Towards adaptive optimal personalization control of robotic hip exoskeleton assistance for individuals with paretic stroke.

Abstract

Neurological disorders, like stroke, can re duce the mobility and quality of life for millions of indi viduals across the world. As promising gait assistive de vices, powered lower-limb exoskeletons have gained wide attention for their potential to improve walking function for individuals with mobility impairments. However, due to the complex human-robot interaction, highly nonlinear dynamics modeling, and person-to-person variability, it is challenging to provide optimal customized assistance from these exoskeletons. In the current study, we showed that a reinforcement learning (RL)-based control framework pro vided adaptive and optimal personalized assistance across the hip joint to improve walking function for individuals with and without mobility deficits. We selected hip kinematic spatial gait symmetry as the control objective for five par ticipants with hemiparetic stroke, and found three optimal control parameters over an average of 201.7±36.1 sec. The kinematic spatial gait symmetry was defined as the joint excursion symmetry on both hip joints within the same gait cycle. By adding optimal assistance, we observed that the joint excursion during gait cycles at the affected hip joint was increased from 37.09° to 48.03° by 29.5% (p = 0.003) and the joint excursion gait symmetry was improved from 7.55% to-0.63% (p = 0.002) compared to transparent mode (no assistance). The findings suggest that the application of an RL-based assistance personalization control scheme to a wearable hip exoskeleton may positively impact gait rehabilitation training outcomes in a clinical population.