The Dance Videos Are Real. The Laundry Robot Is Not Here Yet

Humanoid robots now work real shifts in a few warehouses and car plants, but the home is still unsolved, and a viral video cannot tell you how far off it is.

That is the reading of this paper, and the evidence below supports it. The factory work is genuine and the dance videos are genuine. What a backflip does not show is whether a machine can cope with a kitchen it has never seen, a toddler it was not told about, or a shirt that folds differently from the last one.

The stakes are the size of the promises. Tesla's Elon Musk has said Optimus robots will have "human and then superhuman dexterity," could automate almost all human labor, and might cost as little as $20,000 each, with public sales possible by the end of 2027. Nvidia's Jensen Huang said in January 2026 that humanoids would match human-level ability this year. Morgan Stanley projects nearly one billion of them by 2050, a market worth more than $5 trillion. ^[1]

A humanoid, in the plainest terms, is a robot built on the human plan, with a torso, arms, hands and usually legs, so that it can work in spaces designed for us. The pitch is that the AI breakthroughs behind chatbots will now do for bodies what they did for words. The evidence is that they have not yet.

What the working robots actually do

Start with the good news, because it is real. Agility Robotics says its Digit 4 has logged more than 65,000 hours of commercial operation at customer sites including GXO, Schaeffler, Amazon and Toyota Motor Manufacturing Canada, the longest tenure of any humanoid in industrial work. At a GXO logistics facility near Atlanta, the company says, it handled 100,000 totes with about 98 percent accuracy while on task. ^[5]

Figure AI says its Figure 02 robots ran more than 1,250 hours at a BMW plant in Spartanburg, South Carolina, helping produce over 30,000 X3 vehicles and moving more than 90,000 components. It has now put the upgraded Figure 03 at the same plant for a harder job, picking unsorted parts and arranging them in sequence for the assembly line. ^[6]

Both sets of figures come from the companies themselves, and neither has been independently checked. Even taken at face value they describe a narrow life. Tasks repeat, conditions are controlled, and the cost of failure is a dropped tote.

That 98 percent deserves a second look. On 100,000 totes, two percent is about 2,000, if the figure means what it seems to. What "on task" leaves out, and how often a human steps in, is not disclosed in anything this paper has seen.

The money tells the same story. Agility reported more than $300 million in multi-year customer orders for its next model, Digit 5, as of May 2026. Its net sales in 2025 were $1.8 million. ^[5] Orders are not revenue, and the gap between the two is more than 160 to one.

The humanoid shape is not even settled as the answer for factories. Toyota plans to put 150,000 robots in its own plants and 250,000 across group companies from 2028, and has chosen a wheeled design, betting that fine manual skill matters more in the near term than walking. ^[2] A Hyundai executive said in September 2026 that Boston Dynamics has yet to deploy its Atlas humanoid at scale and remains unprofitable. ^[2]

Why a shirt is harder than a tote

The best robot brains today are called vision-language-action models, or VLAs. They are trained on hours of human teleoperation, meaning a person steers the robot through a task while the system records what happened. A robot running one can pack a lunchbox or fold origami. Ask it for something outside what it was shown, and it is highly likely to fail. ^[1]

Imperial College London's Edward Johns puts it dryly. Today's best models can do "a few things here and a few things there," he said, and are not true generalists. ^[1] It is a kind of memory with no imagination attached. The robot has kept what it was shown, and has nothing to reach for beyond it.

Language models had the whole internet to read. Nothing of the kind exists for the physical world. The ways to collect robot training data, teleoperation, videos of humans, deployment in the real world, are all costly, low quality or unsafe. ^[1] Tesla has felt that last problem. It asked workers in Texas and California to wear motion-capture suits so their movements could train Optimus, and some objected because they knew the robots were meant to replace them. Tesla moved to dedicated data teams. ^[2]

The field is split on whether more data is the cure. Agility cofounder Jonathan Hurst calls the idea that VLAs can reach generality a "fundamentally flawed premise," since it would require "complete data coverage of all of the things that a robot could ever do." ^[1] Yann LeCun, Meta's chief AI scientist, went further at Davos in January 2026. "None of those companies - absolutely none of them - has any idea how to make those robots smart enough to be useful," he said. ^[1] He argues that the methods that worked for language "do not work for high-dimensional, continuous, noisy data," and points to world models, AI trained on video, 3D scans and sensor data to predict physical outcomes. ^[1]

The body is no easier than the brain. Hurst's own line is the best summary: "It's very easy to make a robot that looks like a person. It is dramatically more difficult to make a machine that moves or behaves dynamically or physically like a person." ^[1] Tesla's Optimus V3 hands, according to reporting by The Information relayed by Ars Technica, contain more than 100 small components that need manual assembly. Its touch sensors have proved unreliable, and production equipment struggles to line parts up precisely. The same source says Tesla's AI is still insufficient for general-purpose work. ^[2]

Tesla has stopped making the Model S and Model X at its Fremont factory as of May 2026 and moved line workers and engineers to Optimus, with a target of more than 1,000 robots a week by the end of 2026. ^[2] That is a large bet placed ahead of the evidence.

The home, in the fine print

The consumer machines show the gap in the plainest terms, because their makers have to print the limits. Flourish Robots has announced the Flourish 1 at $3,555, taking preorders for 50 units with delivery expected in December 2026. It can learn a household task in 30 minutes from a phone demonstration. It rides on wheels, cannot climb stairs and cannot lift more than 1.5 kilograms, or 3.3 pounds. ^[3]

Around children and pets it is built to stop. CEO Antoine Marcel said it halts its motion when it detects a person or animal entering its working area. The website adds a line that every parent will understand: "It works around your kids. It doesn't look after them." ^[3]

Others are on preorder too. The 1x Neo costs $20,000 or $499 a month, and Weave's Isaac 1 costs $7,999 or $449 a month. Deliveries are expected from fall 2026 in California and more broadly across the United States in 2027. ^[3] The sources reviewed here give no stair or lifting limits for either of them, so this paper does not say what they can and cannot do. Note that Neo's price is the figure Musk has floated for Optimus, which is not on sale.

A broader accounting comes from Anthropic researchers, who used Claude to assess robot exposure across US job tasks. Robots can perform 74 percent of physical work tasks in the US economy, but they are cost-competitive with human labor for only 0.3 percent. At current rates of price decline, reaching 10 percent would take roughly 40 years. Only two percent of physical tasks can be done by robots in unstructured places such as city streets or private homes. ^[7] The number that matters for your laundry is that last one.

The part nobody puts in the video

Safety is where the demo culture is most exposed. Aaron Ames, a Caltech roboticist who directs its Center for Autonomous Systems and Technologies, said: "Very few of the humanoid companies are really thinking about safe deployment." He describes humanoid demonstrations as "driven solely by performance and making the latest cool video." ^[4]

He points to the failures that shut down Uber's autonomous vehicle program and GM's Cruise as warnings. ^[4] A self-driving car at least stays on a road. A 284-pound walking machine in a hallway is a different problem. That is the weight of Agility's Digit 5. ^[5]

To be fair, some of the industry is moving. Digit 5 is Agility's first robot built to be "cooperatively safe," able to work near people without physical barriers, using AI-based human detection, visual and auditory cues and an independent safety controller. Early access begins in the first half of 2027. ^[5] Ames invented control barrier functions in 2014. These are equations that define safe actions and sit as a guardrail between the AI's decisions and the physical world, and his lab has used them to make humanoids dodge thrown balls and stop short of people. ^[4] The safety tools exist. The open question is whether companies chasing the next viral clip will use them.

The strongest objection

The best case against my reading is that every technology looks narrow before it is not. Digit and Figure are not demos. They are paid work, and each shift is a lesson. Language models were once toys too. And the VLA results, the lunchbox and the origami, are real advances that would have seemed far off a few years ago.

All true. But the scaling story that carried language depended on a pool of data that robotics does not have. The researchers who know the field best, Hurst and LeCun, disagree about the fix, and they agree the present recipe will not do it. Progress in a warehouse is also not progress in a home. A task repeated 100,000 times in one building trains a robot for that building, while a home is 100,000 different buildings. Nobody has shown, in the sources reviewed here, that the first leads to the second on any schedule.

What to ask the next time a robot dances

This paper does not know how many of the year's viral humanoid videos were autonomous and how many were teleoperated, and the dossier shows no audit of them. Companies rarely say. That silence is a fact about the industry, and a reason to ask.

So ask three things. Was that autonomous or remote-controlled? How many times did it fail before this take? And would you let it near your cat?

The tests are already scheduled. Flourish promises its first 50 units in December 2026. Digit 5 opens early access in the first half of 2027. By then we will know whether the robot in the video is the robot in the house.