emg intention decoding

This study addresses the challenges of poor robustness and limited interpretability in surface electromyography (sEMG) signal decoding, which stem from high inter-subject variability and noise sensitivity. The authors propose a physiology-inspired discretization framework for sEMG: by aligning sliding windows to the minimal muscle contraction cycle, they extract ten-dimensional time–frequency features (e.g., RMS, MDF) and apply K-means clustering to generate muscle state tokens, establishing the first physiologically driven sEMG tokenization approach. Evaluated on the newly released multi-action, multi-muscle dataset ActionEMG-43, the method achieves high cross-subject consistency (Cohen’s Kappa = 0.82 ± 0.09) and yields Top-1 action recognition accuracies of 75.5% with a Vision Transformer and 67.9% with an SVM—substantially outperforming raw-signal baselines—while reducing input dimensionality by 96% and enabling interpretable analysis of movement quality.

Surface electromyography (sEMG)-based human–machine interfaces suffer from poor cross-subject generalization, reliance on time-consuming calibration, and high response latency. Method: This paper proposes a zero-shot, low-latency real-time intent detection framework. It introduces a self-supervised masked modeling strategy tailored for sEMG time-series signals, integrated with an online sequence segmentation mechanism to dynamically model muscle activation and achieve fine-grained alignment with user intent—enabling rapid onset detection and stable continuous tracking even during incomplete gesture execution. Results: Experiments demonstrate that, without subject-specific calibration, the method significantly outperforms existing zero-shot transfer approaches in cross-subject gesture recognition, achieving both higher classification accuracy (+8.2% average accuracy) and substantially reduced control jitter (−37% variance). This work establishes a practical, plug-and-play paradigm for intent decoding in wearable robotics and intelligent prosthetics.

This work addresses the insufficient semantic modeling of surface electromyography (sEMG) signals in human activity recognition by proposing LLM-sEMG, a novel framework that introduces large language models (LLMs) into sEMG understanding for the first time. The approach employs a language-guided mapping mechanism to transform continuous sEMG sequences into language-like symbolic sequences, thereby leveraging the rich action-related semantic knowledge embedded in pre-trained LLMs for intent recognition. Experimental results demonstrate that LLM-sEMG achieves high-accuracy activity recognition across multiple public sEMG datasets, confirming the effectiveness and generalization capability of large language models in interpreting non-linguistic physiological signals.

This work addresses the challenges of data heterogeneity, label scarcity, and the absence of a unified representation in electromyography (EMG) signals across users, devices, and tasks. To this end, we propose AEMG—the first large-scale self-supervised representation learning framework tailored for EMG. Our approach uniquely models neuromuscular dynamics as a “physiological language,” leveraging a Neuromuscular Contraction Tokenizer (NCT) to discretize continuous EMG signals into “words” and “sentences,” thereby constructing the largest cross-device EMG vocabulary to date. AEMG introduces a unified self-supervised pretraining paradigm that accommodates arbitrary channel topologies and sampling rates. Experiments demonstrate that AEMG improves accuracy by 5.79–9.25% under zero-shot leave-one-user-out evaluation and achieves over 90% few-shot adaptation performance using only 5% of target-user data.

This study addresses poor generalization in non-invasive surface electromyography (sEMG) gesture decoding, primarily caused by inter-subject and intra-subject signal variability. We propose modeling multi-channel forearm sEMG signals as covariance matrices embedded in the Symmetric Positive-Definite (SPD) Riemannian manifold—thereby explicitly characterizing the spatial synergy topology of sEMG for the first time. Within this geometric framework, unsupervised clustering and supervised classification are jointly realized without domain adaptation or subject-specific calibration. The method is computationally efficient and highly robust, achieving significant improvements over state-of-the-art baselines on mainstream benchmarks—particularly yielding a 12.3% average accuracy gain in cross-subject gesture recognition. Our core contribution is establishing an sEMG-to-SPD manifold mapping paradigm, offering a geometric learning approach to myoelectric decoding that is both interpretable and strongly generalizable.

研究使用深度神经网络从sEMG信号中解码中风后患者的手指运动意图，通过LSTM、CNN和GNN模型处理数据，实现对康复硬件设备的支持。

This work addresses the limited generalization of surface electromyography (sEMG)-based gesture recognition and user authentication across individuals, which stems from inter-subject neuromuscular variability. The authors propose, for the first time, a deep disentanglement model within a multi-task learning framework that explicitly separates task-relevant (gesture) representations from subject-specific (identity) features, thereby uncovering their fundamental differences in cross-day stability and underlying physiological mechanisms. Experimental results demonstrate that the proposed approach significantly improves gesture recognition accuracy in both cross-subject and cross-day scenarios. Furthermore, while the extracted identity-related components can support user authentication, their relatively low cross-day stability corroborates the efficacy and physiological interpretability of the disentanglement strategy.

This study addresses the lack of transferable self-supervised representations for surface electromyography (sEMG)-based continuous hand pose estimation by proposing the EMG-GPT framework. This method introduces, for the first time, a combination of frozen residual vector quantization (RVQ) and a causal Transformer to enable self-supervised pretraining on raw EMG signals. By preserving codebook geometric structures while learning sEMG temporal dynamics through deep autoregressive prediction of future discrete tokens, the framework effectively captures signal characteristics without supervision. Experimental results demonstrate that the proposed model achieves highly competitive performance in both hand pose regression and tracking tasks. These findings validate the feasibility of acquiring cross-task transferable representations relying solely on pure EMG pretraining, offering a promising direction for sEMG-based interaction systems.

本文介绍了emgforge，一种基于MRI的自动化端到端表面EMG模拟管道，用于解决现有模拟器几何固定或不完整的问题。

This work addresses the significant performance degradation in surface electromyography (sEMG)-based gesture recognition caused by electrode repositioning, variations in skin condition, or changes in limb posture across sessions, which hinders real-world deployment. The authors propose the first label-free cross-session sEMG decoding framework that enables transfer without requiring new annotations. It employs a layout-invariant universal encoder combined with unsupervised feature distribution alignment during inference to adapt to new sessions. Trained on a single annotated session, the model generalizes effectively to subsequent unannotated sessions. Evaluated on NinaPro DB6, the method achieves a cross-session macro F1-score of 0.688, substantially outperforming both the conventional per-user LDA baseline (0.540) and existing approaches that rely solely on source-domain data, thereby overcoming the need for repeated calibration.