Event

Gradient-Free Neural HJ Reachability Framework Proposed

1 reports · 1 sources

Zeyuan Feng and other authors propose a discrete-time neural reachability framework for control-disturbance affine systems, which learns backward reachable tubes (BRT) and backward reachable avoidance tubes (BRAT) through Bellman-Isaacs value propagation, and uses gradient-free value probing to construct approximate teacher actions, turning adversarial actor learning into supervised policy learning. The authors claim that on benchmark problems of up to 80 dimensions, the method can learn accurate reachability value functions and is more stable than existing learning-based solvers; in F1-tenth racing, the scalability of the observation space was verified with self-centered inputs of over 16,000 dimensions. The authors also claim that the learned safety filter can generalize zero-shot to unseen tracks and transfer to a physical remote-controlled car, achieving real-time robust obstacle avoidance. All the above results are as reported by the authors.