Imitation Learning for Adaptive Control of a Virtual Soft Exoglove

Area of Science:

• Robotics
• Rehabilitation Engineering
• Biomechanics

Background:

• Wearable robots are common in hand rehabilitation but often ignore individual muscle deficits.
• Customized controllers are needed to address unique patient impairments for effective therapy.

Purpose of the Study:

• To develop a customized wearable robotic controller for hand motor impairments.
• To compensate for patient-specific muscle loss during hand-object manipulation tasks.

Main Methods:

• Utilized reinforcement learning and a biologically accurate musculoskeletal model in simulation.
• Trained a manipulation model using learning from demonstration with subject's video data.
• Simulated motor impairments by weakening muscle forces and compensated with a virtual wearable robotic glove.

Main Results:

• The virtual wearable robotic glove provided shared assistance, supporting hand manipulators with weakened muscles.
• The learned exoglove controller achieved 90.5% of the original manipulation proficiency.
• Demonstrated the controller's ability to address specific muscle deficits and aid object interaction.

Conclusions:

• Customized wearable robotic controllers can effectively compensate for individual muscle deficits in hand rehabilitation.
• This approach enhances functional recovery and manipulation proficiency in patients with neurological motor impairments.