TITLE: Where are the cornflakes? Exploration and Exploitation of Probabilistic Knowledge under Uncertain Sensing
SPEAKER: Marc Hanheide (School of Computer Science, University of Birmingham)

ABSTRACT:

I am presenting Dora, our robot systems that explores its environment in a task-driven way to find objects. Besides taking topological and spatial knowledge into account it is also equipped with a probabilistic ontology that enables it to exploit common-sense knowledge. This probabilistic model comprises information about the likelihood of finding certain objects in particular categories of rooms (e.g. that cornflakes are usually found in kitchens) and observation models of successfully detecting an object if it is present. By reasoning within this probabilistic spatial representation that is incorporates perceived instance knowledge a probabilistic belief state of the world is maintained. Planning within this probabilistic representation using a switching planner, switching between sequential and contingent planning sessions, we can exploit the commonsense knowledge to find objects more efficiently in many cases. However, the system can still cope with violated assumptions, i.e. when common-sense knowledge is not applicable, and sensing errors. I discuss the limitations and caveats of our approach so far and brie%Gfl%@y talk about the challenge to combine task-driven with curiosity driven exploration in our world. The accompanying video can be watched at

http://www.youtube.com/watch?v=0QcmSDZR-c4.