Electromyography-Based Force Estimation in Static and Dynamic Human–Robot Interaction

Abstract

This study presents a force estimation system for the X and Y axes based on surface electromyography (sEMG) signals acquired from eight channels of the Myo armband. The ground truth forces were recorded using an ATI 6-axis force sensor in two degrees of freedom (2-DOF) setup. The proposed system aims to estimate both static directional forces (X+, X-, Y+, and Y-) and dynamic movements, providing a framework for muscle force analysis. A total of five regression models—Fine Tree (FT), Interaction Linear (IL), Medium Gaussian Support Vector Machine (MG-SVM), Exponential Gaussian Process Regression (Exponential GPR), and Wide Neural Network (WNN)—were evaluated across sixty-three case combinations derived from six features in time domain descriptors: Root Mean Square (RMS), Integrated EMG (IEMG), Waveform Length (WL), Zero Crossing (ZC), Variance (VAR), and Willison’s Amplitude (WAMP). This resulted in the assessment of 630 models for both the X and Y axes. The performance of each model was assessed using the Root Mean Square Error (RMSE). Experimental results demonstrated that the Exponential GPR model achieved the highest accuracy in both axes when utilizing the RMS and IEMG feature combination, yielding an average RMSE of 1.0550 N. The WNN followed as the second-best model, with an RMSE of 1.2631 N using IEMG and VAR in both axes. In addition to predictive accuracy, model selection also considered computational aspects such as training time, model complexity, and inference speed. Among all candidates, the WNN offered the most balanced trade-off between accuracy and efficiency, highlighting its potential for real-time force estimation in human-robot interaction applications.