Introduction
In planetary and terrestrial exploration, a significant scale mismatch exists between orbital sensors and ground-based rovers. Orbital platforms provide global context but suffer from low spatial resolution; a single 30-meter pixel often represents a complex mixture of various surface materials. Conversely, rovers provide high-fidelity, "close-up" measurements but are limited by strict time and energy constraints, allowing them to visit only a fraction of the landscape. This paper, originally presented at AAAI 2015, introduces an autonomous exploration framework designed to bridge this gap by using high-dimensional spectral data to guide rovers toward locations that best explain coarse orbital imagery.
Methodology
The core of the proposed framework is a "science-aware" objective function for path planning based on spectral unmixing. The system assumes a Linear Mixing Model, where each orbital pixel is treated as a linear combination of "pure" spectral signatures, known as endmembers, found on the ground.
Instead of traditional exploration goals like reducing statistical uncertainty (e.g., via Gaussian Processes), the rover aims to build a "spectral library" that minimizes the reconstruction error of the orbital map. It identifies waypoints that are likely to contain pure mineral signatures, which act as the "building blocks" for interpreting the larger orbital scene.
The optimization process utilizes Nonnegative Least Squares (NNLS) and Multiple Endmember Spectral Mixture Analysis (MESMA) to ensure physical realism—preventing the model from predicting negative mineral concentrations. Due to the NP-hard nature of the orienteering problem, a greedy planning algorithm is employed to iteratively select the most informative waypoints within a defined energy budget. The system also supports adaptive replanning, allowing the rover to pivot its path if it encounters unexpected mineralogy.
Experimental Evaluation
The framework was validated using data from Cuprite, Nevada, a well-known geological reference site. The researchers used ASTER satellite data (9 bands) as a proxy for orbital measurements and high-resolution AVIRIS airborne data (224 bands) to simulate in situ rover measurements. Simulations involved 256 random trials across three distinct sub-scenes to ensure statistical robustness.
Results and Discussion
The "science-aware" greedy planners significantly outperformed "science-blind" alternatives, such as random or direct paths, with a statistical significance of $p < 0.05$. The adaptive NNLS approach proved most effective, achieving a Root Mean Squared Error (RMSE) of 0.0131 in one test site, compared to 0.0220 for a random path.
Crucially, the algorithms are computationally efficient. While standard Least Squares planning takes less than a second, the more robust NNLS approach requires approximately 25 seconds—well within the operational requirements for real-time planetary rover navigation. These results suggest that incorporating physical priors, such as nonnegativity constraints, provides powerful guidance for AI in scientific exploration.
Conclusion
This research demonstrates that rovers can act as "decoder rings" for orbital data, providing the ground truth necessary to translate global context into specific mineralogical maps. While the current model assumes a flat navigable plane and linear mixing, future work will address complex terrains and non-linear light scattering. Moving from airborne simulations to actual field tests with rover-mounted spectrometers remains the primary next step for this autonomous exploration framework.
Summary updated