A Multi-agent Reinforcement Learning Approach for Seamless and Safe Autonomous Vehicle Platooning

Autonomous driving has been transforming transportation by improving safety, traffic flow efficiency, and fuel economy. Most importantly, autonomous driving has reshaped the transportation landscape by enhancing safety, traffic flow, and fuel savings. In particular, vehicle platooning enables a group of vehicles to travel in a coordinated formation with reduced aerodynamic drag and minimized fluctuation in velocity, lowering the risks of accidents. However, most previous works, such as model predictive control and independent reinforcement learning, usually cannot adapt in real time and optimize safety and efficiency simultaneously. This work is devoted to the development of a scalable and adaptive model for cooperative platoon control. We propose a MARL model based on MADDPG, implemented on the Unity-based simulator, where agents are trained in a centralized way while being deployed in a decentralized manner. Experimental results show that our model can reduce fuel consumption by 15.8%, enhance the stability of velocity, and reduce the collision rate by 2.1%, outperforming traditional methods. It manifests the potential of multi-agent cooperation in practice for the intelligent transportation system and lays the foundation for decentralized, efficient, and safe autonomous vehicle platoons in dynamic real traffic scenarios.