RePO-VLA: Dual-arm robot disturbance recovery success rate raised to 75%

A team from Sun Yat-sen University, South China University of Technology, Pengcheng Laboratory and others proposed RePO-VLA, which segments recovery trajectories, trains a progress-aware semantic value function, and performs value-conditioned fine-tuning, enabling dual-arm vision-language-action (VLA) policies to continue advancing tasks after disturbances. The team also built FRBench, a failure-recovery benchmark; the RoboTwin simulation contains 46 dual-arm tasks and 23,453 trajectories. Under the 1x data setting with 50 recovery demonstrations per task, the full method achieved an average disturbed success rate of 30.0%, versus 20.0% for the π0.5 baseline; the model completing only the RAI stage reached 37.5%. Data-scale experiments were conducted only on two real-robot tasks, pouring water and folding towels: as recovery demonstrations per task increased from 50 to 200, the average disturbed success rates were 30%, 65% and 75% respectively. Current validation mainly covers collected error types such as grasp deviation, slipping and object displacement, and real-robot data still relies on manual collection; the team says recovery capability under severe occlusion, tool misuse and larger perception errors, as well as applicability to a broader range of tasks, still requires further research.

AI 根据报道生成 · 5 days ago更新