Combined Learning of Exoskeleton Mid-level Assistance Profiles and Low-level Control Parameters

Lower limb exoskeletons have the potential to enhance human mobility by reducing metabolic cost and improving rehabilitation outcomes. However, individualized and adaptive control remains a key challenge, as user-specific gait dynamics require exoskeletons to optimize both the mid-level assistance profile and low-level control parameters. Existing approaches often optimize the timing and amplitude of exoskeleton assistance but keep low-level control gains fixed, limiting the system's ability to reduce effort. This work introduces a hierarchical learning and control framework that combines learning of the assistance profile and low-level control parameters for a knee exoskeleton. Results demonstrate that combined optimization improves metabolic efficiency more than optimizing either level separately, achieving an 8.92 reduction in metabolic cost while maintaining stable gait. These findings highlight the importance of integrating both trajectory adaptation and low-level control learning for effective exoskeleton assistance.