Diffusion-Driven Deep Decoding: Advancing EMG-Based Hand Skill Learning for Environment-Free Human–Robot Interaction

Human labor strategies are highly complex, adaptable, and versatile across diverse scenarios, making their decoding essential for developing human-like robotic capabilities. Current optical and wearable-based skill learning methods face limitations due to occlusion, lighting, and interference with natural hand movement, reducing practicality in real work environments. Although electromyography (EMG)-based methods capture rich motion information without such environmental constraints, the nonlinearity and variability of EMG signals, combined with the complexity of hand movements, present significant challenges for accurately decoding both hand motion and object interaction information. In this article, we address the long-range dependencies in EMG signals by constructing an attention-based transformer processing structure, introducing latent space representations, and applying a diffusion strategy to achieve deep decoding of EMG signals. For the first time, this novel perception algorithm achieves synchronous tracking of dexterous finger movements and estimation of objects with similar shapes and sizes, achieving an angular tracking error of 1.599$^\circ$ and an object recognition accuracy of 92.7%. Compared with baseline methods, the proposed estimation algorithms achieved an error reduction of up to 83.1%, validating its feasibility as a next-generation paradigm for decoding human labor strategies and paving the way for more human-like robotic behavior in future applications.