Testing self-driving cars with AI-generated adversarial scenarios

PCASim addresses the challenge of testing autonomous driving systems in complex urban environments by combining adversarial scenario generation with safety agent training. The system builds a repository of adversarial driving behaviors from real data, uses an LLM to generate custom safety-critical scenarios based on user specifications, and trains vehicle behaviors with reinforcement learning to increase scenario diversity while maintaining realism. Experiments show 12% improvement in domain-specific language accuracy, 8% higher success rate for scenario transformations, and 30% better obstacle-avoidance performance compared to baseline approaches. Code and details are available online.