Proc. of the International Joint Conference on
Artificial Intelligence (IJCAI-05).
Abstract
Mobile robot map building is the task of generating a
model of an environment from sensor data. Most existing approaches to
mobile robot mapping either build topological representations or
generate accurate, metric maps of an environment. In this paper we
introduce relational object maps, a novel approach to building metric
maps that represent individual objects such as doors or walls. We
show how to extend relational Markov networks in order to reason about
a hierarchy of objects and the spatial relationships between them.
Markov chain Monte Carlo is used for efficient inference and to learn
the parameters of the model. We show that the spatial constraints
modeled by our mapping technique yield drastic improvements for
labeling line segments extracted from laser range-finders.