Abstract Simulation-based testing can effectively reproduce vehicle behaviors and states in real-world traffic scenarios, thereby enabling more accurate evaluation of autonomous driving systems in terms of decision-making, planning, and control. This paper proposes a Triple Generative Adversarial Imitation Learning-Gated Recurrent Unit (TripleGAIL-GRU)-based trajectory generation method for autonomous driving simulation scenario construction, focusing on three driving behaviors: lane keeping, left lane changing, and right lane changing. Vehicle trajectories and multidimensional features are extracted from the HighD dataset and organized into state-label-action triplets to train a TripleGAIL-GRU model for learning human-like driving policies. The learned policy is then deployed to control environmental vehicles in a static simulation environment, while the autonomous vehicle is controlled by its own driving model, enabling the construction of diverse traffic scenarios. In 1,000 validation runs under identical traffic conditions, the TripleGAIL-GRU model produced only 5 collisions, compared with 22 collisions for the TripleGAIL model, indicating substantially improved trajectory stability and behavioral rationality. In terms of imitation quality, the human-likeness metric decreases from 0.703 for TripleGAIL to 0.557 for TripleGAIL-GRU, indicating closer agreement with expert trajectories. Moreover, in the recurrent-structure ablation study, TripleGAIL-GRU reduces positional RMSE by approximately 32.8% relative to the non-recurrent TripleGAIL baseline. These results demonstrate that the proposed method improves long-horizon trajectory stability and supports the construction of more realistic and risk-revealing simulation scenarios.
Chai et al. (Fri,) studied this question.