Ensuring Safety in Cooperative CAV Lane Change: A Multi-Agent Reinforcement Learning Strategy with Control Barrier Function

Safety is one of the critical considerations for connected autonomous vehicles (CAVs) during lane-changing maneuvers. This study analyzes safety constraints inherent in the lane-changing process. Firstly, we propose a method to calculate safe action spaces given deterministic states for cooperative lane-changing scenarios. Then, the safety constraints are seamlessly integrated into a multi-agent reinforcement learning (MARL) framework. The control barrier function (CBF) enforces safety constraints to mitigate collision risks resulting from abrupt acceleration/deceleration during vehicular maneuvers. Based on this, a control barrier function-enhanced multi-agent actor-critic algorithm (CBF-MAAC) is developed, adding an action filter that maps the output of the actor network to the dynamic safe action ranges. Experimental results show that the proposed method performs better in terms of convergence speed compared to the method without action filters in a typical cut-through scenario consisting of five vehicles. Moreover, it ensures zero-collision training throughout the learning process.