Cooperative Driving at Multiple Unsignalized Intersections in Fully Autonomous Driving Scenarios

The decision-making process for connected and autonomous vehicles (CAVs) at unsignalized intersections is a critical and challenging problem. Previous methods predominantly concentrate on optimizing passage strategies for individual intersections in isolation. However, they often neglect global traffic conditions and task priorities in closed, multi-intersection transportation scenarios, leading to localized congestion. In this work, we propose a method that aims to optimize the passing order of intersections from a global and long-term perspective to enhance overall transportation efficiency. Specifically, we model the coordination of multiple unsignalized intersections as a multi-agent sequential decision problem and solve it through a two-stage method. In the planning stage, we construct fully connected undirected graphs based on vehicle conflict relationships and use the multi-agent proximal policy optimization (MAPPO) algorithm to learn the global priorities. In the scheduling stage, the local vehicle scheduling is formalized as a multi-objective optimization problem. The learned global priorities are soft constraints, while a hybrid filtered beam search determines safe and efficient CAV passing orders. Extensive offline experiments and online tests on real-world and synthetic datasets demonstrate that our proposed method outperforms state-of-the-art approaches in minimizing congestion and enhancing transportation efficiency.