Prediction learning in robotic pushing manipulation

This paper addresses the problem of learning about the interactions of rigid bodies. A probabilistic framework is presented for predicting the motion of one rigid body following contact with another. We describe an algorithm for learning these predictions from observations, which does not make use of physics and is not restricted to domains with particular physics. We demonstrate the method in a scenario where a robot arm applies pushes to objects. The probabilistic nature of the algorithm enables it to generalize from learned examples, to successfully predict the resulting object motion for previously unseen object poses, push directions and new objects with novel shape. We evaluate the method with empirical experiments in a physics simulator.