In order to improve the accuracy of the robotic arm grabbing up objects, a 3D visual identification and grabbing system was proposed based on deep learning. In the visual system, GPU and deep image Open CV function library were combined to perform image pickup, depth data calculation, coordinate transformation, image contour search and convolutional neural network model training respectively.The YOLOv2 algorithm was used to discriminate the type and center point of the target object, and the contour search method was used to find out the angle information of the object as the robot arm operation target point; the camera coordinates were converted to the robotic arm coordinates by using coordinate transformation, and then transferred to the motion control system for object grabbing by TCP/IP communication. The experimental results show that the identification accuracy of all parts in different positions is above 82%, and the grabbing error is within 1~4 mm, which meet the requirements of industrial production.