In 2026, a humanoid robot walking smoothly across a stage stopped being news. Several companies can do it. What still separates them is far less photogenic: whether anyone outside the company has taken delivery of one, paid for it, and kept it running.
That distinction gets lost constantly, partly because announcements and deliveries use the same vocabulary. "Mass production" has meant a thousand units on a customer's factory floor for one manufacturer this year, and an installed but not yet running assembly line for another. Both were reported as production.
The gap between the demo and the delivery
Humanoid robotics in 2026 has two separate stories running at once, and they get reported as though they were the same one. In the first, robots walk with startling fluidity, fold laundry in a research lab and raise money at valuations that would have sounded absurd three years ago. In the second, a much smaller number of machines turn up at a factory, get assigned a task, and are still doing it six months later.
The first story generates most of the coverage because it is visual and it happens on a schedule that suits announcements. The second is slower, duller and considerably harder to fake. If you are trying to work out where this technology actually is, the second story is the one to follow.
The useful question is not which robot is most impressive. It is a much flatter one: how many units went to a paying customer who was not the manufacturer, and are they still switched on?
Who has actually delivered units
By that measure, the clearest case in 2026 is UBTECH's Walker S2. The company began mass production in November 2025, reported roughly ¥1.4bn in cumulative orders across the year, and said it delivered its 1,000th unit in late December on a Dongfeng Liuzhou production line. In January 2026 it announced a cooperation agreement with Airbus covering early-stage testing. A ¥264m border-security contract sat inside that order total.
An order book is not the same as a thousand robots doing productive work, and UBTECH has not broken out how many of those units are past the pilot stage. But named industrial customers and a four-figure delivery count are a different class of evidence from a video.
The design choice that makes the S2 interesting is deliberately unglamorous. It swaps its own battery: walks to a station, pulls the depleted pack off its own back, inserts a fresh one, and stays powered throughout, in about three minutes. A robot that runs for a couple of hours and then needs a person to intervene is something you schedule around. One that handles its own energy is something you can put on a rota.
AgiBot's A2 Lite is worth noting for a different reason: it has a published price. At US$44,560 on the company's international store, it is one of very few full-size humanoids an ordinary buyer can obtain without entering an enterprise sales process. Most of the field quotes on application, which is a polite way of saying the price is whatever the buyer will bear.
What "mass production" has meant elsewhere
Tesla's Optimus is the clearest illustration of how far a launch narrative can drift from a shipping product. As of August 2026 the third-generation robot has still not been shown publicly. Tesla missed an initial reveal target early in the year, said in April that a reveal would come around the middle of it, and set no new date on the July earnings call. A production line converted from the Model S and X floor at Fremont was completed in May, with production described as anticipated later in the year.
Elon Musk has been fairly direct about the difficulty, describing the ramp as likely to be flat and long because the robot uses around ten thousand unique parts, and acknowledging in January that no Optimus units were yet doing commercially useful work at Tesla's own facilities. Public sales have been indicated for around the end of 2027. There has never been a pre-order or a deposit.
None of that makes Optimus a failure — building a humanoid at automotive volume is genuinely hard, and Tesla is one of the few companies with the manufacturing base to attempt it. It does mean that a directory listing it as available, or a comparison table quoting its specifications as though they were fixed, is describing something that does not yet exist.
Figure sits between the two poles. Reporting through mid-2026 put its BotQ facility at roughly one robot per hour, which is a real manufacturing rate rather than a projection, though the company has not published a price or opened general sales.
The money is arriving before the machines
Capital has moved considerably faster than deployment. Venture data indicated robotics startups raised around $23bn during 2026, close to the whole of the previous year with months still to run. In August, XPENG Robotics raised over $900m at a $6.3bn valuation in a round led by IDG Capital, with Tencent and Alibaba participating, for a robot that has no published price and is not on sale.
The XPeng case is a good illustration of what investors are actually buying. IRON has 76 degrees of freedom in the body and 21 in each hand, and runs three of XPeng's own Turing chips. Those are respectable numbers. But the investment case is the production line: XPeng already builds electric vehicles at volume, and the argument is that a humanoid supply chain can borrow from one that already exists. The company targets mass production by the end of 2026 and customer deliveries in 2027.
Whether that transfers is the open question of the sector. A car and a humanoid share suppliers, batteries and some sensing, and share almost nothing about tolerances, actuator counts or how a failure behaves.
Four things worth checking before you believe a launch date
If you are assessing a humanoid announcement rather than enjoying it, four questions do most of the work.
Has it been shown publicly? Sounds basic. A surprising number of specification sheets in circulation describe machines that have never been demonstrated outside a company video, and some describe a design that has already been revised.
What is the runtime, and who changes the battery? Manufacturers publish degrees of freedom far more readily than they publish how long the robot works between charges. That silence is informative. UBTECH has not published a runtime for the S2 either — what it published instead was the swap.
Who has bought one, and for what? A named customer with a named task beats any benchmark. Pilots are worth less than they sound; the interesting number is how many pilots converted.
Is there a price? A published price is a commitment. "Contact for pricing" on a product that has been announced for two years usually means the unit economics are not settled yet.
None of this is scepticism about the technology. The machines are improving quickly, and the battery-swap and manufacturing-rate stories from this year are more significant than most of the walking demonstrations. It is a suggestion about which numbers to write down. Our Bots For Sale directory lists each robot with its actual status and price where one exists, and says plainly when neither is public.