Risk-aware diffusion model for generating extreme weather scenarios and safety evaluation in autonomous driving

Extreme weather scenarios, are likely to cause safety hazards such as perception failures and decision-making anomalies. These scenarios are indispensable for the safety assessment and robustness testing of autonomous driving systems. Most existing scenario generation methods focus on manually defined rules or parameter perturbations, which makes it difficult to simultaneously ensure risk, physical plausibility, and diversity. In this study, a risk-guided diffusion model (RADM) is proposed to automatically generate high-risk and challenging weather test samples. Through multiple rounds of feedback, the generated distribution is optimized so that the samples increase collision risk while maintaining diversity and physical reasonableness. The SafeBench platform is used to validate the effectiveness of RADM in improving collision rates, expanding sample distribution coverage, and ensuring scenario usability. Experimental results show that RADM can efficiently and automatically generate extreme weather test samples for a wide range of autonomous driving testing systems, demonstrating strong potential for real-world applications.