DiffSCTG: Diffusion-based Safety-Critical Trajectory Generation for Autonomous Vehicles

The increasing reliance on autonomous vehicle (AV) systems underscores the importance of evaluating their safety and robustness, particularly in safety-critical driving scenarios. However, the scarcity of rare event data and the limitations of manually generated scenarios constrain the effectiveness of AV evaluations. In this work, we introduce a Diffusion-based Safety-Critical Trajectory Generation (DiffSCTG) framework that leverages the generative capabilities of diffusion models to produce diverse and realistic safety-critical scenarios. DiffSCTG adopts a transformer architecture for diffusion-based probabilistic modeling, which captures complex multi-agent interactions and temporal dependencies in trajectory data. To ensure structured and meaningful trajectory initialization, we propose the truncated diffusion process that begins from an anchored Gaussian distribution derived from slightly disturbed realistic driving scenarios. By progressively generating trajectories from structured noise through a reverse denoising process, DiffSCTG enriches original datasets used to train the AV algorithm. Extensive evaluations in simulation environments like CARLA Town 10 demonstrate its ability to improve AV robustness with significant reductions in collision rates and improved safety metrics, showing its utility for challenging traffic conditions.