The Instance Segmentation and Grasping Sequence Method for Stacked Workpieces

To obtain the grasping sequence of the robotic arm for grasping target objects in a stacked workpiece industrial scenario, this paper proposes an instance segmentation and grasping sequence method for stacked workpieces. Firstly, a Mask R-CNN instance segmentation network based on boundary refinement is designed to obtain more accurate instance mask information. Additionally, a Euclidean clustering method based on smoothness parameters is integrated to correct extreme cases of mis-segmentation by the improved Mask R-CNN instance segmentation network from the perspective of three-dimensional space (where adjacent closely stacked workpieces are mis-segmented as the same workpiece). Secondly, a stacking workpiece grasping sequence filtering mechanism is proposed. This mechanism determines the occlusion status of the target workpiece to be grasped by judging the relationship between visible masks and complete masks obtained through instance segmentation. Furthermore, based on the grasping judgment scores, the workpieces in unoccluded states are arranged in sequence, filtering out target workpieces that satisfy the principle of grasping unoccluded state workpieces first, in a top-to-bottom manner. Experimental results demonstrate that the proposed method can effectively infer the grasping sequence in stacked workpiece environments, thereby driving the robotic arm to perform grasping tasks.