Third-party cyber evaluations involving OpenAI models
openai.com·13h ago
AI Research
arXiv CS.AI·Jun 9·5 min read
ReCoVLA: VLM-Guided Reward Compilation for Failure Recovery in Vision-Language-Action Policies
TL;DR
Vision-language-action (VLA) policies struggle with recovery from unexpected failures. ReCoVLA is a framework that uses a vision-language model (VLM) to identify failure modes and generate structured rewards for recovery training.
✦ Why It Matters
Engineers can leverage ReCoVLA to improve the reliability of robotic systems in unpredictable environments.
Key Takeaways
- Utilize ReCoVLA for robust failure recovery in robotic tasks.
- Incorporate vision-language models to enhance understanding of failures.
- Train residual policies using structured rewards from VLMs.
- Aim for higher success rates in both simulation and real-world applications.
Full Summary
Vision-language-action (VLA) policies are designed for tasks that involve understanding language instructions to perform manipulations, but they often fail in unexpected situations. ReCoVLA is introduced as a failure-conditioned residual recovery framework that keeps the original VLA policy unchanged while leveraging an external vision-language model (VLM) to analyze failure modes and recovery stages.
Instead of generating actions directly, the VLM acts as a semantic reward selector, predicting recovery descriptors and reward masks for training a residual policy in simulation. The results indicate that ReCoVLA improves average success rates from 36.7% to 66.7% in simulations and achieves a 61.7% success rate in physical zero-shot sim-to-real experiments.
This approach effectively decouples high-level understanding of failures from low-level control actions, making it adaptable to various VLA policies. The findings suggest that integrating VLMs can enhance the robustness of robotic systems in real-world applications.