A new framework for self-adaptive artificial intelligence to personalize assistance for patients using robotic exoskeletons and prostheses

Project Abstract

Robotic prostheses and exoskeletons that can personalize assistance to a patient through adaptation are of great value for

individuals with mobility challenges, such as those with amputation or stroke. Studies show mobility is strongly linked to

quality of life, participation and depression, and these technologies have significant ability to enhance human ambulation,

reduce fall risk, and improve overall quality of life. The proposed research aims to create a paradigm shift in the wearable

robotics field by innovating a new artificial intelligence (AI) framework for self-adapting robotic control to personalize assistance to a patient’s unique walking pattern. The overall hypothesis of this work is that AI systems capable of self-adaptive control during dynamic, unstructured community ambulation can improve mobility in patients using robotic prostheses and exoskeletons.

To enable sufficient levels of adaptation in intelligent wearable robotic applications, such as robotic exoskeletons and

prostheses, the team must overcome critical scientific gaps: Challenge 1) Predictive algorithms to determine user intent are

either user-dependent with significant training data requirements or user-independent with high error rates. Challenge 2)

Current human-in-the-loop approaches to adapt control policy are slow due to reliance on metabolic measures and are unable to optimize wearable robotic control outside of a static environment, such as fixed-speed treadmill walking. These technological gaps have impeded the translation of such systems beyond lab settings to real-world community use.

This New Innovator proposal will address these gaps through two primary innovations: Innovation 1) Create self-adaptive

intent recognition systems that learn an individual patient’s gait patterns; Innovation 2) Formulate a human-in-the-loop

(HIL) actor-critic framework that maximizes a multi-objective reward function to self-adapt control policy across users and

environmental states. The concept of a controller framework for wearable robotics that self-adapts both an intent recognition system and control policy to accommodate patient gait across locomotion tasks is novel and has not been previously

investigated. These innovations will initially be validated in able-bodied control subjects using state-of-the-art robotic

exoskeleton technology developed in the PI’s lab. Innovation 1’s concepts of a self-adaptive intent recognition system will

be translated to a robotic knee/ankle prosthesis platform and clinically tested on patients with transfemoral amputation.

Innovation 2’s concepts of actor critical networks for self-adapting control policy will be translated to a hip exoskeleton for

individuals post stroke and validated in clinical experiments. Patient interactions with AI systems deployed to wearable

robotics are critical to accelerate the field and cannot be derived from offline studies or able-bodied control testing. Ultimately, the outcomes will enable self-adaptive wearable robotic technology to improve patient mobility through personalized assistance. AI technology combined with wearable robotics has the potential to increase walking speed,

improve gait quality and stability, and enable more accessibility to diverse locomotion tasks for patient populations with

mobility deficits. This functionality promises to translate to improved community ambulation and enhanced quality of life.