Hardware-Efficient EMG Decoding for Next-Generation Hand Prostheses

TL;DR

The paper addresses the challenge of deploying EMG-based hand prosthesis control on lightweight hardware by introducing DPARS, a hardware-efficient encoder-attention-attractor-refinement architecture. By reducing data dimensionality, focusing on salient temporal information, and combining a coarse attractor-based prediction with a fine-tuning refinement, DPARS achieves about $0.806$ mean $R^2$ accuracy with only $6828$ parameters, outperforming or matching larger CNN/LSTM models. The approach yields substantial hardware gains, including >4× MAC-operation reductions and 50–120× smaller models, enabling true on-chip, energy-efficient real-time control for next-generation RPHs. This work demonstrates the practical potential of compact AI decoders for dexterous, naturalistic prosthetic hand movements in daily use.

Abstract

Advancements in neural engineering have enabled the development of Robotic Prosthetic Hands (RPHs) aimed at restoring hand functionality. Current commercial RPHs offer limited control through basic on/off commands. Recent progresses in machine learning enable finger movement decoding with higher degrees of freedom, yet the high computational complexity of such models limits their application in portable devices. Future RPH designs must balance portability, low power consumption, and high decoding accuracy to be practical for individuals with disabilities. To this end, we introduce a novel attractor-based neural network to realize on-chip movement decoding for next-generation portable RPHs. The proposed architecture comprises an encoder, an attention layer, an attractor network, and a refinement regressor. We tested our model on four healthy subjects and achieved a decoding accuracy of 80.3%. Our proposed model is over 120 and 50 times more compact compared to state-of-the-art LSTM and CNN models, respectively, with comparable (or superior) decoding accuracy. Therefore, it exhibits minimal hardware complexity and can be effectively integrated as a System-on-Chip.

Figures (4)

• Figure 1: Dual Predictive Attractor-Refinement Strategy (DPARS) for real-time EMG decoding. The model includes an encoder layer for dimensionality reduction, followed by an attention network that highlights the most important sample in a data stream. Then, the attractor network generates a coarse prediction ($\hat{Y}_{\mathrm{attr}}$), while the refinement network refines the initial prediction by compensating for errors ($\hat{Y}_{\mathrm{refn}}$). The decoded finger angle ($\hat{Y}$) will be generated by summing the coarse and the refinement signal.
• Figure 2: Impact of Encoding Size on Decoding Accuracy: The blue line represents the mean $R^2$ accuracy against the encoding size, and the shaded area indicates the variance.
• Figure 3: Probability distributions generated by the attractor network. (a) Probability distribution spread among all possible states through minimizing decoder loss only. (b) The network extracts attractors. The probability of non-attractor states equals zero when the entropy regularization term is included in the objective function, leading to >4$\times$ less number of computations.
• Figure 4: The real (blue line) and decoded (orange line) figure angles predicted by the proposed DPARS model.