MRF-based terrain map inference using variational feature projection

Abstract

A method to infer the terrain of unobserved fields as well as navigation paths using vibrations from vehicle-terrain interactions is presented. The relationships between adjacent terrain grids and robot paths are modelled by Markov random field (MRF) for inferring the classes of unobserved field. In order to approximate the complex likelihood of vibration features as a tractable distribution, a variational autoencoder is adopted. By projecting intractable features to the latent space, an approximated likelihood that gives similar performance to a neural network classifier is obtained. Then the numerical maximum a posterior inference for MRF is obtained. The result of inferred terrain maps are compared with that from the feature likelihood as a Gaussian-mixture model.