Physical AI Earnings Index · Founding Edition
How people are paid to teach machines the physical world.
Physical AI has created a new kind of work: people are paid not just to label information, but to physically demonstrate how the world works. This Index tracks the economics of that work — what programs publish about pay, how they pay, what rights they take, and whether contributors share in the value when their data is licensed again.
The headline numbers
- 8programs and roles with enough public detail to profile, across three markets.
- $4the lowest published remote-capture rate — per approved hour of video (TaskVerse; the Philippines, India and Brazil).
- $20–$30per hour for on-site robot operation by exoskeleton (Dexmate, 1099 contract); Tesla’s Optimus data role was advertised at $30/hour plus equity.
- $15Mpaid to Creators to date by Figure’s Index network (Figure, Aug 25, 2026) — the largest public scale signal.
- 0programs that publicly promise contributors ongoing pay when their data is licensed again.
“None found” means we found no such public promise in the materials reviewed — not that no private contract contains one. Firsthand’s published participant release is explicitly “royalty-free.”
The central finding
People are creating reusable assets — and are usually paid like one-time labor.
Transaction A. A person spends an hour producing useful first-person footage and is paid a fixed amount when it’s approved.
Transaction B. That hour becomes part of a dataset licensed into several model-development programs over time.
In every public program we reviewed, the contributor’s economic relationship ends at Transaction A. That doesn’t make the prevailing model unfair — buyers and data companies carry real costs: hardware, recruiting, storage, privacy, annotation, QA, failed footage, sales and legal risk. But it raises a design question the Index will keep tracking: if physical training data becomes more valuable as robotics scales, should the people who create it participate only at collection — or also when it keeps earning?
Three labor markets, not one
There is no honest single hourly rate for physical-AI data work. At least three distinct markets are emerging, each with a different pay architecture.
- Market A
Remote egocentric capture
The worker is the sensor: first-person video of hands, tools and tasks, usually on a phone or wearable camera, often paid only for footage that passes review.
4 programs profiled - Market B
In-lab demonstration & teleoperation
The worker is part of the robot-learning loop: operating robots, exoskeletons or motion capture on-site, usually as an hourly employee or contractor.
2 programs profiled - Market C
Physical-world task networks
Existing labor or creator networks redirected toward AI and robotics data collection, paid per task or per minute.
2 programs profiled
What programs publish about pay
Rates are shown exactly as published. Ranges are never averaged, euro rates are not converted, and approved, usable, session and clock hours are kept apart — a $38 session hour that includes setup and breaks is not the same thing as $4 for an approved hour of footage.
Not plotted — no single published USD hourly rate: EgoVista (€7–€20 per hour depending on task complexity (EgoVista recruiting post, Aug 2026)) · Figure Index (paid by the minute) · DoorDash Tasks (task-specific)
The hidden variables: rejection risk and rights
A $4 “hour” may not be a $4 labor hour.
When pay depends on approved footage, what a contributor actually earns per hour invested is:
effective hourly earnings = cash paid for accepted data ÷ total time to plan, set up, record, upload and fix submissions
Rejected footage, setup, uploads, resubmissions, unpaid onboarding, equipment, travel and payment delays all sit in that denominator. Public postings don’t reveal it, so future editions will publish a stated rate and — once enough contributors can be surveyed credibly — an effective rate.
The same $20 can buy radically different rights.
One program buys a one-time output; another takes a perpetual, sublicensable commercial license; another owns the work outright through employment. Pay without rights context is incomplete, so every profile pairs pay with a rights summary — and future editions will answer this Rights Card for each program:
- Does the contributor retain ownership?
- Is the grant exclusive?
- Can the data be sublicensed?
- Can it be licensed to multiple companies?
- Is the license perpetual?
- Is consent withdrawable?
- What happens to delivered data after withdrawal?
- Are derivative annotations covered?
- Are model weights or learned representations addressed?
- Can voice or likeness be used?
- Does the contributor receive future licensing compensation?
Why a royalty model changes the comparison
Hourly pay is easy to compare because it is certain and immediate. Participation is harder to compare because its value is conditional: a contributor can earn less than an hourly alternative if their data is never licensed, and more if it is valuable and licensed repeatedly. The honest comparison is not “$X an hour vs. royalties” — it’s the shape of the two models over a recording’s life.
- Participation — licensed again and again. A smaller amount at acceptance, then a further payment each time the recording is licensed.
- One-time capture pay. Paid once when the footage is collected or approved; nothing further when it is reused.
- Participation — never licensed. The same smaller amount at acceptance — and no more, because the recording is never licensed.
Benchmark — proposed, not part of the data
Taught’s model
Taught, which publishes this Index, is building the participation model and is included only as a benchmark against the observed market. It has not launched its marketplace and has no observed Teacher earnings, so it is excluded from every figure above, and any future earnings illustration will be labeled hypothetical until real licensing transactions exist.
- Teachers record first-person Lessons on their own phones and are paid a base award for each accepted recording.
- When an eligible accepted recording is licensed and the customer’s payment settles, the Teacher earns a share of the licensing revenue attributed to that recording — 40% to 70%, rising with Teacher Level, which is based on quality.
- A recording licensed again pays again. Buyer-funded Exclusive Lessons instead pay a guaranteed one-time award.
- Teachers keep ownership of their footage, can remove recordings from future licensing, and see every licensing event of their work.
Program profiles
Every profile is limited to what the program’s own public materials say, with sources and the date we last checked them.
8 of 8 programs
- Remote captureRemote
TaskVerse — Everyday Moments
Remote egocentric capture
$4 per hour of approved video
- Paid
- per approved hour of video
- Work
- Everyday activities recorded first-person
- Equipment
- Contributor-owned Android phone and head mount
- Access
- 18+; the Philippines, India and Brazil only
- Rights
- TaskVerse gives participants in such tasks a task-specific privacy consent and notice (Privacy Statement, last modified March 2025)
- Paid again on relicensing?
- Not found in public materials
The lowest published remote-capture rate in this edition, open only outside the U.S. — a marker of global price pressure on commodity household video.
- Remote captureOn-site
Mercor Robotics
Wearable capture inside existing businesses
$12–$18 per approved hour, paid bi-weekly to the participating business; the page does not say what employees receive
- Paid
- per approved hour of video (paid to the business)
- Work
- Hands-on work in restaurants and cafés, repair shops, warehouses, construction, cleaning and manufacturing, among others
- Equipment
- Hardware supplied by Mercor
- Access
- Through participating businesses; each participant signs a consent form before recording
- Rights
- Mercor says it licenses the footage to AI and robotics companies
- Paid again on relicensing?
- Not found in public materials
Turns ordinary work into a licensable robotics-data asset — and pays the business, not the worker, a fixed rate per approved hour.
Mercor — Robotics data-collection partner program ↗Last verified Sep 24, 2026 - Remote captureRemote
EgoVista
Remote egocentric capture
€7–€20 per hour depending on task complexity (EgoVista recruiting post, Aug 2026); paid per hour of usable footage that passes quality check; each mission shows its rate before joining
- Paid
- per usable hour of video
- Work
- Household activities and professional/trade tasks
- Equipment
- Smartphone, GoPro or smart glasses; contributor provides a hands-free setup unless using glasses
- Access
- Mission-based
- Rights
- Compensation described as consideration for a licence over the footage; voice, when collected, licensed for model training only — no voice cloning or speech synthesis
- Paid again on relicensing?
- Not found in public materials
Stronger-than-average public transparency on privacy and licensing; pay remains per mission.
EgoVista — Contributor program ↗EgoVista — recruiting post with rates (Aug 21, 2026) ↗Last verified Sep 24, 2026 - Remote captureVaries
Firsthand
High-spec egocentric capture sessions
$38 per session hour base (includes setup, calibration, breaks and travel over 20 minutes); +$12/hour for clinical, construction and agriculture sessions; not contingent on acceptance, delivery or licensing
- Paid
- per session hour
- Work
- Custom physical-AI datasets
- Equipment
- Multi-sensor rig: head camera, two wrist cameras, depth and IMU, with 3D hand and body pose annotations
- Access
- Managed collection
- Rights
- Sample participant release (v3.2, effective Jan 14, 2026) grants a perpetual, irrevocable, worldwide, sublicensable, royalty-free right; buyers receive a perpetual, buyer-owned commercial license
- Paid again on relicensing?
- Not found in public materials
The clearest trade-off: guaranteed session pay (including setup and breaks) in exchange for broad downstream rights.
Firsthand — Egocentric video data for embodied AI & robotics ↗Last verified Sep 24, 2026 - Task networksRemote
Figure Index
Global record-and-earn creator network
Paid by the minute; Figure does not publish a per-minute rate
- Paid
- per minute of recording
- Work
- Recording everyday tasks; skilled Creators in homes and businesses
- Equipment
- Figure says contributors can receive a recording device
- Access
- App sign-up
- Rights
- Described as a Figure-exclusive pipeline for physical training data
- Scale
- Figure (Aug 25, 2026): $15M paid to Creators to date; over 16M videos; 264,000 app downloads across 100+ (108) countries; over 44,000 weekly active users
- Paid again on relicensing?
- Not found in public materials
The strongest evidence that physical-AI data collection is becoming a mass consumer earning category.
- Task networksVaries
DoorDash Tasks
Physical-world task network
Task-specific; shown upfront and based on effort and complexity
- Paid
- per completed task
- Work
- Body-worn household footage — loading dishwashers, folding laundry, cooking — for robotics and humanoid systems
- Equipment
- Existing Dasher network; available in select U.S. places, excluding California, New York City, Seattle and Colorado
- Access
- Dashers in selected markets
- Rights
- Not detailed in the launch materials reviewed
- Scale
- DoorDash (Mar 19, 2026): more than 8 million Dashers; more than 2 million tasks completed since 2024
- Paid again on relicensing?
- Not found in public materials
Distribution is the advantage: physical-world data from an existing workforce.
- TeleoperationOn-site
Tesla — Data Collection Operator, Optimus
In-house motion capture / humanoid data operation
$30/hour base plus $12,000 equity vesting over four years, +10% PM and +15% overnight differentials (Fremont and Palo Alto; LinkedIn recruiting post, Sept 7, 2026)
- Paid
- per clock hour
- Work
- Humanoid data collection with motion-capture equipment
- Equipment
- Company-provided motion-capture equipment
- Access
- On-site, shift-based
- Rights
- Employee work product
- Paid again on relicensing?
- Not found in public materials
Higher, more predictable pay than most remote capture, with far less flexibility. Pay varies by listing and location.
LinkedIn recruiting post — Data Collection Operator, Optimus (Sept 7, 2026) ↗Last verified Sep 24, 2026 - TeleoperationOn-site
Dexmate — Robotics Data Collection Operator
On-site robot teleoperation
$20–$30 per hour (1099 contract)
- Paid
- per clock hour
- Work
- Operating robots through a wearable exoskeleton, performing structured tasks repeatedly
- Equipment
- Company robots and wearable exoskeleton
- Access
- 1099 contract, on-site at Dexmate's Santa Clara lab; VR/AR experience listed as nice to have
- Rights
- Contractor-generated company training data
- Paid again on relicensing?
- Not found in public materials
The premium paid when workers operate scarce hardware rather than record their own environment.
Dexmate — Robotics Data Collection Operator (Ashby) ↗Last verified Sep 24, 2026
The framework: eight dimensions, no single score
The Index reports pay architecture before the pay number. Each program is described across eight dimensions rather than one reductive rating.
- A
Stated cash compensation
The advertised wage, task rate, video-hour rate, bounty or revenue share.
- B
Payment basis
Clock hour, approved or usable video hour, minute, completed task, accepted submission, salary or licensing event.
- C
Rejection risk
Whether the contributor is paid for time worked, or only for data that survives quality review.
- D
Work flexibility
Remote or on-site, fixed shift or on-demand, geographic limits and minimum commitment.
- E
Equipment burden
Whether the worker must own a phone, buy a mount, travel to hardware, or use company equipment.
- F
Skill & access barrier
Prior skill, physical constraints, technical experience, access to a trade environment or certification.
- G
Rights & reuse
What public materials say about who may use, sublicense, retain, commercialize or repeatedly license the data.
- H
Downstream participation
Whether the original contributor is paid again when accepted data is licensed, reused or commercialized.
What the market is telling us
- Figure says it has paid $15M to Creators through Index, with over 16 million videos uploaded.
- Mercor is pitching ordinary businesses on turning employees’ existing physical work into a revenue stream, and says it licenses the footage to AI and robotics companies.
- DoorDash offers physical-world data collection — including body-worn footage for robotics and humanoid training — through a network of more than 8 million Dashers.
- 1X says the only durable moat in embodied AI will be data, from a mixture of web-scale media, human video, simulation, teleoperated and on-policy robot data.
Human physical behavior is being productized as training infrastructure. The open question is how much of the value stays with the intermediary, and how much with the person whose behavior created the demonstration.
Hypotheses future editions will test
- H1
Capture pay will bifurcate.
Commodity household video will face global downward price pressure, while scarce professional workflows, hard-to-access environments, unusual tools and safety-critical demonstrations command premiums.
- H2
"Approved hour" will matter more than "recorded hour."
As supply grows, buyers will pay for usable signal rather than raw duration.
- H3
Rights transparency will become a competitive variable.
As contributors understand that demonstrations may be reused across commercial training programs, platforms will face pressure to explain licenses in plain English.
- H4
Existing labor networks will enter the market.
DoorDash is unlikely to be the last labor marketplace to monetize its physical footprint.
- H5
The most valuable human data will come from people with access, not AI expertise.
Plumbers, electricians, mechanics, warehouse workers, nurses, cooks, cleaners and tradespeople hold environments and manipulation sequences that are hard to recreate in a lab.
- H6
Revenue participation will emerge as a supply-side differentiator.
This is the hypothesis Taught is explicitly testing.
Methodology
- Research date
- September 24, 2026. Every source was re-checked on that date.
- Universe
- Publicly accessible company pages, contributor pages, employer career pages and recruiting posts, and selected secondary sources where a primary source did not expose the same information.
- Inclusion
- A program is included when its public materials clearly connect the work to robotics, embodied AI or physical-world AI training, and describe at least the work model and the basis of pay.
- Pay handling
- Published rates are preserved as published. Ranges are not collapsed into an average. Euro rates are not converted. Salary, equity and benefits are not converted into hourly equivalents unless the employer publishes the hourly part. Approved, usable, session and clock hours are different units.
- Source hierarchy
- 1. Company-controlled public source · 2. Company recruiting post · 3. University career-board mirror of a named employer · 4. Reputable secondary job board or reporting · 5. Anonymous or unverifiable claims — excluded.
- Rights
- Rights statements are limited to what public materials show. “Not found” is not proof that private contracts never contain such terms.
- Excluded at publication
- Figures from our research draft that we could not verify on the publication date:
- A U.S. egocentric-video project rate ($14 per approved hour) whose university job-board posting was removed before publication.
- A student robotics-video project rate ($8 per usable hour) whose posting was removed before publication.
- A robot-teleoperator role ($30–$55 per hour) whose posting was removed before publication.
- A global freelance rate ($4–$6 per hour of video) and a robotics-operator contract rate ($32 per hour) for which no primary public source could be found.
- Limitations
- Programs change pay frequently; some rates are personalized, geographic or project-specific; public pages may not show full contributor agreements; acceptance rates are rarely public, so effective earnings cannot yet be calculated; postings expire.
- Conflict of interest
- Taught publishes this Index and is building a participation model. It is labeled proposed and excluded from all market figures.
- Corrections
- If a figure is wrong or out of date, email data@taught.life with a source. We correct errors promptly and note the change and date in the next version.
Cite and reuse
Taught (Gunning Industries Corp). “Physical AI Earnings Index 2026 — Founding Edition.” September 24, 2026. https://www.taught.life/resources/physical-ai-earnings-index
The Index text and dataset are published under CC BY 4.0: you may reuse them, including commercially, with attribution and a link to this page. Download the dataset (CSV). Press and research inquiries: data@taught.life.
Sources
- TaskVerse — Everyday Moments project
- TaskVerse — Privacy Statement
- Mercor — Robotics data-collection partner program
- EgoVista — Contributor program
- EgoVista — recruiting post with rates (Aug 21, 2026)
- Firsthand — Egocentric video data for embodied AI & robotics
- Figure — Introducing Index
- Figure — Index app
- DoorDash — Tasks developer site
- DoorDash — Introducing DoorDash Tasks
- LinkedIn recruiting post — Data Collection Operator, Optimus (Sept 7, 2026)
- Dexmate — Robotics Data Collection Operator (Ashby)
- 1X — World Model Lab (Jun 4, 2026)
Teach once. Keep participating.
Taught is testing whether the people who teach robots should share in what their data keeps earning. Request early access — U.S. adults, 18+.