The goal of research in the field of cognitive architecture is to create general intelligent behavior. In contrast, the state-of-the-art in the field of Human-Robot-Interaction is to hand-craft systems that are specially tailored to certain scenarios, thus creating specific rather than general solutions. In this paper I subscribe to the view that we need to understand how robots can learn to interact, rather than to try to hard-code general intelligent social behavior. I further describe new trends in the field on how to implement a knowledge representation for a robot and provide further ideas of how such a knowledge representation could be learned. 1.