日本語Contact us
VisionProductsCase StudyCompanyCareersContact us日本語

What is physical AI?

Physical AI is AI that perceives the real world through cameras and sensors and acts on it through the body of a robot or machine. Where generative AI works on a screen, producing text and images, physical AI works in factories and on streets, grasping, moving, and machining real objects.

The term is used almost interchangeably with embodied AI: AI that has a body and acts on the physical world.

ZREK Inc. | Last updated

Perceive, decide, act

Physical AI works by repeating these three steps. Because it checks the result of every action, it can adapt when conditions change.

  1. Perceive

    • Cameras
    • Depth sensors
    • Force sensors

  2. Decide

    • Trained AI models
    • Robot foundation models (VLAs)

  3. Act

    • Robot arms
    • Hands
    • Mobile bases

Check the result, then perceive again

Figure: How physical AI works

  1. 01Perceive

    Cameras, depth sensors, and force sensors capture the position, shape, orientation, and condition of objects, turning what a person senses with eyes and hands into numbers.

  2. 02Decide

    From what it perceives, the system decides how to move next. Trained AI models do this work, including robot foundation models (VLAs) that handle images, language, and actions together.

  3. 03Act

    Robot arms, hands, and mobile bases apply force to real objects. Sensors then check the result, and the next motion is corrected if anything is off.

How it differs from industrial robots

Conventional industrial robots repeat taught motions with precision. Robots with physical AI look at the situation in front of them and adjust.

ComparedConventional industrial robotsRobots with physical AI
How motions are createdA person teaches positions and sequences one by oneLearned from human demonstrations or simulation data
When a part's position or type changesMust be re-taughtRecognizes the change with a camera and adapts the motion
Best suited forHigh-volume runs of identical partsHigh-mix, low-volume work and parts that vary
Scaling to another siteRebuilt from scratch at each siteLearned motions and models are brought over and refined on site

Industrial AI robots

On one side is physical AI, best known from self-driving cars and humanoids. On the other are industrial robots, which have repeated fixed motions with precision for decades. ZREK works where the two meet.

Physical AISelf-driving carsHumanoidsDronesService robotsIndustrial AI RobotsHigh-mix machine tendingInspecting variable partsWork needing re-teachingIndustrial RobotsConventional armsTeach-and-repeatWeld, paint, assembleFixed, repeated motion

Frequently asked questions

What is the difference between physical AI and generative AI?
Generative AI produces things on a screen: text, images, and code. Physical AI connects that kind of perception and reasoning to a robot's body so it can act on real objects. On a screen, a mistake can simply be undone; in the real world, every action leaves a result, so safety and reliability matter far more.
What are some examples of physical AI?
Self-driving cars, humanoids, drones, delivery and cleaning robots, and industrial robots with built-in AI. In manufacturing, it is starting to be used for work that varies too much for conventional robot teaching to pay off, such as loading and unloading machine tools and inspecting parts that differ from piece to piece.
How does physical AI learn motions?
There are three main ways: imitation learning from human demonstrations, trial and error in simulation, and retraining on data collected on the floor. Not every motion has to be learned. A common setup is to build fixed motions with conventional control and hand only the variable parts to AI.
What is a robot foundation model (VLA)?
A VLA (vision-language-action model) is a single AI model that handles images, language, and actions. Given camera images and an instruction such as "put the part in the box," it outputs robot motions directly. VLAs are pretrained on large amounts of robot data, and the hope is that they can adapt to new tasks with little additional data.
Why is simulation used?
Running many trials on real robots takes time and can damage machines and objects. A virtual environment that reproduces physics lets robots run many trials safely and quickly. Making motions learned in simulation work in the real world, known as sim-to-real transfer, is a major technical challenge.
What is ZREK?
ZREK Inc. is a Japanese startup building physical AI and AI robots for manufacturing. We turn factory work into robot work quickly and make a motion built once reproducible in any factory. Founded in 2021, ZREK is headquartered in Shibuya, Tokyo, with its research lab, ZREK Labs, in Kawasaki, Kanagawa.

Key terms

References

  1. VLAs (robot foundation models)Brohan et al. (2023). RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control. CoRL
  2. VLAs (an open-source model)Kim et al. (2024). OpenVLA: An Open-Source Vision-Language-Action Model. CoRL
  3. Pretraining on data from many robotsOpen X-Embodiment Collaboration (2024). Open X-Embodiment: Robotic Learning Datasets and RT-X Models. ICRA
  4. Sim-to-real (randomizing appearance)Tobin et al. (2017). Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World. IROS
  5. Sim-to-real (randomizing physics)Peng et al. (2018). Sim-to-Real Transfer of Robotic Control with Dynamics Randomization. ICRA
Our vision