Safe Reinforcement Learning for Connected and Automated Vehicle Platooning

Connected and automated vehicle (CAV) platooning is emerging as a promising transportation technology, attributed to its potential for improving traffic efficiency and fuel economy. Meanwhile, this trend has led to increased demands and challenges for the decision-making of CAV platoons. In response, deep reinforcement learning (RL) methods have been recently adopted for CAV platooning, owing to their exceptional capacity to tackle control issues in complex and large-scale autonomous systems. However, the safety concerns associated with RL significantly impede its applicability in safety-critical systems. In light of this challenge, this study develops a novel safe RL framework for CAV platooning. The approach commences by incorporating the actor-critic RL framework to synthesize efficient control policies for CAV platooning. Specifically, a Lyapunov-based critic network is proposed in the actor-critic RL framework, ensuring the driving safety and stability of the CAV platoon. Finally, the effectiveness and computing efficiency of the proposed methodology are substantiated through a series of comprehensive simulation experiments.