Learning dynamic cognitive map with autonomous navigation.

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Robotics

Background:

• Animals navigate complex environments using memory and strategic decision-making.
• Existing computational models often lack adaptability to novel situations.

Purpose of the Study:

• To develop a novel computational model for space navigation and mapping.
• To replicate animal-like navigation using biologically inspired principles and active inference.

Main Methods:

• Incorporating a dynamically expanding cognitive map over predicted poses.
• Utilizing an active inference framework for enhanced generative model plasticity.
• Implementing structure learning and active inference for exploration and exploitation.

Main Results:

• The model demonstrates efficient exploration and exploitation in mini-grid environments.
• It dynamically expands model capacity for novel locations and updates maps with new evidence.
• Achieves rapid environmental structure learning within a single episode with minimal overlap.

Conclusions:

• The proposed model shows robustness and efficacy in navigating intricate environments.
• It learns environmental structures without prior knowledge of observation or world dimensions.
• Offers a promising approach for AI navigation and cognitive mapping.