A Continuous Grasp Representation for the Imitation Learning of Grasps on Humanoid Robots

Models and methods are presented which enable a humanoid robot to learn reusable, adaptive grasping skills. Mechanisms and principles in human grasp behavior are studied. The findings are used to develop a grasp representation capable of retaining specific motion characteristics and of adapting to different objects and tasks. Based on the representation a framework is proposed which enables the robot to observe human grasping, learn grasp representations, and infer executable grasping actions.