Towards a framework to make robots learn to dance

A key motive of human-robot interaction is to make robots and humans interact through different aspects of the real world. As robots become more and more realistic in appearance, so has the desire for them to exhibit complex behaviours. A growing area of interest in terms of complex behaviour is robot dancing. Dance is an entertaining activity that is enjoyed either by being the performer or the spectator. Each dance contain fundamental features that make-up a dance. It is the curiosity for some researchers to model such an activity for robots to perform in human social environments. From current research, most dancing robots are pre-programmed with dance motions and few have the ability to generate their own dance or alter their movements according to human responses while dancing. This thesis explores the question Can a robot learn to dance? . A dancing framework is proposed to address this question. The Sarsa algorithm and the Softmax algorithm from traditional reinforcement learning form part of the dancing framework to enable a virtual robot learn and adapt to appropriate dance behaviours. The robot follows a progressive approach, utilising the knowledge obtained at each stage of its development to improve the dances that it generates. The proposed framework addresses three stages of development of a robot s dance: learning ability; creative ability of dance motions, and adaptive ability to human preferences. Learning ability is the ability to make a robot gradually perform the desired dance behaviours. Creative ability is the idea of the robot generating its own dance motions, and structuring them into a dance. Adaptive ability is where the robot changes its dance in response to human feedback. A number of experiments have been conducted to explore these challenges, and verified that the quality of the robot dance can be improved through each stage of the robot s development.