Transformer-based multi-agent traffic simulation for autonomous vehicle testing in shared urban road segments.

Microscopic traffic simulation plays a crucial role in the development and testing of autonomous driving systems. However, accurately reproducing traffic participant behavior in shared urban road segments remains challenging due to their flexible movement characteristics, particularly for bicycles that potentially interfere with motorized vehicles. This study presents a novel data-driven simulation model that integrates a Transformer-based neural network architecture, trained through imitation learning, with a Markov Decision Process (MDP) formulation. By leveraging a Transformer-based multi-agent policy, the model jointly controls the behaviors of road users, effectively capturing complex multi-agent interactions. The proposed similarity reward function enables comprehensive capture of both trajectory and behavioral features. The MDP-based training ensures consistent and realistic traffic behaviors over long-term simulations. Validation experiments demonstrate the model's effectiveness with a Mean Distance Error (MDE) of 2.123 m over 9.6-second simulations, closely matching real behavioral distributions and achieving an F-1 score of 0.865 for interference scene reproduction. Our proposed scene-centric Transformer policy demonstrates superior computational efficiency, operating 3.8 to 5.4 times faster than agent-centric models, with an inference time of 5.98 ms for 20 agents, meeting real-time processing requirements. Comparative analysis reveals that our MDP-based Transformer model significantly outperforms non-MDP alternatives, reducing MDE by 15.8-45.6% and improving interference behavior reproduction F-1 scores by 10.2-18.7%. Furthermore, validation across diverse road segments demonstrates the model's adaptability to varied urban road environments. This approach enhances the realism of microscopic traffic simulation, improving the reliability of simulation platforms for autonomous vehicle testing.