Physical AI

Physical AI refers to artificial intelligence systems that operate in and interact with the physical world through embodiment, rather than processing text, images, or code purely within computational environments. Where conventional or "digital" AI produces outputs that remain in software — a generated paragraph, a classified image, a recommendation — physical AI must perceive real environments through sensors, reason about objects and forces, and take actions that have physical consequences, all while contending with the continuous, noisy, and unforgiving nature of real-world physics. The term is closely related to, and often used interchangeably with, embodied AI.

Physical AI Versus Embodied AI

The two terms overlap heavily. Embodied AI emphasises the integration of intelligence into a physical body or system — a robot, an autonomous vehicle, a drone, or an automated warehouse — enabling it to sense and act. Physical AI emphasises the broader challenge of intelligence that understands and obeys the laws of the physical world, including the foundation models and simulation tools that make such intelligence possible. In practice, an embodied agent is a physical AI system, and most discussion treats the concepts as two framings of the same frontier: getting AI out of the screen and into the world.

Why It Is Hard

Operating in the physical world introduces challenges absent from purely digital tasks. Sensory input is continuous, high-dimensional, and noisy; the same action can produce different outcomes due to friction, lighting, or wear; and mistakes can be costly or dangerous and cannot simply be undone. Physical systems must also act in real time, integrating perception, planning, and control within tight latency budgets, often on limited onboard compute. Collecting real-world training data is slow and expensive compared with scraping the internet, which has made simulation and synthetic data central to the field.

Enabling Technologies

Several converging advances have driven recent progress in physical AI. Vision-language-action (VLA) models extend the multimodal capabilities of large models to output robot actions, allowing a system to map natural-language goals and visual observations directly to movements. Simulation environments generate vast amounts of synthetic experience, letting agents practise in virtual worlds before transferring skills to real hardware — a process known as sim-to-real transfer. Reinforcement learning and imitation learning provide methods for acquiring control policies, while foundation models for robotics aim to give robots broadly transferable skills rather than narrow, task-specific programming. Specialised onboard compute is required to run these models on the robot itself.

NVIDIA has been a prominent promoter of the concept, framing physical AI around three pillars: simulation platforms for generating training data, foundation models for robots including efforts aimed at humanoids, and edge compute hardware for onboard inference. The framing gained mainstream attention after NVIDIA's leadership described physical AI as a major next frontier at industry events in 2024.

Applications

Physical AI spans a range of embodiments. General-purpose and humanoid robots aim to perform varied manual tasks in warehouses, factories, and eventually homes. Autonomous vehicles apply physical AI to perception and control on roads. Industrial and logistics settings use it for robotic picking, sorting, inspection, and automated material handling. Agriculture, construction, and healthcare are emerging domains, with robots assisting in harvesting, site work, and patient care. In each case the system must close the loop between sensing the environment and acting reliably within it.

Outlook

Physical AI is widely regarded as one of the most significant directions for the field, extending the rapid progress of language and vision models into the physical economy. Substantial challenges remain in safety, reliability, generalisation across environments, energy efficiency, and the cost of hardware, but investment and research activity grew rapidly through the mid-2020s.

References

1. NVIDIA. (2025). What is Physical AI? / What is Embodied AI?. NVIDIA Glossary, nvidia.com.
2. Automate. (2025). NVIDIA on What Physical AI Means for Robotics — 2025 Keynote. automateshow.com.
3. Silicon Valley Robotics Center. (2026). What Is Physical AI? The Complete Guide to Embodied Intelligence. roboticscenter.ai.
4. Embodied Arena. (2025). A Comprehensive, Unified, and Evolving Evaluation Platform for Embodied AI. arXiv:2509.15273.