State of the machines

If you follow robotics communities online, it can feel like humanoids are everywhere.

In reality, there are very few in the world, and even fewer doing real work.

Most of what you see are carefully staged demonstrations. The volume of videos creates the impression that humanoids are already widespread. They are not.

The hardware itself is no longer the main limitation. The intelligence is.

Many humanoids can walk, balance, lift, and move convincingly, sometimes even with superhuman movement, in controlled environments. The real gap is intelligence. These systems still struggle with the variability, unpredictability, and decision making required outside staged settings. While simulation, teleoperation, visual learning systems, and synthetic data are all being explored to help close this gap, real world deployment provides a form of feedback that is difficult to replicate elsewhere.

How many humanoids actually exist

Counting humanoids is surprisingly difficult.

Some counts include dual armed, wheeled mobile manipulators. Others only include true two legged anthropomorphic robots.

The lack of a shared definition makes most headline numbers unreliable.

This is made harder by the fact that not every company shares production or shipment data publicly. Much of what exists is pieced together from announcements, demos, and assumptions rather than verifiable counts.

Companies like Agibot with a wide and varied product line show how blurred these definitions can be.

Source: Agibot product materials and public disclosures

Agibot is one of the most active humanoid producers today and has shared breakdowns of its production across several robot lines. Some are full size bipedal humanoids. Others are smaller agile bipeds. And a significant portion are task focused systems built on a wheeled base with dual arms.

All of these are often grouped into broader “humanoid” counts, despite being very different in form and intended use.

By Agibots own breakdown:

A-Series: full size bipedal humanoids, 1,742 units

X-Series: smaller agile bipedal humanoids, 1,846 units

G-Series: dual arm, wheeled base semi-humanoids, 1,412 units

This is not to diminish the achievement.

It is impressive production at a scale few others have reached.

But it clearly shows how easily very different robot types get mixed into the same “humanoid” bucket.

Unitree publicly pushed back on broader industry estimates. They stated that when counting only pure humanoids, defined as robots with two legs, two arms, and a full humanoid body, they produced over 6,500 units in 2025.

Even among the largest producers, humanoid output today is still measured in low thousands per company.

For an industry that often appears futuristic and ubiquitous online, the real installed base is still remarkably small.

Different paths to the same goal

When you look past the demos and focus on what is actually shipping, a few distinct approaches start to emerge.

Some companies are prioritizing accessibility and wide distribution, getting robots into the hands of researchers, universities, and early adopters to accelerate learning through real-world experimentation.

Others are focusing on controlled industrial deployments, placing robots in structured environments where they can perform specific, repeatable tasks alongside human workers and established processes.

And a third group is planning for massive production scale, leveraging automotive-grade manufacturing capacity to eventually produce hundreds of thousands or even millions of units.

These are not mutually exclusive strategies. Some companies are pursuing multiple paths simultaneously. But the difference in approach shapes how quickly humanoids move from labs into the real world, and what kind of feedback loops drive their improvement.

Getting these machines into varied environments at scale is an important step. Real use leads to failure, data, iteration, and learning. This is a necessary part of humanoids eventually becoming useful.

A humanoid you keep seeing everywhere

The experimentation path

If you spend any time watching the humanoid space, one model appears again and again: the Unitree G1. It is relatively cheap, lightweight, compact, and reasonably accessible. You will see it in research labs, universities, small teams, hobbyist projects, and even robot fighting experiments. It has become one of the default humanoids for experimentation around the world.

Early production at scale

Agibot is producing multiple humanoid lines at notable volume across different form factors, and these units are being deployed in real-world settings. The company recently rolled out its 5,000th mass-produced robot, with units operating in commercial scenarios such as production lines, logistics sorting, security inspection, and scientific research. Agibot has also piloted robots on a production line at Longcheer Technology, handling components within an active manufacturing workflow.

UBTech has marked a production milestone with its Walker S2 robot. As of late 2025, around 1,000 Walker S2 units had been produced, and the company is delivering these industrial humanoids to partners in sectors such as automotive and electronics manufacturing, while also engaging in trials with companies like Airbus and Texas Instruments.

EngineAI has signed a strategic agreement with Duolun Technology to supply at least 2,000 humanoid robots over the next three years for deployment in public security, traffic management, and retail scenarios, signaling a planned production ramp that goes beyond demonstrations and into real operational pipelines.

The industrial deployment path

Boston Dynamics' Atlas has its entire 2026 production run already sold and accounted for, with fleets scheduled to ship to Hyundai's Robotics Metaplant Application Center and Google DeepMind. The production version will be deployed by 2028 at Hyundai's electric vehicle manufacturing facility near Savannah, Georgia, initially focusing on parts sequencing and component assembly. Its majority owner, Hyundai Motor Group, has plans for a robot factory designed to eventually produce up to 30,000 units per year.

Apptronik has moved beyond pure demos with a 2025 pilot collaboration with manufacturing partner, Jabil, where Apollo humanoid robots are being built and tested in live factory workflows on tasks like inspection, sorting, and lineside delivery, a real deployment step toward scaled production.

Agility Robotics’ Digit is already commercially deployed inside warehouse environments. It has moved over 100,000 totes in real operations and has partnered with Mercado Libre, the largest e-commerce and logistics company in Latin America, for further rollout within a structured logistics workflow.

What comes next

Some robots are being distributed widely to encourage experimentation and iteration. Others are being placed carefully into environments where they are expected to perform repeatable work. At the same time, many companies are showcasing impressive demos and ambitious production targets. Not all of these plans will materialize at scale.

That said, the involvement of large automotive manufacturers and major industrial groups puts some of these ambitions on firmer ground than they might first appear. These are companies with existing factories, supply chains, and the ability to ramp production far beyond what most robotics startups could achieve.

Tesla has signaled very ambitious long-term production goals for its Optimus humanoid robot.

In January 2026, the company announced it would wind down production of the Model S and X to repurpose its Fremont factory for Optimus manufacturing.

On the earnings call, CEO Elon Musk stated that Tesla is “really moving into a future that is based on autonomy.”

He added that the company is replacing the Model S and X production line with capacity to produce up to 1 million Optimus units per year, initially targeting Tesla's own factories before eventual external sales.

Tesla is preparing to unveil Optimus V3 in early 2026 as a production-ready design, with sales expected to begin in 2027.

Tesla isn't alone in this automotive-to-robotics shift. XPeng is another example of this shift toward production intent.

“Today, the first unit of our ET1 robot, developed to automotive-grade standards, has successfully rolled off the production line. This marks a critical step toward the large-scale mass production of advanced humanoid robots this year.”

- He Xiaopeng, CEO of XPeng (Jan 2026)

This is notable not because the robots are already widely deployed, but because a major automotive manufacturer is treating humanoids as something that can be built using the same production discipline as cars.

The intent to scale is real. The path to usefulness is still being worked out.