Autonomous Vehicles (AVs) require rigorous safety testing to ensure reliability in real-world environments, but traditional road testing is resource-intensive and fails to efficiently cover high-risk, safety-critical scenarios. In-depth crash data, which captures fine-grained details of pre-crash interactions, environmental conditions, and vehicle trajectories, has emerged as a foundational resource for addressing this gap. This study focuses on developing a unified framework to generate realistic, high-risk testing scenarios for AVs by integrating in-depth crash data with state-of-the-art generative models, aiming to enhance the comprehensiveness and efficiency of AV safety validation.
Xiaoyu Zhang (Thu,) studied this question.