Representations for intelligent navigation in unfamiliar environments

Abstract

The way an agent chooses to represent the world around it is fundamental to its ability to effectively interact with it. The work presented in this thesis is centered around the development of new representations that enable embodied agents to better understand the impact of their actions, so that they may plan quickly and intelligently in a dynamic and uncertain world. In this thesis, we focus on the problem of autonomous navigation in complex, unknown environments. Consider an embodied agent tasked with traveling to an unseen goal in minimum time. In general, effective navigation requires that the agent explicitly reason about portions of the environment it has not yet seen. But the world is intractably complex; exhaustively enumerating all environment configurations is impossible. Instead, we imbue an agent with the ability to more tractably make predictions about uncertainty by changing the way in which it represents its surroundings and the actions it uses to define navigation. This thesis makes three primary contributions. First, we introduce Learned Subgoal Planning, a decision-making paradigm that leverages high-level actions to make tractable predictions about unknown space via supervised learning and enable efficient computation of expected cost. Second, we apply recent progress in image-to-image translation to the task of domain adaptation for image data, allowing an agent to transfer knowledge acquired in simulation to the real world. Finally, we introduce a learned pseudosensor and accompanying probabilistic sensor model that estimates sparse structure in view of an agent from monocular images. Fusing these estimates during exploration of unknown environments, we enable map-building of unfamiliar environments from monocular images suitable for high-level planning with topological constraints.