Cognitive navigation based on nonuniform Gabor space sampling, unsupervised growing networks, and reinforcement...

Area of Science:

• Robotics
• Computational Neuroscience
• Artificial Intelligence

Background:

• Autonomous agents require robust spatial representations for navigation.
• The hippocampus in rats uses place cells for spatial coding, a model for artificial systems.
• Challenges include high-dimensional sensory input and the hidden state problem.

Purpose of the Study:

• To develop an unsupervised learning method for creating spatial representations in autonomous agents.
• To enable goal-oriented navigation in continuous environments using learned representations.
• To validate the model on a physical robot platform.

Main Methods:

• Unsupervised Hebbian learning to construct a state space representation.
• Modeling visual scenes with Gabor filters on a Log-polar graph.
• Path integration to handle self-motion and eliminate sensory aliasing.
• Temporal-difference learning for sensorimotor mapping and reinforcement learning.
• Population vector coding for interpreting neural activity.

Main Results:

• A dense, uniform representation of a 2-D manifold using overlapping place fields was achieved.
• The learned representation mimics hippocampal place cell activity.
• The system successfully addressed the hidden state problem.
• The model enabled goal-oriented navigation for a Khepera robot.

Conclusions:

• Unsupervised Hebbian learning can generate effective spatial representations for autonomous navigation.
• The developed model provides a biologically plausible and computationally efficient approach to spatial learning.
• This work demonstrates the potential for applying such models to real-world robotic applications.