---
title: "What is Physical AI? Robots in factories | Quantlix"
description: "Physical AI is AI that sees, decides and moves things. See what factory robots do today, what the leading labs report, and how to judge a robot for your line."
canonical: "https://quantlix.com/en/blog/what-is-physical-ai-robots-in-factories/"
last-updated: 2026-11-21
---

Source: https://quantlix.com/en/blog/what-is-physical-ai-robots-in-factories/
Language: en
Description: Physical AI is AI that sees, decides and moves things. See what factory robots do today, what the leading labs report, and how to judge a robot for your line.

Perspective Physical AI

# What is Physical AI, and are robots ready for factory work?

Physical AI is AI that sees, decides and moves things. See what factory robots do today, what the leading labs report, and how to judge a robot for your line.

[Ali Muhammad](https://quantlix.com/en/about/leadership/ali-muhammad/)

Founder and Chief Executive Officer

Nov 21, 2026 (2026-11-21) 15 min read

![Orange industrial robots along an assembly line, with the headline “What a robot needs before it can work a shift”.](https://quantlix.com/_astro/insight-what-is-physical-ai-robots-in-factories.TGUip7Qv_Z1sTePo.webp)

01 / Engineering notes

Field notes: Blog

**On this page**

The short answer

Physical AI is artificial intelligence that sees through sensors, decides what to do and moves something in the real world. Robots already pay their way in narrow, repeated, measurable jobs such as camera inspection, palletizing and moving goods. Humanoid robots are improving and handle single jobs at a few sites, but a full, varied shift is still ahead.

In its 25 June 2026 announcement, [BMW Group](https://newsroom.bmwgroup.com/global/bmw-group-advances-the-use-of-physical-ai-in-production-with-figure-03-project-in-spartanburg) said a humanoid robot had spent roughly ten to eleven months at its Spartanburg plant placing sheet-metal parts ahead of welding, and that the pilot supported production of more than 30,000 BMW X3 vehicles. A few years ago that sentence would have belonged in a film. **Physical AI** is the name for AI that sees, decides and moves something in the real world, and that robot is one example of it.

If you run a plant, you have a more practical question than whether this is impressive. Could a machine like that work a shift on your floor, with your parts, your heat and your people? And if not that robot, then what? This guide covers what Physical AI is, why moving is harder than talking, what the leading teams have shown in 2026, where robots already pay their way, and what to ask before one comes through your door.

## What is Physical AI, in plain words?

Physical AI is artificial intelligence that acts in the physical world instead of on a screen. It does three things, always in the same order: it sees through cameras and sensors, it decides, and it moves something.

Most of the AI you have met lives on a screen. It writes emails and summarizes reports, and when it gets something wrong you notice and fix it.

A camera above a bottling line that spots a dented bottle, decides it should not be packed and nudges it off the belt with a small pusher is Physical AI at its simplest. A humanoid robot that walks into a warehouse and does whatever it is asked is the same idea at its most ambitious. Most of what works reliably today sits much closer to the bottle camera, and for many plants that is the better place to start. What is new is how much of the seeing and deciding a machine can now do for itself.

## Why is moving harder for AI than talking?

A robot has to be right far more often than a chat assistant, because its mistakes land in the physical world and cost time, parts or safety. A chat assistant that is right nineteen times out of twenty is a useful colleague: you read its answer and catch the odd slip.

Here is an illustration with round, made-up numbers. A robot that is right nineteen times out of twenty drops one part in twenty. If the line feeds it 2,000 parts a shift, that is about 100 drops, and somebody has to walk over and pick up every one.

The physical world also never holds still. A chatbot is handed tidy text. A robot is handed a carton that arrived slightly crushed, a floor that was mopped ten minutes ago, afternoon sun across the conveyor and a colleague who has just stepped into its path. Any one of those might be something it has never seen.

So a factory asks a different question from a film crew. Not “can it do this?” but “can it do this ten thousand times, in conditions nobody planned for, without hurting anyone?”

## How do robots learn now, instead of being programmed step by step?

Robots are increasingly taught by example: they learn movements from demonstrations and from practice in simulation, instead of an engineer specifying every move. That is a change from the long-established method, where an engineer told an arm exactly where to go, how fast and in what order.

The older method is still everywhere. The International Federation of Robotics (IFR) reported on 25 September 2025 that [542,000 industrial robots were installed in 2024](https://ifr.org/ifr-press-releases/news/global-robot-demand-in-factories-doubles-over-10-years), more than double the number ten years earlier, and that 4,664,000 were in operational use worldwide that year.

Simulation is where much of the new practice happens, because a thousand dropped boxes there cost nothing. On 18 March 2025, [NVIDIA](https://investor.nvidia.com/news/press-release-details/2025/NVIDIA-Announces-Isaac-GR00T-N1--the-Worlds-First-Open-Humanoid-Robot-Foundation-Model--and-Simulation-Frameworks-to-Speed-Robot-Development/default.aspx) released GR00T N1, which it calls an open, fully customizable foundation model for humanoid robots (a large, general-purpose model that robot makers can adapt to their own machines).

The company said it had generated 780,000 synthetic motion trajectories in 11 hours, the equivalent of 6,500 hours, or nine continuous months, of human demonstration. Combining the synthetic data with real data, it reported, improved the model’s performance by 40 percent compared with real data alone. The release does not say which measure that refers to, and all of these figures are company-reported.

The thinking is moving onto the robot, too. On 25 August 2025, NVIDIA announced that its [Jetson AGX Thor](https://nvidianews.nvidia.com/news/nvidia-blackwell-powered-jetson-thor-now-available-accelerating-the-age-of-general-robotics), a computer designed to run AI models at the edge, was available. The company named Agility Robotics, Amazon Robotics, Boston Dynamics, Caterpillar, Figure, Hexagon, Medtronic and Meta as early adopters.

Figure’s CEO said in the release that the company needs to “run powerful AI models directly on the robot.” The same reasoning applies to a factory camera: a machine cannot wait for a data center far away to answer before it closes its hand.

## What have the leading teams actually shown in 2026?

The leading teams report one defined job done many times in real sites, and lab success rates that vary widely by task. Both are useful to read, and the numbers are company-reported.

On 30 July 2026, [Google DeepMind](https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/) released Gemini Robotics 2 and published its success rates. The whole-body and gripper figures are averages across each skill category’s tasks. The multi-finger figures are per task.

| Robot setup | Task | Reported success rate |
| --- | --- | --- |
| Apptronik Apollo 2 humanoid, whole-body skills | Pick up from a shelf | 76.3% |
|  | Pick up from a table | 68.4% |
|  | Pick up from the floor | 45.7% |
| Apollo 2 with five-fingered hands | Unscrew a light bulb | 92% |
|  | Tie a trash bag | 44% |
|  | Use a ziplock bag | 40% |
|  | Screw in a light bulb | 36% |
|  | Use a dustpan | 32% |
| Franka Duo arms with grippers | Precise insertion | 89.6% |
|  | Diverse tool kitting | 78.9% |
|  | General pick and place | 74.2% |

DeepMind itself writes that multi-finger dexterous manipulation “remains challenging,” and that its robots “have more to advance in movement speed.” The same Apollo 2 humanoid picked things up from a shelf 76.3 percent of the time and from the floor 45.7 percent of the time, so a small change in the task can move the result a long way.

Out in working sites, the pattern is one clearly defined job. In a release dated 5 November 2026, [Agility Robotics](https://www.agilityrobotics.com/content/agility-becomes-first-u-s-publicly-listed-pure-play-humanoid-robotics-company-begins-trading-today-on-nasdaq-as-aglt) said its Digit robot had moved more than 100,000 totes (the plastic crates warehouses run on) at a GXO site in Georgia, and reported more than 65,000 hours of operation after nearly three years of real-world deployment. The same release says GXO plans to deploy Digit at two more U.S. logistics facilities in early 2027. BMW’s June 2026 announcement describes a new project at Spartanburg with the newer Figure 03: unsorted parts arrive in bulk containers, and the robot sorts them into a sequencing trolley in the plant’s logistics area.

These examples have a lot in common. Each is one defined job, repeated thousands of times. That is how reliable machines have usually left the laboratory.

## Where do robots already earn their keep?

In our view, robots pay back most reliably where a job is narrow, repeated and easy to measure. These are the places we would look first in a food plant or a factory.

| The job | What the machine does | Why it works today | What to watch |
| --- | --- | --- | --- |
| Checking quality at the end of the line | A camera and a small computer look at every item and reject the bad ones | The light and position can be controlled, and the decision is simple | Glare, dust on the lens, and new packaging nobody told the system about |
| Palletizing | An arm stacks finished cases onto pallets in a set pattern | The cases are the same size and arrive at the same spot | Changeovers to a new case size, and guarding around the arm |
| Moving goods | Self-steering carts carry materials between stores and lines | The routes repeat and can be mapped | Busy aisles, people stepping out, and wet or uneven floors |
| Picking and placing | An arm moves parts that arrive in a fixed position | The part, the grip and the target are always the same | Parts that arrive jumbled, which is where the hard problems begin |
| Welding, painting, dispensing | A programmed arm follows a precise path | A fixed path suits a machine that repeats it exactly | Fixtures that hold each part in exactly the same place |

None of these needs legs. Most need a camera, a sensor, an arm and careful engineering around them. The promise of the humanoid is that one machine could move between jobs like these during a single shift. Until it can do that reliably, a specialist machine usually wins.

## What does a robot decision look like at the end of a biscuit line?

It looks like studying the job before choosing the machine. This is an imagined plant, not a real client or a real result, but the situation is a familiar one.

At the end of the line, sealed cases come off the case packer every few seconds, and two people take turns lifting each one onto a pallet. It is heavy, repetitive work in a hot summer, and the plant struggles to keep people in the job. The owner has seen the humanoid videos and wonders whether one could take over.

The cases are the same size all day, they arrive at the same spot and the stacking pattern never changes. A standard palletizing arm handles that job with very little drama. A humanoid could do it in principle, but the plant would be paying for flexibility the job never uses.

So the arm goes in, inside a guarded area, with a light curtain (an invisible beam that stops the robot the moment someone walks through it) on the open side. The two people move upstream to the case packer, where a jam needs a human eye and a quick hand. Then comes a smaller change that often pays back sooner: a camera over the case sealer that checks every flap is closed and every label is straight, so a bad case is pushed aside before it reaches the pallet. Our [packaging](https://quantlix.com/en/industries/packaging/) page describes the kind of line this applies to, and our [IoT and edge computing](https://quantlix.com/en/services/iot-edge-computing/) service covers connecting cameras like that one to a line’s equipment.

Nothing in that story needs a breakthrough. It needs someone to measure the job first. And when general-purpose robots do arrive at sensible prices, this plant will already know which jobs it has, how often each is done and how well.

## How do you keep people safe around a robot?

Safety around a robot comes from the machine’s own safety design and a risk assessment of the whole installation, not from how clever the AI is. International machine-safety standards cover industrial robots and the cells they work in, and they are revised from time to time. Ask the supplier which standards the installation was designed to and who assessed it. This article is not safety or legal advice: the people responsible for the equipment, and their assessors, decide what applies.

Keep what the robot sees separate from what it does. On our [Physical AI](https://quantlix.com/en/services/physical-ai/) page we describe the approach this way: safety functions stay independent of the AI, so they keep working if the model, the camera or the network fails, and operators have a plain way to stop, override and take over by hand. Google DeepMind says its ER 2 model can bring a robot to a safe stop if someone approaches too closely. That is a developer’s description of a research model, and it sits on top of a plant’s safety design instead of replacing it.

A modern robot is also a connected computer, so who can reach it from outside the building is a safety question as well as an IT one. The [IEC 62443](https://quantlix.com/en/standards/iec-62443/) series is the international set of standards for securing industrial automation and control systems.

## Which five questions should you ask before a robot comes through the door?

Ask about five things: how long the robot runs between interventions, what it does when something is out of place, how people stay safe around it, what an hour of its work really costs, and who repairs it. None of them is about how clever the robot is.

**How long does it run before a person has to step in?** A demonstration is filmed until it works. A shift has no second take. Ask for that figure over days of continuous running, ideally on your parts, in your light and at your temperature.

**What does it do when something is out of place?** A good system copes with small surprises, and stops safely and asks for help with the rest. Ask to see it confused. Ask, too, how it recovers when power drops mid-cycle, and how dust and heat affect its sensors and electronics.

**How does it keep people safe?** Ask for a plain-language answer about the whole installation: guarding, stopping behavior, who assessed it and how outside access is controlled. The section above explains why.

**What does an hour of its work really cost?** Add the price, the engineering to fit it into your line, maintenance, spare parts, power and the person who keeps an eye on it, then divide by the hours it will really run, not the hours in the brochure. Compare that with what the job costs you today. Sometimes the robot wins clearly, and often a simpler machine wins by more.

**Who fixes it at three in the morning?** Ask where the nearest engineer is, how long a spare part takes to arrive and what your own team will be trained to repair. A robot waiting weeks for a part is a very expensive statue.

## Where to start

You can do all of this in a month, without buying anything.

1. Walk your floor with a notebook and list the jobs that are dull, repeated and easy to measure. For each, write down how often it is done, how long it takes, how many people it needs across a day and what tends to go wrong.
2. Pick the one that hurts most: hard to staff, hard on people’s backs, or the source of most of your rejects.
3. Ask what the simplest machine is that could do it reliably. Sometimes that is a robot. More often it is a camera, a sensor, a better fixture or a small change to the conveyor. Our guide to [starting with AI in a factory](https://quantlix.com/en/blog/how-to-start-with-ai-in-a-factory/) walks through that first step.
4. Note the conditions around the job: temperature, dust, light and how steady the power is. They decide what any machine can do there.
5. Measure the job for a few weeks before you change anything. You will need that number later to know whether the change worked.

Robots are improving at a remarkable pace, and it is an exciting time to work with machines. The plants that gain the most will not be the ones that buy the most impressive robot, but the ones that understood their own work well enough to choose the right one.

If there is a job on your floor you have wondered about handing to a machine, tell us about that one machine or one line on our [contact page](https://quantlix.com/en/contact/), and we will reply with a short note on what we would measure first.

## Key takeaways

1. Physical AI is AI that sees, decides and moves something in the real world, from a camera that rejects a bad bottle to a humanoid robot.
2. Robots pay back most reliably on narrow, repeated, measurable jobs: welding, palletizing, camera inspection, moving goods and picking parts that arrive in a fixed position.
3. In Google DeepMind's July 2026 report, robot success rates ranged from 32 to 92 percent depending on the task, so a demo video is not a shift.
4. Before a robot comes in, ask how long it runs between interventions, how it fails, who keeps people safe, what an hour costs and who repairs it.
5. Measure one dull, repeated job for a few weeks before changing anything. That number is what any good supplier will ask for first.

## Questions people ask

### What is Physical AI in simple words?

Physical AI is artificial intelligence that acts in the real world instead of on a screen. It sees through cameras and sensors, decides what to do, then moves something: a robot arm, a self-steering cart or a pusher on a conveyor. A camera that spots a damaged bottle and removes it from the line is a simple example. A humanoid robot is the most ambitious one.

### Are humanoid robots ready for factory work?

For a few narrow jobs, the companies involved say yes. BMW Group reported in June 2026 that a Figure humanoid supported production of more than 30,000 X3 vehicles in a pilot, and Agility Robotics reported in November 2026 that its Digit robots had moved more than 100,000 totes at one warehouse. A full shift of varied work is a different test, and success rates in recent lab reports still vary widely.

### What can robots reliably do in a factory today?

Robots are most dependable on jobs that look the same every time: following a fixed path to weld, paint or dispense, stacking cases in a set pattern, moving parts that arrive in a known position, carrying goods along mapped routes, and checking products by camera under controlled light. When the job rarely changes, the robot rarely meets something it has not seen.

### How do robots learn new tasks now?

Increasingly by example instead of step-by-step programming. A robot can learn from human demonstrations and from practice in computer simulation, where mistakes cost nothing. NVIDIA reported in March 2025 that it generated 780,000 simulated movements for its GR00T N1 model in 11 hours. Results on real hardware still vary by task, so testing on your own floor remains essential.

### Is it safe to have a robot working near people?

It can be when the whole installation is designed and assessed for it, not just the robot. Safety comes from independent safety functions, guarding or sensing around the robot, a clear way to stop it and a risk assessment by qualified people. Ask the supplier which machine-safety standards the installation was designed to, who assessed it, and how outside access to the robot is controlled.

### What should a small plant do about robots now?

Pick one dull, repeated, measurable job and ask what the simplest machine is that could do it well. Often that is a camera, a sensor or a better fixture instead of a robot. Measure the job for a few weeks before and after any change. Knowing your own tasks this well is also the best preparation if general-purpose robots become affordable later.

## Sources

1. [Google DeepMind, Gemini Robotics 2 brings whole-body intelligence to robots (30 July 2026)](https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/)
2. [Agility Robotics, Agility becomes first U.S. publicly listed pure-play humanoid robotics company, begins trading on Nasdaq as AGLT (5 November 2026)](https://www.agilityrobotics.com/content/agility-becomes-first-u-s-publicly-listed-pure-play-humanoid-robotics-company-begins-trading-today-on-nasdaq-as-aglt)
3. [BMW Group, BMW Group advances the use of Physical AI in production with Figure 03 project in Spartanburg (25 June 2026)](https://newsroom.bmwgroup.com/global/bmw-group-advances-the-use-of-physical-ai-in-production-with-figure-03-project-in-spartanburg)
4. [NVIDIA, NVIDIA Announces Isaac GR00T N1, the World's First Open Humanoid Robot Foundation Model, and Simulation Frameworks (18 March 2025)](https://investor.nvidia.com/news/press-release-details/2025/NVIDIA-Announces-Isaac-GR00T-N1--the-Worlds-First-Open-Humanoid-Robot-Foundation-Model--and-Simulation-Frameworks-to-Speed-Robot-Development/default.aspx)
5. [NVIDIA, NVIDIA Blackwell-powered Jetson Thor now available, accelerating the age of general robotics (25 August 2025)](https://nvidianews.nvidia.com/news/nvidia-blackwell-powered-jetson-thor-now-available-accelerating-the-age-of-general-robotics)
6. [International Federation of Robotics, Global robot demand in factories doubles over 10 years (25 September 2025)](https://ifr.org/ifr-press-releases/news/global-robot-demand-in-factories-doubles-over-10-years)

Sources checked on Nov 21, 2026.

- Perspective
- Physical AI

Written by

## Ali Muhammad

Founder and Chief Executive Officer · Quantlix

[See full profile](https://quantlix.com/en/about/leadership/ali-muhammad/)

[Talk to Ali](https://quantlix.com/en/contact/)

[More blog](https://quantlix.com/en/blog/)

## Keep reading

[More blog](https://quantlix.com/en/blog/)

- ![A circular conveyor in a large, bright factory hall, with the headline “The model is the easy part”.](https://quantlix.com/_astro/insight-start-ai-in-a-factory.BsmMIis5_Z1XFbFd.webp)

  Playbook

  Industrial AI Nov 21, 2026 (2026-11-21)

  ### [How should a factory start with AI?](https://quantlix.com/en/blog/how-to-start-with-ai-in-a-factory/)

  Start AI in a factory with one line and one costly problem. Before you buy, check five things: machine data, wrong answers, an owner, internet loss and access.

  15 min read Read
- ![An open electrical control cabinet with wired modules, with the headline “Open AI models, safe machines”.](https://quantlix.com/_astro/insight-open-weight-ai-models-factory-security.DRyiqiUu_1rrXUK.webp)

  Perspective

  Industrial security Nov 21, 2026 (2026-11-21)

  ### [Do open AI models make attacks on factories easier?](https://quantlix.com/en/blog/open-weight-ai-models-factory-security/)

  Do open-weight AI models make attacks on factories easier? Slightly, and the defense is familiar: list what is reachable online, close it, guard the rest.

  16 min read Read
- ![A laboratory workbench with a microscope and instruments, with the headline “When the lab runs itself”.](https://quantlix.com/_astro/insight-self-driving-labs-ai-experiments.xqxXQnT2_Z4wnNE.webp)

  Perspective

  Physical AI Nov 21, 2026 (2026-11-21)

  ### [What is a self-driving lab, and does it work yet?](https://quantlix.com/en/blog/self-driving-labs-ai-experiments/)

  A self-driving lab lets AI choose experiments and robots run them. Here is what has been shown since 2023, what was disputed, and what a factory can borrow.

  14 min read Read

## Need the version tailored to your platform, risk model and roadmap?

Bring us the context. We will turn the patterns in this library into a pragmatic delivery path your team can use.

[Plan a workshop](https://quantlix.com/en/contact/)
