DRL-based path planning and obstacle avoidance of autonomous underwater vehicle

Abstract

Both path planning and obstacle avoidance are important for the navigation safety of autonomous underwater vehicles (AUVs) in unknown environments. In this paper, in order to adjust to the complexity and flexibility of underwater environments, path planning and obstacle avoidance algorithms based on value iterative network (VIN) and deep deterministic policy gradient (DDPG) respectively are proposed to navigate the AUV through an unknown complex area. With a simulation multi-beam sonar equipped to detect obstacles of subsea surroundings, a grid map is constructed online as inputs of VIN and DDPG. Taking advantage of generalization of deep reinforcement learning, methods studied in this paper have demonstrated validity in simulation experiments implemented in Unity3D where dynamic and static obstacles are randomly placed and experiments are conducted.