Goal-directed imitation through repeated trial-and-error interactions between agents

Imitation or learning by demonstration is one of the holy grails in robotics. Wouldn’t it be great to show a service robot how to do the laundry or to show an edutainment robot how to clear up the kids’ toys at the end of the day? However, the community is still at a stage where we can only dream from such applications. Most if not all researchers have been concentrating on sub problems of imitation, such as imitation using logic programming (Kuniyoshi et al. 1994), imitation between agents with different embodiments (Alissandrakis et al. 2004) or the metrics for deciding what to imitate (Billard, this volume); the work presented in this paper is not any different. In the last years insights from cognitive science and developmental psychology have provided new impetus in artificial imitation research (see among others Dautenhahn and Nehaniv (2002)). It is this work that inspired us to build a new model for agents to imitate the intention of observed behaviour. In previous work we have demonstrated how a group of agents can agree on imitating gestures (a class of imitation which Byrne and Russon (1998) have dubbed action-level imitation) and how we avoided to need for an explicit measure of imitation quality. This was done by letting the initiator judge if the imitator’s behaviour corresponded to the demonstrated action (Jansen 2003; Jansen et al. 2004; Belpaeme et al. 2005). In this paper we wish to present some of our ongoing research in goal-level imitation, where agents learn to imitate the goal of an action. This is a natural continuation of action-level imitation, and has been recognised as a notoriously difficult problem as the imitator needs to solve two difficult problems. First, it needs to understand the goal of the observed behaviour, something which is immediately obvious to us, but is hard to grasp if you not do have the necessary world knowledge to deduce the goal. If you see someone sweeping the patio with a broom, is the goal then the sweeping motion itself? Or the swishing noise made by the broom? Or the fact that the