Learning to Generate Critical Trajectory for Autonomous Vehicle Based on Natural Driving Data

Trajectory prediction is an important aspect of path planning and navigation for self-driving vehicles; however, little research has been conducted to assess the robustness of self-driving vehicles under extreme or hazardous traffic conditions by perturbing the prediction module. In this paper, we introduce an innovative criticality scenario generation framework that aims to influence the behavior of self-driving vehicles by generating critical trajectories of the vehicles in the scenarios to maximize the perturbation of the prediction effect of the self-driving vehicles. First, the vehicle state information in the self-driving datasets is clustered and analyzed to extract traffic scenarios with different risk levels. Then, a transformer-based trajectory generation model is used to learn the interaction patterns between vehicles in real critical scenarios, so that critical trajectories can be dynamically generated during testing to enrich the scenarios for self-driving tests. Our experiments on different datasets and models show that the framework is able to assess the risk level of different traffic scenarios and to generate effective and realistic critical trajectories.