Adaptive Sample Selection for Hypothesis Falsification

P. Michael Furlong

Abstract: Current approaches to autonomous exploration focus on collecting observations in the absence of prior knowledge of the phenomena under investigation. However, it is unlikely that robots will arrive at planetary bodies without scientists having formed one or more hypotheses explaining data collected by precursor operations such as satellite images. These exploring robots collect observations to falsify the proposed hypotheses, incorporating those hypotheses can increase the efficiency of observation collection. This paper presents a novel algorithm, formulated in an exploration/exploitation framework, that directs robots to collect samples to determine which of a collection of hypotheses best explain data observed in situ by robots. We simulate a geologic exploration mission with a lander vehicle that can hop between locations of interest. This application is analogous to exploring of, e.g., the Aitken Basin of the south pole of Earth's Moon where sampling sites need to be separated hundreds or thousands of meters. We demonstrate that sampling algorithms aware of the hypotheses under investigation perform statistically significantly better than standard approaches, making more effective use of mission resources.