Recognition and Generation of Leg Primitive Motions for Dance Imitation by a Humanoid Robot

We have attempted to realize a dancing humanoid robot which can imitate dance performances by a human. It is impossible to realize a performance including leg actions on a robot by direct conversion of motion data captured from a dancer. Constrains of robot legs and the dynamic balance must be considered. To solve this problem, this paper proposes a framework based on primitives of leg actions. First, a sequence of primitives is recognized from the captured human motion. Then motion data for a robot are generated from the primitive sequence. The process considers both the constraints and the dynamic balance. This framework brings high flexibility to adapt the performance to various stage conditions or recomposed choreography. Generated robot motions were tested in OpenHRP dynamics simulator. A robot successfully imitated dances for both the upper and lower body.