Adaptive Torque Control of Exoskeletons Under Spasticity Conditions via Reinforcement Learning.

Area of Science:

• Robotics
• Neuroscience
• Biomedical Engineering

Background:

• Spasticity is a disabling symptom in neurological conditions like cerebral palsy, spinal cord injury, and stroke.
• Current wearable robot use for spasticity is limited, especially for higher spasticity levels (above 1+ on the Modified Ashworth Scale).
• The velocity-dependent nature of spastic reflexes complicates the development of safe, personalized robotic controllers.

Purpose of the Study:

• To develop and evaluate a novel adaptive torque controller for knee exoskeletons.
• To address challenges in controlling exoskeletons for individuals with joint spasticity.
• To improve task performance and reduce interaction forces during exoskeleton use.

Main Methods:

• Utilized deep reinforcement learning (RL) to create an adaptive torque controller.
• Developed a digital twin simulation including a musculoskeletal exoskeleton model with joint misalignment.
• Incorporated a differentiable model of spastic muscle reflexes for controller training.
• Simulated knee extension movements to train and test the RL agent.

Main Results:

• The RL agent successfully learned to control the knee exoskeleton across different spasticity levels.
• The proposed controller reduced average maximum joint torques by 10.6% compared to a conventional compliant controller.
• Root mean square values until settling time were decreased by 8.9%.

Conclusions:

• The novel adaptive torque controller demonstrates efficacy in managing spasticity during exoskeleton use.
• Deep reinforcement learning offers a promising approach for personalized robotic assistance in neurological rehabilitation.
• This controller has the potential to expand the application of wearable robots for individuals with moderate to severe spasticity.