正式 GT 标注 · 设计稿 · 2026-08-02 · 未实现，待拍板项已标注

判定流程

数据集普查（100 task × 1 demo 抽样，skill 全量聚合）

高亮行 = 在手相位存在且初判需要 VLM gate 的 skill。"焦点位置"列为草案，待审。

VLM gate prompt 草案（few-shot 例子 = brainstorm，待审）

The robot is executing the subgoal: "<subgoal>".
Held object: "<A>". Other involved objects: "<B>", "<C>"...
Image 1 = full scene. Images 2..k = one close-up crop per involved object (in the order listed).

During THIS subgoal, is the functional / visual-focus region limited to the current hand-contact
points, or does it extend beyond them? Examples of beyond-contact focus:
- place a can into a trash can -> focus = the trash can OPENING (not the grip on the can)
- insert a pen into a pencil case -> the case SLOT, plus the pen TIP (held object's functional end)
- pour from a cup onto a plate -> the cup POURING RIM, plus the plate center
- chop vegetables with a knife -> the knife BLADE EDGE, plus the cut line on the vegetable
- sweep the floor with a broom -> the broom BRISTLES, plus the dirt area
- hang a poster on a nail -> the poster hanging hole / top edge, plus the NAIL
- hand an object to someone -> the receiver's HAND, plus the object's free graspable end
Counter-examples (contact-only, answer none): merely holding or carrying an object; releasing it.

For EACH object, output whether it has a beyond-contact focus region in this subgoal and,
if yes, name that region as a short part phrase.
Output JSON: {"reason": "<one sentence>",
 "objects": [{"name": "<A>", "beyond_contact": true/false, "part_phrase": "..."}, ...]}

在手物 DINO part 的弥散渲染（草案，非高斯但类高斯观感）