Attitude Control Based Autonomous Underwater Vehicle Multi-mission Motion Control with Deep Reinforcement Learning

Abstract

Due to the limitations of classical control strategies for underwater vehicles, the artificial intelligent control technologies have attracted more attention than ever, especially the Deep Reinforcement Learning. However, it is usually confusing to choose a proper Deep Reinforcement Learning algorithm and to configurate an effective reward function according to different autonomous underwater vehicle assignments. To solve the problem, this research explores three different missions that autonomous underwater vehicle might perform, with two Deep Reinforcement Learning algorithms and three reward functions adopted respectively. Deep Reinforcement Learning controllers take the attitude sensor information as input, the control signals of X-rudder blades as output. The simulation experiments results are compared and analyzed, which has an important reference value for the reward function setting and Deep Reinforcement Learning algorithm selection for different autonomous underwater vehicle control tasks.