Neural Network-Based Control of an Undulating Robot for Underwater Pipeline Tracking in Unity 3D

This article presents a novel control system for an underwater robot with undulating propulsion, which is designed for accurate pipeline tracking in complex underwater environments. The system uses the YOLOv8 deep learning model for object detection, enabling the robot to identify pipelines and calculate orientation and cross-track errors using data from a Unity 3D simulation. These errors are processed through a MATLAB control system that integrates Active Disturbance Rejection Control (ADRC) and Fuzzy Logic for adaptive trajectory adjustments. A salient innovation is the implementation of CycleGAN for domain adaptation, thereby bridging the gap between simulated and real-world conditions. The YOLOv8 model demonstrated remarkable performance with a mAP50 score of 0.995 and 0.995 precision. Experimental findings substantiate the system's capacity to ensure pipeline alignment, even through sharp bends. The integration of control strategies and computer vision underscores the system's potential for autonomous underwater pipeline inspection.