Robotics teleoperation and physical-world data

In April 2026 Build AI released one million hours of egocentric factory footage for free, collected via camera glasses worn by Southeast Asian factory workers (DreamVu). Bulk undifferentiated egocentric video had traded at "a few dollars per hour" in 2025. It is now, per the same source, effectively free.

That is one half of this vertical dying in a quarter. The other half — bimanual teleoperation with trained operators on real rigs — still bills the buyer $50–$200 per hour against operator pay of $25–$50 per hour (DreamVu; Robotics Center).

The difference between the two halves is not skill or scarcity of labour. It is that one requires a rig and a room and the other requires a pair of glasses. The physical capital is the moat, and it is the only thing in the labour section of this atlas that structurally blocks the playbook Mercor and the expert-data marketplaces run: sign up ten thousand laptops, route tasks, keep a third.

#Whose budget

The fastest-growing buyer money in the 2026 sweep. Humanoid and VLA programmes are capitalised at $100M+ and name data as a line item, and none of them was going to hire two hundred teleoperators in-house.

Two independent 2026 cost benchmarks converge, which is rare in this atlas. By programme type: simple teleop $15–$30/hr (6–12 demos/hr, 85–95% acceptance); bimanual ALOHA-style $40–$80/hr (2–4 demos/hr, 70–85% acceptance); egocentric wearable $25–$60/hr; full humanoid multi-sensor $80–$150/hr (1–3 demos/hr). Cost per usable humanoid demonstration: $50–$150. A 2,000-demonstration bimanual programme runs about $62,400 all-in (DataXPower).

Per-episode, the second benchmark: simple pick-and-place $8–$15, contact-rich bimanual $25–$35. Vendor rate cards show a $2,500 pilot for 50 filtered demos, $8,000 for a 500-demo single-task campaign, arm leasing at $800/month, and a minimal 500-demo dataset landing at $50k–$200k (Robotics Center).

budget: 5. New line, board-approved, tied to a physical product roadmap rather than to a headcount comparison. The one discount: the very largest buyers — Figure, Tesla, 1X — run proprietary fleets. Your best customers are the ones most able to leave.

#Can you get the supply

Split the question, because the answer is opposite on each side.

Quality teleop: genuinely hard. It needs rigs, floor space, task setup, trained hands and a QA loop. Ramp is documented — 4–8 hours before an operator produces usable output, 20–40 hours before consistent quality (Robotics Center). You cannot do it from a laptop, which is why the marketplace model does not transplant here and why the operator base cannot be spun up in a week by a competitor with more money.

Egocentric capture: trivially easy, and already commoditised. The winning move was to point an existing gig network at it. Awign claims 1,000+ hours/day of 4K first-person video from 1.5 million gig workers across 1,000+ cities — a network built for retail audits and field sales, repointed (Labellerr). Human Archive runs 1,000+ active headsets with heavy India presence; Luel claims 3 million contributors; Lightwheel claims 300k+ hours delivered (DreamVu).

supply: 4, not 5, because the score has to be blended and half the vertical is a race anyone can enter. Note the Where the supply can legally live asymmetry: the rigs sit where the robotics companies are, the glasses sit wherever gig labour is cheapest.

#What the spread looks like

Narrow where it is durable, wide where it is dying — an inversion worth staring at.

Sources: DreamVu, Robotics Center, DataXPower.

spread: 3. The 40–65% teleop gross is better than expert data's 27–33% but it is a gross margin carrying real cost of goods underneath: rig capex amortising at $1.50–$8.00 per demo, plus rejection risk, since acceptance rates run 70–85% on bimanual work. This is a You pay weekly, they pay in sixty days business, not a software one. And the egocentric 70–90% is [UNVERIFIED] — nobody publishes what a gig worker in a Southeast Asian factory is paid per hour of footage — and it belongs to the segment that just went to zero.

Distinguish gross from net here as everywhere (GMV is not revenue): a robotics data company quoting "$X of revenue" is quoting billings that include the operator payroll and the rig depreciation.

#Can you hold it

Moderate, and the answer is unusually physical.

Buyer and seller cannot easily transact around you, because "the seller" is not one person — it is a room, a set of arms, a calibrated capture stack and a supervisor. A robotics company that wants to disintermediate has to build a facility, not sign a contractor. That is a real barrier and it does not exist in Voice, speech and low-resource language data or Expert data for frontier labs.

Against that: the largest buyers vertically integrate anyway. Figure, Tesla and 1X run their own fleets, which is the structural version of the risk One customer is a binary event describes. And Scale AI has 150,000+ robotics data hours plus a dedicated arm; Encord raised a $60M Series C in February 2026 on $110M total for physical-AI data infrastructure (SiliconANGLE). Well-capitalised incumbents are already inside. hold: 3.

#What AI does to it

ai: 4. Better models increase demand for physical demonstrations rather than replacing them — you cannot synthesise contact dynamics you have never observed, and the whole VLA research programme is bottlenecked on real-world trajectories.

The qualifier is that simulation is a live substitute at the margin. Lightwheel is explicitly sim-first. Every hour that sim-to-real transfer improves is an hour of teleoperation nobody buys. This is the same argument as synthetic data in What better models do to each layer, and it is unresolved in both places.

#What would kill it

The teleop half dies three other ways. Sim wins, and demonstrations become a small correction term. The big buyers finish integrating, leaving you selling to the tail. Rig capex eats you — this business needs cash for hardware before it needs cash for people, which is the opposite of every other vertical in the labour section and makes it a poor fit for the venture pattern that funded the rest.

#Who is already there

A cohort that barely existed in 2024.

Also present: Bones Studio (optical mocap), Lightwheel (sim-first), Luel (rights-cleared marketplace), Objectways, Cogito Tech, Labellerr (5,000+ annotators), and Scale AI's robotics arm. Sources: DreamVu, Labellerr on teleoperation, SiliconANGLE.

Note what is absent from that table: revenue. Not one figure. Encord's only revenue estimate anywhere in the atlas is an [UNVERIFIED] third-party guess of ~$12.8M ARR for 2024.

#Where the record is thin

Three specific holes. Rig economics: $1.50–$8.00 per demo of amortisation is a single-source range with no stated assumption about utilisation or asset life, and it decides whether 40–65% gross survives contact with reality. In-house fleet size at Figure, Tesla and 1X is undisclosed, so the addressable market cannot be separated from the integrated one. Acceptance rates — 70–85% on bimanual — come from the vendor selling the service, which has an obvious interest in the number.

speed: 2 for the same reason the moat exists: you cannot invoice until the room is leased, the arms are bought, the operators are trained through their 20–40 hour ramp and the first batch clears QA. Compare Adversarial evals and red-team crowds, where a challenge can be live in a fortnight. Hard supply is good for defensibility and bad for the Which side you build first problem — you carry the capex before the first buyer signs.