In order to solve the problem of robot sorting in the disorderly parts box in the complex industrial production environment, it is necessary to complete the spatial positioning of the workpiece, the recognition of different types of the workpiece and the grasping operation of the robot. The existing vision technology can not meet the random sorting task.Therefore, an intelligent robot sorting system combining binocular stereo vision, deep learning and UR5 robot is proposed.A three-dimensional vision system combining stereoscopic vision and projection structured light was proposed. The energy function of stereoscopic matching was reconstructed to complete the spatial positioning of the workpiece.An instance segmentation method based on deep learning was used to accurately identify the workpiece.A random sorting system based on binocular structured light and deep learning was realized by combining robot hand-eye calibration technology with workpiece spatial positioning and recognition results. By analyzing the success rate of random sorting of screw workpiece in the process of random sorting, the average success rate of all single sorting of screw workpiece of different quantity is 92.8%, and the success rate of random sorting according to the total number of workpiece is 98.8%, which verifies the feasibility of the system.