With the production of bulk parts and a large number of assembly scenarios, most parts are stacked in an unordered manner. This article focuses on the accurate grasping of parts in an assembly line under the guidance of binocular vision. Firstly, based on production needs, this article designs a robot grasping system based on binocular vision, and describes the selection of key equipment such as cameras and robotic arms in the system. Then, in order to achieve the recognition of stacked parts, analysis and research were conducted on point cloud data extraction, point cloud feature recognition, point cloud registration, and part pose estimation in the recognition process. Finally, through the construction of a simulation system, the recognition of stacked parts was achieved, and the recognition accuracy was improved.