What is Physical AI?

Introduction

A robotics team files a ticket asking you to provision a GPU (Graphics Processing Unit) cluster for training a warehouse picking arm. You've allocated sufficient GPU capacity for chatbots and recommendation engines, so scope it accordingly. Then you notice the job spec also calls for a physics simulator, terabytes of sensor replay logs, and a test rig with an actual robot arm bolted to it. This isn't a chatbot workload wearing a different label. It's a different category of AI system, one that works in three-dimensional space instead of in a chat window.

That category has a name: physical AI (Artificial Intelligence). You'll increasingly see it attached to requests from robotics, autonomous-vehicle, and smart-facility teams, and it's worth having a clear, plain-language definition before you're asked to size infrastructure for it. This article gives the definition,  where physical AI already exists, and walks through the specific technical challenges, such as the "sim-to-real gap," that make these workloads appear different from the generative AI systems you're probably more familiar with.

What Makes AI "Physical"?

Generative AI, the large language models and image generators most infrastructure teams have spent the last few years supporting, operates on a narrow slice of the world. A language model (LLM) works in essentially one dimension: it predicts the next word in a sequence of text. An image or video generator adds a second dimension, predicting the next pixel pattern given a prompt. Both are enormously useful, and both stay inside a computer the entire time they operate.

Physical AI is the term for AI systems that do not stay inside the computer. In broad terms, it's the technology that gives autonomous systems (robots, self-driving cars, and smart physical environments among them) the ability to sense their surroundings, make sense of what they're sensing, and carry out complex actions in the real, physical world (for background on this framing, see CSET's "Physical AI" brief Source: CSET, "Physical AI"). Instead of one dimension (text) or two (images), a physical AI system has to reason in three spatial dimensions plus time: where things are, how fast they're moving, and what's about to happen next.

That real-world connection is what researchers call embodiment: the AI's decisions are carried out by a physical body (a robot arm, a vehicle chassis, a drone), and that body's sensors feed back into the next decision. A chatbot that gives you a bad answer only gives you another one when you push back. A warehouse robot arm that misjudges the weight of a box and drops it cannot undo that: the box is broken, and possibly a person nearby gets hurt. Embodiment is what turns AI mistakes from "type again" into "physics happened."

That's the core distinction to carry through the rest of this article: generative AI produces content; physical AI perceives an environment and acts inside it, with real, sometimes irreversible consequences.

Side-by-side comparison showing Generative AI turning a text prompt into text, image, or video output entirely within software, versus Physical AI turning real-world sensor data into a physical action, with the changed environment feeding back into the next round of sensing.

Where Physical AI Shows Up Today

Physical AI isn't a future concept. It's already running in production across a handful of recognizable categories.

Robotics is the most direct example. Amazon has deployed its millionth robot across its fulfillment network Source: Deloitte, "AI goes physical". Joseph Quinlivan, Amazon's VP of Fulfillment Technology and Robotics, describes the underlying engineering work as spanning "robot autonomy, manipulation, sortation, and computer vision" Source: Amazon, "Amazon Robotics VP shares his culture of innovation". Humanoid and mobile robots are moving from pilots into deployment too: Agility Robotics' Digit is purpose-built for logistics and warehouse automation and is already running in pilot deployments, and Figure AI has partnered with BMW to bring a humanoid robot into factory work, Source: humynlabs.ai, "What Is Embodied AI?". Picture a robot arm on a warehouse line: it has to see a box in a bin, figure out where and how to grip it without crushing it or the load stacked on top of it, and place it exactly where the next process step expects it, all while the surrounding scene keeps changing as other bins move through.

Autonomous vehicles are the same idea at higher speeds and stakes. Waymo has passed more than 10 million paid robotaxi rides, and Aurora Innovation is running commercial self-driving trucking routes between Dallas and Houston, Source: Deloitte, "AI goes physical". Think about a car approaching an unfamiliar four-way intersection: it has to track every other vehicle and pedestrian in view, estimate where each of them will be a second or two from now, and choose a path that stays safe under all of those predictions at once, updating that whole picture several times a second as new sensor data comes in.

Smart or physical infrastructure rounds out the picture: sensing and acting on the built environment itself rather than on a single vehicle or arm. Utility company Naturgy uses drones for infrastructure inspection, with plans to expand into higher-risk work like high-voltage line and gas pipeline inspection; GE HealthCare is developing autonomous X-ray and ultrasound systems; and the city of Detroit runs Accessibili-D, an autonomous shuttle service for senior and disabled riders Source: Deloitte, "AI goes physical". None of these systems produce a document or an image as their output; their output is a physical action or an inspection of a physical asset.

A map of three physical AI application categories, Robotics, Autonomous Vehicles, and Smart/Physical Infrastructure, each illustrated with example icons such as a warehouse arm, humanoid robot, self-driving car, delivery truck, inspection drone, and autonomous shuttle.

The Sim-to-Real Gap

Here's where physical AI gets genuinely harder than generative AI to train, and it comes down to a simple problem: data.

A language model can learn from text that already exists on the internet in effectively unlimited quantity, and getting more of it doesn't put anyone at risk. A physical AI system doesn't have that luxury. To learn to pick objects reliably, a warehouse robot arm needs to see thousands of failed grips: boxes dropped, tipped, or crushed. To learn to handle an intersection safely, a self-driving car needs to encounter close calls with pedestrians and other vehicles. Collecting that training data for real, on real hardware, in the real world, is slow, expensive, and in the autonomous-vehicle case, potentially dangerous.

The standard workaround is to train in simulation instead: a physics-based virtual environment where a robot or vehicle can "crash" a million times a day at no real cost and no real risk. A widely cited academic survey on the topic sums up why simulation is used so heavily in this field: it provides "a potentially infinite data source" and addresses the "sample inefficiency and the cost of collecting" real-world data, while sidestepping "safety concerns with real robots", Source: Zhao, Queralta & Westerlund, "Sim-to-Real Transfer in Deep Reinforcement Learning for Robotics: A Survey".

The catch is that simulation is never a perfect stand-in for reality. That shortfall is called the sim-to-real gap (also "the reality gap"): the drop in performance a system shows once it moves from the simulator to real hardware. As the survey puts it, "the gap between the simulated and real worlds degrades the performance of the policies once the models are transferred into real robots", Source: Zhao, Queralta & Westerlund. A grip that works perfectly against a simulated box's idealized friction can slip against a real box with a slightly torn label. A driving policy tuned against simulated pedestrians can misjudge a real pedestrian's gait on a rainy night that the simulator never modeled. One robotics expert, Ayanna Howard, Dean of Engineering at Ohio State University, put it plainly: "visual images in simulated environments are pretty good, but the real world has nuances that look different", Source: Deloitte, "AI goes physical".

Closing that gap, through techniques such as randomizing simulated conditions during training or blending in smaller amounts of carefully targeted real-world data, is one of the central engineering problems in physical AI today. It's also the reason physical AI teams tend to ask infrastructure teams for things a generative AI team never would: physics simulation environments, sensor data pipelines, and sometimes physical test rigs alongside the GPU cluster.

What a Physical AI System Needs to Do

Underneath any specific robot or vehicle, a physical AI system runs a version of the same operating loop, repeated many times per second:

1. Perceive: take in raw sensor data (cameras, depth sensors, force sensors, inertial measurement units) and turn it into an understanding of what's currently happening in the environment.
2. Predict: project the current state forward, asking where that pedestrian will be in one second or whether that stack of boxes will stay balanced once the top one is removed.
3. Reason about physics: weigh the available actions against real physical constraints like gravity, friction, and momentum, not just against a learned pattern.
4. Act: send an actual command to a motor, actuator, or steering system, which changes the environment, thus producing new sensor data, which restarts the loop.

Here's what that looks like in practice, as one technical description of embodied AI puts it: "the system sends motor commands to actuators... the action changes the world state, new sensor data comes in, and the loop repeats in real time", Source: humynlabs.ai, "What Is Embodied AI?". A generative AI model produces one output and waits for the next prompt. A physical AI system never stops looping: it's constantly re-sensing a world it just changed.

Perception alone isn't enough, either. Many computer vision systems can perceive (that is, identify objects in an image) without ever predicting what happens next or reasoning about the physical consequences of an action. Physical AI requires all four stages to work together and run continuously, which is exactly why it takes more than swapping a chatbot's language model for a bigger one.

A continuous four-stage loop, Perceive, Predict, Reason about physics, and Act, with an arrow from Act back to Perceive showing the cycle repeats in real time as the environment changes.

Why This Matters for Foundation Models

Training a separate model for every robot, vehicle, and task does not scale; there are too many body types, sensor configurations, and operating environments. To address this, researchers are applying the same foundation-model approach used for language and image models: train one large model on a broad, pooled dataset, then adapt it to many specific tasks. In one example, researchers combined more than one million real robot trajectories from 22 robot types across 33 academic labs into a shared dataset and trained a generalist model on it. When tested across five research labs, that model outperformed specialist models trained on individual robots by 50% on average, showing that "training a single model on data from multiple embodiments leads to significantly better performance across many robots than those trained on data from individual embodiments" Source: Google DeepMind, "Scaling Up Learning Across Many Different Robot Types".

That's the same foundation-model logic reshaping generative AI, applied to bodies instead of text. Instead of hand-building a bespoke model for every warehouse arm, delivery drone, or vehicle platform, a foundation-model approach aims to learn general physical competence once (perception, prediction, physical reasoning, action) and transfer it across many different physical systems. That's the direction this learning path goes next: a closer look at one specific foundation-model approach purpose-built for physical AI, and how it fits into the broader physical AI workflow you've just been introduced to.

Summary

• Physical AI is AI that perceives, predicts, and acts within physical, spatial, and three-dimensional reality, not AI that only generates text or images.
• Embodiment means the AI's decisions are carried out by a real physical body, so mistakes have sometimes irreversible, real-world consequences.
• Physical AI is already deployed today across robotics (warehouse arms, humanoid robots), autonomous vehicles (robotaxis, self-driving trucks), and smart/physical infrastructure (inspection drones, autonomous shuttles).
• The sim-to-real gap (the performance drop between simulated training and real-world deployment) is a central challenge, because collecting enough real-world training data directly is slow, costly, and often unsafe.
• A physical AI system needs more than perception: it must run a continuous perceive → predict → reason → act loop, reasoning about physical constraints like gravity and friction before it ever sends a command to a motor.

References

All sources cited in this article, in order of first citation:

1. Center for Security and Emerging Technology (CSET), Georgetown University. "Physical AI." CSET Publications. https://cset.georgetown.edu/publication/physical-ai/. Accessed July 2026.
2. Deloitte. "AI Goes Physical: Navigating the Convergence of AI and Robotics." Deloitte Tech Trends 2026. https://www.deloitte.com/us/en/insights/topics/technology-management/tech-trends/2026/physical-ai-humanoid-robots.html. Accessed July 2026.
3. Amazon. "Amazon Robotics VP Shares His Culture of Innovation." About Amazon. https://www.aboutamazon.com/news/operations/amazon-robotics-culture. Accessed July 2026.
4. Zhao, W., Queralta, J.P., and Westerlund, T. "Sim-to-Real Transfer in Deep Reinforcement Learning for Robotics: A Survey." arXiv. https://arxiv.org/abs/2009.13303. Accessed July 2026.
5. HumynLabs. "What Is Embodied AI? The Complete Guide to AI That Lives in the Real World." https://humynlabs.ai/blog/what-is-embodied-ai-the-complete-guide-to-ai-that-lives-in-the-real-world. Accessed July 2026.
6. Google DeepMind. "Scaling Up Learning Across Many Different Robot Types." Google DeepMind Blog. https://deepmind.google/blog/scaling-up-learning-across-many-different-robot-types/. Accessed July 2026.