Aligning Perception, Reasoning, Modeling and Interaction: A Survey on Physical AI

Area of Science:

• Artificial Intelligence
• Robotics
• Physics Simulation

Background:

• Growing interest in integrating physical laws into AI systems.
• Prior surveys examined embodied intelligence and world models separately.
• Need for a unified developmental pathway from observation to physical comprehension.

Purpose of the Study:

• To present a systematic framework for physical AI development.
• To connect embodied intelligence and world models as a unified progression.
• To analyze advancements in physical AI through four interconnected stages.

Main Methods:

• Analysis of architectural innovations, training methodologies, and causal inference.
• Synthesis of how physical understanding emerges through cumulative integration.
• Examination of embodied systems and environmental feedback loops.

Main Results:

• A four-stage framework: perception, reasoning, modeling, and embodied interaction.
• Demonstration of how each stage enables and enhances the next.
• Evolution from task-specific solutions to integrated architectures for causal reasoning and prediction.

Conclusions:

• The framework provides foundations for next-generation physical AI.
• Implications for safe, generalizable, and interpretable AI deployment in robotics and scientific discovery.
• Highlights the cumulative integration of AI capabilities for physical understanding.