Enhancing Autonomous Science: Efficient Robotic Informative Sampling and Environment Hypothesis Testing

Abstract

Robotic systems can be crucial in automating information gathering tasks in natural environments. However, sample collection can be expensive and damaging to the environment, making data sparse and resource-to-information efficiency of utmost priority. Moreover, effective data collection as well as appropriate models can fundamentally improve scientific understanding. This thesis uses hypothesized models to accelerate information gathering while simultaneously validating those models.

First, a Multi-Quantity Gaussian Process is introduced that links multivariate Gaussian distributions and linear models, allowing to derive parametric models between multiple quantities from sparse, non-collocated data. The technique also yields coefficients of determination, providing goodness-of-fit measures for those hypothesized models. The learned inter-quantity relationships adapt the model's predictions such that good hypotheses are utilized and poor ones are ignored.

Second, an objective function is employed that balances exploration and exploitation to reduce both the mean and maximum model errors. New sample locations are optimally chosen based on global and local uncertainties through the Gaussian Process variance and a nearest-sample prediction difference. Sample spread is maintained while actively choosing regions that test the proposed quantity relationships.

Third, the informative objective function is scaled by the travel-distance to reduce travel costs and maximize energy-to-information efficiency. This naturally self-adjusts to different search areas and quantity variations without any manual parameter tuning. Furthermore, computation is saved by dropping poor hypotheses from consideration in future learning iterations.

Finally, the full algorithm is showcased directing an autonomous ground vehicle in a farming application characterizing pasture abundance, demonstrating collection and mapping of plant heights and evaluation of correlation with prior elevation data.