Michael Felsberg

1
The COSPAL Project

• Michael Felsberg
• Computer Vision LaboratoryLinköping
  UniversitySWEDEN

2
Fact Sheet

• COSPAL acronymCOgnitiveSystems using
  Perception-Action Learning
• STREP IST-2003-004176
• Project Start1. July 2004

3
Consortium

• Linköping University, SwedenComputer Vision
  Laboratory (coordinator)Gösta Granlund
• Christian-Albrechts-University of Kiel,
  GermanyCognitive SystemsGerald Sommer
• University of Surrey, UKCentre for Vision,
  Speech, Signal ProcessingVaclav Hlavac
• Czech Technical University in Prague, Czech
  RepublicCenter for Machine PerceptionJosef
  Kittler

4
Consortium
5
Objectives

• Architectures for autonomous robots with
  cognitive capabilities
• Vision based systems
• Stationary manipulator

6
Approach

• Action precedes perception
• Incremental learning and exploration
• Bidirectional interface to reasoning unit

7
Early Demonstrator

• Parallel progress in WPs
• Concentrate on architectural aspect
• Fully automatic learning and evaluation

8
Final Demonstrator I

• Embedding into real world
• Scalability of the architecture
• Stability and robustness

9
Final Demonstrator II
10
Just a toy example?
Fact this scenario can easily be solvedby a
traditional system design.
but ...

• Is it robust?
• Can it generalize?
• Is it scalable?

11
Robustness/Generalization
?
12
Challenge of COSPAL
is NOT to solve the puzzle problem
is NOT to do model registration and fitting
but to develop a system architecture whichis
able to semi-autonomously,step-by-step build up
knowledgeand solution strategies
13
Architecture

• Homogeneous network structures
• Localized representations (channels)
• Associations between channels to be learned

14
Architecture II

• Inhomogeneous network structures
• Compact representations
• Sub-net learning controlled through super-net

15
Hardware

• PCs
• Firewire cameras
• RX90
• 3D turntable

16
Software Interfaces

• Heterogeneous code modules
• C
• Matlab
• Python
• Loosely coupled processes
• Controlling process
• XML files
• Exception servoing (PVM)

17
Summary

• Autonomous cognitive system
• Action-precedes-perception learning
• Bidirectional interface to symbolic unit
• Incremental learning at all levels
• Demonstrator as proof of architecture
• Architectural building blocks for systems in more
  serious applications