A soft exoskeleton for hip extension and flexion assistance based on reinforcement learning control

Area of Science:

• Robotics
• Biomechanics
• Artificial Intelligence

Background:

• Traditional soft exoskeleton force control struggles with environmental and individual gait variations.
• Adaptive control is crucial for optimizing exoskeleton assistance in real-world scenarios.

Purpose of the Study:

• To develop and validate a reinforcement learning (RL)-based control method for soft exoskeletons assisting hip flexion-extension.
• To enhance the adaptability and effectiveness of soft exoskeletons for human gait assistance.

Main Methods:

• System identification was used to model the exoskeleton's motor control system.
• An adaptive controller was trained using the TD3 RL algorithm in a simulated environment.
• The controller adjusts Pulse Width Modulation (PWM) values to match desired assistive forces.

Main Results:

• The RL-based controller demonstrated effective tracking of desired assistive force profiles.
• Metabolic experiments showed significant reductions in net metabolic cost.
• Wearers experienced a 12.9% ± 3.3% decrease on uniform surfaces and 10.7% ± 3.7% on slopes compared to power-off mode.

Conclusions:

• Reinforcement learning offers a robust and adaptive solution for soft exoskeleton control.
• The proposed RLPID method effectively reduces metabolic demand during walking, showcasing practical benefits for users.