Event

RePO-VLA Releases Recovery-Driven Policy Optimization Framework

1 reports · 1 sources

Weijia Liufu, Xiaoyu Guo, and others proposed and released the RePO-VLA recovery-driven policy optimization framework, which assigns different roles to successful, recovery, and failure trajectories for long-horizon, contact-intensive manipulation tasks. The method first uses recovery-aware initialization to split corrective segments into new episodes, then learns a progress-aware semantic value function, and uses value-conditioned refinement to make the policy prefer high-progress actions. The team also released the Failure-Recovery Benchmark (FRBench), which provides standardized error injection and recovery-focused evaluation. The authors report that in simulated and real dual-arm tasks, the method raises the success rate in adversarial scenarios from 20% to 75% on average, and up to 80% in scaled-up real-world trials; at deployment, a single fixed high-value token guides actions toward the learned success manifold, without needing an online failure detector or heuristic retries. These results are reported by the authors and await independent verification.