A data-driven grasp planning method based on Gaussian Process Classifier

Abstract

This paper presents a grasp planning method for grasping novel objects from point clouds provided by the Kinect camera. By applying machine learning, the planning method can generate two points which represent the contact point and direction of grasp. This method is based on three components: 1) grasp configuration which can present the location of contact points and the direction of grasp, 2) features which take force closure and grasp stability into account, and 3) Gaussian Process Classifier which is used to calculate the grasp quality by using the features of each grasp configuration. Two experiments are carried out to verify our method. The results demonstrate that the robot using this approach can successfully grasp objects with partial point clouds.