Statistical models and decision making for robotic scientific information gathering

Abstract

Mobile robots and autonomous sensors have seen increasing use in scientific applications, from planetary rovers surveying for signs of life on Mars, to environmental buoys measuring and logging oceanographic conditions in coastal regions. This thesis makes contributions in both planning algorithms and model design for autonomous scientific information gathering, demonstrating how theory from machine learning, decision theory, theory of optimal experimental design, and statistical inference can be used to develop online algorithms for robotic information gathering that are robust to modeling errors, account for spatiotemporal structure in scientific data, and have probabilistic performance guarantees. This thesis first introduces a novel sample selection algorithm for online, irrevocable sampling in data streams that have spatiotemporal structure, such as those that commonly arise in robotics and environmental monitoring. Given a limited sampling capacity, the proposed periodic secretary algorithm uses an information-theoretic reward function to select samples in real-time that maximally reduce posterior uncertainty in a given scientific model. Additionally, we provide a lower bound on the quality of samples selected by the periodic secretary algorithm by leveraging the submodularity of the information-theoretic reward function. Finally, we demonstrate the robustness of the proposed approach by employing the periodic secretary algorithm to select samples irrevocably from a seven-year oceanographic data stream collected at the Martha's Vineyard Coastal Observatory off the coast of Cape Cod, USA. Secondly, we consider how scientific models can be specified in environments - such as the deep sea or deep space - where domain scientists may not have enough a priori knowledge to formulate a formal scientific model and hypothesis. These domains require scientific models that start with very little prior information and construct a model of the environment online as observations are gathered. We propose unsupervised machine learning as a technique for science model-learning in these environments. To this end, we introduce a hybrid Bayesian-deep learning model that learns a nonparametric topic model of a visual environment. We use this semantic visual model to identify observations that are poorly explained in the current model, and show experimentally that these highly perplexing observations often correspond to scientifically interesting phenomena. On a marine dataset collected by the SeaBED AUV on the Hannibal Sea Mount, images of high perplexity in the learned model corresponded, for example, to a scientifically novel crab congregation in the deep sea. The approaches presented in this thesis capture the depth and breadth of the problems facing the field of autonomous science. Developing robust autonomous systems that enhance our ability to perform exploratory science in environments such as the oceans, deep space, agricultural and disaster-relief zones will require insight and techniques from classical areas of robotics, such as motion and path planning, mapping, and localization, and from other domains, including machine learning, spatial statistics, optimization, and theory of experimental design. This thesis demonstrates how theory and practice from these diverse disciplines can be unified to address problems in autonomous scientific information gathering.