Learning to Catch Moving Objects with Reduced Impulse Exchange

Abstract

The paper presents a learning from demonstration approach to the catching of moving objects with a robot manipulator. The work explicitly reduces the impulse exchange by minimizing the relative velocity during the contact phase by learning relative relation between the object and the catcher instead of learning separate forward model of the object and trajectory generator for the robot. This contributes to the damage prevention on both, the object and the robot. The demonstrated catching movements are modelled by a Gaussian mixture model (GMM), which describes the probability distribution over the demonstrated data set. Gaussian mixture regression (GMR) is employed for motion prediction. A timing controller is designed to trade-off the catch location and the catching time. The learning scheme comprises the relative position and velocity profile between the object and the catcher. The approach allows to generate the movement of the catcher with guaranteed position and velocity convergence to the reference inferred from the position and velocity of the object. Experimental results with a five degree of freedom arm and a Photonic-Mixing-Device (PMD) camera for object detection validate the approach.