Action-Conditioned Convolutional Future Regression Models for Robot Imitation Learning

Alan Wu, A. J. Piergiovanni, Michael S. Ryoo

Abstract: This paper presents convolutional neural network (CNN) architectures for robot action policy learning, and compares them for the robot imitation learning from unlabeled example videos. These not only include standard behavioral cloning but also action learning models with explicit future frame/representation regression. Our objective is to make the robot learn to visually imagine the future consequences of taking an action from a number of example videos, and take advantage of it to learn the optimal behavior. We introduce the approach of decomposing an action model into a convolutional encoder-decoder as well as an action-conditioned future regressor, and present an approach to train them jointly. Our real-time experiments with a ground mobility robot explicitly compare different CNN models for imitation learning, and the results confirm that the use of the action-conditioned future regression benefits the robot. We show both the qualitative results of future frame regression and the quantitative evaluation of robot actions.