視覚と行動のend-to-end学習による経路追従行動の模倣

―データセットを収集してオフラインで訓練する手法の検討―

抄録

We investigate a method for offline learning of vision-based path-following behavior using pre-collected images and actions. Our previous method has learned such behavior online. The feature of the method is that it imitates the behavior generated by self-position estimation using a LiDAR sensor as an input to the behavior using vision as an input. However, it has been a problem that imitation learning requires a long training time. Therefore, we try to shorten the training time by using offline learning. Furthermore, we will clarify how much visual information around the path is required for applying the method to a real robot. As a result, we verified that the method shortens the training time. We also clarified the required visual information through experiments.