Accurate and agile control of a pneumatic robotic actuator by GP-based feedforward learning

Abstract

This paper proposes the combination of PD-feedback and learning-based feedforward control to solve reference tracking tasks in pneumatic actuators and soft-robotics with nonlinear dynamics and complex hysteresis characteristics. The feedforward control consists of a static gain and a hysteresis compensation, which are predicted by Gaussian Process models. The proposed method is validated on a pneumatic actuator, and the experimental results demonstrate the method’s capability to solve the reference tracking tasks, despite requiring only 22 seconds of training data. The results further demonstrate the potential of learning-based control for pneumatic actuators and soft-robotics, by superseding the need for laborious manual tuning and controller design.