Route-Guided CBF-QP Repair for Safe Vision-Language-Action Models

This paper introduces Route-Guided CBF-QP Repair, a post-hoc repair module designed to enhance the safety and reliability of vision-language-action (VLA) models in obstacle-rich manipulation scenarios. By incorporating short-horizon action chunks predicted by the VLA into a route-guidance term within a Control Barrier Function Quadratic Programming (CBF-QP) framework, the method ensures that unsafe actions are corrected without deviating excessively from the learned task trajectory. The approach maintains hard CBF safety constraints while using soft guidance to navigate around obstacles, improving collision avoidance, task success, and SafeSuccess metrics compared to the AEGIS baseline on the SafeLIBERO benchmark. The project page is available on GitHub.