Deep Reinforcement Learning for Autonomous Vehicles Collaboration at Unsignalized Intersections
Jian Xin Zheng, Kun Zhu, Ran Wang
Abstract
As conservative intersection management, signalized intersection has a significant bottleneck in improving traffic efficiency when it comes to connected autonomous vehicles (CAVs). In this paper, to make the intersection management more fine-grained, a decentralized conflict-free coordination scheme is tailed for CAVs at intersections without traffic signals. First, the problem of multiple vehicles navigation through an unsignaled intersection is formulated as a Partially Observable Stochastic Game (POSG). Second, we propose a cooperative multi-agent proximal optimization algorithm (CMAPPO) to make driving-decision for each CAV agent and achieve collaboration in a distributed manner. Finally, simulations are carried out on SUMO to evaluate the proposed method. The results show that the CMAPPO has significant effectiveness in solving the multi-vehicle coordination at intersections.