Learning and reproduction of valence-related communicative gesture

This paper proposes a robotic system capable of learning and reproducing robot gestures based on the Learning by Demonstration (LbD) approach. We focused on those gestures that are used for communicative purposes in human-human interaction. These gestures appear in various motions and this variation causes a delicate difference in the meaning and feeling that is delivered. While some (psychology and ethology) studies have shown that these variations are related to factors such as emotion, intimacy, and intensity, the best way to achieve robotic learning of these variations to allow for the reproduction of these motions remains unclear. With this motivation, we used the term `valence´ from psychology as a causal factor and tried to build a system capable of representing and learning relations between `valence´ factor and motion variation. Though there are many variations, we especially focus on the number of repetitions in this work. The system can segment a given motion into a set of unit motions by using states constructed by Gaussian Mixture Model(GMM) and Bayesian Network(BN) model is used to represent transition probabilities between states. In the model, transition probabilities are affected by `valence´ value and appropriate motion corresponding to given `valence´ value can be reproduced. Proposed system is applied to waving-hand motion of humanoid robot DARwIn-OP and we evaluate the validity of the system.