Deep Reinforcement Learning for Overtaking Decision-Making and Planning of Autonomous Vehicles

The safe overtaking of autonomous vehicles has become an important focus in recent robotics and AI research. Considering the scenario of overtaking with oncoming vehicles, this paper proposes a hierarchical framework based on deep reinforcement learning to interact with the traffic environment and learn to overtake safely and efficiently. At the high level, an optimized actor-critic network with TRPO deep reinforcement learning algorithm is used to make safe overtaking decisions. At the low level, a reliable lane-changing path planning strategy is employed for motion control. A well-designed reward function is introduced to guide the agent in learning efficient overtaking behaviors. The effectiveness of the proposed framework is demonstrated through simulation experiments.