Active Learning of Joint Attention

Joint attention is the skill of attending to the same object another person is looking at. The acquisition of this skill is crucial in children for the development of many social and communicative abilities, and has been proposed as a critical social capability for interactive robots. Although recent attempts to model the acquisition of this skill on a robot have been moderately successful (Nagai et al., 2003; Triesch et al., 2006), they all assume that the robot remains passive during the learning process. Infants, on the other hand, have already acquired some rudimentary sensorimotor skills by the time they start to learn joint attention. We believe that these sensorimotor skills can jumpstart and considerably accelerate the learning of joint attention. In this paper we demonstrate on a humanoid robot how to use pointing and reaching to accelerate the learning of joint attention. We show that a robot can acquire this skill with a 95 % accuracy after a total of only 220 training samples compared to 85% accuracy after totals of 10,000+ samples in other approaches.