Robust active binocular vision through intrinsically motivated learning.

Area of Science:

• Computational Neuroscience
• Robotics
• Machine Learning

Background:

• The efficient coding hypothesis explains sensory neuron responses based on signal redundancy.
• Extending efficient coding to active perception involves learning efficient sensory codes and optimizing sensor movement via reinforcement learning.

Purpose of the Study:

• To investigate the robustness and self-calibration capabilities of an intrinsically motivated learning approach for active perception.
• To test the system's ability to recover from perturbations in a robotic binocular vision system.

Main Methods:

• Implemented an active perception system on a robot using intrinsically motivated learning for efficient coding.
• Introduced perturbations to the stereo cameras to assess system robustness and recovery.
• Evaluated system performance based on motor compensation, sensory encoding adaptation, and behavioral policy adjustments.

Main Results:

• The system fully recovered from perturbations compensable by motor actions.
• Performance degraded gracefully when motor compensation was not possible.
• Recovery was enhanced when both sensory encoding and behavior policy adapted to perturbations.

Conclusions:

• The intrinsically motivated learning approach yields a robust, self-calibrating active perception system.
• The system demonstrates significant adaptability and resilience to unmodeled disturbances.
• This method holds promise for developing autonomous perceptual systems capable of self-correction.