Neural Network Model of Continuum Soft Robots Based on WaveNets Equipped with Adaptive IIR

Soft robots have attracted attention from many areas, such as robotics, materials, automatic control, and artificial intelligence. The particular class of continuum soft robot (cSR) is the most challenging given its highly nonlinear dynamics from the hyperviscoelasticity and varying stiffness, which arise from the modeling of an infinite conglomerate of particles enclosed in a deformable finite-dimensional body based on the lumped parameters assumption and the D’ Alambert principle. The cSR is a complex model that presents a significant challenge in the field. Its nonintuitive structure, variable natural frequency, and parametric uncertainties, including its input matrix and lack of reliable velocity measurements, make it difficult to approximate its dynamics using only input and partial output data. This paper proposes a wavelet-based neural network (WaveNet) with an adaptive infinite impulse response (IIR) output filter. The WaveNet sweeps over modal frequencies, while the IIR enforces feedback and forward loops to propagate current and past neural outputs. The input shapes a controllable canonical output form. Simulations for a 2-degree-of-freedom CSR show that a bounded approximate modeling error is guaranteed at a low computational cost.