CAREER: Reinforcement-Learning Assist-As-Needed Control For Robot-Assisted Gait Training

点击翻译按钮获取中文摘要

This Faculty Early Career Development (CAREER) grant will develop adaptive, assist-as-needed controllers for a powered ankle brace (orthosis). Such systems may promote the learning of desired gait patterns during physical rehabilitation. Reinforcement learning will be used to shape person-specific control policies that balance movement error and user effort. The approach accounts for each user’s ability to learn and their existing patterns of inter-limb coordination. The control methods will be tested by people walking on a treadmill at constant speed, and over-ground at self-selected speeds. Research participants will include healthy people and a small group of stroke survivors with gait deficits. This project will promote the progress of science and advance the national health by developing a new intelligent ankle-foot orthosis controller. This new controller promises to improve gait retraining outcomes in stroke survivors. The project includes a plan to advance engineering education. It will also spread knowledge of wearable robots at a STEM camp offered to underrepresented middle school children.This project will develop novel assist-as-needed control methods for powered ankle orthoses. There are two main research objectives in this project. The first will establish new reinforcement learning (RL-based) control strategies capable of self-adapting the control policy of a robotic orthosis to optimally balance the tradeoff between movement error and user effort in order to promote human learning of target gait trajectories. The performance of four different RL-based controllers will be compared to that of the most common form of adaptive assist-as-needed controller during treadmill walking by neurologically intact individuals and by a small cohort of stroke survivors. The second objective will extend these novel control strategies to situations wherein the desired target motion is unknown to the controller and must be computed on-line. This step is necessary in order to encourage desired patterns of interlimb coordination during over-ground walking, where terrain can be uneven and obstacles are to be avoided. Both objectives will be pursued by investigating model-free on-line RL control methods for optimal policy search, using adaptive frequency oscillators coupled with kernel-based nonlinear filters for on-line estimation of desired, time-varying gait trajectories. The research promises to advance a fundamental understanding of the bidirectional adaptations that can arise when adaptive human and machine intelligences interact through physical channels, here in the context of gait rehabilitation.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.