Biologically inspired grasp planning using only orthogonal approach angles

Abstract

One approach to robotic grasping is to compute hand configurations, including contact locations, which would maximize a variety of measures of grasp quality. There is evidence that human-guided robotic grasps exhibit orthogonality of the wrist as a key feature. Orthogonality of the hand to the object frame or surfaces is often included in state-of-the art grasp synthesis algorithms, but here we present a systematic study of the efficacy of orthogonality alone. The orthogonality-alone planner works with a surprisingly good success rate on a variety of physical objects using only a single exposure from a depth camera and no object models whatsoever. Principal axes of an object are identified from the point cloud, the approach angle is determined from the second principal axis, and the orientation of the hand is vertical from the first axis. When this technique is applied to 19 novel objects presented in front of a physical robot, utilizing automatic object segmentation, the grasp success rate is 98.4%.