Characterization of Human Balance through a Reinforcement Learning-based Muscle Controller.

Area of Science:

• Biomechanics
• Computational Neuroscience
• Rehabilitation Engineering

Background:

• Clinical balance tests lack objectivity and are subjective.
• Existing computational models (e.g., center of pressure tracking, inverted pendulum) fail to capture multi-joint and muscle contributions to whole-body balance.

Purpose of the Study:

• To propose a novel musculoskeletal modeling and control methodology for investigating human balancing capabilities in the center of mass (COM) state space.
• To objectively quantify balance capability and explore the limits of dynamic balance.

Main Methods:

• Integration of a musculoskeletal model with a balance controller trained via reinforcement learning (RL) using Proximal Policy Optimization (PPO).
• Exploration of balance recovery from random initial COM states to define a balance region (BR).
• Investigation of the effects of muscle weakness and neural excitation delay on BRs.

Main Results:

• A balance region (BR) was obtained, enclosing successful state-space trajectories for COM balance.
• Comparison with linear inverted pendulum models showed similar trends but reduced recoverable areas.
• Muscle weakness and neural excitation delay were found to reduce balancing capability.

Conclusions:

• The developed RL-based musculoskeletal model provides an objective method for quantifying human balance.
• This approach offers a promising avenue for personalized balance assessments and evaluating balance in individuals with health conditions.
• The study highlights the importance of multi-joint and muscle contributions in dynamic balance control.