Summary

ReCoVLA is a failure-conditioned residual recovery framework that keeps a pretrained VLA frozen and uses an external VLM as a semantic reward selector — predicting a recovery descriptor and reward mask that drives in-simulation residual-policy RL training, followed by zero-shot sim-to-real deployment. It raises average simulation success from 36.7% to 66.7%, and outperforms baselines by 18.3 percentage points in physical Fetch robot experiments.

Key Contributions

  • Keeps the base VLA entirely frozen and learns only a residual recovery policy, which limits the risk of catastrophic forgetting or capability regression from RL fine-tuning of the full policy.
  • Uses a VLM as a semantic reward selector to generate task-relevant recovery descriptors and reward masks automatically, avoiding hand-engineered reward functions for each failure type.
  • Demonstrates zero-shot sim-to-real transfer of the residual recovery policy, with a physical validation on a real Fetch robot rather than simulation-only results.

Strengths

  • The frozen-base-policy-plus-residual-recovery design is architecturally conservative and practical: it doesn’t risk degrading the base VLA’s general competence to fix specific failure recovery behavior.
  • Real physical robot validation (Fetch), not just simulation, and an 18.3pp improvement over baselines gives a concrete, checkable claim beyond simulation-only results.
  • Automating reward specification via VLM-generated descriptors and masks reduces the reward-engineering burden that has historically limited RL applicability to diverse failure types.

Weaknesses

  • Reliance on the VLM’s ability to correctly identify failure states and generate appropriate recovery descriptors means errors in VLM judgment propagate directly into the RL reward signal — the paper’s handling of VLM misjudgment is not detailed in available coverage.
  • Simulation success rate of 66.7% after improvement, while a large jump from 36.7%, still leaves roughly a third of simulated recovery attempts failing, suggesting substantial headroom remains.

Open Questions

  • How does ReCoVLA’s residual recovery policy generalize to failure modes not represented in its simulation training distribution?
  • Does the VLM reward selector need to be re-tuned or re-prompted per robot embodiment, or does it transfer across platforms with the same base VLA?

Significance

A well-scoped contribution to the growing body of VLA failure-recovery work in this vault (FAR, HAVE, CoRe), notable for combining frozen-base-policy conservatism with automated VLM-driven reward generation and real hardware validation.