Toward Load-Robust Motion Estimation Using an EMG-Driven State-Space Model With a Variable Stiffness Musculoskeletal Model

Accurate and robust human motion estimation is essential for enabling effective electromyography (EMG) signal-driven neural-machine interfaces in daily activities. Variation in loading weights is one of the critical factors affecting the performance of EMG-based interfaces. Although the robustness of EMG-driven musculoskeletal models (MMs) has been verified under diverse loads, the impact of load-induced changes in muscle co-contraction levels remains largely unaddressed, often leading to performance deterioration. To address these limitations, we proposed an EMG-driven state-space model to estimate hand and wrist movements without requiring additional training across different loading conditions. In this model, we developed an MM with variable joint stiffness as the state model. For the observation model, a back-propagation neural network was employed to map state variables to EMG features. The observation was a set of EMG features exhibiting robustness to load variations. Comprehensive experiments were conducted under four distinct loading conditions. The results demonstrated that the proposed method, trained exclusively with zero-load data, achieved estimation performance comparable to conventional MMs trained on load-specific data, while significantly outperforming conventional MMs trained with zero-load data. The outcomes validated the effectiveness of our method in improving the robustness and accuracy of EMG-based interfaces across varying loading conditions.