Stewart Jamieson, S.M., 2020
Yogesh Girdhar, Advisor
Contemporary scientific exploration most often takes place in highly remote and dangerous environments. These environments are very hostile to humans, which makes robotic exploration the primary means of exploration, but they also impose restrictive limits on how much communication is possible, making remote command and control of the robot inefficient. We propose an approach to more efficient autonomous robot-based scientific exploration of remote environments despite limits on human-robot communication. We find this requires the robot to have a spatial observation model that can predict where to find various phenomena, a reward model which can measure how relevant these phenomena are to the scientific mission objectives, and an adaptive path planner which can use this information to plan high scientific value paths. Our first contribution is enabling general-purpose spatial observation modelling through spatio-temporal topic models, which are well suited for unsupervised scientific exploration of novel environments. Our next contribution is an active learning criterion which enables learning an image-based reward model during an exploration mission by communicating with the science team efficiently. We show that using these together can result in a robotic explorer collecting up to 230% more scientifically relevant observations in a single mission than when using lawnmower trajectories.