Abstract: In the automation industry, robot object grasping highly depends on machine vision for positioning analysis. Aiming at the problems of low positioning accuracy and poor effect of existing binocular vision robots, an improved SGBM stereo matching algorithm which can accurately locate in parallel robots is proposed. The gradient information is integrated into sad algorithm as the similarity measure function of cost calculation; Then the adaptive bilateral filtering is used as the cost aggregation, which makes up for the cumbersome defect of artificially setting parameters to find the optimal process. At the same time, the noise is removed and the matching accuracy is improved, so as to complete the reconstruction of 3D point cloud. Finally, the hand eye calibration algorithm is used to transfer the camera coordinate data to the robot coordinate system, and then guide the robot to complete the positioning and grasping of the workpiece. Experiments show that only when the workpiece is placed arbitrarily in the field of vision of the binocular camera, it can complete accurate grasping, and the average error rate is less than 5%. Compared with the traditional method, it has higher real-time and accuracy, and has a good application prospect in the automation industry.