Characterizing Motor Control of Mastication With Soft Actor-Critic

Area of Science:

• Biomechanics
• Computational Neuroscience
• Robotics

Background:

• Direct measurement of human masticatory muscle activation is limited by ethical and safety concerns.
• Estimating muscle activations and forces is crucial for understanding the masticatory system's complex dynamics.
• Reinforcement learning (RL) offers a method to learn system dynamics through adaptive exploration and reward-based training.

Purpose of the Study:

• To develop a deep reinforcement learning (RL) model for learning the inverse dynamics of a simulated masticatory system.
• To create a parametric neural model capable of controlling jaw movements with high accuracy.
• To establish a performance analysis platform for evaluating masticatory system functionalities.

Main Methods:

• Utilized a deep RL algorithm, Soft Actor-Critic, to train a neural network on a simulated masticatory system.
• Trained the model to learn activation patterns for achieving desired jaw positions.
• Integrated a force regularization term into the RL reward to promote metabolic efficiency.

Main Results:

• The trained neural model achieved sub-millimeter accuracy in navigating the 3D envelope of jaw motion.
• A performance analysis platform was introduced, assessing range of motion, metabolic efficiency, agility, symmetry, and accuracy.
• The model demonstrated the ability to learn more metabolically efficient policies and showed an inverse correlation between efficiency and agility/range of motion.

Conclusions:

• The developed RL-based masticatory model provides a valuable tool for biomechanical analysis and simulation.
• The framework has potential applications in surgical treatment planning and predicting post-operative rehabilitation outcomes.
• This approach offers a novel method for studying complex biological systems where direct experimentation is challenging.