A vision-based docking planning network for robust recovery of autonomous underwater vehicles

Autonomous Underwater Vehicles (AUVs) are core intelligent platforms in fields such as marine resource exploration and offshore engineering operations. Stable underwater autonomous recovery technology is crucial for ensuring the continuous safety and efficient operation of AUVs. During close-range docking, AUVs rely heavily on optical beacons for localization. However, the loss or occlusion of partial optical beacons can significantly degrade localization accuracy or even cause localization failure, resulting in erroneous docking guidance trajectories. To address this issue, this paper proposes an end-to-end vision-based docking planning network (VDockNet), which is developed based on imitation learning. The proposed network directly maps docking images to guidance paths, maintaining accurate path planning even when optical beacons are incomplete in the camera field of view. In addition, to address the problem of robust recovery, a hybrid docking strategy is introduced, which incorporates network uncertainty and sensor noise into consideration. The proposed network demonstrates robustness and accuracy by generating precise docking paths despite water current interference. VDockNet achieved a 96.34% docking success rate in simulations with current perturbations and an 82.68% planning success rate in zero-shot evaluations on offline real-world underwater images. Nevertheless, safety constraints for actual deployment on embedded platforms require further validation.