Recoleta
GitHub
Source note

RePO-VLA: Recovery-Driven Policy Optimization for Vision-Language-Action Models

Authors
Weijia Liufu, Xiaoyu Guo, Ruiyi Chen, Jingzhi Liu, Kaidong Zhang, Xiwen Liang, et al.
Source
arXiv
Published
Collection
Embodied AI
Vision Language ActionRobot RecoveryBimanual ManipulationRobot Data ScalingGeneralist Robot Policy
Open arXivMarkdown source

Summary

RePO-VLA trains VLA robot policies to recover after contact-rich manipulation drift by using success, failure, and recovery rollouts with different labels. It reports adversarial success rising from 20% to 75% on average, with up to 80% in scaled real-world trials.

Problem

  • Long-horizon bimanual tasks can fail after small grasp, contact, or timing errors, even when the robot could still recover.
  • Success-only imitation gives weak supervision for drift and often discards failed rollouts that contain useful early behavior.
  • Naive imitation of recovery episodes can mix the actions that caused drift with the actions that repaired it.

Approach

  • Recovery-Aware Initialization slices recovery segments out of full episodes and resets observation history, so the policy learns correction from the current adverse state.
  • A Progress-Aware Semantic Value Function uses frozen V-JEPA trajectory features, language embeddings, and successful reference trajectories to assign dense progress labels.
  • Failed trajectories keep useful prefixes through a reliability decay label, with alpha = 3.0, while terminal drift gets low value.
  • Value-Conditioned Refinement trains a flow-matching pi_0.5 policy with a value token, using success, raw recovery, and failure data together.
  • At deployment, the policy uses normal rolling history and a fixed value condition v = 1.0, with no online failure detector or hand-coded retry rule.

Results

  • The paper claims average adversarial success improves from 20% to 75%, with up to 80% success in scaled real-world trials.
  • FRBench-Sim contains 23,453 simulated bimanual episodes across 46 tasks: 17,061 nominal success episodes and 6,392 verified failure-recovery episodes.
  • FRBench uses 4 injected error types: premature close, grasp slip, grasp position offset, and grasp orientation mismatch.
  • Error-injection coverage includes 8,022 premature-close cases, 3,516 grasp-slip cases, 4,686 position-offset cases, and 688 orientation-mismatch cases before final filtering.
  • The simulation protocol runs 50 rollouts per task and injects a grasp disturbance by holding the gripper open for 30 frames, about 1 second.

Link