RL Improves CAV Platoon Joining Safety in Mixed Traffic.

Connected and Automated Vehicle (CAV) platooning holds significant promise for enhancing road safety and traffic flow. However, integrating CAV platoons into real-world mixed traffic, where human-driven vehicles exhibit unpredictable behaviors, presents substantial control challenges. This research explores the application of Reinforcement Learning (RL) to manage CAV platoon joining maneuvers in such complex environments. The study introduces a generic modeling and simulation framework, coupled with the SUMO simulator, to investigate and compare various Deep Reinforcement Learning (DRL) algorithms, including Deep Q-Network (DQN), Double Deep Q-Network (DDQN), and Proximal Policy Optimization (PPO). The objective is to achieve safe and efficient joining, either by penalizing risky behaviors within the learning process or by employing an external safety controller. Results indicate that PPO, particularly when incorporating risk-related penalties into its reward function, significantly outperforms DQN and DDQN, achieving a 98% joining success rate and less than 1% collision rate. This enhanced safety, however, often requires more decision steps to complete the maneuver, highlighting an inherent trade-off between safety, efficiency, and the speed of decision-making. The research underscores the critical need to balance these factors when designing RL-based controllers for autonomous systems in dynamic, mixed-traffic settings.