Wearable Robotic Device Design and Emulation of Control Strategies Using Machine Learning

Chapter 1 includes the introduction. Chapter 2 introduces the development of a two-degree-of-freedom robotic ankle exoskeleton designed to assist with both plantarflexion and inversion-eversion. This device established the technical foundation for later studies by demonstrating precise torque control and improving the design of a robotic ankle-foot prosthesis (AFP) emulator for subsequent human-in-the-loop (HIL) experiments. Chapter 3 presents the first study on prosthetic personalization, applying human-in-the-loop Bayesian optimization (BO) to tune ankle stiffness in individuals with simulated amputation. The findings demonstrate that BO-driven tuning reduces the metabolic cost of walking more effectively than conventional weight-based tuning, motivating the search for alternative, clinically viable cost functions. Chapter 4 builds on the previous study by introducing the symmetric foot force-time integral (FFTI) as a pressure-based cost function for HIL optimization. The results show that FFTI correlates with metabolic cost and enables rapid, real-time tuning of prosthesis stiffness, improving gait symmetry and reducing walking effort. Chapter 5 extends this optimization approach to clinical prosthesis fitting by incorporating socket-interface pressure as a reward function for optimization. This study demonstrates that pressure-based tuning significantly improves comfort and gait mechanics, providing an alternative to traditional prosthetic fitting methods. Chapter 6 shifts focus to gait adaptation, analyzing how transtibial amputees adjust to prosthesis parameters over time. The results show inter-individual differences in adaptation trajectories, suggesting that prosthesis control should be dynamic rather than static, continuously evolving to accommodate long-term motor learning. Chapter 7 concludes with a discussion on the broader implications of real-time prosthesis personalization, synthesizing the findings from prior chapters and proposing future directions for adaptive prosthetic control. The dissertation culminates in a framework that integrates biomechatronic design with machine learning-based optimization, paving the way for the next generation of intelligent, user-responsive prostheses.