Event Location: NSH 1109

Abstract: There is considerable interest in having robots continue to perform autonomously over time in human environments despite changes in their environment. Most state-of-the-art robot localization and mapping approaches assume that the world can be represented as a static map. However, real human environments are not static – some areas like corridors might be mostly static, but areas like atria and cafes have numerous moving obstacles, as well as objects like chairs, tables and bins that are frequently moved around. When most of the observations made by a robot consist of such moving and movable objects, the limitations of the static world assumption become apparent.

In this thesis I propose to model varying human environments as consisting of three distinct classes of objects: permanent, movable, and moving. I introduce hybrid Markov / Non-Markov localization to reason about observations of objects from different classes, and to account for correlations between observations due to unmapped objects. Furthermore, I propose to build long-term static maps of the permanent objects, and probabilistic object maps of the movable objects by processing sensor logs collected from a team of deployed robots over time. The long-term static maps will consist of features that persist over time, even if temporarily occluded by movable objects. The probabilistic object maps will include the observed shape and time-dependent distribution of poses of movable objects. I propose to evaluate the contributions of this thesis over a “1000K” challenge: to have the robot autonomously traverse at least 1000 km in varying human environments while performing tasks. I expect to demonstrate the robustness of hybrid Markov / Non-Markov localization in changing environments over such deployments.

Committee:Manuela Veloso, Chair

Reid Simmons

David Wettergreen

Dieter Fox, University of Washington