EMG-Based Simulation for Optimization of Human-in-the-Loop Control in Simple Robotic Walking Assistance

Exoskeletons offer promising solutions for enhancing human mobility; however, personalizing assistance parameters to optimize physiological outcomes remains challenging. Human-in-the-loop (HIL) optimization has emerged as an effective strategy for tailoring device control, often using electromyography (EMG) as a real-time proxy for metabolic cost. This study simulates HIL optimization using surrogate models built from the average root mean square of the muscles’ activations (EMG-RMS) derived from treadmill walking trials with a robotic waist tether. Nine surrogate models were evaluated for prediction accuracy, including gradient boosting (GB), random forest, support vector regression, and Gaussian process variants. Seven global optimization algorithms were compared based on convergence time, EMG-RMS at optimum, and efficiency metrics. GB achieved the highest predictive accuracy (1.57% RAEP). Among optimizers, the gravitational search algorithm (GSA) produced the lowest EMG-RMS value (0.17 normalized units) and the fastest convergence (0.32 s), while particle swarm optimization (PSO) achieved 0.36 EMG-RMS in 1.61 s. These findings demonstrate the value of EMG-based simulation frameworks in guiding algorithm selection for HIL optimization, ultimately reducing the experimental burden in developing personalized exoskeleton assistance strategies.