Learning an object-grasp relation for silhouette-based grasp planning

Abstract

This work addresses the problem of experienced-based grasping of unknown objects. A relation between objects and grasps is learned based on example grasps for a set of given objects. This machine-learning approach is applied to the problem of visual determination of grasp points based on a given silhouette of an object. The approximated function allows computing a grasp quality for any objectgrasp combination. For the dimension reduction of the object silhouettes, a combination of image normalization and principal component analysis is used. A Support Vector Regression is used to learn the object-grasp relation. For the evaluation, the objects are grasped with a two-finger gripper and an imprint of a planar object on a tactile sensor matrix is used as an imaging method. © 2009 Springer-Verlag Berlin Heidelberg.