Evaluating Optimality in the Control of Wearable Robotic Devices

Abstract

Lower-limb wearable robots, such as prostheses and exoskeletons, have the potential to fundamentally transform the mobility of millions of able-bodied and disabled individuals during work, recreation, and/or rehabilitation. Researchers have developed high-performance robotic devices that are lightweight and autonomous, yet we have not seen their widespread adoption as clinical or commercial solutions. One major factor delaying the dissemination of this technology is that our control of these devices remains limited, especially outside controlled laboratory settings. While many adaptive control strategies have been proposed and tested for wearable robotic systems, it is still an open question as to how to provide users with optimal assistance. This is a challenge because a clear definition of 'optimal assistance' has not yet been established, due to the various goals of assistive devices and the complex interactions between robotic assistance and human physiology. Furthermore, it has been difficult to move these systems outside the laboratory environment due to constraints imposed by tethered hardware and sensing equipment, as well as the difficulty associated with quantifying non-steady-state activities.

To address this challenge, the goal of the work presented in this dissertation is to further our understanding of how to provide users with optimal assistance from robotic exoskeletons and prostheses outside the laboratory environment. The first theme of this dissertation is the translation of experimental methods (e.g., human-in-the-loop optimization) and measurement tools outside the laboratory environment. The second theme is the evaluation of context-specific objectives for the optimization of wearable robotic devices (e.g., metabolic cost, user preference) in clinical and research settings. This dissertation comprises four primary projects.

First, I evaluated the impact that changing the power setting of the BiOM commercial powered ankle prosthesis has on the metabolic cost of transport of individuals with transtibial amputation. This work informs clinical tuning and prescription practices by revealing that, to minimize their metabolic energy consumption, individuals require a higher power setting than the setting chosen by the prosthetist to approximate biological ankle kinetics. Second, I investigated an alternative method for estimating instantaneous metabolic cost using portable wearable sensors. The goal of this project was to accurately estimate instantaneous metabolic cost without relying on indirect calorimetry, and thus enable the use of human-in-the-loop optimization algorithms outside the laboratory environment. I utilized linear regression algorithms to predict instantaneous metabolic cost from physiological signals, and systematically compared a large set of signals to determine which sensor signals contain the highest predictive ability, robust to unknown subjects or tasks. In the third project, I built upon this work and demonstrated that including the derivatives of physiological signals in the linear regression algorithm could improve both the speed and accuracy of instantaneous metabolic cost prediction from portable sensors. Finally, I investigated user preference as an objective for the control of lower-limb robotic exoskeletons. In this work, I demonstrated that exoskeleton users can quickly and precisely identify unique preferences in the characteristics of their ankle exoskeleton assistance in two dimensions simultaneously, and that preference changes with speed, exposure, and prior experience. These results provide insight into how users interact with exoskeletons, and establish important benchmarks for researchers, designers, and future consumers. Together, these four projects evaluate the notion of optimality in the control of wearable robotic systems and lay the foundation for translating optimization protocols outside the laboratory environment.