VOMMI: Portable demonstrations connect to VLA post-training

Yutian Zhang and other authors released VOMMI, a portable demonstration collection and learning framework that connects portable RGB demonstrations to VLA post-training through offline trajectory reconstruction and online visual-motor conditioning. Among these, R2-VO uses sparse geometric anchors to refine trajectories, reducing the average absolute trajectory error of body and hand streams by 24.6%. The authors state that a policy post-trained using only portable demonstrations has a base velocity error 18.2% lower than a policy trained with robot-collected demonstrations, with comparable end-effector translation accuracy; the complete system's average success rate on three real robot tasks is 8.3 percentage points higher than OpenPI 0.5. In the experimental setup, each task uses 500 portable trajectories, of which 75 are held out for RGB-VO evaluation, with 200 robot demonstrations as a reference.

AI 根据报道生成 · 2 days ago更新