2A1-Q09 Learning and Control Musculoskeletal Nonlinearity of Tendon Driven Musculoskeletal Humanoid

Abstract

In this paper, we propose a novel learning method to find relation from muscle-length to joint-angle and vise verse in a musculoskeletal robot, based on nonlinear bidirectional mapping between muscle-length and joint-angle. We show that mapping from joint-angle to muscle-length can be learned as linear combination of multiplications of joint-angle. This formulation enable us to extract not only muscle-jacobian but also EKF for posture identification in a musculoskeletal robot. Finally we applied the method to tendon driven musculoskeletal robot "Kenshiro" and verified its validity.