Emergent Gait Strategies Defined by Cluster Analysis When Using Imperfect Exoskeleton Algorithms

In operational settings, lower-limb exoskeletons may experience errors where an expected torque is not applied, impacting a user's gait strategies. In this study, we introduced an exoskeleton control algorithm with five different fixed error rates up to 10% error (90% accuracy). Participants (
N
= 22) walked with a bilateral ankle exoskeleton while completing a targeted stepping task. We assessed the impact of exoskeleton error rates on joint kinematics, muscle activation, and task performance using a k-means clustering algorithm (
$k=5$
) to define gait strategies, which were interpreted in the context of human-exoskeleton fluency. One of the five emergent strategies was considered fluent, where users minimized muscle activation and aligned with the exoskeleton's goal of reducing metabolic cost while also maintaining acceptable task accuracy. Three strategies had acceptable task error, but involved increased muscle activation about the hip or ankle, thus negatively impacting human-exoskeleton fluency. One strategy minimized muscle activity, but had unacceptable task performance. Some users transitioned from fluent to non-fluent gait strategies after using the controller with higher error rates. Understanding emergent gait strategies can inform the development of exoskeleton algorithms that support appropriate gait strategies and system use.