Human communicative skills rely on a complex mixture of social and spatio-temporal perception-action processes. Autonomous virtual agents that interact with people could benefit greatly from socio-communicative skills; endowing them with such skills, however, could mean decades of person-years in manual programming. Enabling the agents to learn these skills is an alternative solution the learning mechanisms for this, however, are also well beyond the state of the art. Our goal is to develop new cognitive architectural principles to allow intelligent agents to learn socio-communicative skills by observing and imitating people in dynamic social situations on the job.

Skill development, in our approach, is a fundamental architectural feature: Learning happens through unique architectural constructs specified by developers, coupled with the ability of the architecture to automatically reconfigure itself to accommodate new skills through observation. As an appropriate and challenging demonstration the resulting system will control a virtual humanoid television host, capable of taking interviews with users and conducting a 10-minute TV program. By observing and imitating humans the TV-host will acquire increasingly complex socio-communicative skills; by the end of the project it will have reached socio-communicative skills comparable to an average human television show host. The agent's skills will be evaluated using a thorough, formal method.