Reciprocal Learning of Intent Inferral with Augmented Visual Feedback for Stroke

Area of Science:

• Robotics
• Human-Computer Interaction
• Neurorehabilitation

Background:

• Classical intent inferral methods for controlling wearable robots rely on unidirectional biosignal inputs, limiting user adaptability.
• Existing machine learning models for intent inferral lack direct user observability of their internal states.
• Effective control of assistive robotic devices requires intuitive user interaction and adaptation.

Purpose of the Study:

• To introduce reciprocal learning, a novel bidirectional paradigm for human adaptation to intent inferral classifiers.
• To enhance the intuitive control of wearable robots by facilitating user adaptation to machine learning models.
• To improve the performance of robotic hand orthosis for stroke patients through adaptive intent inferral.

Main Methods:

• Developed a reciprocal learning paradigm involving iterative stages of machine learning model updates and human adaptation guided by augmented visual feedback.
• Implemented the paradigm for a robotic hand orthosis, inferring intents (open, close, relax) from electromyographic (EMG) signals.
• Utilized LED progress-bar displays to provide users with visual feedback on classifier predictions.

Main Results:

• Reciprocal learning demonstrated performance improvement in a subset of stroke subjects (two out of five).
• The paradigm did not negatively impact the performance of other subjects.
• Hypothesized that subjects learned to generate more distinguishable and separable biosignals through reciprocal learning.

Conclusions:

• Reciprocal learning offers a promising bidirectional approach to enhance human adaptation to intent inferral classifiers in wearable robotics.
• The proposed method shows potential for improving the control of assistive devices like robotic hand orthoses for neurorehabilitation.
• Further research is warranted to explore the mechanisms of user adaptation and optimize reciprocal learning strategies.