A GPU-accelerated cortical neural network model for visually guided robot navigation.

Area of Science:

• Computational Neuroscience
• Robotics
• Vision Science

Background:

• Human navigation in cluttered environments relies on vision, yet brain representations of steering and self-motion perception remain unclear.
• Existing research on neural circuitry for self-motion lacks connection to active steering control.

Purpose of the Study:

• To develop a cortical neural network model for visually guided navigation.
• To investigate how neural signals, particularly from cortical area MT, can guide active steering.
• To demonstrate the role of embodiment in visually guided navigation.

Main Methods:

• Developed an embodied cortical neural network model on a physical robot.
• Incorporated a motion energy model for V1 and a spiking neural network for MT.
• Model processed optic flow, object positions, and goal locations to generate motor commands.

Main Results:

• The model successfully steered the robot around obstacles toward a goal in a real-world environment.
• Generated robot trajectories closely matched human behavioral data.
• Demonstrated that MT neural signals can provide sufficient motion information for steering.

Conclusions:

• Neural signals in a model of cortical area MT can guide a physical robot on human-like paths.
• Embodiment is crucial, linking brain function models with physical constraints for behavior.
• The study bridges computational neuroscience, robotics, and vision science for navigation research.