Predictive Exploration for Autonomous Science

Abstract

Often remote investigations use autonomous agents to observe an environment on behalf of absent scientists. Predictive exploration improves these systems ’ efficiency with onboard data analysis. Agents can learn the structure of the environment and predict future observations, reducing the remote exploration problem to one of experimental design. In our formulation information gain over a map guides exploration decisions, while a similar criterion suggests the most informative data products for downlink. Ongoing work will develop appropriate models for surface exploration by planetary robots. Experiments will demonstrate these algorithms on kilometer-scale autonomous geology tasks. On Remote Autonomous Science In general today’s planetary exploration robots do not travel beyond the previous day’s imagery. However, advances in autonomous navigation will soon permit traverses of multiple kilometers. This promises significant benefits for planetary science: rovers can visit multiple sites and survey vast areas of terrain in a single command cycle. Long autonomous traverses present new challenges for data collection (Gulick et al. 2001). These rovers will travel over their local horizon so that scientists will not be able to specify targets in advance. Moreover, energy and time constraints will continue to limit the number of measurements; sampling density will decrease as mobility improves. Finally, bandwidth limitations will preclude transmission of most collected data. These resource bottlenecks beg the question: is it possible to explore efficiently, with long traverses and sparse sampling, while preserving our understanding of the explored environment? One strategy to improve sampling efficiency involves onboard data understanding (Pedersen 2001). Pattern recognition technologies enable robots to place instruments and take measurements without human supervision. These robots can autonomously choose the most important features to observe and transmit (Castaño et al. 2003). This document suggests that these agents must learn and exploit structure in the explored environment. In