Robust object exploration and detection

Abstract

In this thesis we explore models and algorithms used by an autonomous agent to find objects in the real world. We begin by tackling the problem of determining the existence and location of an object robustly given the sensors employed on an agent. Our major contribution lies in modeling the spatial correlations between the sensor and object. Next, we develop models and algorithms used to explore the world in order to find all of the objects. We develop a model with tractable inference which reasons about the locations of all the objects, seen and unseen, by reformulating our problem into one of assigning objects to particular clusters. Along the way we analyze theoretical properties relating to the number of un-informative decision any agent must make in order to find all the objects in the world using the theory of random graphs, particularly percolation theory. The developed systems improve upon the state-of-the art in both simulation and real-world experiments.