Bridging Human-Robot Co-Adaptation via Biofeedback for Continuous Myoelectric Control

Abstract

This letter proposes a novel human-robot coadaptation framework for robust and accurate user intent recognition, specifically in the context of automatic control in assistance robots such as neural prosthetics and rehabilitation devices empowered by electrophysiological signals. Our goal is to incorporate user adaptability early in the training phase to facilitate both machine recognition and user adaptability, rather than relying solely on brute-force machine learning methods. The proposed framework is featured by applying biofeedback-based user adaptive behavior into model training, while the machine can adapt to those changes through online learning. Specifically, this study focuses on the recognition of two-degree-of-freedom simultaneous and continuous wrist movement intentions based on surface electromyogram (sEMG) array signals, and the performance is tested on twelve able-bodied subjects. The coadaptive evaluation experiment demonstrates the robust control of this method by introducing sEMG electrode displacement as perturbations. Experimental results show that this method improves the completion time of centre-out tasks by 13% compared to conventional methods (Cohen's d=0.637), and debias 86% of the effect of electrode shift perturbations. This study provides insights into the potential for incorporating human adaptability into machine intelligence to improve user intent recognition and automatic robot control.

Gesture Classification in Electromyography Signals for Real-Time Prosthetic Hand Control Using a Convolutional Neural Network-Enhanced Channel Attention Model

“…Specifically, each convolutional layer is followed by a batch of normalization (BN) layer and rectified linear unit (ReLU) activation function to speed up convergence and address the problem of gradient explosion. To handle channel data from different muscle positions, the sizes of the convolutional kernels in the first, second, and third layers are set to (3,3), (4,3), and (5, 3), respectively. By sliding these kernels along the temporal and channel dimensions, the network can capture dynamic patterns in the temporal sequence and correlations among different channels simultaneously.…”

“…To address these issues, surface electromyography (sEMG) gesture recognition has been implemented into prosthetic hands, which has become one of the most promising technologies for smart prosthetics [2]. The sEMG gesture recognition is a non-intrusive technique that can accurately decode intended hand movements by measuring and analyzing electrical activities generated by muscular contractions and relaxations near the skin's surface [3], thus enabling the wearer to autonomously control myoelectric prosthetics [4]. However, to achieve a natural and stable autonomous control of myoelectric prosthetic hands to perform various fine hand movements such as grasping, pinching, and fingering, appropriate signal processing methods and the resulting fast, accurate, and reliable gesture recognition are crucial.…”

Predicting and Shaping Human-Machine Interactions in Closed-loop, Co-adaptive Neural Interfaces

“…In a neural interface, signals from the user are translated via a decoder algorithm to control a device. Decoder algorithms that adapt to users during interface operation can improve performance and enable personalization (Taylor et al, 2002; Danziger et al, 2009; DiGiovanna et al, 2009; Jarrassé et al, 2012; Orsborn et al, 2014; Shenoy and Carmena, 2014; Hahne et al, 2017; Brandman et al, 2018; Silversmith et al, 2020; Rizzoglio et al, 2021; Gigli et al, 2022; Hu et al, 2023). Users also adapt as they learn to control the interface because they receive real-time feedback (Carmena et al, 2003; Ganguly and Carmena, 2009; Fetz, 2007; Jackson and Fetz, 2011; Albert and Shadmehr, 2016).…”