In order to solve the problems of slow calculation speed and poor anti-noise ability when underwater robots perform underwater target recognition, an image recognition algorithm based on SIFT and FLANN was proposed and verified on a self-developed robotic arm platform. The improved algorithm was used to improve the accuracy of target recognition; the manipulator model was established based on the D-H method for kinematic analysis to achieve system optimization control. Multiple-group grasping experiments were performed by using the manipulator. The results show that by using the proposed method, not only the sea cucumber can be accurately identified, but also the position of the recognition frame can be automatically adjusted, and the success rate of the robotic arm grasp can be improved, which verifies the effectiveness of the proposed optimization algorithm.