Evaluation metric for instance segmentation in robotic grasping of deformable linear objects

Abstract

Automating the assembly and handling of deformable linear objects requires their robust detection. This paper introduces a new evaluation metric for the results from instance segmentation. The metric enables estimating the proportion of valid grasp poses and graspable objects for specifc gripper models. The results demonstrate that masks with similar scores in area-based metrics can have different grasp pose validity outcomes. In addition, it is indicated that when handling deformable linear objects with a vacuum gripper, it is possible to achieve a grasp precision and grasp recall of about 90 %.