Traversability estimation in partially occluded and deformable terrain for planetary rovers

Abstract

Terrain traversability estimation is a fundamental requirement for autonomous planetary rovers and their ability to conduct long-term missions. This is particularly important in exploration scenarios as scientifically interesting sites are often located in challenging terrain. This thesis proposes an approach to address two fundamental challenges for terrain traversability estimation techniques. First, representations of terrain data typically built by the rover’s exteroceptive sensors are often incomplete due to occlusions and sensor limitations. Second, rover-terrain interaction can cause terrain deformation, which may significantly alter the difficulty of traversal. The proposed approach in this thesis comprises of two distinct but interconnected components. The first component, named Rigid-Terrain Traversability Estimation (R-TTE), computes an initial prediction, given incomplete terrain data, under the assumption that the terrain is rigid. This component first learns a kernel function that best represents the evolution of the rover’s configuration from experimental training data. In operation, this function is used in a GP regression framework to predict the terrain traversability. The second component, named Rigid-to-Deformable Terrain Traversability Estimation, refines the initial estimate computed by R-TTE to account for the effects of potential terrain deformation, using a near-to-far learning approach based on multi-task GP regression. This component exploits the local variations in the rover’s configuration angles and the rover driving speed to build a probabilistic model of terrain deformation effects on the rover’s configuration. The proposed estimation process enables the rover to anticipate the effects of terrain deformation, including in areas with incomplete terrain data. This allows the rover to accurately estimate of its attitude and chassis configuration angles on any queried location on a given terrain representation.