視覚と行動のend-to-end学習により経路追従行動をオンラインで模倣する手法の提案

―経路への復帰行動の解析と復帰行動を強化する教師データ収集法の検討―

抄録

We have proposed an online imitation method of path-tracking behavior based on end-to-end learning of vision and action. In recent years, many studies of autonomous movement using end-to-end learning have been reported. However, these studies have also observed deviations from the target path. One of the possible reasons for this is the lack of training data for returning to the path. In this paper, we perform end-to-end learning to follow a route generated by a map-based navigation system. The dataset was collected in two ways, one is to learn only the area around the route and the other is to learn the state away from the route, and the generated path-tracking behaviors were analyzed. In addition, we proposed a new method of collecting teacher data to reinforce the behavior of returning to the path, and verified the effectiveness of the method by experiments using a simulator.