# Physical AI companies: who builds it, and who pays for video of real work

> Physical AI companies in 2026: robot makers, model labs and data firms, what real-world video they pay for, and how a business with a crew gets paid.

Canonical: https://origindatapartners.com/physical-ai/physical-ai-companies
By: Origin Data Partners
Status: Published
Published: 2026-11-15
Updated: 2026-11-15
Sources checked: 2026-11-15

## Quick answer

Physical AI companies fall into three groups. Robot makers such as Figure, Tesla and Agility Robotics build the machines. Model labs such as Physical Intelligence, Google DeepMind and NVIDIA build the software that controls them. Data companies such as Scale AI, Appen, micro1 and Memories.ai collect and label the footage both groups train on, and they are the ones that pay for video. Household footage is already cheap. Skilled commercial work, recorded with consent, is scarcer. A business with a crew can tell Origin what it does and find out which programs fit.

## Key points

- Startups that sell physical AI training data raised $2.1 billion across 55 deals in 2026 to date, up from $1.4 billion across 51 deals the year before (PitchBook, November 2026).
- Robot makers and model labs rarely buy from a small business directly. Data companies and business recording programs sit in between and pay the business.
- Public business rates read on 15 November 2026 run from $12 to $18 per approved hour to a flat $35 per recorded hour. They are statements by the programs, not a promise of pay.

## What do "physical AI" and "embodied AI" mean?

Physical AI is AI that acts in the real world instead of on a screen. [NVIDIA's glossary](https://www.nvidia.com/en-us/glossary/generative-physical-ai/) puts it this way: physical AI "enables autonomous machines, such as robots in factories and robotaxis, to perceive, understand, and perform complex actions in the physical world." Embodied AI is the older research term for the same idea: a model that has a body, senses its surroundings and moves in them. In 2026 the two are used almost interchangeably, and "physical AI" is the phrase investors and chip makers prefer.

What makes it a market for ordinary businesses is data. A language model could learn from text that was already on the internet. A robot has no such archive of people doing physical work, seen from where the worker stands. One founder told [MIT Technology Review](https://www.technologyreview.com/2026/04/01/1134863/humanoid-data-training-gig-economy-2026-breakthrough-technology/) that "you need to give lots and lots of variations for the robot to generalize well." Those variations have to be recorded, in real kitchens, shops and job sites, by people doing the work.

Source: [NVIDIA glossary: What is physical AI?](https://www.nvidia.com/en-us/glossary/generative-physical-ai/)

## Which companies make the robots?

Robot makers build the hardware and, increasingly, their own models. Figure says it trained its robot to navigate a home "using 100% egocentric human video data", with no robot demonstrations ([Figure, Project Go-Big](https://www.figure.ai/news/project-go-big)). Fortune reported in November 2026 that Figure also pays people to record everyday tasks for its models ([Fortune](https://fortune.com/2026/11/10/exclusive-in-the-race-for-robot-training-data-ex-meta-researchers-startup-turns-to-gig-workers-with-cameras/)). Agility Robotics' Digit has moved more than 100,000 totes at a GXO warehouse in Georgia ([The Robot Report](https://www.therobotreport.com/agility-gxo-logistics-plan-expanded-deployment-digit-humanoid/)). Business Insider reported in 2025 that Tesla would train its Optimus humanoid mainly by recording videos of workers performing tasks; the job behind that is covered in [Tesla data collection operator](https://origindatapartners.com/physical-ai/tesla-data-collection-operator).

Keep the scale in proportion. The [International Federation of Robotics](https://ifr.org/ifr-press-releases/news/global-sales-of-professional-service-robots-surge-24-percent) counted about 7,000 full-size humanoids sold worldwide in 2025 and says most current uses "remain specialized and often require human teleoperation". Investors spent over $6 billion on humanoid robots in 2025, according to MIT Technology Review. The money is ahead of the robots, and data is what the money is trying to buy.

Robot makers do not usually sign up a plumbing company or a machine shop one at a time. They buy footage in bulk from data companies, or run their own apps for individuals. That is why the third group below matters most to a business owner.


## Which labs build the models?

Model labs build the software "brain" and license or share it across many robots. Fortune describes Physical Intelligence as a startup that builds AI software that can control a variety of different types of robots, and says it has tested combining human video with teleoperation data. Google DeepMind calls its Gemini Robotics model "an advanced vision-language-action (VLA) model" ([Google DeepMind](https://deepmind.google/discover/blog/gemini-robotics-brings-ai-into-the-physical-world/)): it takes in camera images and an instruction and outputs movement. NVIDIA sells the chips most of the field trains on, and software platforms for robotics under the names Cosmos and Isaac.

Labs train on three kinds of data. Teleoperation, where a person drives a robot through a task, is the closest match to what the robot will do and the most expensive to collect. Simulation is cheap but only as good as its physics. Human video, shot from the worker's point of view, is the cheapest way to show a model thousands of real places and ways of doing a job. [How robots learn from video](https://origindatapartners.com/physical-ai/how-robots-learn-from-video) explains the methods.


## Which data companies collect and structure the footage?

This is the group that pays. Data companies recruit people or businesses to record, check the footage, label hands, objects and steps, and sell the result to labs and robot makers. [PitchBook](https://pitchbook.com/news/articles/physical-ais-data-boom-is-a-race-to-get-out-of-the-data-business) reported in November 2026 that startups in this segment had raised $2.1 billion across 55 deals in the year to date, up from $1.4 billion across 51 deals the year before.

The table lists examples with what a public source says about each. It is not a list of programs to join, and inclusion is not a recommendation. Company pages and news reports were read on the dates in the sources.

| Company | What a public source says | Source |
| --- | --- | --- |
| Scale AI | Announced it had gathered more than 100,000 hours of footage for robotics | [MIT Technology Review](https://www.technologyreview.com/2026/04/01/1134863/humanoid-data-training-gig-economy-2026-breakthrough-technology/) |
| micro1 | Has hired thousands of contract workers in more than 50 countries to record tasks | [MIT Technology Review](https://www.technologyreview.com/2026/04/01/1134863/humanoid-data-training-gig-economy-2026-breakthrough-technology/) |
| Appen | Says it could collect around 60,000 hours of video a month from as many as 5,000 contributors | [ABC News (Australia)](https://www.abc.net.au/news/2026-09-20/egocentric-training-gig-workers-helping-robots-learn-tasks/107135216) |
| Memories.ai | Says its network recorded more than 1 million hours of first-person video in its platform's first month in beta | [Fortune](https://fortune.com/2026/11/10/exclusive-in-the-race-for-robot-training-data-ex-meta-researchers-startup-turns-to-gig-workers-with-cameras/) |
| Maxinsights | Has 4,000 people in eight countries wearing cameras and gloves while they cook, clean and fold laundry | [PitchBook](https://pitchbook.com/news/articles/physical-ais-data-boom-is-a-race-to-get-out-of-the-data-business) |
| Build AI | Released Egocentric-10K, which it describes as the first dataset collected exclusively in real factories | [Hugging Face dataset card](https://huggingface.co/datasets/builddotai/Egocentric-10K) |


## What do they pay for, and what is losing value?

They pay for video of a person doing a real task, with the hands and the work in frame. Most is first-person, from a phone or small camera on the head or chest. Some buyers add a second, third-person camera. One data company's public robotics page, read on 15 November 2026, describes "Structured POV and third-person capture of domain experts performing real-world tasks" and names repair, assembly, industrial workflows, logistics and field operations.

Not all footage is worth the same. Household chores were the first thing collected at scale, and the price shows it. [ABC News](https://www.abc.net.au/news/2026-09-20/egocentric-training-gig-workers-helping-robots-learn-tasks/107135216) reported a recorder in Indonesia earning between US$3 and US$10 an hour, and said Appen was moving away from footage collected at industrial scale in lower-cost countries because it "had become commoditised and de-valued". PitchBook's report opens with the line that there is "an expiration date on the lucrative business of selling data to robot makers", and quotes an investor who expects the most valuable data to come from robots doing real work.

What holds value is what is hard to get: skilled work in commercial settings, done by people who do it for a living, with consent from the workers and the site. A trained tech replacing a compressor, a line cook working a rush or a welder running a bead cannot be staged in a rented flat. That is the footage a business with a crew can supply and an individual with a phone cannot.

Three business recording programs publish figures on their own pages, read on 15 November 2026. One pays $12 to $18 per approved hour, depending on the kind of work. One pays a flat $35 for every recorded hour in a booked session. One gives its own estimate of $500 to $2,000 a month per recording employee, assuming 12.5 to 50 approved hours a week. These are the programs' statements, not a promise of pay: the agreement you sign and the hours a program accepts decide what arrives. For individuals the range is wider. A public listing for an on-site video recorder in San Francisco, read on 15 November 2026, offered $35 to $40 an hour with devices provided, while MIT Technology Review found a recorder in Nigeria paid $15 an hour. [The business programs compared](https://origindatapartners.com/physical-ai/business-video-programs-compared) and [the pay rates tracker](https://origindatapartners.com/physical-ai/pay-rates-tracker) set the figures side by side.


## Which workplaces are wanted?

Programs publish lists of the work they accept, and the lists look alike: hands-on work with tools, parts, food or stock, repeated in many variations. Each page below covers what counts in that kind of business, what is never recorded and an illustrative sum for a typical crew.

- [Plumbing, HVAC and electrical](https://origindatapartners.com/physical-ai/plumbing-hvac-electrical), [construction](https://origindatapartners.com/physical-ai/construction) and [carpentry workshops](https://origindatapartners.com/physical-ai/carpentry-workshops)
- [Auto shops](https://origindatapartners.com/physical-ai/auto-shops), [manufacturing and fabrication](https://origindatapartners.com/physical-ai/manufacturing) and [food manufacturing](https://origindatapartners.com/physical-ai/food-manufacturing)
- [Warehouses](https://origindatapartners.com/physical-ai/warehouses), [retail and grocery](https://origindatapartners.com/physical-ai/retail-grocery) and [commercial cleaning](https://origindatapartners.com/physical-ai/commercial-cleaning)
- [Restaurants and commercial kitchens](https://origindatapartners.com/physical-ai/restaurants-kitchens), [hotels](https://origindatapartners.com/physical-ai/hotels), [event venues](https://origindatapartners.com/physical-ai/event-venues) and [salons and barbershops](https://origindatapartners.com/physical-ai/salons-barbershops)
- [Farms and nurseries](https://origindatapartners.com/physical-ai/farms-nurseries) and [landscaping and lawn care](https://origindatapartners.com/physical-ai/landscaping-lawn-care)
- [Laboratories](https://origindatapartners.com/physical-ai/laboratories), where interest is real but no program publishes a rate yet


## How does a business with a crew get paid?

The business is the party that signs and the party that is paid. In outline: the business agrees which tasks, areas and people may be recorded. Workers who opt in wear a head-mounted phone or a small camera rig during normal jobs. The program reviews the footage and counts the hours it accepts. The business is paid for those hours and decides what to share with the people who recorded.

Origin is an independent data partner. Tell us your trade, crew size and where you work, and we'll email you within one business day about recording work that fits your crew. It costs your business nothing. Before anyone records, settle consent: [recording employees at work](https://origindatapartners.com/physical-ai/recording-employees-at-work) sets out the state rules on sound, notice and private rooms.


## Your records can be worth more than your footage

Video is one thing AI companies pay a business for. The other is the written record of how the work was done: job files, service histories, inspection reports, estimates set against actual costs, maintenance logs. Those train and test models that plan and diagnose, and they are licensed, not recorded. If your company has several years of them, [see what your company records are worth](https://origindatapartners.com/check-fit).


## Next step

If your crew does hands-on work, [check whether your business qualifies](https://origindatapartners.com/physical-ai#qualify). It takes about a minute and shows an estimate straight away.

Origin Data Partners is independent and does not represent Amelia Robotics. Origin may earn a reward from Amelia Robotics if a business it introduced joins and gets paid; Amelia pays that reward itself. Origin charges your business nothing. [How Origin is paid](https://origindatapartners.com/referral-disclosure).


## Questions and answers

### What is a physical AI company?

A company that builds AI for machines that act in the real world: robot makers, the labs that build the models that control robots, and the data companies that collect and label the footage those models learn from. NVIDIA's glossary describes physical AI as enabling autonomous machines to perceive, understand and perform complex actions in the physical world.

### Which physical AI companies pay for video of real work?

Mostly the data companies, not the robot makers. Scale AI, Appen, micro1, Memories.ai and Maxinsights have all been reported collecting first-person video at scale, and a few programs pay businesses for footage of their crews. Which ones are open, and for what work, changes often. Origin can tell a business which programs fit it.

### How much do physical AI companies pay per hour of footage?

For businesses, public pages read on 15 November 2026 state $12 to $18 per approved hour at one program and a flat $35 per recorded hour at another; a third estimates $500 to $2,000 a month per recording employee. For individuals recording household tasks, reports range from US$3 to US$15 an hour. None is a promise of pay.

### Can a small business sell footage to physical AI companies?

Yes, through a business recording program. A small business does not need to deal with a robot maker or a lab. The program agrees the tasks, supplies or specifies the equipment, reviews the footage and pays the business for the hours it accepts.

### Is it legal to record employees for physical AI training data?

Silent video of normal work in work areas is generally lawful in the US with notice, but sound brings in wiretap laws and several states need everyone's consent. Private rooms are off limits. Get each worker's written, opt-in consent. See [recording employees at work](https://origindatapartners.com/physical-ai/recording-employees-at-work); this is not legal advice.

### Is embodied AI the same thing as physical AI?

Near enough. Embodied AI is the research term for AI that has a body and learns by acting in an environment. Physical AI is the newer industry term and also covers self-driving vehicles and factory automation. Both depend on real-world data.

## Sources

- [NVIDIA glossary: What is physical AI?](https://www.nvidia.com/en-us/glossary/generative-physical-ai/)
- [PitchBook: Physical AI's data boom is a race to get out of the data business (Nov 2026)](https://pitchbook.com/news/articles/physical-ais-data-boom-is-a-race-to-get-out-of-the-data-business)
- [Fortune: In the race for robot training data, a startup turns to gig workers with cameras (10 Nov 2026)](https://fortune.com/2026/11/10/exclusive-in-the-race-for-robot-training-data-ex-meta-researchers-startup-turns-to-gig-workers-with-cameras/)
- [MIT Technology Review: the gig economy training humanoid robots (Apr 2026)](https://www.technologyreview.com/2026/04/01/1134863/humanoid-data-training-gig-economy-2026-breakthrough-technology/)
- [International Federation of Robotics: service robot sales, 2025](https://ifr.org/ifr-press-releases/news/global-sales-of-professional-service-robots-surge-24-percent)
- [ABC News (Australia): gig workers helping robots learn tasks (Sep 2026)](https://www.abc.net.au/news/2026-09-20/egocentric-training-gig-workers-helping-robots-learn-tasks/107135216)
- [Figure: Project Go-Big](https://www.figure.ai/news/project-go-big)
- [The Robot Report: Agility and GXO expand Digit deployment](https://www.therobotreport.com/agility-gxo-logistics-plan-expanded-deployment-digit-humanoid/)
- [Google DeepMind: Gemini Robotics](https://deepmind.google/discover/blog/gemini-robotics-brings-ai-into-the-physical-world/)
- [Hugging Face: Egocentric-10K dataset card](https://huggingface.co/datasets/builddotai/Egocentric-10K)

## Related reading

- [/physical-ai](https://origindatapartners.com/physical-ai)
- [Companies that pay businesses and people to record work for AI and robot training (2026)](https://origindatapartners.com/physical-ai/companies-that-pay-to-record-work)
- [What AI companies pay for recording real-world work: pay rates tracker (Nov 2026)](https://origindatapartners.com/physical-ai/pay-rates-tracker)
- [How robots learn from video of people working (egocentric data, teleoperation, explained)](https://origindatapartners.com/physical-ai/how-robots-learn-from-video)
- [Recording employees at work for AI training: consent and notice rules by state](https://origindatapartners.com/physical-ai/recording-employees-at-work)
- [Which programs pay businesses to record work? Six compared (2026)](https://origindatapartners.com/physical-ai/business-video-programs-compared)
