Beating-Time Gestures Imitation Learning for Humanoid Robots

Abstract

Beating-time gestures are movement patterns of the hand swaying along with music, thereby indicating accented musical pulses. The spatiotemporal configuration of these patterns makes it difficult to analyse and model them. In this paper we present an innovative modelling approach that is based upon imitation learning or Programming by Demonstration (PbD). Our approach-based on Dirichlet Process Mixture Models, Hidden Markov Models, Dynamic Time Warping, and non-uniform cubic spline regression-is particularly innovative as it handles spatial and temporal variability by the generation of a generalised trajectory from a set of periodically repeated movements. Although not within the scope of our study, our procedures may be implemented for the sake of controlling movement behaviour of robots and avatar animations in response to music.