Research on Collision Avoidance of Unmanned Surface Vehicle Based on Improved DDPG

This paper addresses the collision avoidance of the underactuated unmanned surface vehicles (USVs) in the environment of random static obstacles based on an improved deep deterministic policy gradient (DDPG) algorithm. An improved algorithm based priority experience replay (PER) is developed to control the USV. Furthermore, the design of a marine environment includes obstacles with random initial positions. Moreover, the USV is modeled with the Norrbin model and the reward function set complies with the model and requirements of collision avoidance. The advantages of the developed collision avoidance scheme are that first, it could effectively prevent the emergence of paths that do not satisfy the actual navigation requirements of the USV and the PER mechanism can improve the USV collision avoidance accuracy rate; second, the designed environment is highly random, which is crucial for solving the problem that the method lacks practical usefulness. In addition, the proposed reward function optimizes the USV's trajectory so that it satisfies the requirements for both collision avoidance and appropriate USV motion. These features make the algorithm more useful. Finally, simulation and comparative experiments verify the effectiveness of the algorithm.