A Vision-Based Recognition and Navigation Method for Autonomous Underwater Robot Grasping

The development of underwater robotic systems with autonomous grasping capabilities is challenging due to the complexity of the operation environment, limited sensing performance, and computation load. This paper proposes a fast recognition and navigation method for underwater robot grasping. In detail, we theoretically demonstrate the feasibility of underwater robots that can quickly identify and localize targets at a long distance through a monocular camera, thus navigating the robot to approach the target in real-time and then generate accurate grasping points at short-range through stereo-vision cameras to help the robot complete autonomous grasping. Experiment results show that our constructed system can recognize and determine the target's position within 0.03 seconds, then detect and generate an accurate grasping point within 1 second at a distance of 50 cm from the target with the error of 2~3 cm. Our research proves that underwater robots equipped with low-cost devices such as commercial stereo-vision cameras and edge computing devices can identify targets and generate accurate grasping points in real-time, promoting research in the field of autonomous underwater grasping.