A review on reinforcement learning-based highway autonomous vehicle control

摘要

Autonomous driving is an active area of research in artificial intelligence and robotics. Recent advances in deep reinforcement learning (DRL) show promise for training autonomous vehicles to handle complex real-world driving tasks. This paper reviews recent advancement on the application of DRL to highway lane change, ramp merge, and platoon coordination. In particular, similarities, differences, limitations, and best practices regarding the DRL formulations, DRL training algorithms, simulations, and metrics are reviewed and discussed. The paper starts by reviewing different traffic scenarios that are discussed by the literature, followed by a thorough review on the DRL technology such as the state representation methods that capture interactive dynamics critical for safe and efficient merging and the reward formulations that manage key metrics like safety, efficiency, comfort, and adaptability. Insights from this review can guide future research toward realizing the potential of DRL for automated driving in complex traffic under uncertainty.