Bridging scenario generation and regulatory standards: A physically plausible generative AI framework for automated vehicle safety certification

The transition from mileage-based to scenario-based validation is emerging as a critical requirement for the regulatory certification of automated driving systems. However, establishing a standardized testing framework is hindered by the scarcity of safety-critical scenarios in naturalistic data and the tendency of existing methods to sacrifice physical plausibility for high collision rates. To address this gap, this study proposes TraSafeDiff, a conditional diffusion-based framework designed to generate safety-critical traffic scenarios that are both adversarially challenging and physically consistent. The framework integrates multimodal scene information through a deep encoder, while differentiable kinematic, boundary, and angular-motion energy penalties guide generated trajectories toward greater physical plausibility. In an open-loop evaluation on nuScenes, TraSafeDiff achieves a higher collision rate and a lower off-road rate than the evaluated scenario-generation baselines, while its acceleration and jerk distributions show closer alignment with naturalistic data. By balancing adversarial severity with kinematic plausibility, this study provides a candidate-scenario generation approach that may complement broader evidence-based safety evaluation and regulatory-assurance workflows.