Physical Artificial Intelligence: The Concept Expansion of Next-Generation Artificial Intelligence

Yingbo Li, Yucong Duan, Anamaria-Beatrice Spulber, Haoyang Che, Zakaria Maamar, Zhao Li, Chen Yang, Yu LeiarXiv 2021 · arXiv

Introduction

Current advancements in Artificial Intelligence (AI), such as natural language processing and image recognition, primarily reside within the domain of "Digital Artificial Intelligence" (DIAI). These systems function as "brains without bodies," processing data and signals within isolated computational environments. This paper argues for a paradigm shift toward Physical Artificial Intelligence (PAI), an expansion where intelligence is deeply integrated into physical materials and distributed systems, bridging the gap between digital logic and physical reality.

Theoretical Framework and Taxonomy

The transition to PAI requires a robust conceptual framework. The authors categorize PAI into two primary subdomains:

  1. Integrated Physical AI (IPAI): Intelligence and physical components are localized within a single entity, such as a humanoid robot.
  2. Distributed Physical AI (DPAI): Intelligence and physical actuators are spread across a wide environment, exemplified by Smart Cities or Industrial Internet of Things (IIoT) networks.
The architecture of Distributed Physical AI (DPAI) illustrating the flow from perception in the physical world through sensor clusters to central processing and execution.

Furthermore, the paper applies the Data-Information-Knowledge-Wisdom (DIKW) hierarchy to PAI. It posits that Knowledge Graphs are critical for PAI to evolve from simple data processing to "Wisdom"—the capacity to act effectively and autonomously within complex, unpredictable physical environments.

Multidisciplinary Convergence

Creating PAI is not merely a software challenge; it necessitates the fusion of five core scientific disciplines: Computer Science, Materials Science, Mechanical Engineering, Chemistry, and Biology. This convergence is essential for developing autonomous systems that mimic natural organisms.

The multidisciplinary integration required for PAI, showing the convergence of Computer Science, Biology, Chemistry, Material Science, and Mechanical Engineering.

A key insight derived from this integration is the "Cannikin Law" of PAI. This principle suggests that the development of PAI is limited by its weakest link. For instance, if materials science cannot provide a sufficiently flexible sensor, the entire system's capability is constrained, regardless of the sophistication of the underlying AI algorithms.

Key Insights: AI as a "Whole Body"

While DIAI focuses on mimicking human logical reasoning, PAI aims to replicate the entire human body’s interaction with the world. This involves multi-modal fusion, where the system processes diverse data streams—such as material properties, temperature, vision, and sound—to make real-time physical decisions. This shift moves AI from passive observation to active, embodied participation in the physical world.

Governance and Industry Impact

The transition to PAI introduces significant governance challenges. The authors identify four primary "existence" problems:

  • Security: Protecting both digital data and physical integrity.
  • Fake Data: Addressing physical-context deepfakes.
  • Social Privacy: Managing ubiquitous tracking in physical spaces.
  • Bias: Mitigating biases inherited from training data that manifest in physical actions.

In terms of industry, PAI is poised to transform logistics (autonomous delivery droids), agriculture (automated harvesting), and healthcare (proactive elderly care via biological sensors).

Conclusion and Future Directions

The primary hurdle for PAI remains the organizational friction of coordinating across vastly different scientific fields. Future research must focus on developing robust Knowledge Graphs that address the "5 Ws" (Who, What, When, Where, Why) of physical interaction, moving PAI from simple automation toward true autonomous wisdom.

Summary updated

BibTeX
@misc{li2021physical,
  author = {Yingbo Li and Yucong Duan and Anamaria-Beatrice Spulber and Haoyang Che and Zakaria Maamar and Zhao Li and Chen Yang and Yu Lei},
  title = {{Physical Artificial Intelligence: The Concept Expansion of Next-Generation Artificial Intelligence}},
  year = {2021},
  eprint = {2105.06564},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2105.06564},
}
Figure