History of Robotics Research and Development of Japan2005Integration, Intelligence, etc.Reproducing Human Dance Motions by a Biped Humanoid Robot

Shin'ichiro NakaokaNational Institute of Advanced Industrial Science and Technology (AIST)
Atsushi NakazawaKyoto University
Fumio KanehiroNational Institute of Advanced Industrial Science and Technology (AIST)
Kenji KanekoNational Institute of Advanced Industrial Science and Technology (AIST)
Mitsuharu MorisawaNational Institute of Advanced Industrial Science and Technology (AIST)
Hirohisa HirukawaNational Institute of Advanced Industrial Science and Technology (AIST)
Katsushi IkeuchiThe University of Tokyo
A possible way of generating motions of humanoid robots is to mimic target motions demonstrated by humans. This can be an efficient and intuitive solution to the complex motion generation problem in which a number of joints of humanoid robots should move in a coordinated way. The solution is also useful when a robot should move like a human or the original motion itself has a significant value as dances. This study achieved the above way of the motion generation for the whole body motions of a biped humanoid robot. A biped-type robot must keep the dynamic balance of the whole body in order not to fall down. This is not generally achieved if the robot just tries to follow the original human motion trajectory, because the structure and dynamic balance of the robot body is not necessarily same as the original human body. The problem was solved by applying the “learning from observation” paradigm to the whole body motion of a biped humanoid robot. Leg task models are designed based on the contact state between the robot feet and the floor, and adapting human motions to the robot is done by two steps consisting of the task recognition and generation based on the task models. First, a sequence of leg tasks such as both-leg standing, left-foot step, right-foot step, and squat is recognized from captured human motion trajectories. Then the lower-body motion of the robot is recomposed on the robot model by referring to the task sequence. Since the dynamic balance and mechanical constraints of the robot is taken into account in the recomposition process, the resulting motion becomes dynamically stable one and the robot can perform it without falling down. The characteristics of the original motion are also preserved by keeping the key timings and positions represented by the task models. We developed the software system that process the above method, and it enabled biped humanoid robot “HRP-2” to successfully reproduce a motion of Japanese "Aizu-banndaisan" folk dance performed by human dance master. This was the first achievement of reproducing human whole body motion including leg motions by a biped humanoid robot. 22th RSJ Best Paper Award in 2008. 画像の認識・理解シンポジウム (MIRU) 2005 インタラクティブセッション優秀賞 in 2005.
Aizu-Bandaisan Dance Performance by a Dance Master and HRP-2
Overview of the Developed Method

Related Article