TrafCopAgent: synergizing reinforcement learning and multi-agent collaboration for adaptive emergency traffic control

Rapid passage of emergency vehicles (EMVs), such as ambulances, is crucial for saving lives and minimizing property loss. However, existing traffic control methods, including mainstream reinforcement learning (RL) approaches, often rely on fixed reward strategies that struggle to handle the dynamic and high-stakes nature of emergency scenarios effectively, leading to suboptimal coordination between traffic lights, regular vehicles(REVs), and EMVs. In this work, we propose a Scenario-Aware Multi-Agent Emergency Traffic Control framework, which dynamically adapts its decision-making strategy based on the traffic context. In regular traffic scenarios, the system employs an optimized RL policy to maximize global traffic efficiency. Upon the detection of EMVs, the framework automatically switches to a LLM-based Multi-Agent Decision module, which leverages the advanced semantic understanding and reasoning capabilities of Large Language Models (LLMs) to handle complex situations that are beyond the generalization capability of traditional RL models. Extensive experiments on multiple public traffic scenarios show that our method provides clear advantages over existing methods. Specifically, our approach further reduces the travel time and waiting time of EMVs while concurrently decreasing the travel time of REVs, achieving a better balance between emergency response priority and overall traffic flow efficiency.