Characterizing Human-Exoskeleton Fluency for Co-Adaptive Control of Ankle Exoskeletons

Abstract

The performance of assistive and augmentative lower-limb exoskeletons has been tested in laboratory settings and can achieve goals of improved walking economy and metabolic cost reduction. However, it has been shown that even in controlled laboratory environments, exoskeletons may experience errors of improper torque assistance due to misalignments between the measured and actual state of the human-exoskeleton system. If the torque assistance repeatedly hinders a user's actions due to errors, the user may begin to anticipate errors and resist the exoskeleton. In order for exoskeletons to be adopted for everyday use, it is important to understand the immediate and long-term effects of exoskeleton errors on user motion and trust in the system, as errors are likely to occur in operational settings with complex, changing environments. It is also necessary to understand the strategies that people utilize when interacting with exoskeletons to design control methods to support collaboration between humans and exoskeletons. This work used the Dephy bilateral powered ankle exoskeleton, which applied an assistive plantarflexion torque during push-off to minimize energy expenditure during walking. The main aims of this thesis are to characterize the (1) immediate effects and (2) residual effects of exoskeleton errors on human gait strategies, as well as (3) develop a co-adaptive exoskeleton controller to support collaboration between the user and exoskeleton.

In Chapter 2, immediate compensatory hip behavior was identified in response to pseudo-random exoskeleton errors (loss of exoskeleton assistance) as users maintained acceptable task performance on a targeted stepping task. Quantitative measures of human-exoskeleton fluency---the alignment of the user and exoskeleton's goals---were developed using joint kinematics and muscle activity metrics. Emergent gait strategies were identified in Chapter 3 using k-means clustering as users walked with imperfect exoskeleton algorithms with fixed error frequencies (0-10% error in all strides) and were characterized as fluent or non-fluent. In Chapter 4, we designed and modeled a co-adaptive control algorithm that supports human-exoskeleton fluency by adjusting torque assistance in response to measures of muscle activity and joint kinematics along the lower limbs. The proposed co-adaptive algorithm successfully modulated peak torque in response to various fluent and non-fluent behaviors compared to an ankle-only controller that did not account for hip and knee compensatory strategies.

These results inform future exoskeleton controller design and evaluation metrics for human-exoskeleton collaboration. For example, developers may utilize measures of fluency during system development and testing. They also contribute to current literature on adaptation to exoskeletons, co-adaptive algorithms, and human-robot interaction metrics for the field of human-exoskeleton research.