View-based programming with reinforcement learning for robotic manipulation

In this paper, we study a method of robot programming with view-based image processing. It can achieve more robustness against changes of task conditions than conventional teaching/playback without losing its general versatility. In order to reduce human demonstrations required for the view-based robot programming, we integrate reinforcement learning with the method. First we construct an initial neural network as a mapping from images to appropriate robot motions using human demonstration data. Next we train the neural network with actor-critic reinforcement learning so that it can work well even in task conditions that are not identical to those in the demonstrations. Our proposed method is successfully applied to pushing and pick-and-place tasks in a virtual environment.