Prior knowledge-assisted reinforcement learning for vehicle platoon control in cut-in scenarios

In mixed traffic environments, vehicle platoons are inevitably exposed to cut-in maneuvers performed by human-driven vehicles (HDVs), which introduce substantial uncertainty and degrade platoon control performance. To address this challenge, this study proposed a prior knowledge-assisted reinforcement learning (PKARL) control framework for vehicle platoons. A distributed model predictive control (DMPC)-based optimization controller was first developed to provide structured guidance for deep reinforcement learning (DRL) policy learning. On this basis, the state space, action space, reward function, and training scenarios were systematically designed to enable effective responses to HDV cut-in disturbances. Furthermore, a safety filter was incorporated to enforce stability and safety constraints on the control actions generated by the DRL policy. By integrating prior control knowledge, the proposed framework improved training efficiency and enhanced safety compared with conventional DRL methods. Driver-in-the-loop experimental results showed that, relative to the DMPC approach, the proposed PKARL method reduced the cumulative speed error by 23.0% and the cumulative spacing error by 20.1%. In addition, compared with the DRL method, PKARL achieved further reductions of 10.5% in cumulative speed error and 1.1% in cumulative spacing error. These results demonstrate the effectiveness of the proposed method in mitigating HDV cut-in disturbances and improving platoon tracking performance.