Online learning of optimal robot behavior for object manipulation using mode switching

This paper presents an optimal robot motion learning method that involves object manipulation where dynamics of robots and environment are unknown. The dynamics of the environment is acquired by the robot's experience through online learning. A reinforcement learning framework which incorporates model identification is proposed. Based on the learning framework, an idea of effective motion acquisition is proposed through decomposing the task by detecting `switching of dynamics', which is called mode-switching. Object manipulation is divided into two modes, approaching to the object and pushing it toward the goal. This enables the robot to learn motions while reducing number of trials and to behave more dexterously by integrating modes, each of which was learned separately. The proposed learning method is evaluated in simulation of a wheeled robot. It was shown that appropriate motion for re-approaching and re-pushing to accurately move the object to the goal can be realized using the proposed idea of planning with mode switching.