Learning and Predicting Center of Mass through Manipulation and Torque Sensing

Abstract

Autonomous manipulation of unknown objects is very challenging. With current computer vision techniques, new objects can be detected and tracked based on appearance, but manipulation often requires physical information of the objects, which remains hidden. This paper introduces a learning approach that includes reinforcement learning and feedforward learning based on torque sensing and manipulation to learn and predict the centers of mass (COMs) of unknown objects. After reinforcement learning is applied to learning the COMs of multiple objects, the data collected are further used to train a neural network to predict the COMs of new objects given their different shapes. Such a prediction can be used as an initial guess to speed up further reinforcement learning of the COM of a new object. Results demonstrate that the approach works with a generic robot manipulator and is robust to sensing uncertainty without the need of high-precision sensors and gripper.