Humanoid Robots Explained: Cost, Makers and What Works in 2026

What a Humanoid Robot Is and Why 2026 Is Different

A humanoid robot is a machine built on the human body plan: a torso, two arms, two legs, hands and a head full of sensors. Nobody builds one to look like a person. They are built because doors, stairs, shelves and tools are already sized for a human body, so a robot with the same shape can work in a building nobody has to rebuild.

Between 2023 and 2026, foundation models improved how a robot turns camera images and spoken instructions into movement, Chinese manufacturers pushed actuator and battery costs low enough to sell small humanoids at used-car prices, and BMW and Hyundai began paying for pilots in real plants. A humanoid still cannot do any job you hand it, but it is now worth building in volume.

The label covers three quite different products. The table groups what is on sale or in pilots in late 2026.

The label also covers wheeled upper-body robots and legs-only research platforms, which is one reason shipment counts disagree.

Humanoids vs the robots already in factories

Factories already run on robots, and nearly all of them are fixed arms. The International Federation of Robotics (IFR) counted 542,076 conventional industrial robots installed in 2024. Against that, the roughly 7,000 full-size humanoids sold in 2025 are about one for every 77 (different years, but the scale is clear). A humanoid does not compete with a welding arm that repeats one motion for years. Its case is the work between the fixed machines: carrying parts, feeding stations and moving totes where a conveyor was never built. For a wider look at how factories use AI, see our guide to AI in manufacturing.

How a Humanoid Robot Works: Body, Brain and the Data Center Behind It

A humanoid robot runs on three layers: a body of motors and sensors, an onboard computer that handles balance and perception in milliseconds, and a trained model that decides what to do. The first two live inside the robot. The third is trained in data centers and simulation, then copied onto the machine.

• Unitree's G1 stands 1,320 mm tall, weighs about 35 kg and has 23 joints in its base version (Unitree, 2026). The G1 EDU research version goes up to 43.
• Cameras and depth sensors handle sight, joint encoders report position, and the hands carry force sensors.
• A 9,000 mAh quick-release battery gives the G1 about 2 hours. 1X lists NEO at 842 Wh and 4 hours, with 6 minutes of charging per hour of run-time.
• An onboard computer handles walking, balance and collision avoidance locally.
• A vision-language-action model turns images and instructions into motor commands.

The onboard chip: NVIDIA Jetson Thor

NVIDIA's Jetson Thor is the robot computer that 1X lists as the brain of NEO. According to NVIDIA's Jetson Thor page, the top T5000 module and the developer kit deliver 2,070 FP4 TFLOPS (sparse) with 128 GB of memory at 273 GB/s on a Blackwell GPU, in a power range of 40 to 130 watts. The smaller T4000 is rated at 1,200 FP4 TFLOPS and 40 to 70 watts. The developer kit launched at $3,499 (Hackster, 2025). The 2,070 figure assumes sparse FP4, so read it as a best case for that number format.

For scale, a data center H100 GPU draws up to 700 watts. Thor works within a fifth of that or less, and that efficiency is the point of edge hardware. Our guide to edge AI covers the trade.

A quick power check

1X lists NEO's battery at 842 Wh and its run-time at 4 hours. That works out to an average draw of about 210 watts for the entire robot. If the Thor module ran at its 130-watt ceiling the whole time, compute alone would be about 62 percent of the energy budget. Even at its 40-watt floor it would be about 19 percent. Real draw sits somewhere between, and 1X does not publish it. That is my arithmetic from two vendor specifications. So in a robot, efficiency matters more than peak speed. Every watt spent thinking is a watt not spent moving, and it shortens the shift.

Where the brain is trained

The model that tells a robot what to do is trained far from the robot. NVIDIA released Isaac GR00T N1 as an open humanoid model in March 2025. It pairs a vision-language model that reasons about the scene with a diffusion transformer that generates motor actions, and later versions (N1.5 and N1.6) followed. Training mixes real demonstrations, which people record by remotely operating a robot, with synthetic data generated in simulators such as NVIDIA's Isaac Sim. All of it runs on GPU clusters like the ones described in our guide to AI training vs inference.

Tesla is taking the same idea to the factory floor. On its Q3 2026 earnings call on October 21, 2026, Musk said Tesla would need roughly 10,000 to 20,000 robots to generate training data through self-play, which he called an Optimus Academy.

A robot also cannot wait for a network round trip to decide whether to catch a fall. Balance and collision avoidance run on the robot, and slower planning can lean on the cloud. Cloud-only control would mean a dropped Wi-Fi signal drops the robot, literally.

How Many Humanoid Robots Are Sold: Three Counts That Disagree

Somewhere between 7,000 and 25,000, depending on who is counting. The three most cited numbers measure different things, so do not add them or average them.

The IFR, in its World Robotics 2026 release, counted about 7,000 full-size humanoids sold in 2025, and said the sector is still moving from labs to pilots. It also identified 174 humanoid manufacturers worldwide, which is roughly 40 full-size sales per maker if you divide one by the other, and noted that most applications still need human teleoperation. IDC's higher figure includes robots that stand around 1.2 to 1.3 meters. Later in 2026 IDC revised its 2025 baseline to about 18,500, and the first half of 2026 alone came in about 35 percent above that.

Who ships them

IDC's first-half 2026 data shows how concentrated the market is. Chinese vendors hold more than 95 percent of it. AgiBot shipped about 8,600 units, roughly 35 percent of the total, Unitree about 5,000 and UBTech about 900. For 2025, IDC ranked AgiBot first with about 5,200 units, of which 1,300 were full-size, and Unitree second. Unitree's IPO prospectus, as reported, puts its own 2025 shipments above 5,500.

Where they go

Where the robots go matters more than how many there are. IDC says 69 percent of first-half 2026 shipments went to research and education, exhibitions, entertainment and government data centers. The other 31 percent covers factory and warehouse work and other uses. A shipment is a robot delivered. It is not a robot doing a paid job, and for most of 2026 the two have been different things.

IDC also projects more than 750,000 humanoid shipments in 2030, a figure that depends on how the US restrictions in section 6 play out.

Who Makes Humanoid Robots and What They Cost

Chinese makers sell the cheapest and most numerous humanoids, while US makers run the best documented factory pilots. Almost no full-size industrial robot has a public price. The table lists the makers with a verified status as of November 2026.

"We've started what I'd call early production on our high-volume line in Fremont." (Elon Musk, Tesla Q3 2026 earnings call, October 21, 2026)

On the same call Musk said it would take at least a year, possibly longer, to reach a rate of 20,000 units a week, which is about one million a year. Pricing was not discussed.

Public prices

Prices are public only for the smallest and the most consumer-facing robots.

The Number Most Guides Don't Show

IDC reports humanoid revenue of more than $740 million in the first half of 2026 on nearly 25,000 units. Divide one by the other and the average robot sold for about $29,600, which matches IDC's own average. A year earlier it was about $37,000. For full-year 2025, IDC's $440 million over about 18,000 units works out to roughly $24,400. The division is mine, from IDC's published totals.

The average humanoid sold today already costs about what Tesla says Optimus will cost at scale. The market is not waiting for a $20,000 to $30,000 robot. It is buying them now, and they are small research machines, not factory workers. The average also fell from $37,000 to $29,600 in a year, so the mix is tilting toward cheaper units from Chinese makers. Full-size humanoids made up 41.6 percent of IDC's 2025 revenue, so each full-size robot moves the average far more than a small one, and the Western full-size robots in pilots have barely begun to ship.

What Humanoid Robots Can Actually Do Today

In 2026 a humanoid can do narrow, repeated material-handling jobs inside a mapped, supervised area, and little else without help. Most of the 25,000 shipped in the first half are used for research, shows and collecting training data.

The best documented factory case is BMW's plant in Spartanburg, South Carolina. Figure AI's Figure 02 ran an 11-month pilot there that helped build more than 30,000 BMW X3 vehicles by inserting sheet-metal parts for welding. In June 2026 BMW said Figure 03 would move into logistics, picking unsorted parts from large containers and placing them on trolleys in the order the assembly line needs.

"Our 11-month deployment of Figure 02 proved that humanoids are no longer lab experiments." (Brett Adcock, Figure AI, June 2026)

Agility's Digit moves totes at logistics sites, and Hyundai plans to put Atlas on its Georgia Metaplant line for parts sequencing from 2028, with assembly tasks later. Every named deployment is a defined task in a structured space, not a general-purpose worker.

What the demos hide

Remote operators are the first thing a demo leaves out. The IFR notes that most humanoid applications still need human teleoperation. 1X says NEO works autonomously by default, but a 1X Expert can remotely supervise tasks it does not know. Companies rarely publish how many minutes of human help each task takes.

Batteries come second. A runtime of 2 to 4 hours means swaps or charging breaks inside a shift. The 6 minutes of charge per hour of run-time that 1X lists for NEO is quick for a battery that size, and it still adds up.

Safety rules lag behind. The humanoid-specific standard, ISO 25785-1, was still a working draft in early 2026 according to standards trackers, with estimates for publication ranging from 2027 to 2028. Until then the industrial robot standard, ISO 10218, is the baseline, which is why pilots keep robots in defined zones.

And a robot delivered is not a worker employed. With 69 percent of IDC's first-half shipments going to research, exhibitions and similar uses, the number of humanoids doing paid work is far smaller than 25,000.

Why Humanoid Robots Matter for AI Infrastructure

Humanoids put AI hardware under a tougher constraint than a data center does: the compute has to fit in a robot's weight, heat and battery budget while also keeping it upright.

At CES on January 5, 2026, NVIDIA CEO Jensen Huang called the ChatGPT moment for physical AI "nearly here," according to Fortune's report of the keynote. NVIDIA's press release that day was more definite, saying the ChatGPT moment for robotics "is here." The difference between "nearly here" and "is here" is roughly the distance the technology still has to cover.

Compute at both ends

On the robot, chips like Jetson Thor compete on tera-operations per watt. In the data center, training a vision-language-action model and generating synthetic demonstrations in a simulator both consume the same GPU hours as any other model. Each new robot fleet is also a data source, which is why Tesla talks about building 10,000 to 20,000 robots for self-play. For how accelerators differ by job, see our guide to AI accelerators.

The supply chain split

Chinese vendors hold more than 95 percent of humanoid shipments by IDC's count, while the most capable AI chips and models come largely from US companies. On July 28, 2026 the US Federal Communications Commission added "foreign-produced advanced robotic devices" to its Covered List, which blocks new models from receiving the equipment authorization needed to import and sell them in the United States. Models authorized before that date can still be sold, and some software and firmware updates are allowed through at least January 1, 2029. Vendors can seek conditional approval after a national-security review and commitments to move production to the US. The rule applies by where a device is made, not by the nationality of the company.

IDC's trade model projects US humanoid revenue 58 percent below its pre-restriction baseline by 2030, a loss of more than $6 billion, because humanoids have short product cycles that force frequent recertification. Treat that as a projection.

Compare the power budgets from section 2. A data center GPU can spend 700 watts on a single chip, while NEO's whole body averages about 210 watts. The accelerators that matter most for humanoids are not the biggest ones. They are the ones that deliver the most decisions per joule.

What to Watch Through 2027

Expect more robots before you see more useful work. These dated milestones will show whether that changes.

• Musk said on October 21, 2026 that outside companies should receive Optimus around the end of 2027, that the AI5 chip should go into Optimus around April 2027, and that reaching 20,000 units a week would take at least a year.
• Figure 03 begins logistics sequencing at BMW Spartanburg. The first published results for a fleet, not a single pilot, are what to look for.
• Atlas production started in 2026, with the first Metaplant use planned from 2028 and a planned US robot factory with 30,000 units of annual capacity.
• ISO 25785-1 moves from working draft toward publication, which decides where humanoids can work beside people without fences.

IDC's forecast of more than 750,000 shipments in 2030 is about 15 times the annualized pace of the first half of 2026 (nearly 25,000 in six months is roughly 50,000 a year). The arithmetic is simple. Whether the factories exist to absorb that many robots is not.

My read: units will arrive before usefulness. The numbers to ask vendors for are the ones nobody publishes today: how often a human has to step in, how many hours a robot runs between failures, and what a task costs once remote help is counted. Until those appear, treat every shipment total as a count of robots built, not work done.

Frequently Asked Questions

Related Articles