Humanoid robot companies: who ships and who only demos

Every week another one of the humanoid robot companies posts a slick video. Very few of them can show a customer who pays every month for a robot that is in service. I read the filings and trade press before I believe a demo. This article sorts ten names into three groups: shipping, paid pilots, and demo only. It builds on our wider coverage in The Robotics Revolution.

The sorting uses public evidence as of 15 October 2026. Where a number comes from the company itself, I say so.

How to tell shipping from demoing

I use five tests. A company does not need to pass all five, but the more it fails, the more skeptical you should be.

1. A named, paying customer site. A logo on a slide does not count. A plant or warehouse does.
2. Repeat orders. One trade-press columnist put it well: the true test is the second purchase order, not the demonstration.
3. Units in service. Not units built, not units planned. Robots doing work.
4. A published safety approach. If a robot works near people, the buyer needs to know how it fails.
5. Price or RaaS terms. If nobody will say what it costs, the product is not ready for a purchase order.

The same article names mean time between interventions as the metric that matters most. Almost nobody publishes it. Keep that in mind when you read the table.

The scorecard

Agility (Digit)

Agility is the closest thing the West has to a humanoid in service. In its SPAC announcement it listed Schaeffler, GXO, Toyota Motor Manufacturing Canada and Mercado Libre as customers, with nine facilities and more than 65,000 operating hours. These are company numbers, but they sit in a document Agility will be held to.

The headline is more than $300 million of multi-year Digit v5 orders. Read the small print: the orders are subject to contractual milestones. IEEE Spectrum estimates that this is comfortably under 1,000 robots, about a tenth of what its Oregon factory could build.

Agility also has the clearest safety story. IEEE reports that Digit 5 lifts 23 kg to 2.1 m, works near people without barriers, and is built to avoid, stop or sit down rather than fall. Illustrative RaaS pricing is about $8,500 a month, and EU and UK buyers come in 2027. Shipping, but small.

Unitree

Unitree ships more humanoids than almost anyone, and by most accounts few of them do factory work. Its listing documents say it sold more than 5,500 humanoids in 2025, on revenue of 1.699 billion yuan. Counterpoint puts its first-half 2026 shipments at more than 7,000.

A secondary report on the prospectus says the filing admits that robotic hands are not yet precise or durable enough for sustained use, and that industrial-use revenue is small. So the units are real, but many go to research, education and showcase work. I rate it Shipping, with that caveat.

UBTech

UBTech has the best paper trail for factory sales in China. In November 2025 it said it had begun mass production and delivery of several hundred Walker S2 robots, with Walker orders above 800 million yuan and customers that include BYD, Geely and Foxconn.

Its half-year results show the follow-through. Full-size humanoid revenue rose 1,445% to RMB 590.3 million, and 921 units were sold. The company is still loss-making, and the figure is not broken out by model. Still a real sales line, so Shipping.

AgiBot

AgiBot is the volume leader by one count. Counterpoint ranks it first in the first half of 2026 with about 9,700 units and over 43% share, out of more than 22,000 shipped worldwide. The five largest vendors by shipments are all Chinese.

That is an analyst estimate, not an audited delivery count, and I could not verify named customer sites on a primary page. I list it as Shipping on that estimate alone, but I would ask any AgiBot seller what the robots actually do all day.

Figure

Figure has the best demo-to-pilot story in the US. BMW says Figure 02 supported production of more than 30,000 X3 vehicles over an 11-month deployment, and Figure 03 is now going to Spartanburg for a logistics sequencing job.

What is missing matters. The coverage gives no unit count, no price and no order size. Figure says its BotQ factory can make up to 12,000 robots a year, but capacity is not sales. One named customer and one long pilot is a Paid pilot, not a fleet.

Boston Dynamics (Atlas)

Boston Dynamics says all 2026 Atlas deployments are already committed, with fleets going to Hyundai's application centre and Google DeepMind. More customers are planned for early 2027. No price was given.

In September the company said the first phase of that Hyundai centre is operating, starting with parts logistics, with a Hyundai plan for 25,000 units over the next few years. That is a plan. Note that the first customer is also the owner. On top of that, Rohit Prasad took over as CEO in October, replacing an interim CEO. I call this Paid pilots, with an in-house buyer.

Apptronik (Apollo)

Apptronik closed a Series A of over $935 million in February 2026, and named partners: a 2024 commercial agreement with Mercedes-Benz, a GXO proof of concept, and Jabil as a production partner.

The most recent deployment news is Robot Park, where Apollo fleets collect data to train the next generation of intelligence. That is work, but it is training work. I found no unit count, price or repeat order. It earns Paid pilots on the strength of the Mercedes agreement, and I am less sure of that call than the others.

1X (NEO)

1X sells a home robot. A trade-site report says its Hayward factory has started full-scale production, with capacity for 10,000 a year, early access at $20,000 or $499 a month, and first customer shipments planned for 2026. Early robots go to its own lab and internal home testing.

The order page still shows a $200 deposit. I found no page confirming that a customer has received a NEO. Until one does, it is preorders and a factory, so Demo only.

Tesla (Optimus)

Tesla reported in April that Fremont production would start in late July or August. Musk said the 2026 rate was impossible to predict. In January he conceded that no Optimus robots were doing productive work in Tesla's factories, and the article discusses no outside customers.

I could not confirm that production has begun. Tesla reports earnings on 21 October, which may settle it. Today it is Demo only.

Persona AI

Persona is aiming at shipyard welding. IEEE Spectrum describes a humanoid prototype that is still learning to weld, with partnerships in Korea with HD Hyundai and POSCO. There is no unit count and the approach has yet to be proven at scale. A sensible target market, but Demo only.

What the scorecard says

Three things stand out.

First, the volume is in China. Counterpoint's first-half count has all top five vendors based in China. Shipments are not the same as productive work, though.

Second, the Western leaders are still at a handful of sites, not thousands. Agility's evidence is the deepest, and even there the orders are conditional.

Third, capacity announcements are not sales. Factories rated for 10,000 or 12,000 units a year are common. Evidence of that many customers is not.

If you are buying

Ask for the same five things from every seller: the site name, the repeat order, units in service, the safety method, and the monthly price. Ask for mean time between interventions in writing.

Start with a paid pilot on one task you already understand. Do not buy a platform. The same columnist notes that specialized, non-humanoid robots usually win on unit economics in new facilities. A humanoid earns its place where the building was made for people and cannot be changed.

I track which robots are actually earning their keep every Tuesday in the Physical AI Briefing. Get it by email.