新智元 (微信公众号)T2·72 pts
- Open-vocabulary 3D affordance prediction constrains grasp selection so that later placement or hanging steps retain feasible space
- Natural-language task descriptions drive scene and task organization, with motion planning producing executable demonstrations and cutting real-robot teleoperation cost
- A benchmark of 50 tasks, 7 skill types and 5 robot arms over 500+ rigid and articulated objects is released alongside open-source code