Efficient Mapping of Environmental Phenomena with Autonomous Robotic Systems

Abstract

In this dissertation, we investigate methods utilized by single and multi robotic systems to map potentially nonstationary, time-varying environments. Key to modeling the environment from collected sensor data is the Gaussian process, used for its concise mathematical representation of the spatial field of interest. Towards realtime mapping, we develop a framework in which spatially varying hyperparameters of the Gaussian process kernel can be trained online while remaining computationally manageable, and demonstrate the advantage of our method in accurately mapping spatial phenomena with changing local variability. In decentralized settings, we propose a communication criteria that maximizes mutual information to facilitate collaborative multi-agent mapping in low communication bandwidth environments. Simulation experiments explore the trade-off between model similarity and joint information gain. Next, we examine the environmental monitoring problem in the Bayesian optimization framework. We propose a Gaussian process pure exploration algorithm with easily computable theoretical bounds on the simple regret, that delineate the relationship between number of samples and solution accuracy. Furthermore, an automatic stopping criterion is proposed for terminating the optimization process with accuracy guarantees under model uncertainty.