Multi-Agent Reinforcement Learning for Collaborative Decision-Making at Unsignalized T-Intersection

Connected autonomous vehicles (CAVs) come to provide new avenues for urban traffic management. Towards unsignalized T-intersection in mixed traffic, this paper proposes a centralized decision-making framework for CAVs with right-of-way guidance. We systematically model multi-vehicle driving at T-intersection through connected cooperation and design a parameter-sharing multi-agent reinforcement learning (MARL) collaborative decision algorithm. Additionally, a right-of-way cooperator (RWC) based on priority is also designed for the MARL algorithm to reduce conflicts among vehicles. The traffic simulation experiments in SUMO show that the proposed approach enhances traffic safety and efficiency.