A Deep Reinforcement Learning Based Approach for Autonomous Overtaking

Autonomous driving is concerned to be one of the key issues of the Internet of Things (IoT). In this paper, we propose a deep reinforcement learning scheme, based on deep deterministic policy gradient, to train the overtaking actions for autonomous vehicles. In contrast to conventional autonomous driving systems that require expensive LiDAR or visual cameras, our method uses low-cost sensors like ultra-wideband antenna arrays and inertial measurement units to obtain easily available measurements such as distance, angle, and speed information. Our method directly projects raw sensory measurements into continuous control signals for overtaking maneuvers. Simulation results demonstrate that our method outperforms two other state-of-the-art algorithms. Besides, the technique can be generalized to other areas of autonomous driving such as merging, platooning and formation changing, by modifying the parameters and conditions of the reward function under the same framework.