Vision-Based Topological Mapping and Navigation With Self-Organizing Neural Networks.

Area of Science:

• Artificial Intelligence
• Robotics
• Cognitive Science

Background:

• Autonomous agents require robust spatial mapping and navigation for real-time environmental interaction.
• Current vision-based navigation models face challenges with memory scalability and indirect path following.

Purpose of the Study:

• To present e-TM, a self-organizing neural network framework for incremental topological mapping and navigation.
• To address limitations of existing models by improving memory efficiency and navigation directness.

Main Methods:

• e-TM utilizes episodic memory to model exploration trajectories and extract landmarks as events.
• A memory consolidation process transfers environmental layout knowledge into spatial memory using Fusion Adaptive Resonance Theory (ART) networks.
• Transfer learning integrates human demonstrations for smoother locomotion control.

Main Results:

• e-TM demonstrates significant reductions in navigation time costs compared to state-of-the-art models like SPTM.
• The framework learns significantly sparser topological graphs, indicating improved memory efficiency.
• Fusion ART networks enable compact spatial representations and discovery of navigation shortcuts.

Conclusions:

• e-TM offers an efficient and scalable solution for vision-based mapping and navigation in autonomous agents.
• The proposed framework enhances navigation performance through effective memory management and knowledge transfer.