Policy Design for an Ankle-Foot Orthosis Using Simulated Physical Human-Robot Interaction via Deep Reinforcement...

Area of Science:

• Robotics
• Biomechanics
• Machine Learning

Background:

• Designing effective robotic orthosis controllers is challenging due to the complexity of human-robot interaction (pHRI).
• Computer simulations offer a time- and cost-efficient alternative to physical prototyping for controller development.
• Deep reinforcement learning (deep RL) provides a powerful framework for optimizing control policies in complex dynamic systems.

Purpose of the Study:

• To propose a novel two-stage deep reinforcement learning framework for designing a robotic orthosis controller.
• To leverage human-robot dynamic simulation for efficient controller development.
• To validate the controller's ability to assist human walking, particularly in cases of muscle impairment.

Main Methods:

• A two-stage deep RL approach was employed, starting with imitation learning on a healthy musculoskeletal model (OpenSim-RL).
• Subsequent stages involved creating weakened muscle models and training the orthosis policy using an elastic foundation model for pHRI prediction.
• The framework utilized deep reinforcement learning for both gait generation and orthosis control policy optimization.

Main Results:

• The musculoskeletal model accurately imitated human walking patterns, validated against experimental data.
• The trained robotic orthosis controller demonstrated effectiveness in assisting a weakened soleus muscle.
• Simulation results closely matched experimental gait data, confirming the approach's validity.

Conclusions:

• The proposed two-stage deep RL framework successfully designs robotic orthosis controllers for human assistance.
• The simulation-based approach effectively models physical human-robot interaction and optimizes assistive torque.
• This method offers a viable strategy for developing personalized robotic assistance for gait rehabilitation.