Shared Autonomy Locomotion Synthesis With a Virtual Powered Prosthetic Ankle.

Area of Science:

• Robotics
• Artificial Intelligence
• Biomechanics

Background:

• Virtual environments offer safe testing for wearable robotic devices.
• Simulating walking requires modeling both hardware and user interaction.
• Predictive locomotion synthesizers generate virtual user movements for device evaluation.

Purpose of the Study:

• To implement a Deep Reinforcement Learning (DRL) controller for a powered prosthetic leg within a shared autonomy framework.
• To simulate a Human-Machine Interface (HMI) using continuous user intent representation for prosthetic control.
• To evaluate the controller's effectiveness in non-steady state walking scenarios.

Main Methods:

• Implemented a DRL-based motion controller in the MuJoCo physics engine.
• Shared control autonomy between a simulated human user and the active prosthesis policy.
• Utilized a data-driven, continuous representation of user intent to drive the HMI.
• Tested the system in simulated non-steady state walking conditions.

Main Results:

• The DRL agent produced realistic torque profiles and ground reaction forces without explicit optimization.
• Continuous user intent representation reduced compensatory gait patterns.
• The rate of simulated tripping was halved with the proposed intent representation.
• Co-adaptation between user and prosthesis emerged as a training challenge.

Conclusions:

• The developed framework enables effective training and evaluation of prosthetic controllers in virtual environments.
• Continuous user intent representation improves prosthetic control and user stability.
• The approach facilitates the transition of advanced prosthetic controllers from simulation to real-world application.