From Particles to Perils: SVGD-Based Hazardous Scenario Generation for Autonomous Driving Systems Testing
For autonomous driving system testers, PtoP provides a plug-and-play method to generate diverse and realistic failure scenarios in high-dimensional traffic spaces, outperforming existing search-based seeding methods.
PtoP uses Stein Variational Gradient Descent (SVGD) to generate diverse, failure-inducing initial conditions for autonomous driving system testing, improving safety violation rate by up to 27.68%, scenario diversity by 9.6%, and map coverage by 16.78% over baselines.
Simulation-based testing of autonomous driving systems (ADS) must uncover realistic and diverse failures in dense, heterogeneous traffic. However, existing search-based seeding methods (e.g., genetic algorithms) struggle in high-dimensional spaces, often collapsing to limited modes and missing many failure scenarios. We present PtoP, a framework that combines adaptive random seed generation with Stein Variational Gradient Descent (SVGD) to produce diverse, failure-inducing initial conditions. SVGD balances attraction toward high-risk regions and repulsion among particles, yielding risk-seeking yet well-distributed seeds across multiple failure modes. PtoP is plug-and-play and enhances existing online testing methods (e.g., reinforcement learning--based testers) by providing principled seeds. Evaluation in CARLA on two industry-grade ADS (Apollo, Autoware) and a native end-to-end system shows that PtoP improves safety violation rate (up to 27.68%), scenario diversity (9.6%), and map coverage (16.78%) over baselines.