Robot Grasping Based on Stacked Object Classification Network and Grasping Order Planning

In this paper, the robot grasping for stacked objects is studied based on object detection and order planning. Firstly, a novel classification network (SOCN) proposed to realize recognition. The takes into account visible volume of further adjust its inverse density parameters, which makes training process faster smoother. At same time, SOCN adopts transformer architecture has self-attention mechanism feature learning. Subsequently, planning method investigated, depends security score extracts geometric relations dependencies between objects, it calculates relation, classification, size. evaluated by using depth camera UR-10 complete tasks. results show that our high accuracy effectively successfully executes safely.