Generalizing pouring actions between objects using warped parameters

One of the key challenges for learning manipulation skills is generalizing between different objects. The robot should adapt both its actions and the task constraints to the geometry of the object being manipulated. In this paper, we propose computing geometric parameters of novel objects by warping known objects to match their shape. We refer to the parameters computed in this manner as warped parameters, as they are defined as functions of the warped object's point cloud. The warped parameters form the basis of the features for the motor skill learning process, and they are used to generalize between different objects. The proposed method was successfully evaluated on a pouring task both in simulation and on a real robot.