Spatial relation learning in complementary scenarios with deep neural networks.

Area of Science:

• Artificial Intelligence
• Cognitive Science
• Robotics

Background:

• Cognitive agents require environmental concept learning, including spatial relations, for real-world interaction.
• Existing methods for learning spatial relations have limitations in scope and experience diversity.

Purpose of the Study:

• To propose and evaluate three complementary approaches for a cognitive agent to learn spatial relations.
• To conceptualize a cognitive architecture integrating these approaches for robust spatial relation learning.

Main Methods:

• Embodied learning using a model for left-right relation instruction following.
• Geometric feature learning from a visual dataset for directional relation understanding.
• Knowledge base utilization for disembodied spatial relation reasoning.

Main Results:

• The embodied model demonstrated learning of basic spatial relations.
• Reasoning models successfully learned directional relations from visual data across different frames of reference.
• The integration of multiple approaches is shown to be complementary.

Conclusions:

• A combined cognitive architecture effectively integrates embodied, visual, and knowledge-based learning for spatial relations.
• This integrated approach enhances a cognitive agent's ability to understand and interact with its environment.