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

Area of Science:

• Artificial Intelligence
• Traffic Simulation
• Autonomous Driving Systems

Background:

• Accurate microscopic traffic simulation is vital for autonomous driving system development.
• Reproducing flexible traffic participant behavior, especially bicycles interacting with vehicles, is challenging.
• Existing models struggle with complex multi-agent interactions in urban environments.

Purpose of the Study:

• To develop a novel data-driven simulation model for realistic traffic behavior reproduction.
• To enhance the reliability of simulation platforms for autonomous vehicle testing.
• To improve the capture of complex multi-agent interactions and interference scenarios.

Main Methods:

• Integrated a Transformer-based neural network with imitation learning and a Markov Decision Process (MDP).
• Employed a Transformer-based multi-agent policy for joint control of road users.
• Utilized a similarity reward function for trajectory and behavioral feature capture, with MDP-based training for long-term consistency.

Main Results:

• Achieved a Mean Distance Error (MDE) of 2.123 m over 9.6-second simulations, closely matching real behavioral distributions.
• Demonstrated high effectiveness in reproducing interference scenes with an F-1 score of 0.865.
• Showcased superior computational efficiency (3.8-5.4x faster than agent-centric models) with 5.98 ms inference time for 20 agents.

Conclusions:

• The proposed MDP-based Transformer model significantly outperforms non-MDP alternatives in accuracy and realism.
• The scene-centric Transformer policy offers enhanced computational efficiency, meeting real-time processing needs.
• The model's adaptability across diverse urban road segments validates its potential for improving autonomous vehicle testing platforms.