Teaching and reinforcement learning of robotic view-based manipulation

The authors study “view-based teaching/playback”: a method of robot programming with view-based image processing which can achieve more robustness against changes of task conditions than the conventional teaching/playback scheme. Reinforcement learning was integrated with this method to make necessary human teaching minimal and it was applied to robotic manipulation in a virtual environment in our previous study. However, the process of online reinforcement learning requires huge computation time. Moreover, the low success rate of reinforcement learning makes the expected time to learning success longer. Thus we accelerate the process of reinforcement learning in this paper to make our proposed method more practical. We also try to improve the success rate of reinforcement learning for further learning acceleration.