EMGBlend: Heterogeneity-Aware Self-Supervised Pretraining for Gesture and Force Decoding

A self-supervised framework that harmonizes heterogeneous surface EMG datasets through geometry-aware attention, band-limited spectral targets, and source balancing.

Yuwei Jia; Cheng Zhong; Jinyang Yu; Zhe Cui· 2026· DOI 10.48550/arXiv.2609.25582

The core problem

Surface electromyography (EMG) is a cornerstone signal for human-machine interfaces, enabling gesture recognition, force estimation, and prosthetic control. The research community has accumulated numerous public EMG datasets, but these resources are highly heterogeneous: they differ in electrode layout, channel count, frequency support, and overall size. A naive approach of simply mixing datasets for pretraining can misalign channel semantics, introduce spectral targets that some devices cannot observe, and allow large or high-channel-count datasets to dominate the learning process. EMGBlend is introduced as a self-supervised framework designed around these differences. Rather than concatenating data blindly, it combines shared channel patches with geometry-aware attention, restricts spectral targets to each recording's supported frequency band, and balances exposure across data sources. The authors pretrain a 109M-parameter model on 11 public EMG sources and evaluate it on gesture recognition, continuous-force regression, and contact classification. The central hypothesis is that explicit mechanism design—not simple concatenation—is the key to leveraging heterogeneous EMG corpora

Innovation

EMGBlend was evaluated on three downstream tasks: gesture recognition, continuous-force regression, and contact classification. Across these tasks, it consistently outperformed two control conditions: matched random initialization and waveform reconstruction (i.e., pretraining without the heterogeneity-aware mechanisms). **Gesture Recognition.** Multi-source pretraining with EMGBlend improved gesture recognition accuracy compared to random initialization and waveform reconstruction controls. Fixed-budget source controls confirmed that the gains are not merely due to increased data volume but to the heterogeneity-aware design. **Continuous-Force Regression.** For force decoding, EMGBlend remained competitive, though the improvements were less pronounced than for gesture recognition. The authors note that cross-person NinaPro force estimation remains difficult, indicating that generalization across individuals is still a challenge. **Contact Classification.** EMGBlend also showed advantages in contact classification, further demonstrating the transferability of the learned representations. **Ablations.** Ablation studies confirmed that each of the three mechanisms—geometry-aware

Surface electromyography (EMG) is a cornerstone signal for human-machine interfaces, enabling gesture recognition, force estimation, and prosthetic control. The research community has accumulated numerous public EMG datasets, but these resources are highly heterogeneous: they differ in electrode layout, channel count, frequency support, and overall size. A naive approach of simply mixing datasets for pretraining can misalign channel semantics, introduce spectral targets that some devices cannot observe, and allow large or high-channel-count datasets to dominate the learning process. EMGBlend is introduced as a self-supervised framework designed around these differences. Rather than concatenating data blindly, it combines shared channel patches with geometry-aware attention, restricts spectral targets to each recording's supported frequency band, and balances exposure across data sources. The authors pretrain a 109M-parameter model on 11 public EMG sources and evaluate it on gesture recognition, continuous-force regression, and contact classification. The central hypothesis is that explicit mechanism design—not simple concatenation—is the key to leveraging heterogeneous EMG corpora for transfer learning.

EMGBlend operates as a self-supervised pretraining framework with three core mechanisms tailored to EMG heterogeneity.

Why it matters

The results position EMGBlend as a principled approach to leveraging heterogeneous EMG datasets. The key insight is that naive concatenation is insufficient and can even be harmful due to misaligned channel semantics, unobservable spectral targets, and source imbalance. By explicitly designing mechanisms to handle these issues, EMGBlend achieves better transfer. The geometry-aware attention mechanism allows the model to share knowledge across datasets with different electrode layouts, effectively learning a unified representation space. The band-aware spectral targets prevent the model from chasing frequencies that are not present in certain recordings, which is a common pitfall when mixing datasets with different sampling rates. Source balancing ensures that smaller datasets are not drowned out by larger ones, promoting diversity in the learned features. However, the study also highlights limitations. Cross-person NinaPro force estimation remains difficult, suggesting that inter-individual variability in EMG signals is a significant barrier. Future work could explore personalization or domain adaptation techniques to address this. Additionally, while the model has 109M parameters, the computational cost of pretraining on 11 sources may be substantial, though the authors do not provide detailed efficiency analysis in the abstract. The broader implication is that as the number of public EMG datasets grows, frameworks like EMGBlend will be essential for harnessing their collective power. The code is available at https://github.com/tamanano/EMGBlend, facilitating reproducibility and further research. The work aligns with trends in self-supervised learning for biosignals, where heterogeneity is the norm rather than the exception.

Who should read this

CS practitioners and researchers

Opening member content…