深層強化学習によるサッカーヒューマノイドロボットのドリブル動作獲得

抄録

The goal of this research is to use deep reinforcement learning to acquire dribbling skills for a soccer humanoid robot in an obstacle environment. We also address the problem of learning time, which is one of the challenges in reinforcement learning. Using the constructed simulator, we designed a state transition model that represents the dribbling of the humanoid robot with a Neural Network. By using this state transition model as an environment for reinforcement learning, we succeeded in reducing the learning time and acquiring obstacle-aware dribbling. We confirmed that the learned policy can be transferred to a simulated soccer environment used in RoboCup 2021 Humanoid League. We also tested the trained policy using a real humanoid robot and observed that the robot was able to dribble the ball towards the goal.