Reproduce “Diffusion Models for Safety Validation of Autonomous Driving Systems” on OpenResearch

About this paper

Juanran Wang, Marc R. Schlichting, Harrison Delecki, Mykel J. Kochenderfer

Researchers at Stanford University developed a framework utilizing conditional denoising diffusion models with a multi-stage self-improvement training algorithm to efficiently generate diverse and realistic failure scenarios for autonomous driving systems. This method overcomes reliance on external failure data or prior system knowledge, achieving high failure rates and broader coverage of failure modes compared to the Cross-Entropy Method.

Key insights

• Diffusion models, when combined with a self-improvement training strategy, can effectively learn and generate samples from extremely rare and complex failure distributions without explicit supervision from failure data.
• The multi-stage refinement process, by progressively lowering robustness thresholds and retraining on accumulated elite samples, enables efficient discovery and accurate modeling of diverse failure modes.
• The ability to condition failure generation on initial traffic states is crucial for practical applicability, allowing targeted stress testing of autonomous driving systems in varied real-world contexts.

Summary by alphaXiv.