Vision-Based Grasp Planning Based on Grasp Quality Metrics and Its Hardware Implementation

Abstract

This work presents grasp planning on everyday objects using vision. The hand considered is a one degree-of-freedom parallel jaw gripper of Mitsubishi Movemaster robot. Candidate grasping points are chosen on the object and a grasp matrix is computed for the grasp. The grasp matrix can be used to computationally determine a force-closure grasp feasibility. For selecting the candidate grasping points, image of the object is used. Three quality metrics based on different physical notions of quality of grasp are computed. The first quality measure tells how far a grasp is from violating the friction limits, the second gives the worst case performance of the force-closure for all external wrenches, and the third tells how well the object is enclosed from all directions. The main contribution of the paper is to compare grasps based on different quality measures and understand their physical interpretation.