Position Estimation for Autonomous Underwater Vehicles (AUV) Docking Based on Reinforcement Learning

Autonomous underwater vehicle (AUV) docking technology is currently a hot topic in underwater robot research, and precise and reliable navigation is a key technology for AUV docking. Position problem is key for the AUV docking, which is solved by fusing information from different sensors, such as the nertial navigation (INS) and the ultra-short baseline positioning system (USBL). The Kalman filter is a conventional fusion method, but it suffers the difficulty in deciding a nice Kalman filter gain. In this paper, a reinforcement learning-based extended Kalman filter compensation method is proposed, which learns the filter gain from historic data including sensor measurements and estimations. Simulation shows that the reinforcement learning method can effectively compensate for the filter gain and accurately estimate the position which verifies the feasibility and effectiveness of the proposed method.