Grasp motion learning with Gaussian Process Dynamic Models

We propose an online method for grasp motion learning using the Gaussian Process Dynamic Model (GPDM). Given human grasp motion data (in the form of position and orientation trajectories of the fingertips and palm), from approach to final grasp pose, a GPDM is trained with this data, and then used to generate new grasping motions even when the path to the object is partially blocked by obstacles. Variance tubes are applied to ensure that collision avoidance and other physical constraints are satisfied. Case studies reporting on the efficiency and naturalness of our grasp motions are presented.