Physics-informed reservoir computing with autonomously switching readouts: a case study in pneumatic artificial muscles

We introduce an approach based on physics-informed neural networks to predict the length of a McKibben pneumatic artificial muscle (PAM) from a series of pressure measurements. We implemented an echo state network, which is a type of recurrent neural network with autonomously switching readouts corresponding to the different physical states of the PAM. The physical state we focus on in the current study is the direction of motion affected by hysteresis. The switching is realized by introducing gate architecture, whose states are also controlled by using the same recurrent network that outputs the length of the PAM. We demonstrated that handling the different physical states of the PAM by switching readouts will robustly yield performance in predicting the length of the PAM. We also demonstrated that Gaussian mixture models as a classifier for clustering the reservoir state autonomously and the results in classification are consistent with the physical state of the PAM.