Explainable Machine Learning and Uncertainty Quantification in Robotic Lunar Surface Traversal

With autonomous or semiautonomous system teams, trustworthiness and transparency are key factors in humans granting or ceding autonomy. Surface exploration of the Moon and other planetary bodies with rovers depends on the ability to accurately assess the terrain and therefore the traversability of potential hazards. We explore the use of Explainable Basis Vectors (EBVs) and Uncertainty Quantification to explain deep learning models in the domain of terrain classification from image data. These methods are incorporated into a hybrid machine learning model architecture, combining an autoencoder and supervised model to produce EBVS. In this paper, we introduce the Explanations in Lunar Surface Exploration (ELSE) system. We explore how we will make use of the ELSE system to improve traversability judgment in lunar rover exploration, both for the machines making automated judgments and remote human operators.