強化学習による走行動作獲得に基づく人混みで自律走行する移動ロボット

抄録

The paper introduces a study on acquisition of action models to avoid crowds by reinforcement learning. There are many dynamic obstacles such as crowds avoid for autonomous mobile robots in real worlds. Since it is dangerous to conduct training of action models in the real environments, simulation environments are used for training. In this paper, training of action models was performed by deep reinforcement learning in simulation environments where there are many pedestrians walking at random or along flows in various densities. Then, trained action models were evaluated in a viewpoint of arrival to the destinations. The results indicated that training in realistic human density is effective for traveling to destinations avoiding pedestrians, compared with humans in high density.