Physical AI: What It Is and Is

0 1 7 minutes read

Physical AI is .

NVIDIA is talking about it, consulting firms are talking about it, and so are investors and robotics startups. They all talk about Physical AI. Suddenly, this word is everywhere.

But what exactly is Physical AI? And most importantly, what is missing?

In this post, let's narrow down the wording. We will not only explain the concept itself but also try to distinguish it from the adjacent terms that are often confused with it — global models, integrated AI, physics AI, and digital twins.

1. What is Physical AI

A working definition that I find useful is this [1]:

Physical AI is that AI it closes the loop between seeing and acting in the real physical world.

Most AIs today live on screens. They perform tasks such as classifying images, summarizing documents, writing emails, or recommending which movie to watch next. They are all really useful activities, but they all take place in the digital world.

Physical AI is different than that. It comes out of the digital world and into reality interacts with the real world. It takes in what is happening around it through the senses, works out what needs to be done, and finally does it. The action can be accomplished by a robotic arm, a humanoid, a drone, a self-driving car, or an industrial machine on a factory floor.

Physical AI output is no longer just text on the screen; it is a real movement in the world.

Think about picking up a cup. A chatbot that just explains how to do it is not Physical AI. But a robot that can recognize a cup, and adjust the grip to move the cup where you want, that's very close to Physical AI.

Therefore, Physical AI is not a single model that does one thing. It is a whole system that should:

Figure 1. Physical AI: A system that can sense, understand, predict, plan, act, and receive feedback. (Photo by author)

With this definition in mind, we can now separate Physical AI from the adjacent terms that it is often confused with.

Let's start with the World Model.

2. Physical AI vs World Model

Name World Models [2] it often comes up in similar discussions of robots, autonomous agents, simulations, artificial environments, etc. If we're not careful, it's very easy to slip into treating it as another way to say Physical AI.

But that is not true.

World Model is a model, as the name implies. Practically, this model is an internal representation of how the environment is changing. What naturally arises from it is prediction strength. It allows the agent in that location to wait for what's next. If I move forward, what do I encounter? When I push this object, does it slide or roll? If that car changes lanes, where will it be in two seconds? All those questions can be answered with that predictive ability.

A robot might use a world model to imagine what would happen if it approached the cup from different angles. But the important thing to note here is that the world model itself is not a robot, not a gripper, not a car controller, and not a full system that moves the cup.

That's actually the key relationship here: The world model can live inside Physical AI, especially for simulation, training, and programming. But by itself, it only predicts the possible future; it doesn't. Without sensors/controllers/actuators that connect to the real world, it can only think.

Figure 2. World model predicts, Physical AI Actions.

To summarize: world models predict possible. Physical AI actions in the real world.

3. Physical AI vs Integrated with AI

This is a very mixed couple. You see that many people and companies use these two terms interchangeably. In fact, in a recent post from Cambridge Consultants [3]Capgemini's chief innovation officer literally called it “physical AI, or hybrid AI as some call it.”

But I believe there are useful distinctions to keep in mind, and the key is about what they emphasize.

Physical AI emphasizees what the system doesand this includes seeing, deciding, and acting in the real world. Embedded AI, on the other hand, is about what constitutes intelligence.

The word “embodied” in its name comes from the embodiment hypothesis [4]. That theory believes that intelligence doesn't just reside in the software you load into a machine. It grows with the body. Physically, it can feel, try things, and learn from what happens. The body here is not an output device at the end of the process; it's part of how he learns to think in the first place.

In practice, however, Physical AI and Embodied AI often go hand in hand. This is especially true in practical robotics discussions, as most real robots are a combination of AI and Physical AI.

Figure 3. Virtual AI in the world. Integrated AI learns from the body. (Photo by author)

To summarize: Physical AI is about simulation in the world. Embedded AI is about artificial intelligence and its interaction with the environment.

4. Physical AI vs Physics AI

This couple is easy to confuse if you are not careful in reading their names. One is Physicalanother one Physics.

When people say Physics AI, what they are usually referring to are AI methods combine the laws of physics. A good example is the physics-informed neural network, or PINN [5].

Technically, PINN is a regular neural network. What makes them different is the composition of the loss. On top of the double loss of the ratio between prediction and observation, the model is also penalized whenever its prediction violates the observed statistics or constraints.

Take battery cooling for example. Say we want to predict the temperature evolution inside the battery pack under different battery operating conditions. A pure data-driven neural model will learn the pattern directly from the sensor readings. PINN can learn from the same data, but in addition, it is also required to respect energy balance measurements and known initial/boundary parameters by injecting those physics-informed loss terms.

This is very useful because the model can produce reasonable predictions instead of falsifications. Also, it has the potential to allow the model to achieve the same level of accuracy but with a reduced data size. An additional advantage is that it can adapt better under conditions that cannot be covered by the training set.

However, on its own, physics AI is still only a model. It can predict or simulate a physical process, but it does not act in the real world.

And that is the main difference in Physical AI.

Basically, Physics AI can cross over into Physical AI if it is connected to an action loop. For our battery cooling example, if the intelligent cooling control system learns from the predictions generated by the temperature model and adjusts the fan speed accordingly in real time, the physics-aware model is now part of the Physical AI system.

Figure 4. Physics AI is about the laws of nature. Physical AI is about physical action. (Photo by author)

To summarize: Physics AI is about the laws of nature. Physical AI is about physical action.

5. Physical AI vs The Digital Twin

Digital twins are another confusing term, especially in industrial settings such as factories, warehouses, or private systems.

Here is the widely accepted definition of a digital twin [6]: I am a visual representation of a real physical object, process, or system. A digital twin is often connected to data from the physical system, and can update itself when the physical system changes.

On the factory floor, a digital twin production line is often used. A factory production line usually consists of machines, conveyors, sensors, products, quality control stations, etc., and the corresponding digital twin is the physical counterpart of that line. Imports various sensor readings, machine status, and maintenance records to understand the current state of the line. As the actual production line changes, it updates the visual view in real time.

Now, with this virtual twin, production engineers can monitor the health of the line and potentially identify performance issues. Another common use of a digital twin is to simulate scenarios to develop operational policy before implementing it in practice.

But none of this makes a digital twin a Physical AI, since a digital twin doesn't decide or create an actual production line by itself.

Figure 5. A digital twin represents a physical system. Physical AI works on a physical system. (photo by author)

To summarize: A digital twin represents a physical system. Physical AI works on a physical system.

6. Putting it together

Now that we look back at the different terms we've put together, we can see one pattern stand out: global models, hybrid AI, physics AI, and digital twins are all about understanding or representing the physical world. Physical AI is the only one defined by it to play in it.

In a real system, those different concepts can (and probably should) work together: to support a single robot operating in a virtual world, digital twins may be used to assist the robot system, and a world model or a physics-aware model may be used to generate predictions to guide the robot's action.

Hopefully, the next time you come across these words in an article, paper, or product brochure, you'll be able to separate them instead of letting them blur together.