This study develops a YOLO (You Only Look Once)-based 3D perception algorithm for UVMS (Underwater Vehicle-Manipulator Systems) for precise object detection and localization, crucial for enhanced grasping tasks. The object detection algorithm, YOLOv5s-CS, integrates an enhanced YOLOv5s model with C3SE attention and SPPFCSPC feature fusion, optimized for precise detection and two-dimensional localization in underwater environments with sparse features. Distance measurement is further improved by refining the SGBM (Semi-Global Block Matching) algorithm with Census transform and subpixel interpolation. Ablation studies highlight the YOLOv5s-CS model’s enhanced performance, with a 3.5% increase in mAP and a 6.4% rise in F1 score over the base YOLOv5s, and a 2.1% mAP improvement with 15% faster execution than YOLOv8s. Implemented on a UVMS, the algorithm successfully conducted pool grasping experiments, proving its applicability for autonomous underwater robotics.
Underwater robots conducting inspections require autonomous obstacle avoidance capabilities to ensure safe operations. Training methods based on reinforcement learning (RL) can effectively develop autonomous obstacle avoidance strategies for underwater robots; however, training in real environments carries significant risks and can easily result in robot damage. This paper proposes a Sim-to-Real pipeline for RL-based training of autonomous obstacle avoidance in underwater robots, addressing the challenges associated with training and deploying RL methods for obstacle avoidance in this context. We establish a simulation model and environment for underwater robot training based on the mathematical model of the robot, comprehensively reducing the gap between simulation and reality in terms of system inputs, modeling, and outputs. Experimental results demonstrate that our high-fidelity simulation system effectively facilitates the training of autonomous obstacle avoidance algorithms, achieving a 94% success rate in obstacle avoidance and collision-free operation exceeding 5000 steps in virtual environments. Directly transferring the trained strategy to a real robot successfully performed obstacle avoidance experiments in a pool, validating the effectiveness of our method for autonomous strategy training and sim-to-real transfer in underwater robots.
Accurate identification of hydrodynamic parameters is essential for high-fidelity modeling and control of unmanned underwater vehicles (UUVs). Compared with towing tank experiments and computational fluid dynamics simulations, system identification based on free-running trial data offers a cost-effective and scalable alternative. However, in real ocean environments, unmodeled lumped disturbances—such as shear currents, stratification-induced buoyancy variations, and wave-induced drift forces—strongly couple with the vehicle’s intrinsic dynamics. Conventional least-squares estimators and physics-informed neural networks tend to absorb environmental effects into the physical parameters, leading to physically inconsistent estimates. To address this challenge, this paper proposes a physics-guided hybrid network (PG-HyNet) with input-domain structural decoupling. The architecture explicitly separates the intrinsic rigid-body dynamics from spatially varying environmental disturbances by assigning dynamics-related states to a physics-constrained branch and position-dependent variables to a residual disturbance branch. A staged training strategy is introduced to stabilize identification and suppress parameter drift during optimization. The framework is validated using high-fidelity simulations incorporating shear currents, density stratification, and wave drift effects, as well as real-world lake trial data. The results demonstrate that PG-HyNet significantly improves robustness against disturbance-induced parameter compensation, enabling physically consistent hydrodynamic parameter recovery while accurately capturing spatially varying environmental disturbance effects.