左：三路相机的重放录像（腕部 / 肩上 / 对面）· 中：agent 每一轮的推理摘要、代码，以及下发的动作和它写的说明，逐字原样 · 右：focus（这一轮 agent 收到的 mesh）与这一轮合爪的仿真真值 · 下：产物架与成对时间轴，再下面是每次合爪的三维视图。点 ▶ 播放，或点时间轴上任意一点跳转；键盘 ← → 逐轮、空格 播放/暂停。
▸ 本组条件：两组提示词的唯一差别
加 focus 组的任务说明里多出这一段（原文）：
## Focus tool (required)
`r.focus(center, radius, note)` returns the triangle mesh of the surfaces your cameras see in the **current** frame within `radius` m (at most 0.15) of `center` ([x, y, z] m, env-local world frame), with the robot's own body removed: `{"vertices": (N,3) m, "faces": (M,3), "path": PLY file, "meta": per-camera counts}`. It is built only from the current depth and camera calibration (the same data as `r.get_obs(depth=True)`); surfaces no camera sees are absent. `note` is required: what this focus is on and what you want to learn from it.
**Before every gripper close you must call `r.focus(...)` on the region the fingers will touch**, and base the approach and the close on that mesh.