Integrated Target Recognition and Grasp Control for the Underwater Vehicle-Manipulator System

Underwater vehicle-manipulator system (UVMS) is regarded as a promising tool for humans to autonomously grasp target in marine industry. However, the complex ocean environment and the strong coupling of the UVMS make it challenging to realize the aforementioned goal. This article is concerned with an integrated target recognition and grasp control issue for the UVMS. Particularly, the binocular cameras are employed to train the YOLOv4-tiny network, through which an adaptive multicriterion selection (AMCS) strategy is developed to realize online target recognition. Along with this, we adopt the integral reinforcement learning to codesign the reaching and grasping controllers, such that the UVMS can simultaneously pursuit and grasp the target in finite time. Note that the following advantages of our solution are highlighted: the AMCS-based recognition strategy can avoid the repeated training of the target classifier over the traditional kernel correlation filters; and the codesign mode can improve the grasp efficiency and decrease the execution time as compared with the separate design. Finally, simulation and field tests are both presented to verify the effectiveness of our solution.