Using the soft actor-critic algorithm to generate rapid eye movements with an unconstrained 6DOF biomimetic robotic...

Area of Science:

• Computational neuroscience
• Biomimetics
• Ophthalmology

Background:

• Understanding the neural control of rapid eye movements (saccades) is crucial.
• Biomimetic models offer valuable insights into complex biological systems.
• Previous models have limitations in capturing the full dynamics of saccadic eye movements.

Purpose of the Study:

• To develop an accurate, six-degree-of-freedom biomimetic eye model.
• To explore the neural control signals governing saccade generation.
• To test the hypothesis that saccade generation optimizes multiple costs.

Main Methods:

• Developed a six-degree-of-freedom computational model of a biomimetic eye.
• Employed a model-free reinforcement learning algorithm with biologically inspired inputs.
• Trained the model in an open-loop manner to generate saccadic eye movements.

Main Results:

• The model successfully replicated human-like saccadic characteristics without explicit enforcement.
• Observed nonlinear main sequence relationships and compliance with Listing's Law.
• Demonstrated normometric pulse-step-like controls and antagonistic muscle pairing.

Conclusions:

• Saccade generation appears to be governed by optimizing multiple costs like accuracy, duration, energy, and tendon tension.
• The biomimetic model provides a powerful tool for investigating neural control of eye movements.
• Reinforcement learning can uncover biologically plausible control strategies for complex motor tasks.