Exploring planetary surfaces: Active learning for resource mapping with autonomous rovers

Abstract

The exploration of planetary surfaces presents unique challenges that necessitate autonomous decision-making to optimize scientific data collection. Such data collection is necessary to facilitate technological advancements and increase our knowledge of planetary bodies. This thesis investigates the use of active learning algorithms to enhance the efficiency of planetary surface exploration, with a particular focus on Gaussian Processes (GPs) and cost-aware query policies that ensure efficiency and minimization of annotation costs. By employing active learning techniques, autonomous robotic agents can select informative sampling locations, reducing the number of samples and distance required while maximizing scientific findings. Through a combination of simulation studies and real-world experiments, this research evaluates the performance differences of GPs and Bayesian Neural Networks (BNNs) in constrained trajectory exploration. The findings indicate that GPs consistently achieve faster convergence, require fewer samples, and result in shorter travel distances compared to BNNs. Despite the flexibility of BNNs in modeling complex spatial distributions, they exhibit higher computational demands and reduced reliability in sparse-data environments. A field demonstration conducted on Mauna Kea, a recognized lunar analog site on the island of Hawai'i, further validates the applicability of active learning algorithms in real-world planetary exploration settings. The results highlight the potential of GP-based active learning in reducing mission duration and optimizing energy efficiency for autonomous robotic explorers. Along with this, in-situ testing was completed in a sand court at the University of Hawai'i at Manoa where various query policies were tested on their efficiency at minimizing costs associated with data collection. This work contributes to the development of intelligent exploration strategies for planetary surface missions. A future research direction includes additional real-time deployment of active learning algorithms, particularly with fully autonomous rovers in lunar regolith testbeds with scientific instrumentation planned for use on a space mission.