About two minutes · no cost · no commitment
CorpusMarket Robotics
What is physical AI?
Physical AI is artificial intelligence that acts in the real world through a machine, such as a robot arm, a humanoid or a self-driving car. It is the same kind of learning that produced chatbots, pointed at a much harder problem. The thing it is short of is examples of physical work.
By Linus Näslund, Co-Founder
- No cost to apply
- Nothing shared before we talk
- Reply within 2 business days
Contents07
What is physical AI, in plain words
Physical AI is software that sees the world through sensors and acts on it through a machine. A camera shows a cluttered table. The model decides which object to pick up and how. Motors carry out the motion. Then the loop repeats, many times a second.
The word "AI" matters here. The behavior is learned from examples. Nobody writes out each movement by hand. The word "physical" matters too. The output is a motion with consequences, such as a gripped cup, a turned wheel or an opened door.
That is the whole physical AI definition most people need. The rest of this page covers what sets it apart, what it looks like in practice, who is building it, and why progress depends on robot training data.
How it differs from a chatbot and from factory automation
A chatbot takes in text and produces text. If it gets a sentence wrong, you ask again. Physical AI takes in camera images and produces movement, and the world does not offer a second try. A dropped plate stays broken.
The other difference is where the knowledge came from. Chatbots learned from the enormous amount of writing already on the internet. There is no comparable record of what hands do when they wire an outlet or plate a dish.
Physical AI is also different from the robots that have worked in factories for decades. A classic industrial robot repeats one programmed path with great precision. It welds the same seam on the same part in the same position, behind a fence, and it stops if anything is out of place.
| Classic industrial automation | Chatbot | Physical AI | |
|---|---|---|---|
| Takes in | Fixed positions, simple sensors | Text | Camera images, sensors, instructions |
| Produces | The same motion every time | Text | Motion that adapts |
| How it is made | Programmed by engineers | Learned from internet text | Learned from recordings of physical tasks |
| Handles surprises | No | In words only | That is the goal, and it is unfinished |
The promise of physical AI is a machine that copes when the part is crooked, the box is torn or the kitchen is one it has never seen. That promise is only partly met today.
Physical AI vs embodied AI
The two terms describe nearly the same thing, and you will see both.
Embodied AI is the older phrase from research. It means intelligence that lives in a body and learns by acting, as opposed to a program that only reads and writes. Some robot makers use it today. Figure, for example, writes about embodied AI when it describes its humanoid work.
Physical AI is the newer label and the one industry now uses most. NVIDIA uses it broadly, for autonomous machines that perceive, understand and act in the physical world, and it counts factory robots, robotaxis, warehouse robots and humanoids under it.
If there is a difference in use, it is one of emphasis. Embodied AI stresses the body and the learning. Physical AI stresses the whole field, including the chips, the simulators and the vehicles. For a reader trying to understand the topic, they can be treated as the same.
Physical AI examples by category
- Humanoids. Two arms, two hands, usually two legs. The shape is chosen so the robot can use spaces and tools built for people. Figure describes its Figure 03 as a general purpose humanoid, and Agility Robotics builds a humanoid called Digit for industrial environments.
- Warehouse arms. Arms that pick items out of bins and place them in boxes. This is one of the most mature uses, because the setting is controlled and the task repeats. See robots in warehouses.
- Self-driving vehicles. Waymo describes its service as autonomous ride-hailing, with no human driver in the car. Driving was the first large physical AI effort, and it shows how long the last stretch can take.
- Drones. Aircraft that hold position, avoid obstacles and inspect or map without a pilot steering every move.
- Household robots. Robot vacuums are the familiar case. Home robots with arms are newer. 1X sells a home robot called NEO, and its own description is candid: for chores NEO does not know yet, a 1X expert can supervise it remotely.
That last detail is typical of the field. The machines exist, and much of what they do still leans on people.
Physical AI companies: who does what
It helps to sort physical AI companies by role, since no single company does everything.
| Role | What they do | Examples |
|---|---|---|
| Model builders | Train the models that turn camera images and instructions into motion | Physical Intelligence publishes its π0 models with open code. Google DeepMind develops the Gemini Robotics models |
| Robot and vehicle makers | Build the machines, often with their own models | Figure, 1X, Agility Robotics, Waymo |
| Chip and simulation platforms | Supply the computing, virtual worlds and starter models others build on | NVIDIA offers Isaac Sim for simulation and synthetic data, Cosmos for world models, and the open GR00T models for humanoids |
| Data suppliers | Record or collect the examples the models learn from | Data-collection companies, and programmes that pay businesses for recordings of their crews |
The roles blur. Robot makers train models, and platform companies release models of their own. The models at the center of all this are usually vision-language-action models, explained in our guide to VLA models. Simulation leans more and more on world models.
Why data is the bottleneck
All of these systems learn from examples, and they need a great many. How robots learn goes through the methods. The short version is that a model gets good at what it has seen done, in settings like the ones it has seen.
Labs have three ways to get examples. They can steer robots by hand and record the result, which is accurate and slow. They can generate examples in simulation, which is fast and never quite matches reality. Or they can use video of people doing the work.
Video of people is the only source that exists at anything like the needed scale and variety. It has to be the right kind, though: filmed from the worker's own point of view, with the hands in frame, in real workplaces. That is called egocentric video, and very little of it has been recorded so far.
Where businesses come in
The places where hands-on work happens belong to businesses. A plumber's van, a restaurant line and a framing crew's job site are exactly the settings robots have seen least.
That is why some programmes now pay businesses whose crews record their ordinary work. CorpusMarket takes enquiries from those businesses and introduces the ones that fit to a recording programme run by our partner Amelia Robotics. If your people work with their hands, tell us about your business. It takes two minutes, and we reply within 2 business days.
Questions and answers
Frequently asked questions
What is physical AI in simple terms?
AI that controls a machine in the real world. It takes in what cameras and other sensors pick up, decides what to do, and moves a body, such as a robot arm or a vehicle, to do it.
What is embodied AI, and is it the same as physical AI?
Embodied AI means AI that has a body and learns by acting with it. In practice the two terms overlap almost completely. Embodied AI is the older research term, and physical AI is the label industry has adopted more recently.
What are some physical AI examples?
Humanoid robots, robot arms that pick and pack in warehouses, self-driving vehicles, drones that fly and inspect on their own, and household robots. What they share is that a learned model, and not a fixed script, decides the motion.
Which kinds of companies work on physical AI?
Four broad groups. Labs that build the models, companies that build the robots and vehicles, platforms that supply chips and simulation software, and suppliers of training data.
Why is physical AI harder than a chatbot?
A chatbot learned from text that already existed online, and a wrong word costs little. A robot has no ready-made archive of physical work to learn from, and a wrong motion can break something. Both problems come back to data.
More pages
Read next
How robots learn: imitation learning and diffusion policy explained
Most robots that do something new today learned it by imitation learning, which means copying recorded examples of the task. This guide explains that method and the others around it without the math, and ends with what each one needs in data.
Read moreWhat is a VLA model?
A VLA model is the kind of AI model that drives most new general-purpose robots. It looks through the robot's cameras, reads an instruction in ordinary language, and outputs the movements to carry it out. How well it does that depends heavily on what it was shown during training.
Read moreRobotics training data
A language model learned from text that was already on the internet. A robot has no such archive to learn from. Robotics training data has to be made on purpose, and most of the cost and effort in teaching a robot goes into getting enough of it.
Read moreNext step
Tell us about your business
If your crew does hands-on work, a robotics lab may pay for video of it. Send your details and we will tell you whether it fits and what happens next.
- No cost to apply
- Nothing shared before we talk
- Reply within 2 business days