Learning nonlinear muscle-joint state mapping toward geometric model-free tendon driven musculoskeletal robots

To control a musculoskeletal tendon-driven robot we propose a novel method to learn musculoskeletal nonlinear bidirectional mapping between muscle length and posture (joint angle) from a real musculoskeletal robot. We show the nonlinear musculoskeletal mapping from joint angle to muscle length can be learned as a linear combination of simple nonlinear functions. This formulation can be extended to posture estimation (mapping from muscle length to joint angle) by EKF (Extened Kalman Filter) and torque estimation by differentiation in a musculoskeletal robot. In this paper, we applied the method to tendon driven musculoskeletal robots and verified the validity.