Feedforward neural network for controlling qbmove maker pro variable stiffness actuator

Abstract

This paper presents an application of neural networks in control of cutting-edge robotic actuators. Namely, the latest version of robot actuators is inevitable multivariable and distinctively non-linear system in order to enable accurate but at the same time safe robot behavior. Therefore, non-linear elastic elements in transmission are desired to enable actuator stiffness variation. Inconsistency in serial production of such non-linear elastic elements as well as challenges in control of such complex mechanism, are solved using neural networks. Thus, efficient computation necessary for real time control is achieved. Superiority of feedforward control using neural networks over model-based feedforward are validated by experiments on the laboratory setup — qbmove maker pro actuator.