Autonomous motion control of USVs, especially in complex marine conditions, is always a fundamental problem. Conventional methods consider the USV hydrodynamic model and the influences of environmental disturbance separately. However, due to the randomness of wind, wave and current, the accumulative error of each model can be large. To address this issue, this paper presents an end-to-end USV tracking control method via deep reinforcement learning, where a modern Reinforcement learning algorithm Actor-Critic is adopted. Given no prior knowledge of the dynamical system, the proposed method takes as input the information of environment (e.g., speed of wind and flow, etc.), ship and target trajectory, then produces the ship control signal (i.e., rudder angle and forward momentum) directly. We further propose a customized reward function to appraise the performance of ship agent. The presented simulation results demonstrate that this novel algorithm performs well in tracking tasks under complex marine conditions which is designed to change constantly.
展开▼
