Over the last two decades there has been a increase in sophistication in research and edutainment robots. From Honda’s Asimo series to Aldebaran's Nao, all are amazing feats of mechanical and electrical engineering. These robots can navigate around the house, mow the lawn and even walk. However, no robot recognizes its owner, nor is it able to read moods, or strike up a basic conversation or learn about its environment: current state-of-art robots lack cognitive and social capabilities. At the Centre for Interactive Intelligent Systems we investigate how robots could learn to imitate humans or even other robots, how they can learn about new concepts and how they can learn to communicate with humans.
We are studying the frontiers of cognitive and developmental robotics in the EU FP7 ITALK and CONCEPT projects.
Humans are a symbolic species, meaning that they cut up their perception of the world into manageable chunks. We are specifically interested in how humans handle this chunking of the very basic senses. I investigate the nature of perceptual categories, and in particular, the nature of colour categories. Colour categories exhibit some almost universal properties; for example, almost every culture has a word like "red", designating that almost everyone possesses a category for red; but few people have a word and category for "turquoise". There are several hypotheses why this is so. Some claim colour categories are in some way innate, others believe colour categories are learnt through interacting with the environment, while others believe that culture plays a role in shaping colour categories. I use artificial intelligence techniques to test drive these hypotheses, and the results lead me to believe that culture is the main catalyst for shaping colour categories (long abstract and PhD thesis).
Using genetic programming we evolved task-specific visual feature detectors, which were then used as the first input stage for visual discrimination algorithms. What basic functionality should be present? How can we generate this basic and more complex functionality? What kind of computational methods (genetic programming, genetic algorithms, or other optimisation techniques) can we use? How can evolutionary and self-organizational principles be applied to the generation and selection of visual feature detectors?
I have built and programmed several autonomous robots. My specific interest is in behaviour-based robots, robots that exhibit life-like behaviour with only a minimal amount of computation. I had a special interest in heterogeneous groups of autonomous systems, where we studied how robots with different bodies can cooperate and coexist in a shared environment.
I joined the AI Lab Robocup98 and RoboCup99 teams, headed by Andreas Birk. I constructed the vision system for our five-robot soccer team. The vision accurately tracked and predicted the position of ten robots (our team and the opponents) and a ball in real-time.
In the Talking Heads experiment is a large-scale demonstration of theories on the evolution of language. In the experiment robots observe the world and construct a language to describe the world to each other. The experiments demonstrated concepts on lexicon formation, meaning creation, and the self-organisation of language.