Aiming at the requirements of recognition accuracy and grabbing accuracy of target object for robot based on visual assistance, a vision grabbing system based on deep learning target detection method is proposed. Firstly, based on the improved depth learning-based target detection algorithm, the recognition of the captured object is realized. Secondly, according to the method of Zhang Zheng-you's camera calibration, the end gripper flexible jaw and the depth camera are used for hand-eye calibration. Finally, the gripping experiment of the object is completed on the pneumatic bionic arm experimental platform. Experimental results show that compared with the traditional visual grabbing system, the grabbing system based on deep learning target detection can not only improve the recognition accuracy of the robot on the captured object, but also can adapt to the capture of multiple categories of objects in various complicated environments, and has achieved satisfactory results.