RLMP: A Sim-to-Real Reinforcement Learning Manoeuvre Planner Validated on Public Highway

Overview & Synopsis

Decision-making for automated vehicles in dynamic highway environments remains a significant challenge, as traditional rule-based planners often fail to generalize across complex traffic scenarios. To address this, this paper presents a Reinforcement Learning Manoeuvre Planner (RLMP) that employs a Deep Reinforcement Learning (DRL) policy for high-level lateral decision-making. A key contribution of our work is a multi-stage methodology designed to bridge the critical simulation-to-reality gap. We first demonstrate the feasibility of our approach by training an agent in the RL-based simulator MetaDrive. The policy is then validated in a high-fidelity digital twin, using CarMaker-RTMaps co-simulation framework, which serves as an essential intermediate step. Finally, we deploy the trained agent in a real-world vehicle, confirming the effectiveness of our DRL-based planner and highlighting the challenges of physical implementation. Our results confirm that a DRL agent can learn robust and generalizable lane-changing policies, validating our framework as a viable pathway for developing and deploying learning-based autonomous driving systems.