12.6 N，机器人最大倾斜约 14.6°。固定版本新运行仅 1/3 成功，早期批次 0/3；原始动作开环回放 0/1，于 4.86 s 摔倒。这是可行解见证，不是稳健控制器。The selected witness succeeds completely in 136.66 s with success score 100 and passes independent three-level mapping audit; peak fixed-structure contact is about 12.6 N and maximum robot tilt about 14.6 degrees. Fresh runs of the frozen version succeed only 1/3 and an earlier cohort 0/3; the open-loop raw action replay fails 0/1, falling at 4.86 s. This is a feasibility witness, not a robust controller.