Vector-based navigation using grid-like representations in artificial agents.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Computational Neuroscience

Background:

• Deep neural networks excel in many tasks but struggle with navigation.
• Mammalian navigation relies on grid cells in the entorhinal cortex for spatial representation and path integration.
• Current artificial agents lack the sophisticated spatial cognition seen in mammals.

Purpose of the Study:

• To develop a deep reinforcement learning agent with mammal-like navigation abilities by leveraging grid cell functions.
• To investigate if grid-like representations can improve agent performance in challenging environments.
• To explore the computational benefits of emergent grid-like representations for navigation.

Main Methods:

• Trained a recurrent neural network to perform path integration, observing the emergence of grid-like representations.
• Utilized these emergent grid-like representations as a basis for a deep reinforcement learning navigation agent.
• Evaluated agent performance against expert humans and comparison agents in unfamiliar and changeable environments.

Main Results:

• The recurrent network developed representations similar to grid cells and other entorhinal cells.
• Agents with grid-like representations significantly outperformed human experts and other agents in navigation tasks.
• Emergent grid-like units provided metric quantities for vector-based navigation and enabled shortcut behaviors.

Conclusions:

• Emergent grid-like representations provide agents with a Euclidean spatial metric and vector operations, crucial for proficient navigation.
• This approach supports neuroscientific theories on the role of grid cells in vector-based navigation.
• Combining path-based and vector-based strategies using grid-like representations enhances navigation in complex, dynamic environments.