Learning Grasp Configurations for Novel Objects from Prior Examples

We present a new approach to learning grasp configurations for a novel object from prior example objects. We assume the novel object and the example object belong to the same category (i.e., objects share the same topologies and have similar shapes). Both the novel and the example objects are segmented into the same semantic parts. We learn a grasp space for each part of the example object using a combination of optimization and learning algorithms. We perform shape warping between the corresponding parts of the example object and the novel object and then compute the corresponding grasps based on the results. Finally, we assemble the individual parts and the associated grasps on the novel object and use local replanning to adjust grasp configurations to satisfy the stability and physical constraints (e.g., that the grasp are penetration-free). Our algorithm can automatically handle a wide range of object categories and a variety of robotic hand grasps.