Learning an AUV docking maneuver with a convolutional neural network

Abstract This paper proposes the use of a convolutional neural network (CNN) to guide and control an autonomous underwater vehicle into the entrance of a docking station by mapping camera input to an error signal. An external positioning system synchronized with the internal sensors of the vehicle is to obtain a dataset of images matched with the position and heading of the vehicle. By using a guidance map, each position is converted into a desired heading that guides the vehicle into the docking station. The CNN is then trained to estimate, for each frame, the error between the desired vehicle heading and the actual heading of the vehicle. After the training period, the camera input and the CNN are used to control the vehicle towards the docking station, achieving autonomous docking. To enhance the stability of the docking, the paper proposes a transformation of the polar coordinates that avoids large angular errors when reaching the entrance of the docking station. The proposed framework is implemented, and experimental data from a mission are presented.