FailBank is a four-stage framework that converts runtime safety feedback into permanent policy improvements for Vision-Language-Action models. It uses a CBF-based safety module to generate counterfactual corrections that are then used to update the underlying policy.
HOW THIS AFFECTS YOU
●
builderYou can implement a system where your robots learn from their mistakes and safety interventions in real-time.
●
researcherThis addresses the persistent policy-shield mismatch in robotic manipulation tasks.