Diffusion-Based Corner Scenario Generation Method for Autonomous Driving with a Dynamics-Based Decoder

Ensuring the safety of autonomous driving systems requires the generation of high-risk traffic scenarios. However, existing methods often lack physical plausibility, diversity, and controllability. This paper proposes a logical corner scenario generation framework integrating rule extraction, a risk-aware conditional diffusion model, and a dynamics-based decoder. Cut-in scenarios with annotated risk levels are extracted from naturalistic data using rules based on minimum time to collision (TTC) and vehicle states. A diffusion model generates control sequences under varying risk levels, which are then decoded into smooth, physically plausible trajectories via a bicycle model. A comprehensive evaluation metric system covering physical plausibility, diversity, and risk is developed. Experiments on the NGSIM dataset demonstrate that the proposed method outperforms rule-based, control-perturbation, and non-decoder diffusion baselines. The method preserves physical plausibility while simultaneously enhancing diversity and the coverage of high-risk behaviors, thereby providing an effective tool for corner scenarios testing and robustness validation in autonomous driving systems.