A Gait-Conditioned Human-in-the-Loop Optimization Framework Based on Gait–Assistance–Response Relationship Model for Exoskeleton Assistance

The effectiveness of wearable exoskeletons critically depends on assistance patterns, yet their determination is challenging because physiological responses vary across individuals and gait conditions. Human-in-the-loop (HIL) optimization can personalize assistance, but its high experimental cost limits its application across diverse tasks. In this study, we developed a gait-conditioned HIL optimization framework based on a hierarchical gait–assistance–response model. The model characterizes assistance–response relationships using parameterized surfaces and predicts gait-dependent surface parameters through polynomial regression. The experiments focused on knee exoskeletons with extension assistance, where the physiological response was quantified as muscle activation based on normalized sEMG RMS from three major knee extensors on the assisted side. The model was built on data collected under predefined gait conditions and used to guide HIL optimization in validation experiments involving five healthy adults under unseen gait conditions. The validation results showed that, with the same CMA-ES optimizer but different configurations, two-generation model-guided HIL optimization showed a mean additional 6.4 percentage-point reduction in muscle activation compared with two-generation empirical HIL optimization, approaching four-generation empirical HIL optimization with only a mean 1.3 percentage-point difference. The preliminary results suggest the potential of the proposed strategy to reduce the number of optimization generations while maintaining comparable assistance performance by providing guidance for optimization configurations.