A Comprehensive Analysis of Data Driven NN Models for EMG-Based Force Estimation in Rehabilitation Robotics

This work introduces data-driven estimators to enable EMG-based control of upper-limb rehabilitation robots. Surface electromyography (EMG) signals from key arm muscles are preprocessed to extract amplitude envelopes and synchronized with force to form a structured dataset. Three neural network models - NARX, RBF, and LSTM - are trained to estimate joint position and force from EMG, with NARX and LSTM capturing temporal dependencies and RBF approximating instantaneous mappings. Hyperparameters are optimized via a focused grid search to ensure stable, efficient performance suitable for real-time control. This work highlights the potential of neural network-based EMG decoding for subject-specific, intelligent human-robot interaction in therapeutic applications. the NARX model offers a superior balance of performance and computational efficiency suitable for real-time applications, over the LSTM and RBF NN based estimators.