Abstract
Probabilistic methods, such as generative networks, are being explored to synthesize robust input spaces for optimal control strategies with minimal computational overhead. These methods can be used to compensate for nonlinear couplings between biological movement control and robot nonlinear dynamics, such as in soft robotic mechanisms for lightweight, low-cost rehabilitation exoskeletons. In our prior study, we presented NeuroGAIN — Neuromechanical Generative Autoencoding Impedance Network — an offline learning approach integrating Long-Short Term Memory (LSTM) and Variational Autoencoder (VAE) to forecast neuromuscular demand and the corresponding movement kinematics of a wearer using a neuromuscular measurement model conditioned on surface electromyography (sEMG) signals. In this work, we implement NeuroGAIN in an online sequential quadratic programming-based model-predictive control (MPC) framework for robust HRI optimization. We evaluate NeuroGAIN-MPC in an experiment with five healthy participants performing functional hand tasks while wearing a soft hand exoskeleton to find that NeuroGAIN-MPC enables robust optimal control of HRI dynamics with minimal analytical modeling, a key step toward developing individualized interventions.
Funding Information
  • Alabama Space Grant Consortium (NNH19ZHA001C)

This publication has 40 references indexed in Scilit: