This paper proposes a deep learning based visual detection and grasping method to solve the problems of the existing robotic grasping systems, including high hardware costs, difficulty in adapting to different objects, and large harmful torques. The channel attention mechanism is used to enhance the ability of the network to extract image features, improving the effect of target detection in complex environments using the improved YOLOV3. It is found that the average recognition rate is increased by 0.32% compared with that before the improvement. In addition, to address the discreteness of estimated orientation angles, an embedded minimum area bounding rectangle (MABR) algorithm based on VGG-16 backbone network is proposed to estimate and optimize the grasping position and orientation. The average error between the improved predicted grasping angle and the actual angle of the target is less than 2.47°, significantly reducing the additional harmful torque applied by the two-finger gripper to the object in the grasping process. This study then builds a visual grasping system, using a UR5 robotic arm, a pneumatic two-finger robotic gripper, a Realsense D435 camera, and an ATI-Mini45 six-axis force/torque sensor. Experimental results show that the proposed method can effectively grasp and classify objects, with low requirements for hardware. It reduces the harmful torque by about 75%, thereby reducing damage to grasped objects, and showing a great application prospect. © 2023 Beijing University of Aeronautics and Astronautics (BUAA). All rights reserved.
