Interactive scenario generation based on Generative Adversarial Imitation Learning integrated with trajectory...

Autonomous driving systems, being safety-critical applications, require rigorous testing throughout their entire development lifecycle. Scenario-based testing has become the dominant method due to its efficiency and safety advantages, with recent advancements focusing on enhancing interactivity to improve test authenticity. In this paper, we propose an interactive scenario generation method that incorporates a trajectory matching (TM) module with Generative Adversarial Imitation Learning (GAIL), thus reducing the necessary training time for GAIL and improving the efficiency of generating dynamic interactive scenarios. Firstly, the driving intention recognition component, based on a self-attention mechanism, is proposed to separately convert the driving intentions of the agent (background vehicle) and the expert (human driver traces) into sequences containing intention information. Secondly, the similarity analysis component based on Dynamic Time Warping (DTW) was proposed to analyze the similarity between driving intention sequences. By combining these two components, we propose a trajectory matching module that analyzes the agent's driving intentions and selects the most behaviorally relevant human driving traces from the Naturalistic Driving Data (NDD) to serve as expert demonstrations during the training process. In the GAIL baseline considered in this study, expert demonstrations are selected from the available human driving trajectories without explicitly considering the behavioral intention of the current agent trajectory. In contrast, the proposed TM-GAIL introduces intention-conditioned trajectory matching to preferentially select human driving traces with behavioral intentions similar to those of the current agent trajectory. This approach achieves a reduction in training time by providing behaviorally relevant expert demonstrations for discriminator training. Validation in a cut-in scenario demonstrates that background vehicles trained with this approach achieve human-level interaction capabilities within fewer training iterations, significantly enhancing testing realism.