A Centralized Intelligent Transportation System for Cooperative Vehicles and Dynamic Resource Allocation Based on Multi-Agent Transformer

Abstract

Autonomous vehicles (AVs) are expected to improve traffic efficiency and safety through automated perception, decision-making, and control. When equipped with vehicle-to-everything (V2X) communication capabilities, AVs further evolve into connected and automated vehicles (CAVs), enabling cooperative decision-making and system-level traffic optimization. Despite this potential, coordinating large numbers of CAVs in complex and dynamically changing traffic environments remains a major challenge.

This thesis proposes a centralized intelligent transportation system for cooperative control of CAVs. The centralized command centers coordinate multiple CAVs by dynamically allocating road resources and sending maneuver-level commands in real-time. The proposed framework is built based on a Multi-Agent Transformer (MAT). To support flexible traffic management, a customized map without predefined lane directions or traffic signals is designed using RoadRunner, allowing lane usage and intersections to be dynamically reconfigured based on real-time traffic conditions.

To enhance safety and learning stability in dense traffic scenarios, a collision risk evaluation model is introduced to provide dense safety-related feedback during training, effectively reducing collision frequency among CAVs. The proposed system is validated in the CARLA simulator under variable traffic densities.

Experimental results show that traffic efficiency gradually decreases with the increasing of vehicle' density. The learned policies exhibit emergent cooperative driving maneuvers among CAVs, including platooning, early braking, anticipatory acceleration, and adaptive lane reallocation, which collectively smooth traffic flow and mitigate collision risks.

Overall, this thesis proves that the centralized coordination combined with MAT provides an effective solution for managing cooperative maneuvers of CAVs, offering valuable insights for the design of future ITS.