Unsupervised Model Learning of Pneumatic Musculoskeletal Robots with Inadequate Sensors

This paper provides an unsupervised model learning method for pneumatic musculoskeletal robots without torque sensors. In particular, we focus on learning the inverse dynamics model and employ multilayer perceptrons (MLPs) to capture the nonlinear and time-varying behaviors of pneumatic artificial muscles. To overcome the inadequacy of measurement devices that hinder the process of collecting the necessary input-output data for supervised training, a novel way to update MLPs is derived by coupling the process of parameter tuning with a model-based controller design scheme and solving for the condition that guarantees concurrent performance improvements of the controller and the learning model. The derived method applies to MLPs with any number of hidden layers and the word unsupervised refers to the fact that it is not built upon back-propagating a cost function. The effectiveness of our method is verified on a self-made musculoskeletal robot joint with one degree-of-freedom and the experiment results show that the learned model is accurate enough for both model-based trajectory tracking control and future state prediction.