Human-in-the-Loop Optimization for Knee Exoskeleton Flexion Assistance

Human-in-the-loop optimization (HILO) has been used to identify subject-specific assistive strategies and improve the performance of wearable exoskeletons. However, there is still a gap in research on HILO regarding knee exoskeleton flexion assistance. We present a HILO methodology that optimizes the flexion torque delivered by the knee exoskeleton. The cooperation mechanism of flexor and antagonist muscles assisted by the exoskeleton is first analyzed through the dynamic model. Furthermore, the online rapid metabolic evaluation function, including the electromyographic (EMG) signals of the semitendinosus (SEM) and vastus medial (VM), is designed based on the cooperative working mechanism of the antagonist muscle. The upper and lower controllers are designed respectively to realize the construction of the HILO closed-loop control system. Finally, we experimentally demonstrate the effectiveness of the HILO methodology proposed in this letter through EMG and metabolic indicators. In particular, for the knee joint, the time required to identify assistance strategies is significantly reduced by our protocol. This research will contribute to the design of fast convergence HILO methods from the perspective of human natural driving.

I. Introduction

With the development of technology, wearable lower limb exoskeletons are increasingly used in rehabilitation training and light mobility [1], [2], [3]. The higher the level of customization of the exoskeleton's assistance, the better its ability to adapt to changes in the participants' physiology [4], [5], [6]. The HILO is influenced by factors such as force loading accuracy, assist profile, optimization method, and objective function [7], [8]. Research on knee exoskeletons focuses primarily on control methods to ensure force loading accuracy [9], [10]. In terms of assist profile, the assist profiles to be optimized for the hip and ankle joints are similar to the statistical profiles [11], [12]. Due to frequent power conversion in the knee joint, the energy storage and release mechanism was initially used to construct the flexion-assisting profile [13]. Optimization strategies are widely used for HILO [7], [14]. Gradient descent mainly depends on the signal-to-noise ratio of the measured metabolic value [15]. Compared to the gradient descent, the convergence time of Bayesian optimization for a hip exoskeleton was reduced by more than half [11]. On the other hand, the covariance matrix adaptive evolution strategy was used to calculate the next generation of control law for the knee exoskeleton [16]. Metabolic consumption in the above studies is measured using indirect calorimetry, which still has limitations in speed, comfort, and generalization [17].