Learning eye vergence control from a distributed disparity representation.

Area of Science:

• Robotics
• Computational Neuroscience
• Computer Vision

Background:

• Accurate vergence control is crucial for robotic systems mimicking human visual perception.
• Existing methods often rely on explicit disparity computation, which can be computationally intensive or prone to errors.
• Neural models offer a biologically plausible alternative for complex visual tasks.

Purpose of the Study:

• To introduce two novel neural network models for closed-loop vergence angle control in robotic heads.
• To demonstrate that these models can determine vergence angles without explicit disparity calculations.
• To develop a robust model capable of operating under general conditions, removing prior constraints on gaze and stimulus orientation.

Main Methods:

• Development of two neural models: a simplified and a complex version.
• Implementation of a closed-loop control system utilizing post-processed responses from disparity-tuned complex cells.
• Integration of actual gaze direction and vergence angle as inputs for the models.
• Testing the models under restricted (frontoparallel plane) and general (unrestricted gaze and stimulus) conditions.

Main Results:

• Both neural models successfully controlled the vergence angle of the robotic head.
• The models achieved vergence control by processing neural responses, gaze direction, and current vergence angle, bypassing explicit disparity computation.
• The complex model demonstrated reliable performance even when assumptions about gaze direction and stimulus orientation were removed.

Conclusions:

• Neural network models can effectively control robotic head vergence without explicit disparity computation.
• The proposed models offer a flexible and robust approach to vergence control, particularly the complex model which handles general visual scenarios.
• This work contributes to the development of more sophisticated and human-like visual systems in robotics.