深層生成モデルを用いた自己位置推定とサッカーロボットを題材とした予備実験

抄録

Localization is a key component for a variety of mobile robot tasks. Nowadays, probabilistic approaches are widely used for achieving robust localization. This paper reports an attempt to train a neural-network-based probabilistic robot localization model in an end-to-end manner. We constructed a conditional variational autoencoder (CVAE) to fit the posterior probability of a 2D robot position and orientation given a camera image captured by the robot. We trained and evaluated the CVAE model using the data collected by a robot soccer simulator for Robocup Humanoid League. The accuracy and the limitations of the trained localization model are discussed in this paper.