Control-Theory Taxonomy of Physical AI (dreossi.github.io)

The term "Physical AI" has evolved to encompass a broad range of applications that intersect with the physical world, including robotics, materials science, and biotechnology. This article introduces a taxonomy grounded in control theory to categorize the methods and characteristics of Physical AI, which respects the laws of physics, manages scarce and costly data, and operates continuously in dynamic environments. Through this framework, methods can be organized into four categories: Perception, Model, Control, and Optimize, each playing a specific role in building responsive and adaptive AI systems. This taxonomy is significant for the AI/ML community as it clarifies the distinctions between Physical AI and traditional AI models, highlighting the importance of integrating multiple functionalities into cohesive systems. By elucidating relationships between methods, such as blending perception and control in Vision-Language Architectures (VLAs) or merging predictive and optimization tasks in Physics-Informed Neural Networks (PINNs), the article points toward the potential for developing more advanced and unified models. The concept of a "four-column model" is introduced, suggesting a future where a single foundation model could autonomously perceive, model, control, and optimize within real-world environments, pushing the boundaries of what Physical AI can achieve.