PaperlayerView exact revision on arXiv

Source-linked AI summary

Data-Centric Neuromotor Interfaces for Portable Human-Machine Interaction

Jiaxuan Li, Di Wu, Jianhua Liu, Yuxin Zhao, Jinnuo Li, Xiao Zhang, Zhenzhi Ying, Changsheng Dai, Xiang Li, Liming Shu

arXiv:2608.30301v1cs.RO

TL;DR

Constrained edge devices make complex neuromotor-interface models difficult to deploy, motivating a data-centric approach that improves physiological signal quality and separability. The paper combines wireless high-bandwidth acquisition with muscle-specific sEMG electrodes and an ultra-compact classifier. The resulting 2,210-parameter model reaches 94.36% accuracy across 34 gestures, while the authors identify limited participant diversity and controlled laboratory conditions as important scope boundaries.

  • Problem

    Complex model-centric architectures create computational overhead that conflicts with resource-constrained edge deployment, while practical links between signal quality and algorithmic complexity remain insufficiently established.

    Source: Introduction · pp. 2–3

  • Method

    The study develops a data-centric forearm sEMG interface using high-quality wireless acquisition, muscle-informed electrode arrays, and compact linear gesture-recognition models.

    Source: Introduction · pp. 3–4, HMI Offline Performance Evaluation · p. 11

  • Results

    94.36% accuracy across 34 gestures was achieved with a 2,210-parameter, 17.68 kB model, compared with approximately 45% under traditional data configurations.

    Source: Discussion · p. 22, Conclusion · p. 24

  • Takeaways & Limitations

    The findings support data-centric neuromotor-interface design as a validated pathway for compact, edge-deployable dexterous decoding.

    Source: Discussion · p. 22, Conclusion · p. 24

    The findings are based on a small cohort of healthy, right-handed young adults in a controlled single-day laboratory environment, leaving broader populations and real-world perturbations for further validation.

    Source: Limitations · pp. 23–24

Abstract

from arXiv · show

Dexterous human-machine interaction requires intuitive and expressive interfaces that can be efficiently deployed on constrained edge devices. Flexible material-based neuromotor interfaces hold considerable promise, as they decode human movement intention into natural control. Although emerging flexible electronic skins enable wearable high-fidelity data acquisition, practical deployment inevitably involves trade-offs between computational resources and portability. We present a data-centric paradigm where physiological features yield fundamental separability, providing sufficient discriminative cues for recognition. A wireless, high-bandwidth system developed for collecting various electrophysiological signals, when integrated with muscle-specific electrodes, forms a surface electromyography-based interface. Exploiting highly separable data, a 2,210-parameter model achieves 94.36% accuracy across 34 gestures and can be rapidly deployed on edge devices, establishing a new thousand-parameter benchmark for dexterous decoding. The underlying data-algorithm interactions in the data-centric paradigm are further clarified, demonstrating its feasibility in real-world scenarios. This study provides a principled and validated pathway for practical deployment of reliable neuromotor interfaces.

Introduction

The study addresses edge-deployment barriers caused by model-centric complexity by exploring whether higher-quality physiological data can provide separability for compact neuromotor-interface models. It develops and validates a data-centric forearm sEMG interface for gesture recognition.

Research motivation

  • Model-centric approaches pursue accuracy with increasingly intricate architectures, creating computational overhead that conflicts with constrained edge hardware.The passage links massive parameter scales and sophisticated computation to limited hardware capacity and deployment difficulties.Source: Introduction · pp. 2–3
  • Data-centric design improves data quality, consistency, and coverage, while this study emphasizes intrinsic spatiotemporal structure and separability rather than large-scale acquisition.The proposed pathway targets improved quality of original physiological data as an alternative to costly large-scale collection.Source: Introduction · p. 3
  • Systematic references for electrode-array electrical parameters and the relationship between data quality and algorithmic complexity remain insufficient.These gaps motivate practical, portable, high-precision neuromotor-interface designs.Source: Introduction · p. 3

Study approach

  • The interface uses forearm sEMG to test whether physical spatiotemporal fidelity simplifies feature geometry enough for ultra-compact linear recognition models.The authors distinguish this contribution from proposing another high-density sensor or lightweight algorithm alone.Source: Introduction · pp. 3–4
  • High-quality signals provide fundamental separability that supports reliable gesture recognition on resource-constrained edge devices.The paper defines fundamental separability as inherent discriminability among motion-category feature distributions in raw high-fidelity physiological signal space.Source: Introduction · pp. 4–5

Materials and Methods

The authors build a wireless, multi-channel physiological acquisition system with two flexible forearm electrode layouts and compare their signal characteristics for gesture-interface development. The design combines muscle-informed coverage with a distributed D64 array and a conventional R64 reference.

Acquisition system

  • The wireless unit captures sEMG, ECG, EEG, and EOG through 64 channels at 2000 Hz using a pluggable electrode connection.It uses common-ground differential sampling with a wristband reference electrode.Source: sEMG Acquisition Unit · p. 5

Interface architecture

  • The interface wirelessly forwards 64-channel sEMG from the electrode array to edge hardware, where features are classified into gestures and converted into actuator commands.The processing pipeline connects muscle activity acquisition with device control.Source: Data-centric HMI-oriented neuromotor… · p. 6

Electrode design

Gesture set

  • 34 gestures span relax, single-finger, multi-finger, wrist, and thumb groups, covering both single-degree-of-freedom and higher-degree-of-freedom movements.The set includes daily functional movements with varying recognition difficulty.Source: Gesture Group and Electrode Design · p. 6

Electrode design

  • Electrode placement balances deep-muscle signal extraction with broad coverage because forearm muscles differ in size, depth, and spatial overlap.The passage highlights both large-signal muscles near the elbow and smaller, overlapping distal muscles involved in thumb and finger motion.Source: Gesture Group and Electrode Design · pp. 6–7

Signal characterization

  • D64 captures localized muscle activation with prominent inter-channel variation, while R64 produces more evenly distributed signals with a narrower activity range.The layouts therefore differ in spatial differentiation and signal-amplitude consistency.Source: Gesture Group and Electrode Design · pp. 9–10

Data Preprocessing and Feature Extraction

The edge-device pipeline filters raw sEMG, extracts windowed RMS features, normalizes them, and maps them to gesture categories. Session-based validation and participant protocols are used to evaluate recognition without data leakage.

Preprocessing

  • Raw sEMG is bandpass-filtered from 20–500 Hz, followed by notch filtering to attenuate 50 Hz power-line interference and harmonics.The preprocessing removes task-irrelevant noise before feature extraction.Source: Data Preprocessing and Feature… · p. 10

Feature extraction

  • RMS features are extracted from every channel using a 200 ms sliding window and 100 ms step size.The features summarize windowed signal magnitude across channels.Source: Data Preprocessing and Feature… · p. 10

Classification

  • The resulting features are Z-score normalized and classified into predefined gesture categories with Logistic Regression before control commands are transmitted.Classification outputs are used to operate interaction devices.Source: Data Preprocessing and Feature… · p. 10

Evaluation protocol

  • One independent session was reserved for validation, while the remaining 10 sessions used 10-fold session-based leave-one-out cross-validation.This design was used to monitor hyperparameters and prevent data leakage.Source: Data Collection · p. 11

HMI Offline Performance Evaluation

The evaluation examines how sEMG data properties and low-complexity classifiers support gesture recognition, with accuracy assessed across 34 gesture classes and tested in online HMI scenarios. The study also evaluates edge deployment using a lightweight Logistic Regression model on ESP32-S3 hardware.

Data and evaluation design

  • Five temporal sampling rates and varied spatial sampling rates and channel layouts were systematically evaluated for their effects on sEMG data quality and decoding performance.The baseline used eight circularly arranged channels at 200 Hz, while raw 2000 Hz signals were downsampled to 1000, 500, 400, and 200 Hz.Source: HMI Offline Performance Evaluation · p. 11
  • Seven low-computational-overhead classifiers were compared to test whether high-quality data supports accurate recognition with low-complexity algorithms.The classifiers included DT, KNN, RF, LDA, SVM, MLP, and LR.Source: HMI Offline Performance Evaluation · pp. 11–12
  • Accuracy was calculated for multi-class recognition across all 34 gesture classes, with Wilcoxon signed-rank tests used for statistical significance analyses.The evaluation counts correctly classified samples across the total gesture categories.Source: HMI Offline Performance Evaluation · p. 12

Online HMI evaluation

  • Two online experiments evaluated feasibility: 34-gesture recognition and 20-gesture control of a dexterous robotic hand.The 20-gesture robotic-hand task was limited by the hand's degrees of freedom.Source: HMI Online Performance Evaluation · p. 12

Edge deployment

  • The Logistic Regression model was trained on a workstation and deployed to the ESP32-S3 edge device for HMI operation.The ESP32-S3 integrates Wi-Fi, Bluetooth 5, dual-core processing, and 512 KiB SRAM.Source: HMI Online Performance Evaluation · pp. 12–13

Statistical analysis

  • A multiple linear regression model estimated main effects, two-way interactions, and three-way interactions among data properties and subjects, with ω2 quantifying explained variance.Higher ω2 values indicate stronger effects on classification accuracy while accounting for residual error.Source: Analysis of Variance · p. 13

Results and Discussion

Improved signal quality increased feature separability, enabling compact models to maintain high gesture-recognition accuracy and operate on constrained edge hardware. Online experiments further validated the interface for real-time recognition and robotic-hand control.

High-Quality Data Empowers Low-Complexity Algorithms

  • Higher TSR and SSR improved decoding, whereas classifier responses to channel layout depended on the SSR condition.TSR improvements produced comparable gains across classifiers, while SSR affected different algorithms unequally.Source: High-Quality Data Empowers… · p. 14
  • TSR and SSR produced the strongest effects on classification accuracy, while channel-layout variation explained approximately 4% of the variance.The results identify temporal sampling rate, spatial sampling rate, and individualised electrode configuration as distinct contributors to decoding performance.Source: High-Quality Data Empowers… · p. 15

Data Separability Determines Decoding Capability

  • Improved data quality produced clearer gesture-group boundaries and progressively lower DBI values in feature space.The reported changes were associated with stronger intra-class clustering and richer spatial or temporal physiological information.Source: Data Separability Determines Decoding… · pp. 16–17
  • FSI tracked classification accuracy more closely than raw DBI, with linear-regression R^2 = 0.985 and Spearman correlation ρ = −0.992.The fifth-root channel-scaling factor was empirically selected, and FSI decreased as data configurations improved while accuracy increased.Source: Data Separability Determines Decoding… · p. 18

Deployment Validation

  • 99.14% online accuracy was obtained for 34 gestures, while robotic-hand control reached 98.99% overall accuracy.The 34-gesture test used post-processing, and the robotic-hand task evaluated 20 gestures on an ESP32-S3 platform.Source: Deployment Validation · pp. 20–21
  • On an ESP32-S3 at 160 MHz, the 2,210-parameter model required 4.39 kFLOPs, achieved 0.28 ms single-sample latency, and reached 3,563.8 Hz throughput.The device has 512 KiB of SRAM, and the full model occupied 17.68 kB of memory.Source: Deployment Validation · p. 21

Discussion

Conclusion

The study shows that improving hardware-level sEMG signal quality enables highly compact models to decode dexterous gestures on resource-constrained edge devices. Its data-centric framework links high-fidelity acquisition with practical, resource-efficient neuromotor interface design.

  • 94.36% offline decoding accuracy was achieved for a dexterous multi-intent task after elevated TSR, enhanced SSR, and optimized CL improved sEMG data quality.The ultra-compact model improved from approximately 45% to 94.36%.Source: Conclusion · p. 24
  • 2,210 parameters and a 17.68 kB memory footprint enabled initial real-time online validation on resource-constrained edge devices.The authors note potential further compression through quantization or reduced precision.Source: Conclusion · p. 24
  • High-fidelity spatio-temporal acquisition captured steadier signal fluctuations, subtle spatial differences, and large-scale muscle activation patterns.Source: Conclusion · p. 24
  • The data-centric paradigm shifts algorithmic requirements toward leveraging intrinsic separability in high-quality physiological signals for lightweight neuromotor interfaces.Source: Conclusion · p. 25

Authors’ contributions

The authors’ contributions are distributed across study conception, analysis, device development, data acquisition, support, supervision, funding, and project administration.

Availability of data and materials

The data supporting the findings are available from the corresponding author upon reasonable request, alongside hardware design specifications.

AI and AI-assisted Tools Statement

The authors used Gemini 2.0 for language polishing, grammar refinement, and assistance generating specified graphical elements. They reviewed and edited the AI-generated content and accepted responsibility for the published material.

Financial support and sponsorship

The study received support from two Chinese funding bodies, which had no involvement in study design, data processing, or manuscript composition.

Declarations

The study received ethics approval, followed the Declaration of Helsinki, and obtained participant consent for participation and publication of identifiable facial images.

Loading 2608.30301v1…