Self-supervised Dock Pose Estimator for Unmanned Surface Vehicles Autonomous Docking

Unmanned Surface Vehicles (USVs) are increasingly applied to marine operations such as environmental monitoring. It faces a notable challenge in achieving precise autonomous docking at ports and stations still depending on remote human control for accuracy and safety. To unlock the full potential of fully unmanned maritime deployment, it is pivotal to achieve visual servoing to re-dock the USV. This study introduces a novel monocular camera-based self-supervised learning pipeline for autonomous docking. Through careful label design, the Self-supervised Dock Pose Estimator (SDPE), achieves the data collection and neural network training processes by eliminating the need for conventional manual labeling, hand-crafted feature engineering, and camera calibration. The SDPE can accurately predict the dock pose, facilitating the implementation of position-based visual servoing (PBVS) for efficient and autonomous docking. The effectiveness of our proposed solution is tested and validated in a Virtual RobotX (VRX) simulation environment, reflecting its capability to handle the autonomous docking task. Experimental results show the precision of SDPE and the overall feasibility of our comprehensive framework in autonomous docking scenarios. Experiment videos are available at: https://youtu.be/QEPLOOC1ce0.