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

―データセット収集密度の動的調整による学習の効率化―

抄録

We have proposed a method for path-tracking of autonomous mobile robots by end-to-end learning of vision and action. Our method aims to diversify the navigation methods for these robots. It involves collecting a dataset while the robot moves and using this for learning. However, the density of data collection has not been discussed, and the density is uniform. This may result in excessive data collection in certain route areas. Addressing this, we first check for any bias in the collected dataset. We then propose a method to adjust the data collection density by varying the robot’s translational speed based on the training data. The effectiveness of our proposed method is confirmed through experiments in a simulator.