Jul 2026· 2026 IEEE/ASME International Conference on Advanced Intelligent Mechatronics (AIM)· pp. 1-6· 0 citations· 26 references
Abstract
Electromyography (EMG) serves as a robust signal source for developing intuitive human-machine interfaces (HMIs). With the proliferation of commercial and prototype EMG armbands, there is a growing need to balance high-accuracy gesture recognition with hardware constraints. While deep learning architectures have achieved remarkable performance, they often require significant computational resources and high-end sensors. This study investigates an algorithmic-centric perspective, exploring whether sophisticated feature extraction (FE) can compensate for reduced hardware specifications. Using a 20-gesture dataset collected from 10 subjects, we evaluated performance across three hardware configurations: 500 Hz/8-bit (C1), 500 Hz/12-bit (C2), and 1000 Hz/8-bit (C3). We compared traditional FE methods against modern paradigms designed to capture spatial and spatio-temporal signal dynamics, specifically PHASOR, Myoelectric Temporal Patching (MTP), and the proposed WaveLSTM (Wave long short-term memory) framework. WaveLSTM embeds deep-learning-inspired processing schemes, i.e., utilizing short- and long-term memory components, into a computationally efficient pipeline with significantly lower overhead than standard deep learning models. Our results demonstrate that WaveLSTM achieves superior accuracy, exceeding 95% in a leave-one-trial-out validation scheme. Notably, while classical FE approaches exhibited lower overall recognition rates, WaveLSTM demonstrated hardware invariance, maintaining high performance regardless of changes in sampling frequency (p=0.275) or ADC resolution (p=0.941). These findings suggest that advanced spatio-temporal feature engineering can enable high-performance, complex gesture recognition on low-cost, low-power wearable devices without sacrificing accuracy.
Wearable electromyography (EMG) interfaces enable natural human-robot interaction, and low-density configurations are preferred for practical deployment due to their simplicity and cost-effectiveness. However, limited spatial information challenges reliable and timely pattern recognition. This study compares three segmentation strategies for three-channel EMG gesture recognition - fixed sliding window, event-gated adaptive thresholding, and Teager-Kaiser energy operator (TKEO) - evaluated on six gestures from eight subjects under leave-one-subject-out cross-validation with BiLSTM+Attention classification. Mean classification accuracies ranged from 59.17% (fixed window) to 64.39% (adaptive), with no statistically significant differences among methods (p > 0.18). Temporal evaluation showed mean onset-to-decision latencies of 113.9 ms (TKEO), 137.7 ms (fixed window), and 149.7 ms (adaptive) - all within the 300 ms real-time feasibility threshold. Cross-subject variability (26-88%) substantially exceeded inter-method differences, suggesting that subject-specific factors are a more critical performance determinant than segmentation strategy. These results are intended as a direction-confirming study for temporal processing design in low-channel wearable EMG systems; statistical generalization to larger populations is reserved for future work.
Electromyography (EMG)-controlled orthoses and wearable assistive systems require classifiers that combine accurate motion recognition, efficient embedded deployment, and support for local model formation. Long-term EMG deployment is affected by signal drift and inter-session variability, motivating architectures that can support post-deployment model updates. Local processing can reduce dependence on external computation and data transfer. To address these challenges, this work presents a hardware-targeted Hyperdimensional Computing (HDC) classifier for trainable EMG classification on the resource-constrained GateMate A1 FPGA from Cologne Chip. The proposed architecture performs on-device HDC model formation and inference directly on FPGA. Linear Discriminant Analysis (LDA) serves as a conventional offline-trained and FPGA-inferred baseline. The evaluation uses eight anonymized single-session EMG recordings from seven healthy participants and one participant with spinal cord injury with 32 channels and five motion classes and includes accuracy, robustness, and routed FPGA implementation analysis. Across all recordings, the per-recording best-case HDC configurations reach 95.20% classification accuracy, while the LDA baseline achieves 98.54% overall accuracy. Under controlled input perturbation with a standard deviation of σ=0.10, HDC retains 92.98% mean accuracy compared with 80.98% for LDA. A first board-level power measurement indicates an energy cost of approximately 0.40 mJ per inference sample and 0.87 mJ per training sample. The experiments demonstrate single-session on-device HDC model formation and inference, while longitudinal validation, fatigue robustness, electrode-shift robustness, and inter-session adaptation remain future work. The results indicate that HDC provides an architectural foundation for trainable wearable edge–AI systems with local model updates.
Surface electromyography (sEMG) signals enable intuitive human-machine interaction by capturing muscle activation patterns associated with hand gestures. However, the accurate recognition of hand gestures using sEMG signals still remains challenging due to the complex nature of the signal variations. This work proposes a wavelet-based deep learning framework for EMG gesture recognition using time-frequency representations. The segmented sEMG signal is converted into the Continuous Wavelet Transform (CWT) spectrogram to obtain the multi-scale muscle activation. Two lightweight deep learning frameworks are developed to effectively learn gesture representations from these wavelet features. The Wavelet Attention Convolutional Neural Network (WA-CNN) integrates scale and channel attention to emphasize informative frequency bands and EMG electrodes. The Wavelet State-Space Model (WSSM) incorporates efficient temporal modeling of gesture dynamics. Experimental results on the Ninapro DB2 dataset demonstrate that the proposed models outperform conventional machine learning and deep learning baselines. In particular, WSSM achieves 95% accuracy while maintaining low model complexity suitable for real-time wearable EMG-based interaction systems.
Mainak Ghosh, Anup Nandy· International Conference on...· 0 citations
Objective. Restoring intuitive and natural hand control remains a key challenge in neuroprosthetics. A promising approach is to decode continuous finger-joint angles from electromyography (EMG) signals, enabling dexterous interaction. Deep neural networks show strong potential for this task, but are often too computationally demanding for embedded deployment and are rarely validated in real-time. Approach. This study presents a unified two-phase framework for real-time decoding of 11 finger-joint angles from medium-density EMG signals. The framework integrates preprocessing selection, architecture optimization, and deployment-related constraints. A convolutional neural network (CNN) baseline and an adapted version of the dual predictive attractor-refinement strategy (DPARS) were evaluated both offline and under closed-loop real-time control, with online data further used to refine the models. Main results. DPARS achieved real-time performance comparable to CNN ( R2 = 0.750 vs 0.777), while reducing model size by 7 × (227.8 KB) and forward pass latency by 120 × (0.25 ms), resulting in approximately 6 × lower energy usage (0.0208 Wh) in embedded execution. Significance. These findings highlight the importance of integrated system design for multi-degrees of freedom EMG decoding, where performance emerges from the interplay between preprocessing, model design, and real-time constraints, enabling efficient embedded control and more intuitive, dexterous neuroprosthetics.
Oscar Osvaldo Soto Rivera, Vincent Alexandre Mendez, Aiden Xu et al.· Journal of Neural Engineerin...· 0 citations
Human Activity Recognition (HAR) using wearable sensors is relevant to rehabilitation, assistive robotics, and mobile health applications. This study presents (i) SDALLE, a publicly available multimodal dataset integrating surface electromyography (EMG) and inertial measurement unit (IMU) signals acquired using a DELSYS Trigno wireless system; (ii) a harmonized, subject-dependent comparison with the public ENABL3S dataset; and (iii) an exploratory Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) augmentation strategy for class balancing. SDALLE contains recordings of nine healthy male subjects walking, jogging, climbing stairs and descending stairs. EMG and IMU signals were synchronized, standardized, and segmented using 75%-overlapping windows. In the implemented pipeline, feature standardization preceded segmentation, after which the windows were randomly divided into 70% training and 30% testing subsets without participant- or trial-level grouping. Consequently, overlapping windows, including windows from the same participant and trial, could occur in both subsets. Under this window-level protocol, attention-enhanced CNN–LSTM models recorded approximately 98–99% accuracy on ENABL3S and approximately 99% on SDALLE. These values are protocol-specific and may be optimistic; they do not measure generalization to unseen participants or establish a portable model ranking. GAN augmentation increased the SDALLE row count by approximately 12% and changed average testing accuracy from 96.9% before augmentation to 96.5% afterward under the implemented protocol. The study therefore provides a dataset resource and an exploratory subject-dependent comparison of predictive and computational behavior. Participant-independent evaluation with training-fold-only normalization, model selection, and GAN training is required before drawing conclusions about cross-subject robustness or deployment readiness.
M. Farouk, M. F. El-Khatib, M. Awad et al.· Scientific Reports· 0 citations
This paper investigates wearable-based recognition of human activities and gestures to support Human-Robot Interaction (HRI) in object-handover and assembly-like scenarios. Electromyography (EMG) and Inertial Measurement Unit (IMU) signals were collected using a Myo armband, culminating in a novel dataset introduced as MAGIC-HRI (Multimodal Activity, Gesture and Intention Collection) with a large taxonomy of 53 movement classes, including Brazilian Sign Language (LIBRAS) numbers (0-9), hand gestures, object/tool handover actions (pick up/give/hold), tool-manipulation tasks, and generic assembly/idle motions, collected from 11 participants with 10 samples per class (530 samples per participant). Signals are segmented by detecting muscle activation via an EMG energy envelope, then processed using sliding windows; time- and frequency-domain features are extracted. Multiple classical classifiers are tuned via cross-validated grid search, with Random Forest as the strongest baseline. A Leave-One-Subject-Out (LOSO) protocol reveals a large generalization gap, indicating substantial subject dependence. A personalized adaptation experiment suggests that injecting a small number of samples from a new user can markedly improve recognition. Overall, the study contributes a broad, HRI-driven multimodal dataset, a rigorous evaluation emphasizing generalization, and practical evidence that personalization is likely required for robust deployment in practical HRI.
Ruan Rithelle Chagas de Faria Carminati, Giovanni Braglia, L. Biagiotti et al.· 0 citations