Reproduce “ReCoVLA: VLM-Guided Reward Compilation for Failure Recovery in Vision-Language-Action Policies” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Haodi Hu, Chung-Ta Huang, Jing Liu, Ye Wang, Kei Suzuki, Matthew Brand, Toshiaki Koike-Akino

ReCoVLA helps a vision-language-action robot recover from known failures by freezing its base policy and training a corrective policy in simulation. A vision-language model identifies failures and selects task-grounded rewards; a deterministic compiler binds entities and gates rewards by recovery stage. Across three Fetch tasks, M4 raises simulation success from 36.7% to 66.7%, with 61.7% physical success, for catalogued failures.

Key insights

  • Decoupling high-level semantic failure understanding using a VLM from low-level corrective control through a residual policy improves recovery robustness.
  • Utilizing VLMs to guide structured reward compilation for reinforcement learning is more stable and reliable than direct action generation.
  • Stage-gated reward activation, which ensures reward components are active only under specific recovery preconditions, is critical for achieving reliable and efficient recovery.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.