Learning from Observation Paradigm: Leg Task Models for Enabling a Biped Humanoid Robot to Imitate Human Dances

Abstract

This paper proposes a framework that achieves the Learning from Observation paradigm for learning dance motions. The framework enables a humanoid robot to imitate dance motions captured from human demonstrations. This study especially focuses on leg motions to achieve a novel attempt in which a biped-type robot imitates not only upper body motions but also leg motions including steps. Body differences between the robot and the original dancer make the problem difficult because the differences prevent the robot from straightforwardly following the original motions and they also change dynamic body balance. We propose leg task models, which play a key role in solving the problem. Low-level tasks in leg motion are modelled so that they clearly provide essential information required for keeping dynamic stability and important motion characteristics. The models divide the problem of adapting motions into the problem of recognizing a sequence of the tasks and the problem of executing the task sequence. We have developed a method for recognizing the tasks from captured motion data and a method for generating the motions of the tasks that can be executed by existing robots including HRP-2. HRP-2 successfully performed the generated motions, which imitated a traditional folk dance performed by human dancers.