Threat-guided Policy-aware Scene Perturbation Enhances Safety in Autonomous Driving RL

A new arXiv preprint (2608.10403) introduces Threat-guided Policy-aware Scene Perturbation (TPSP), a method to improve safety in online reinforcement learning (RL) for autonomous driving. The paper addresses the challenge of ensuring safety in RL policies due to insufficient exposure to safety-critical driving scenes, which are rare in real-world traffic due to their long-tailed nature. Existing methods synthesize challenging scenes but often optimize scene generation separately from the evolving policy, failing to model how perturbations relate to the current policy's weaknesses. TPSP incorporates a policy-aware scene encoder to capture the interaction between generated perturbations and the policy's learning needs, guiding scene generation to target specific weaknesses. The method is designed for online RL, where policies are updated in real-time, and aims to enhance training diversity and robustness. The paper is authored by researchers and posted on arXiv, a preprint server, indicating it has not yet undergone peer review. The work contributes to the field of safe autonomous driving, a critical area in AI and robotics.