From Particles to Perils: SVGD-Based Hazardous Scenario Generation for Autonomous Driving Systems Testing

2026-04-20

AI Summary

This work addresses the challenge that existing search-based seed generation methods struggle to discover diverse and realistic failure cases for autonomous driving in high-dimensional traffic scenarios, often converging to a limited set of failure modes. To overcome this limitation, the study introduces Stein Variational Gradient Descent (SVGD) into test scenario generation for the first time, integrating adaptive stochastic seed generation with a reinforcement learning-based tester. By leveraging the attraction–repulsion dynamics among particles in high-risk regions, the approach jointly optimizes risk awareness and scenario diversity. The method seamlessly integrates into existing online testing frameworks. Evaluations on the CARLA platform across Apollo, Autoware, and end-to-end driving systems demonstrate up to a 27.68% increase in safety violation rates, a 9.6% improvement in scenario diversity, and a 16.78% gain in map coverage.

Technology Category

• Search and Optimization: Sampling/Simulation-based Search
• Planning, Routing, and Scheduling: Replanning and Plan Repair
• Reasoning under Uncertainty: Stochastic Optimization

Application Category

• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines
• Security and Privacy: Large-scale security measurements
• Web Mining and Content Analysis: Web data generation and simulation

Abstract

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.

Problem

• autonomous driving systems testing
• hazardous scenario generation
• high-dimensional search
• failure diversity
• simulation-based testing

Innovation

• SVGD
• hazardous scenario generation
• autonomous driving testing
• failure mode diversity
• adaptive seeding