RA5 Spatial cognition and multimodal situation awareness

… how can an embodied system come to a conceptualisation of sensory and sensory-motor data, generate plans and actions to navigate and manipulate in typical home settings...

The work at the UvA will mainly concentrate on the research activities RA1,2 and 3.:x

5.1 Hierarchical representations of space

Because metric maps do not scale well and because a cognitive robot should be instructed at a `conceptual’ level we need to integrate topological maps and metric maps. In WP5.1 we study probabilistic methods for this.
figuresCogniron2

5.2 Categorization of spaces and objects

In the hierarchical representation an important issue is the categorization of sensory information. In WP 5.2 we will study feature representations which are invariant for small rotations and translations of the robot, but in which a good discrimination between different spaces can be found. Since we expect that categories can be characterized by both their appearance as well as their spatial features we may use depth information and visual information.

Can we use graphs on local salient points for that? (Graph-Theoretical Methods in Computer Vision A. Shokoufandeh and S. Dickinson,Springer-Verlag Heidelberg Lecture Notes in Computer Science, Volume 2292, 2002, pp 148--174.)
Can we use other local features? (Jim Little: Mobile Robot Localization and Mapping with Uncertainty using Scale-Invariant Landmarks, Intl. Journal of Robotics Research, Vol. 21, No. 8, August 2002, pp. 735--758.)

5.3 Learning concepts by interaction with humans

Instead of a bottom-up, data driven approach, concepts can also be learned by interaction with humans. The robot therefore has to recognize and track the user and have a simple speech dialogue. Also gesture recognition needs to be implemented.

We will study different learning methods, such as supervised learning and reinforcement learning.

Some of the work carried out by Stephan ten Hagen: (Stephan ten Hagen. Concepts and navigation targets. In Proceedings of the 15th Dutch-Belgian Artificial Intelligence Conference, BNAIC'03, Nijmegen, The Netherlands, October 2003.

5.4 Models of temporal relations

•Develop and study models which are well suited to deal with uncertainties in data on temporal relations.
Dynamic Bayesian Networks will be used to model the movements of humans in an environment. These models will be studied in the context of the home tour scenario

•

Partners:

•LAAS:

EPFL:

KTH

UvA:

UniKarl

IPA: