A Semi-supervised Learning Based on Variational Autoencoder for Visual-Based Robot Localization

Robot localization, the task of determining the current pose of a robot, is a crucial problem of mobile robotic. Visual-based robot localization, which using only cameras as exteroceptive sensors, has become extremely popular due to the relatively cheap cost of cameras. However, current approaches such as Bayes Filter based methods and Visual Odometry need knowledge of prior location and also rely on the feature points in images. This paper presents a novel semi-supervised learning method based on Variational Autoencoder (VAE) for visual-based robot localization, which does not rely on the prior location and feature points. Because our method does not need prior knowledge, it also can be used as a correction of dead reckoning. We adopt VAE as an unsupervised learning method to preprocess the environment images, followed by a supervised learning model to learn the mapping between the robot's location and processed images. Therefore, one merit of the proposed approach is that it can adopt any state-of-the-art supervised learning models. Furthermore, this semi-supervised learning scheme is also suitable for improving other supervised learning problems by adding extra unlabeled data to the training data set to solve the problem in a semi-supervised manner. We show that this semi-supervised learning scheme can get a high accuracy for pose prediction using a surprisingly small number of labeled images compared to other machine learning methods.