Policy Transfer from Simulation to Real World for Autonomous Control of an Omni Wheel Robot
Yuto Ushida, Hafiyanda Razan, Takuto Sakuma, Shōhei Kato · 2020
We aim to develop an autonomous mobile robot which supports workers in warehouse to reduce their burden. The robot acquire state-action policy to avoid obstacles and reach a destination by reinforcement learning using LiDAR sensor. In case of real-world application of reinforcement learning, the policy learned previously under simulation environment are generally diverted to real robots because of uncertainties that is unexpected under simulation environment, for example, friction, sensor noise and so on. In this paper, we proposed a method to refine action control of an omni wheel robot by transfer learning on real environment to deal with this problem. We conduct the experiment of searching the route for reaching a goal on real environment using transfer learning's results and verify the effectiveness of the policy acquired.