An ocular biomechanics environment for reinforcement learning.

Area of Science:

• Computational Neuroscience
• Robotics
• Biomechanics

Background:

• Reinforcement learning (RL) models human movement, offering insights into neural control and aiding prosthetic/robotic design.
• Ocular biomechanics, particularly rapid eye movements like saccades, presents a complex control challenge.
• Existing RL applications have not extensively explored ocular systems.

Purpose of the Study:

• To extend reinforcement learning applications to control an ocular biomechanical system.
• To train an agent to perform saccades, a fundamental rapid eye movement.
• To establish a framework for using deep RL in ocular biomechanics research.

Main Methods:

• Developed a simulated ocular environment.
• Employed the Deep Deterministic Policy Gradients (DDPG) algorithm for agent training.
• Trained an agent to control ocular movements for saccade generation.

Main Results:

• The trained agent successfully performed saccades, matching desired eye positions.
• Achieved a mean deviation angle of 3.5°±1.25° in saccade accuracy.
• Demonstrated the feasibility of RL for controlling ocular biomechanical systems.

Conclusions:

• Deep reinforcement learning offers a powerful tool for investigating ocular biomechanics.
• This framework represents a significant first step in applying advanced RL to understand eye movement control.
• Future work can leverage this approach to further elucidate neural control mechanisms and enhance ocular technologies.