画像入力型移動ロボットに対する強化学習

抄録

In simulations, when a mobile robot acts, an action yields the corresponding state transition in deterministic environments. Whenever the robot selects the same action at a state, the position and the orientation of the robot are assumed to be identical after transitions. However, in a real environment, the same action at the same state can lead the robot to different states. The problem called the state-action deviation problem. To conquer this problem, a vision-based reinforcement learning system using Hough transformation is proposed. In the proposed method, state transitions of mobile robot are conducted considering its position and orientation. The effectiveness of the proposed method is shown by experiments in a real environment.