In the last ten to twenty years, there has been a growing interest among artificial intelligence researchers in the construction of so-called autonomous agents. Autonomous agents should be able to operate in largely unknown and unstructured dynamical environments without any need for human intervention. Previous techniques in AI were found to be inappropriate in the construction of such agents. This led to a rethinking of the underlying metaphors, and to the emergence of alternative perspectives on intelligent behavior. While the importance of this extended view on cognition developed in autonomous agent research is now widely recognized in cognitive science, autonomous agents themselves have received little attention as possible models of cognitive systems. This might be due to the fact that most of the autonomous agents that were developed until now are related to natural agents (animals, humans) only in a very broad, metaphorical sense. Autonomous agents are still rarely used as empirical models which aim at reproducing empirical data gathered from animals or humans. Starting from the successful example of biorobotics, where autonomous agents are used as empirical models in biology, we propose a new methodology to use autonomous agents as empirical models in cognitive science, called comparative cognitive robotics (CCR). The key features of comparative cognitive robotics are:

As a collaboration partner in constructing autonomous agents, we choose the comparative psychology of learning and adaptation.

We are mainly interested in those phenomena of learning and adaptation that can be found in a wide variety of species.

Empirical research with the model animal and the construction of the robot model work hand in hand. Work on the model inspires new experiments and new theories, and the empirical findings are used to update the model.

The model animal and the robot model are tested in the same or similar environments commonly used in comparative psychology, with the same or similar means of analysis. The match between those measurements is taken as an indication of the quality of the model.