The best-documented humanoid deployment is seven units

TL;DR. Humanoid robots are the explicit bet that task redefinition is unnecessary: a machine that works in environments built for people, without modification. The clearest documented industrial deployments are seven commercial Digit units at Toyota after a year-long pilot, and single-digit units at BMW, where a Figure robot spent eleven months on the Spartanburg line. On Tesla's Q4 2025 earnings call in January 2026, Elon Musk said Optimus was "not in usage in our factories in a material way" and that units were primarily for learning. Unitree shipped roughly 5,500 units in 2025, the global volume lead, at $16,000 for a G1, and its Q1 2026 net profit fell 52%. And where humanoids are deployed, they are moving totes in warehouses, which is the bounded task in the engineered environment where a non-humanoid robot already works.

---

Status: contested, with company statements separated from circulated figures. The Musk quotation is from Tesla's Q4 2025 earnings call of 28 January 2026. Unit counts at Toyota and BMW come from industry reporting rather than company disclosure. Shipment rankings are disputed between AgiBot and Unitree. Widely repeated figures of 50,000 Optimus units or 10,000 Figure deployments do not originate with the companies they describe, which is the article's subject.

---

What the companies have said

Start with statements that carry legal weight, because they are the smallest number and the most reliable.

On Tesla's Q4 2025 earnings call, 28 January 2026, Elon Musk said Optimus was "not in usage in our factories in a material way", and that units were primarily for learning rather than productive tasks.

Tesla has never published an Optimus production count. It had targeted 5,000 units for internal factory use in 2025; reporting put actual output at a few hundred, below ten percent of the target.

Boston Dynamics began Atlas production in January 2026, with all 2026 units committed to Hyundai Motor Group and Google DeepMind.

Unitree shipped approximately 5,500 units in 2025 across its line, is targeting 10,000 to 20,000 for 2026, and prices a G1 at $16,000 and an R1 at $5,900. Its Q1 2026 net profit fell 52% year on year.

What is circulating

That Tesla has passed 50,000 cumulative Optimus units. That Figure has surpassed 10,000 deployments. That more than a thousand Optimus robots are working Tesla production lines.

None of those figures originates with the company it describes.

They appear in coverage, get repeated, acquire the appearance of established fact, and then sit alongside a chief executive's own statement that the robots are not in material use. Both circulate simultaneously and only one of them is attributable.

This is citation decay operating in real time on a live subject rather than on a decade-old case, and it is worth watching for that reason.

What is actually documented

The numbers that survive checking are small and they are real.

Agility Robotics: seven commercial Digit units at Toyota, following a year-long pilot. Digit is also reported working at GXO and Amazon logistics sites moving totes.

Figure AI: single-digit units at BMW. A Figure robot spent eleven months on the BMW Spartanburg production line, in a period during which the plant built more than 30,000 vehicles.

Apptronik's Apollo is in testing with Mercedes-Benz, and is the hardware platform for Google DeepMind's Gemini Robotics work.

These are genuine. Paid work, on a clock, at named sites, with a customer who chose to renew. Eleven months on a production line is not a demonstration.

They are also seven units, and single digits.

The volume is elsewhere and it is cheap

China accounted for roughly 90% of 2025 global humanoid shipments. AgiBot reports 5,168 units for 2025 at a 39% share on one analyst's count; Unitree disputes the ranking with its own 5,500-plus figure. The dispute is between two companies claiming the same few thousand units, which itself indicates the size of the market.

At $16,000 for a G1 and $5,900 for an R1, these are research platforms, education units and consumer devices rather than industrial labour.

And the economics are not yet working. The volume leader's quarterly profit halved while shipping more units than any Western competitor. Cost leadership at these volumes is not producing margin, which is the pattern of an industry buying market share ahead of a product-market fit it has not established.

The teleoperation question, handled fairly

Multiple named outlets have documented patterns suggesting some prominent public demonstrations involved teleoperation.

This is not disqualifying and should not be presented as a scandal. Every manufacturer uses teleoperation during development. It is how demonstrations get filmed, how data gets collected, and how a system is tested before its autonomy is trusted.

It is relevant context for interpreting footage, and it matters for one specific reason: the intervention rate is the number separating autonomy from supervision, and no humanoid manufacturer publishes it.

A robot performing a task in a video and a robot performing that task unaided are different claims, and the current disclosure practice does not let an outsider tell them apart.

What the framework predicts, and what happened

Humanoids are the wager that accommodation is unnecessary. A human-shaped machine in a human-shaped world, no workspace rebuild, no domain narrowing, no target standardisation, no deleted sub-task, no widened tolerance.

So look at where the deployments are.

Warehouses, moving totes. A bounded, repeated task, on a flat floor, in a facility already designed around material handling robots, which account for the majority of North American robot orders.

A production line, for eleven months. The most engineered environment in industry, with fixed stations and known part presentation.

The successful humanoid deployments are doing what a non-humanoid robot already does, in the environment where the non-humanoid already works.

Which means the accommodation happened anyway. The form factor changed; the redefinition did not go away. And that is not a criticism of the robots. It is an observation that the hardest version of the claim has not yet been tested at any scale, because the deployments that exist did not require it.

Three things this establishes

Company statements and circulated figures differ by orders of magnitude on a live subject. The gap between "not in material use" and "50,000 units" is not a rounding error, and the second is repeated far more often than the first.

Volume and viability are separate. Shipping 5,500 units at a halved profit is a real achievement and not yet a business, and coverage that reports the first without the second describes momentum rather than an industry.

And the deployments that exist do not test the humanoid thesis. Totes in a warehouse and stations on a line are the environments where accommodation already succeeded. The claim that a human-shaped robot works in an unmodified human environment remains, at meaningful scale, untested.

What it does not establish

That humanoids will not work. Capital, state industrial policy and serious engineering are all committed, manufacturing costs are reportedly falling quickly, and the deployed capability is genuinely better than three years ago.

That the small numbers are the ceiling. Production capacity is being built at Fremont, at Agility's dedicated factory, and across several Chinese manufacturers. Building capacity is not filling it, and it is also not nothing.

That any specific circulated figure is false. They are unattributed rather than disproven, which is a different claim and the one made here.

And nothing about the teleoperation content of any particular demonstration. The point is that intervention rates are unpublished, not that a given video was faked.

What is unresolved

Intervention rates. No manufacturer publishes them, and until one does, autonomy claims cannot be checked from outside.

Whether serial production works. A humanoid has on the order of 10,000 components and no precedent exists for producing one at volume. Tesla's own Model 3 ramp took more than eighteen months to stabilise on a far simpler product.

Whether unit economics close. Falling manufacturing cost against a halving profit at the volume leader is not yet a resolved picture.

And whether a general-purpose deployment appears. The test is a humanoid doing materially different tasks in an environment nobody adapted, with disclosed intervention rates. Nothing published meets it.

The counter-argument

Seven units is what year one of a hardware category looks like. Automobiles, aircraft and personal computers all had single-digit deployments before they had thousands, and citing the current number as a verdict confuses a starting point with a limit.

The framework may be unfalsifiable as applied here. Any environment can be described as somewhat engineered, so pointing out that a warehouse floor is flat risks explaining away every deployment in advance. A fair test needs a definition of "unmodified" specified before the evidence, and this article has not supplied one.

The circulated-figures criticism targets journalism, not the technology. That coverage repeats unattributed numbers says something about coverage. The robots are unaffected by how badly they are described.

And Musk's statement may be conservative rather than deflating. A chief executive telling investors robots are not yet material is managing expectations under legal constraint, which is the opposite of the incentive to exaggerate, and may understate internal progress.

The short version

Humanoids are the explicit bet that accommodation is unnecessary: a machine that works in a human environment without modifying it.

The clearest documented deployments are seven commercial Digit units at Toyota after a year-long pilot, and single-digit units at BMW, where a Figure robot spent eleven months on the Spartanburg line. On Tesla's Q4 2025 earnings call, Elon Musk said Optimus was "not in usage in our factories in a material way." Tesla has never published a production count and missed its 2025 internal target by more than ninety percent.

Circulating alongside those: 50,000 cumulative Optimus units, 10,000 Figure deployments, a thousand robots on Tesla lines. None of those figures comes from the company it describes.

Volume exists and it is cheap. China took roughly 90% of 2025 shipments, Unitree shipped about 5,500 units with a G1 at $16,000, and its Q1 2026 profit fell 52%. Two companies dispute which of them leads on a few thousand units, which indicates the market's size.

And where humanoids are deployed, they move totes in warehouses and stand at stations on production lines. Bounded tasks, flat floors, fixed part presentation: the environments where non-humanoid robots already work.

So the accommodation happened anyway. The form factor changed and the redefinition did not disappear. Which means the hardest version of the claim has not been tested at scale, because the deployments that exist never required it.

Common questions

How many humanoid robots are actually deployed in industry? The clearest documented figures are seven commercial Agility Digit units at Toyota following a year-long pilot, and single-digit units at BMW, where a Figure robot spent eleven months on the Spartanburg production line. Digit is also reported working at GXO and Amazon logistics sites, and Apptronik's Apollo is in testing with Mercedes-Benz. These come from industry reporting rather than company disclosure.

What has Tesla said about Optimus? On the Q4 2025 earnings call of 28 January 2026, Elon Musk said Optimus was not in usage in Tesla's factories in a material way, and that units were primarily for learning rather than productive tasks. Tesla has never published an Optimus production count. It had targeted 5,000 units for internal factory use in 2025, and reporting put actual output at a few hundred, below ten percent of that target.

Where do the much larger figures come from? Not from the companies. Claims that Tesla has passed 50,000 cumulative Optimus units, that Figure has surpassed 10,000 deployments, or that more than a thousand Optimus robots work Tesla production lines do not originate with the companies they describe. They appear in coverage, get repeated, and acquire the appearance of settled fact while sitting alongside a chief executive's own contrary statement. Being unattributed is not the same as being false, and it is the claim made here.

Who ships the most humanoids? Chinese manufacturers, accounting for roughly 90% of 2025 global shipments. Unitree shipped approximately 5,500 units in 2025 with a G1 priced at $16,000 and an R1 at $5,900, while AgiBot reports 5,168 units at a 39% share on one analyst's count and the two dispute the ranking. That two companies contest leadership over a few thousand units indicates the size of the market.

Is the business working? Not yet, on the available evidence. Unitree leads on volume, ships at prices no Western competitor matches, and its Q1 2026 net profit fell 52% year on year. Manufacturing costs are reportedly falling quickly, and cost leadership at current volumes is not producing margin, which is the pattern of an industry buying share ahead of an established product-market fit.

Are the demonstration videos teleoperated? Multiple named outlets have documented patterns suggesting some prominent public demonstrations involved teleoperation. This is not disqualifying: every manufacturer uses teleoperation during development, for filming, for data collection and for testing before autonomy is trusted. It matters because the intervention rate is the number separating autonomy from supervision, and no humanoid manufacturer publishes it, so an outsider cannot distinguish a robot performing a task from a robot performing it unaided.

Do the current deployments prove the humanoid case? No, and this is the central point. Humanoids are the bet that a human-shaped machine works in a human environment without modification. The deployments that exist involve moving totes in warehouses and standing at stations on production lines: bounded, repeated tasks on flat floors in facilities already designed around material handling robots. The form factor changed and the accommodation did not go away, which means the hardest version of the claim has not been tested at scale because these deployments never required it.

What would count as evidence that it works? A humanoid performing materially different tasks, in an environment nobody adapted for it, with published intervention rates per hour of operation, sustained over months, at more than single-digit unit counts. The first three are the parts nobody currently reports. And a fair test needs "unmodified environment" defined in advance, which this article has not supplied and which is the strongest objection to its own framework.