Task Learning Based on Reinforcement Learning in Virtual Environment

As a novel learning method, reinforced learning by which a robot acquires control rules through trial and error has gotten a lot of attention. However, it is quite difficult for robots to acquire control rules by reinforcement learning in real space because many learning trials are needed to achieve the control rules; the robot itself may lose control, or there may be safety problems with the control objects. In this paper, we propose a method in which a robot in real space learns a virtual task; then the task is transferred from virtual to real space. The robot eventually acquires the task in a real environment. We show that a real robot can acquire a task in virtual space with an input device by an example of an inverted pendulum. Next, we verify availability that the acquired task in virtual space can be applied to a real world task. We emphasize the utilization of virtual space to effectively obtain the real world task.