Wearable Robots Personalization Using Physiological Feedback with Machine Learning Approach

Wearable robots have made signicant progress in recent years, expanding their application domain to gait rehabilitation, human augmentation, and injury prevention. However, due to the complex human-robot interaction between the device and subjective behavior, the device's eectiveness is inconsistent. One approach to addressing this inconsistency is online personalization of the device, such as human-in-the-loop optimization. However, long optimization times, non-portable setups, and a lack of adaptation knowledge have limited optimization to lab-based short-term setups. I mainly address this limitation using a machine learning-based approach in this work. One of the main limitations of HIL optimization is the time it takes to complete the process, which is caused by the slow cost estimation process. The metabolic cost, which is the main cost function in the optimization process, is still limited due to its low signal-to-noise ratio and slow dynamics. Addressing this limitation would reduce the overall optimization time and make optimization accessible to a wider range of people. In this work, I have developed a data-driven model-free cost acquisition system that uses novel formulations, such as phase plane representation, to estimate the steady-state metabolic cost. I have also tested this fast estimation with the new HIL setup for squatting applications. The new setup uses a 2-degree-of-freedom ankle exoskeleton to assist squatting using an oboard control system. The optimizations used the new proposed metabolic cost estimation process and were able to predict the optimal condition of subject squatting. To develop an optimization system compatible with outdoor applications and multiple activities, I also developed an activity recognition system that uses a single accelerometer to identify the activity the subject is performing and provide corresponding assistance. In addition to this, I have also developed an open HIL optimization framework. This optimization framework is compatible with multiple cost functions, acquisition functions, and varied exoskeleton and prosthesis devices. In this work, I have also tested portable cost functions, such as ECG, and tested the compatibility of the ECG as the cost function in the optimization procedure. I have developed a continuous optimization system to address another limitation of personalization: adaptation. Given the importance of adaptation while the subject is wearing the exoskeleton and the importance of optimizing subject-specic speed, I have developed a continuous optimization framework that uses previous experience and meta-learning to address the subject's adaptation dynamics. In summary, this work addresses the three main limitations of the current exoskeleton personalization methods: optimization speed, using the fast metabolic cost estimation method; portability of the optimization, using the activity recognition system, portable cost function, and congurable portable optimization framework; and adaptation dynamics of the subject, using continuous optimization.