Training an Unconstrained 6 DOF Biomimetic Robotic Eye With Deep Reinforcement Learning

Area of Science:

• Computational neuroscience
• Ophthalmology
• Robotics

Background:

• Understanding the neural control of saccades is crucial for explaining eye movement disorders.
• Existing models often lack biological plausibility or fail to capture the full spectrum of human saccadic behavior.

Purpose of the Study:

• To develop a biologically plausible computational model of a biomimetic eye (6 DOFs) to investigate saccade control signals.
• To compare model-generated movement characteristics with human saccadic behavior.
• To test the hypothesis that saccade generation optimizes a reward function balancing various costs.

Main Methods:

• Development of a six degrees-of-freedom (DOFs) biomimetic eye model.
• Utilizing a model-free deep reinforcement learning algorithm trained under a reward optimization constraint.
• Analysis of control strategies, cost evolution, and the impact of noise on movement characteristics.

Main Results:

• The model's emergent control strategies replicated key human-like saccadic characteristics.
• Observed behaviors included nonlinear main sequence relationships, adherence to Listing's and Donders' Laws, and straight oblique trajectories.
• The model demonstrated normometric pulse-step-like controls and antagonistic extraocular muscle pairing without explicit enforcement.

Conclusions:

• Reinforcement learning optimizing a reward function can explain the generation of complex human-like saccades.
• The biomimetic eye model provides a novel framework for studying ocular motor control.
• This approach offers insights into the neural computations underlying rapid eye movements.