New Method Enhances Autonomous Driving Safety Training

Ensuring the safety of autonomous driving systems, particularly those trained with online reinforcement learning (RL), is challenging due to the rarity of dangerous real-world scenarios. A new method, Threat-guided Policy-aware Scene Perturbation (TPSP), addresses this by intelligently generating safety-critical training scenes. Unlike previous approaches that create challenging scenes without considering the evolving policy, TPSP explicitly models the interaction between the policy's behavior and its environment. TPSP uses a policy-aware scene encoder to identify and selectively perturb critical objects within a scene, rather than applying uniform modifications. It then employs a threat-guided optimization strategy, evaluating perturbed scenes based on the difference in threat levels between original and perturbed policy rollouts. This guides the generation of highly informative, safety-critical experiences, significantly improving safety learning efficiency and performance in simulated driving environments.