Personalized ML-based wearable robot control improves impaired arm function

Abstract

Abstract Portable wearable robots offer promise for assisting people with upper limb disabilities. However, movement variability between individuals and trade-offs between supportiveness and transparency complicate robot control during real-world tasks. We address these challenges by first developing a personalized ML intention detection model to decode user’s motion intention from IMU and compression sensors. Second, we leverage a physics-based hysteresis model to enhance control transparency and adapt it for practical use in real-world tasks. Third, we combine and integrate these two models into a real-time controller to modulate the assistance level based on the user’s intention and kinematic state. Fourth, we evaluate the effectiveness of our control strategy in improving arm function in a multi-day evaluation. For 5 individuals post-stroke and 4 living with ALS wearing a soft shoulder robot, we demonstrate that the controller identifies shoulder movement with 94.2% accuracy from minimal change in the shoulder angles (elevation: 3.4°, depression: 1.7°) and reduces arm-lowering force by 31.9% compared to a baseline controller. Furthermore, the robot improves movement quality by increasing their shoulder elevation/depression (17.5°), elbow (10.6°) and wrist flexion/extension (7.6°) ROMs; reducing trunk compensation (up to 25.4%); and improving hand-path efficiency (up to 53.8%).