A fundamental view of the robotics landscape

00 — First stepsThe machine with the mind of a toddler

In the autumn of 1973, in a laboratory at Waseda University in Tokyo, a machine the size of a grown man lifted one foot and set it down. Then it paused, gathered itself, and did it again. Each step took the better part of a minute. The machine was called WABOT-1, and it was the first full-size humanoid robot anyone had built.

It could walk, after a fashion. It could grip an object with two hands and feel its weight. It could hear a simple question and answer in Japanese through an artificial mouth. The team that built it, led by Professor Ichiro Kato, reckoned it had roughly the mind of a year-and-a-half-old child.

One of those engineers, Hiromichi Fujisawa, drew the job of teaching it to speak. Decades later the BBC asked him how it felt at the time. His answer was four words. "We were not confident."1

Thirteen years later, four thousand miles away, Honda was doing the same thing in secret. Inside a research center in Wako, engineers built a pair of legs called E0 and taught it to walk in a straight line. It took about fifteen seconds to complete a single step. The company hid the work for a decade. When it finally showed the descendant of those legs — the P2, in 1996 — the field was stunned that a robot could balance on two of them at all. The line ran on to ASIMO, the friendly white astronaut that bowed and jogged and climbed stairs and, for a while, looked like the future.

This past spring, the future looked different. A company called Figure put three of its robots on a warehouse floor in front of a live camera and promised an eight-hour shift with nobody at the controls. The robots ran for eight hours, and then kept going. They ran past a full day. People in the chat asked the founder to name them, so he printed labels and stuck them on the chests: Bob, Frank, Gary. By the time the stream crossed twenty-six hours the three of them had moved more than thirty-three thousand packages, reading the world straight from their own cameras.2

Between Fujisawa's "we were not confident" and a robot named Gary working through the night lies the whole story of this industry. It took fifty-three years. Most of the ground was covered in the last three.

What follows is a map of where things stand in the middle of 2026: what actually changed, who is building what, and why Japan, the United States, and China are chasing the same machine for completely different reasons.

01 — Why nowThe body was never the hard part

For most of those fifty-three years, the legs were not the problem. By 2017 Boston Dynamics had a robot doing backflips. The trouble was that every one of those motions was scripted in advance. The robot was a puppet with no puppeteer once it met a situation nobody had programmed. It had a body and no general mind.

Then the same kind of model that powers a chatbot turned out to work on a robot. Show it enough examples of a task and it learns to map what the cameras see to what the motors should do. The industry calls these vision-language-action models. Figure's is named Helix. Tesla folds it into the same neural network that drives its cars. A handful of startups sell nothing but the brain. This is the actual unlock, and it is recent: the bodies were ready years before anything smart enough to run them showed up.

The catch is data, and the catch is large. Text was lying around the internet for free, so a chatbot could read essentially all of it. There is no internet of robots folding laundry. Every useful example has to be manufactured — by paying people to wear motion-capture rigs, by puppeting robots through chores by hand, or by training in simulation and hoping the lessons survive contact with the real world.

That single fact explains most of what comes next. It is why a robot ran a 26-hour livestream — the run was generating training data and public proof at the same time. It is why a home-robot company is happy to sell you a machine that a human in another country quietly drives when it gets stuck. And it is why China believes it has already won the part that matters. Add the duller tailwinds — factory labor is short, Japan and others are aging, governments want to reshore manufacturing, and the electric-car boom made motors and batteries cheap — and you have the conditions for the rush that is now on.

02 — JapanThe country that invented the future, then put it down

Japan got there first, and for a long time got there alone. WABOT in the seventies. ASIMO at the turn of the century. And underneath the famous humanoids, a quieter dominance: two of the four companies that build most of the world's factory robot arms — Fanuc and Yaskawa — are Japanese, and have been for decades.

Then, right before the moment it had spent forty years building toward, Japan flinched. Honda retired ASIMO in 2022. SoftBank had already stopped making Pepper, its friendly storefront robot, in 2021, after building around 27,000 of them.3 The companies looked at what a general-purpose humanoid would cost and what it might do wrong, and decided the world wasn't ready for it yet.

The most telling fact in Japanese robotics this year sits in a single press release. When Honda's trading arm wanted humanoids for warehouse work in the spring of 2026, it did not reach for a Honda robot. It signed a deal to use Chinese ones, made by UBTech.

Japan's bet now is quieter, and in its way shrewd. It is betting capital and parts. SoftBank's Masayoshi Son, who is rarely accused of thinking small, agreed to buy ABB's robotics division for about $5.4 billion and now calls physical AI "the next trillion-dollar opportunity." Japanese firms still make the precision gears, motors, and force sensors that almost everyone else's robots depend on. And in April 2026 the IEEE gave Honda's old P2 a formal award for being the first robot to walk like a person. It was a fitting honor and a slightly mournful one — a prize for the past, handed over while the present moved abroad.

03 — United StatesThe brain, the money, and the show

The American bet is that intelligence wins, and that whoever builds the smartest brain and raises the most money gets there first. It is a bet made loudly.

Figure — the headline act

In February 2024, Figure raised $675 million at a $2.6 billion valuation. Nineteen months later it raised more than a billion dollars again, this time at $39 billion — a fifteen-fold jump, one of the fastest in the history of hardware.4 Its robots spent eleven months on a BMW line in South Carolina, loading more than 90,000 parts. Its brain, Helix, runs on the robot itself rather than in the cloud.

And then there was the stream. The 26-hour run with Bob, Frank, and Gary was a real engineering achievement and a piece of theater at the same time, which is the American style exactly. The honest footnote belongs in the open: "zero failures" meant the system never crashed, not that the robots never erred. They dropped boxes and reset themselves, on camera. Competitors pointed out that their robots have been doing the same work in actual warehouses for years, off camera. Both of those things are true at once.

Tesla — the scale bet

Optimus is the volume play. Tesla started building its third-generation robot at Fremont in January 2026 and talks about a million units a year, then ten million. The reality in mid-2026 is a few hundred robots, mostly doing chores inside Tesla's own factories and gathering data.5 But the edge is real: Optimus runs on the same vision system that drives Tesla's cars, backed by a company that knows how to make machines by the million and sell them cheap. Its target price — $20,000 to $30,000 — is one no rival will say out loud.

Boston Dynamics — the old master

Owned by Hyundai since 2020, Boston Dynamics spent thirty years making the robots in everyone's viral videos. Its new Atlas is fully electric and frankly overbuilt next to its rivals: 56 joints, a 50-kilogram payload, and the party trick of walking to a station and swapping its own battery so it never stops. The first fleets ship this year to Hyundai's car plants and to Google DeepMind. The company's tagline for the new machine is plain: "the best robot we have ever built."

Agility, Apptronik, 1X — the workers

Beneath the giants sit the companies quietly doing the unglamorous version. Agility's Digit has held down a full-time job at a GXO warehouse for over a year, and the company's entire pitch is a rebuke to the hype: "No vaporware. No hype." Apptronik's Apollo is the body Google chose for its robot brain, Gemini Robotics, and works lines at Mercedes-Benz and GXO. And 1X went the other way entirely — it sells a home robot, Neo, for $20,000 or $499 a month, in three tasteful colors, that folds laundry and puts away dishes. Sometimes a human operator in Norway takes the controls to finish the job. It sold out its first year of production in five days.6

The brains for hire

Two of the most valuable companies in the field build no robot at all. Physical Intelligence and Skild make the brain and sell it to everyone — the Android-for-robots play. Skild raised $1.4 billion in January 2026 at a $14 billion valuation; Physical Intelligence sits around $5.6 billion, founded by some of the researchers who built Google's robot-learning work. And behind all of them stands Nvidia, selling the chips and the simulators — the one company guaranteed to make money no matter which robot wins.

What America does not have yet is revenue. Figure alone is valued at more than Goldman Sachs thinks the entire humanoid market will be worth in 2035. These are not prices. They are wagers on a future that appears on no income statement yet.

04 — ChinaThe body, the price, and the volume

China is making the opposite bet, and for now it is the one paying off in the only currency that counts at this stage: robots out the door. In 2025, Chinese companies built somewhere between 80 and 90 percent of the humanoid robots in the world.7

Unitree — the price weapon

The Hangzhou company sells a humanoid, the G1, for about $16,000, and a smaller one, the R1, for under $6,000 — numbers the Americans cannot get near. It is also, almost uniquely in this field, profitable: it turned a net profit on roughly $165 million of revenue in the first nine months of 2025. It became famous overnight when sixteen of its robots danced a folk dance, spinning handkerchiefs, on the national New Year's broadcast watched by hundreds of millions. It is now racing to list on Shanghai's stock market at a valuation of about $6.2 billion.

UBTech — the factory hand

UBTech is the industrial one. Its Walker S2 was the first humanoid that can swap its own battery, which in plain terms means it can work around the clock. It is already on the lines at BYD, Geely, and Foxconn, and at the electric-car maker Zeekr a team of them worked a real coordinated shift, running on a model built from China's own DeepSeek. Walker orders passed $112 million in 2025. The company is listed in Hong Kong, where its shares more than doubled over the year.

AgiBot — the data play

AgiBot was founded in 2023 by a thirty-something ex-Huawei engineer who is internet-famous in China for posting his builds to Bilibili. It is valued somewhere between $10 and $15 billion. Its bet is brains and data as much as bodies: it released one of the largest public datasets of robots doing manual work, because it understands the data problem as clearly as anyone in California does.

The real advantage is the country

China's edge isn't any single company. It's the place. The Pearl River and Yangtze River deltas can turn a robot-arm design into a prototype in days. The same factories that built the world's electric cars make the motors, batteries, and sensors a humanoid needs — at 30 to 50 percent less than the West pays. And the state has its thumb firmly on the scale: a national plan to put 100,000 humanoids to work by 2027, a 2026 program to train them on real factory jobs, the first national safety standards, and billions in financing behind all of it. A Carnegie report put the contrast plainly. While Silicon Valley poured its energy into chatbots, Beijing decided the real prize was AI with a body, and organized a country around making it.8

05 — The moneyA market priced entirely on faith

Put the wager in numbers. More than $13 billion flowed into embodied AI in 2026 alone. The banks, meanwhile, have produced forecasts for the eventual market that are less estimates than declarations of belief. Goldman Sachs sees a $38 billion humanoid market by 2035. Morgan Stanley sees $5 trillion by 2050, with more than a billion robots at work. Citigroup says $7 trillion. They are describing the same technology, in the same few decades, with answers that differ by a hundredfold.

That spread is the honest summary of the whole field. Nobody knows. The robots can clearly do more than they could three years ago, and clearly far less than the valuations assume. The money is a long way ahead of the revenue. The demos are a long way ahead of the deployments. What you are watching is a real technology and a speculative mania braided together so tightly that, right now, no one can fully separate them.

06 — Three betsThe same machine, three different reasons

Strip away the noise and the three players are betting on three different things. Japan is betting on craft, components, and patience — sell the picks, sit out the gold rush. America is betting on intelligence, capital, and story — build the smartest brain, raise the most money, and trust that scale will follow. China is betting on manufacturing, price, and sheer volume — build them cheap, put them to work, and let a million imperfect robots teach you how to make a good one.

The uncomfortable possibility is that all three are partly right, and the prize ends up split: American brains, Chinese bodies, Japanese parts. That is the one arrangement nobody in any of the three capitals actually wants.

07 — Where this standsThe children are working nights

So, fundamentally: where are we? Closer to WABOT than the marketing suggests, and closer to science fiction than the skeptics will admit. The robots can hold down a real shift now, which they genuinely could not do three years ago. They also still drop the box, freeze, and wait for a human to step in. In a sense they are still the year-and-a-half-old children Kato's team measured in 1973 — only now the children work nights, on camera, for an audience.

The honest watchlist is not the next viral clip. It is the boring numbers: units actually shipped, hours worked with no human nearby, cost per robot, and whether anyone pays for the work rather than the promise. Those four lines will tell you who is winning long before the valuations do.

Fujisawa and his team were not confident in 1973. Fifty-three years on, the confidence has arrived — a little early, and unevenly spread across three countries that each think they hold the winning piece. The machines are taking their first real steps. We are about to find out how far they can walk.