Learning by demonstration for a dancing robot within a computational creativity framework

The paper presents a system that learns a set of movements for a creative dancing robot. A human user only dances in front of an 3D camera, and automatically the acquisition system segments the acquired sequence of postures depending on the detected music beat and rhythm. A clustering phase allows the system to group the identified actions in 20 classes, defining the set of movements that is typical of a given person. Analysis of the k-mean algorithm outcomes using different distances is reported. The human postures are translated in the corresponding robot joints configurations and are used to compose dance choreographies creatively. A cognitive architecture developed in previous works drives the process of dance creation. Experimentation shows the sets of movements derived from human users with different dance skills. Audience evaluates the robot performances based on these sets, and results are coherent with the quality and richness of the acquired movements.