Treffer: Intelligent docking control of autonomous underwater vehicles using deep reinforcement learning and a digital twin system.

This study proposes an intelligent docking control system for an autonomous underwater vehicle (AUV), integrating deep reinforcement learning (DRL) with a digital twin (DT) simulation platform. A You Only Look Once (YOLO) object detection model enables real-time visual recognition of the docking station, and the Deep Deterministic Policy Gradient (DDPG) algorithm governs depth and trajectory control under nonlinear hydrodynamic conditions. A customized digital twin system was developed using Python, MATLAB, and Unity to generate simulated sensor and image data for training the DDPG and image-based DDPG (I-DDPG) controllers. Compared to traditional and supervised control strategies, the proposed actor-critic DRL approach offers superior adaptability and continuous control for visual-based underwater docking. Experimental validations in a towing tank demonstrate that the trained controllers successfully transferred from simulation to reality, maintaining robustness across static and dynamic docking scenarios. The results validate the feasibility of combining DRL and digital twins for reliable, precise control in unstructured underwater environments, demonstrating the system's potential in intelligent marine robotics applications. [ABSTRACT FROM AUTHOR]