Efficient and Effective Grasping of Novel Objects through Learning and Adapting a Knowledge Base

This paper introduces a new approach to establish a good grasp for a novel object quickly. A comprehensive knowledge base for grasping is learned that takes into account the geometrical and physical knowledge of grasping. To automate the learning process as much as possible, learning happens in a virtual environment. We used the GraspIt! [16] simulation environment with the Barrett hand for this work. As only approximate features of objects are used for training the grasping knowledge base (GKB), the knowledge gained is rather robust to object uncertainty. Based on the guidance of the GKB, a suitable grasp for a novel object can be found quickly. The newly gained grasping information of the new object can also be feedback to the GKB so that the knowledge base continues to improve as it is exposed to more grasping cases. The GKB serves as the "experience" of the robotic gripper to make grasping more and more skillful. We implemented the approach and tested it on a wide variety of objects. The results show the effectiveness of this approach to achieve quick and good grasps of novel objects.