Generating synthetic gait patterns based on benchmark datasets for controlling prosthetic legs.

Area of Science:

• Biomedical Engineering
• Robotics
• Machine Learning

Background:

• Prosthetic legs aim to restore locomotion for amputees.
• Deep neural network (DNN) controllers offer advanced prosthetic leg functionality.
• A key challenge is acquiring sufficient training data, especially for diverse ambulation modes.

Purpose of the Study:

• To develop a method for generating synthetic gait patterns using generative adversarial networks (GANs).
• To enable DNN-based controllers to execute ambulation modes not present in the initial training datasets.

Main Methods:

• A conditional GAN was trained on existing gait datasets.
• Synthetic gait patterns (vertical load, joint angles) were generated for various ambulation modes.
• A DNN controller was trained using these synthetic patterns to generate impedance parameters.

Main Results:

• The GAN achieved high fidelity in generating synthetic gait patterns (R²=0.97, SSIM=0.94).
• A DNN controller was successfully trained for level-ground walking, sit-to-stand, and stand-to-sit motions.
• Bypass testing with four participants validated the generated control parameters for knee and ankle.

Conclusions:

• Synthetic gait data generated by GANs can effectively train DNN models for prosthetic leg impedance control.
• Conditional GANs provide a reliable method for creating gait data for underrepresented ambulation modes.
• This approach facilitates the efficient development of advanced prosthetic leg controllers.