Dynamic Target Following Control for Autonomous Vehicles with Deep Reinforcement Learning

摘要

Target following control is an operating condition for autonomous driving in vehicle platooning and service robots. But the model-based target following control is challenging due to the nonlinearity and uncertainty of the vehicle dynamics model. To address this problem, this paper proposes a dynamic target following control method based on deep reinforcement learning for autonomous vehicles. Its primary purpose is to make the car follow dynamic targets smoothly and reasonably. Only collected state data sets are used to learn control policy. The policy learning consists of two steps. Firstly, we describe a specific following situation in which we design a dense reward function for realizing following control. We use empirical data to pre-train with deep neural networks. Then the deep deterministic policy gradient (DDPG) algorithm is used to train the control policy to achieve continuous optimization. We conducted experiments with Carla simulation software, including the comparison with the supervised learning-based approach. The results show the effectiveness of our approach.