Publication

Autonomy Artificial Intelligence Robotics (AAIR)
2022

GAUSSIAN PROCESS MODELING OF TERRAIN SLOPE FOR GROUND VEHICLE LOCALIZATION

by Jesse Pentzer; Karl Reichard

Abstract

This paper presents a Gaussian process model of terrain slope for use in a GPS-free localization algorithm for ground robots operating in unstructured terrain. A wheeled skid-steer robot is used to map the terrain slope within an operational area of interest. The slope data is sampled sparsely and used as training data for a Gaussian process model with a two-dimensional input. Three different covariance functions for the Gaussian process model are evaluated with hyperparameters selected through maximizing the log marginal likelihood. The resulting Gaussian process model is used in the measurement update function of a localization particle filter to generate expected slope values at particle positions. Preliminary localization testing shows sub-ten meter accuracy with no initial knowledge of position. However, the overall performance of the filter is highly dependent on the variability of the terrain that the robot traverses.