Foraging, Prospecting, and Falsification – Improving Three Aspects of Autonomous Science

Abstract

Robots exploring the subsurface ocean of Europa, for example, will not have reliable communications with scientists on Earth. Robots exploring with unreliable communications must conduct scientific exploration autonomously. Automation of scientific exploration requires both opportunistic and deliberative decision-making algorithms.

Opportunistic decision making chooses to investigate events or phenomena which could not have been anticipated from prior knowledge but which may yield valuable scientific data. Deliberative decision making uses prior knowledge to plan observations that increase information. Approaches to deliberative and opportunistic science autonomy that work in the laboratory may not work in the field.

This thesis presents three algorithms designed to improve the performance of robots conducting autonomous science investigations. The first algorithm, foraging, improves opportunistic responses by deciding between sampling immediately available objects or searching for better options. Foraging moves science autonomy beyond simply responding to matched templates or anomalous data. The work recognizes that robots may not get to choose which objects they can to sample, but must deal with what they encounter. Our approach has increased performance in selecting objects to sample when sampling costs are high, without neglecting opportunities when sampling costs are low.

The second algorithm addresses how to effectively conduct prospecting without relying on either arbitrary thresholds or responding to anomalies. Threshold-based algorithms cannot distinguish between anomalies and true changes in the distribution driving sensor readings. We present an algorithm that directly poses the question of whether or not a change has occurred. The change detection algorithm developed in this thesis encodes a level of confidence that a change has occurred, based on data collected by the robot. This can improve the efficiency of the investigation.

The third algorithm represents a new approach to information gathering based on falsification. Recognizing scientists come to missions with hypotheses formulated, the algorithm uses those hypotheses to choose sampling actions that help determine which hypothesis is most credible. Prior information gathering approaches consider one or fewer hypotheses, and focus mainly on sampling the hypotheses' domain. We show that our approach biases the belief in hypotheses towards the most credible one.

The thesis proven in this work is that accounting for operational and environmental context improves science autonomy algorithms. Each of these algorithms improve components of autonomous science, and thereby the process as a whole.