Safe Lyapunov-Based Deep Neural Network Adaptive Control of a Rehabilitative Upper Extremity Hybrid Exoskeleton

Hand cycling and reaching activities are rehabilitative exercises for individuals with movement disorders. For those with insufficient strength to exercise by themselves, electricity can be carefully applied to a muscle to generate force. This application of electricity is called functional electrical stimulation (FES) and FES has been shown to have many health benefits.

Prior research has shown that rehabilitation is improved by 1) repetition of the exercise, and 2) active effort including from FES. For some individuals, weakness and fatigue limit the effectiveness of rehabilitation therapy. Another limitation of FES-based exercise is that FES causes fatigue to occur at a faster rate than normal.

Fatigue can be reduced by using a combination of FES and robotics (e.g., a powered cycle, or a robot arm), called hybrid exoskeletons. For example, applying FES only when it is most efficient and having the robot help only when needed will reduce fatigue while encouraging active effort. Fatigue can be further reduced by adaptively changing how much the FES and robot help in the exercise.

The goal of this project is to develop safe adaptive methods for controlling hybrid exoskeletons that have the potential to significantly transform the rehabilitation of individuals with movement disorders. Throughout this project, the project team will invite middle and high school students to participate in lab tours and/or experiments that evaluate the designed methods to motivate the students to seek out advanced education in science, technology, engineering, and math (STEM) fields.

The intellectual merit of this project arises from the design, analysis, and experimental demonstration of safe saturated deep neural network (DNN)-based FES controllers with real-time closed-loop (Lyapunov-based) DNN weight update laws, which can approximate the complex dynamics of upper extremity hybrid exoskeletons and guarantee overall system stability. Objective 1 will develop a saturated, concurrent learning-inspired, and DNN-based FES control law that updates the DNN in multiple timescales and develop an adaptive DNN- and admittance-based motor controller to improve participant safety.

Objective 2 will develop real-time and Lyapunov-based adaptive update laws for both the inner- and output-layer DNN weights, while the exoskeleton's motor controller will include barrier functions to constrain the exoskeleton within a user-defined safe set. Objective 3 will experimentally evaluate the proposed controllers in populations with and without movement disorders, survey participants for user feedback, identify the most promising control architectures, investigate the FES controllers' potential to reduce motor power requirements, and develop new exoskeleton design guidelines.

Successful completion of this project could transform the rehabilitation industry by enabling more personalized and energy-efficient control of a hybrid exoskeleton. Moreover, another outcome is to acquire experimental data to enable the future development of an untethered upper extremity hybrid exoskeleton that uses FES to lower the weight and cost of the exoskeleton.

This project is jointly funded by the Disability and Rehabilitation Engineering Program and the Established Program to Stimulate Competitive Research (EPSCoR).

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.