Learning to Control of an Under-actuated Autonomous Surface Vehicle Based on Model-based Deep Reinforcement Learning

Control of autonomous surface vehicles is a challenging task due to their nonlinearities, imposed disturbances, strong couplings, under-actuation, and various constraints. Prevailing methods are basing on an explicit mathematical model. The paper presents a learning-based motion controller for an autonomous surface vehicle without any mathematical models. A model-based deep reinforcement learning approach is propose for achieving the trajectory following task. At first, a deep neural network is trained for approximating the dynamical model of the autonomous surface vehicle by using recorded input and output data only. Then, a model predictive controller based on the learned neural network together with a reward function is presented for the autonomous surface vehicle to follow arbitrary trajectories. It is shown that after learning with random data collected from the autonomous surface vehicles, the proposed learning-based controller is able to follow trajectories with excellent sample efficiency. Simulation results are given to illustrate the proposed model-based deep reinforcement learning method for trajectory following of an autonomous surface vehicle.