NeuroGAIN: Neuromechanical Generative Demand Forecasting Toward Optimal Control of Soft Hand Exoskeletons

Soft robotic interfaces are being investigated for safe, comfortable long-duration hand assistance and therapy, as rigid devices can be prohibitively large, heavy, power consumptive, and costly. However, optimal control objectives are difficult to satisfy on-line when neuromechanical and robot nonlinearities are coupled. In prior studies, we developed a neuromechanical measurement model of the hand, and segmented soft robot control into low-dimensional data-driven sub-problems seeking to address this issue. However, substantial training was required to solve them, with only near-optimal convergence due to the variable topology of human-robot interaction (HRI) during movement. There remains a need for robust ‘demand forecasting’ to enable robust HRI optimization during movement. In this work, we develop Neuromechanical Generative Autoencoding Impedance Networks (NeuroGAIN), a hybrid learning approach integrating Long-Short Term Memory (LSTM) and Variational Autoencoder (VAE) networks enabling simplification of the sub-problems by forecasting the topology of the state space before computation. We evaluated NeuroGAIN using data collected from five healthy adults performing hand poses with and without robot interaction. We find that the proposed method can predict topological features of the HRI space during movement with up to 84% accuracy and can generate latent features with minimal loss. NeuroGAIN enables robust prediction of HRI dynamics advancing on the state of art in neuromuscular modeling and estimation for soft rehabilitation robotics.