[1Q4-GS-11] Robot and real worlds: Machine learning

[1Q4-GS-11-04] Autonomous Control of an Omni Wheel Robot to Close Gap Between Simulation and Real Environments Using Transfer Learning

〇Yuto Ushida¹, Shunta Ishizuya², Razan Hafiyanda³, Shohei Kato^1,4, Takuto Sakuma¹ (1. Computer Science Program, Dept. of Engineering, Graduate School of Engineering, Nagoya Institute of Technology, 2. Dept. of Computer Science, Graduate School of Engineering, Nagoya Institute of Technology, 3. Dept. of Computer Science and Engineering, Graduate School of Engineering, Nagoya Institute of Technology, 4. Frontier Research Institute for Information Science, Nagoya Institute of Technology)

Keywords:Transfer Learning, Maze Search, Reinforcement Learning, Omni Wheel Robot

Recently, the spread of online shopping is increasing handling amount of baggage, but workers are in short supply in the distribution industry. Anyway, we aim to develop the autonomous omni wheel robot in warehouse to reduce worker’s burden. We acquire the rules for autonomous action control by reinforcement learning using the sensor data to avoid obstacles and reach the destination. In the case of application of reinforcement learning to real machine, the rules learning previously under simulation system are generally diverted to actual machine. However, real environments have uncertainties that is not unexpected under simulation system. In this article, we aim to refine action control of robot by transfer learning on actual environment to deal with this problem. We conduct the experiment of searching the route for reaching the goal on real environment using transfer learning’s results and verify the effectiveness of the policy acquired by transfer learning.