China solved the robot body. The brain is still up for grabs

In Shanghai’s Jing’an district, a shop has opened where the sales staff are also the merchandise. In May 2026, Unitree Robotics opened what is being described as Asia’s first embodied intelligence experience store. The space spans just over 100 square meters and sells a range of end-user robotics products directly to consumers, including the company’s G1 and R1 humanoid robots and its Go2 robotic dog. The robots double as sales associates: they guide visitors, shake hands, and pose for short videos. As a piece of retail theater, it works. The store has been drawing steady crowds since it opened. Now, the other half of the picture. The robots in that store cannot fold your laundry, unload a dishwasher without supervision, or put in a reliable eight-hour shift doing anything open-ended in a home. The distance between those two paragraphs is the most important thing happening in robotics in 2026, and it is the part that shipment charts and viral videos keep leaving out. China has, by most measures, won the race to build the robot body. Nobody — in Shanghai or in Silicon Valley — has cracked the brain. I have been testing that claim against a live sample. These months, I am in Palo Alto for a Stanford executive program, and the robot question comes up at nearly every founder dinner — with American founders, Chinese founders based in the Valley, and several who build on both sides of the Pacific. Some are convinced humanoids will be working in our homes within three years; others say 10 years. The pattern in the answers is remarkably consistent, though. The Silicon Valley timelines run longer; the Chinese ones run shorter. Yet almost nobody disagrees on two points: China’s robot hardware and supply chains are, in the words I keep hearing, “unbeatable.” And on the software, training, and models, America remains ahead. Not by a wide margin, but ahead. The body is solved, and it is increasingly Chinese The hardware scoreboard is lopsided. China’s leading humanoid robot maker, Unitree, sold roughly 5,500 units last year, more than any other company in the world. Shanghai-based AgiBot followed closely with just over 5,100 units. By comparison, leading American competitors such as Figure AI and Tesla each sold only around 150 units over the same period. China’s humanoid output is expected to nearly double again in 2026, with Morgan Stanley raising its forecast for shipments from 28,000 to 50,000 units. Entry prices are collapsing too: consumer-facing models now start under $10,000, with Unitree's flagship humanoids listed around $13,500. Consumer distribution is forming around the hardware. UBTech’s consumer brand has launched pre-orders on JD.com for what it describes as the world’s first full-size, hyper-realistic humanoid robot, targeting emotional companionship rather than industrial labor. Chery-incubated Aimoga sells its humanoid directly to ordinary shoppers on JD.com at roughly 285,800 RMB ($42,200). Capital has followed. In the first quarter of 2026 alone, total financing across China’s humanoid robotics sector reached 68.1 billion RMB ($10 billion), according to China Media Group, a figure that already exceeds the full-year financing total recorded in 2025. Unitree’s application to list on Shanghai’s STAR Market has passed the exchange’s review, with plans to raise 4.2 billion RMB ($620 million). The supply chain, the prices, the shops, the IPO pipeline — the body business is real and it is scaling. The half nobody has solved Look closely at the most shared humanoid moments of the past 18 months — the Spring Festival Gala dancers, the boxing matches, the backflips — and almost all of them are feats of motion control. They showcase extraordinary choreography, executed by systems that do not decide anything. Much of what looks like autonomy in commercial demos still involves teleoperation, with a human pilot standing offstage. This is not China’s embarrassment; it is everyone’s. When American startup 1X began delivering its Neo home robot, the company acknowledged that remote human operators would handle many tasks while the AI learns on the job. The reason is a data problem with no easy fix. Language models got smart by reading an internet that already existed. There is no internet of touch, grip strength, friction, and failure for robots to read; every hour of real manipulation data has to be produced, expensively, one robot at a time. Robots have trained in simulated worlds for years. The classic kind are hand-built virtual factories running video-game physics, and they are too small, too few, and too clean, which is why skills learned there tend to break on contact with a real kitchen. The phrase that has taken over the industry this year, “world models,” describes the AI-generated upgrade: systems that learn how the physical world behaves from enormous volumes of video and sensor data, then generate unlimited, varied, realistic practice environments on demand. In effect, the industry is trying to manufacture the missing internet synthetically. For executives, the practical translation matters more than the terminology: every robot “deployment” you read about is also a data-collection operation. The three ways robots learn today — human tele-operators recording demonstrations, world-model simulation, and supervised trial-and-error on the job — all improve with access to real workplaces. This is why robot makers are suddenly eager to place machines into factories, warehouses, and stores on generous terms. The host company’s floor is the classroom. Any leader offered a pilot in 2026 should treat operational data as part of the price and negotiate it: who owns the recordings, whether your workflows end up training a machine your competitor later rents, and what discount your data earns you. Silicon Valley’s uneven year The American side of the race is more turbulent than valuations suggest. Tesla’s Optimus program, the sector’s loudest promise, reportedly paused production over hand-and-arm reliability problems shortly after program lead Milan Kovac departed. Only a few hundred units were built last year, while the long-awaited V3 redesign has been delayed again. Analysts now place consumer availability in 2027 at the earliest. Figure AI is the strongest counterexample. Its Helix vision-language-action model ran an 11-month pilot at BMW’s Spartanburg plant, loading sheet metal on an active production line. The company now assembles about one robot per hour at its own factory. Meanwhile, Nvidia has positioned itself as an arms dealer to everyone, releasing Cosmos 3 in June — a world foundation model that generates synthetic physical environments — with robot makers on both sides of the Pacific building on it. The American bet is the inverse of China’s: solve the brain first, and assume the body can be bought later. China’s world-model scramble Chinese labs spent the past year racing to close the software gap, and the efforts are uneven in revealing ways. AgiBot currently fields the most complete stack. Its Genie Operator-2 model posts the strongest published scores on standard robot-manipulation tests; its Genie Sim 3.0 simulation platform launched at CES as open source; and its training datasets deliberately record failures — missed grasps, collisions, dropped objects — because failure is what teaches a model physics. Notably, AgiBot also builds on Nvidia’s Cosmos; even China’s embodied-AI leader borrows the American world model where it helps. Unitree, the hardware king, is the software chaser. It has open-sourced UnifoLM, its world-model architecture, and in February, it released footage of a G1 assembling motor components in a factory under its first production AI model. Open-sourcing reads two ways: as ecosystem building and as an admission that no single company can generate enough physical data alone. At the same time, the giants have arrived. ByteDance’s Seed lab shipped GR-3, a robot model that learns new tasks from minimal demonstration data, alongside its own robot to run it on. Robot data-collection and training centers are also multiplying across Chinese cities with municipal backing. Here’s an honest scorecard: the body race produced a clear winner, and the brain race has none yet. Chinese teams are closing the research gap faster than Western firms are closing the manufacturing gap — but closing is not closed, and a benchmark victory is not laundry folded. That is also, almost word for word, the consensus of the founders around my dinner table in Silicon Valley. The American software lead is real but narrowing, while the Chinese hardware lead is not narrowing at all. What to watch over the next year For anyone trying to position for the next year, the useful signals are the boring ones. Watch which pilots survive. The robots still employed after six months — BMW’s line workers, AgiBot’s factory changeover crews — matter more than any launch video. Watch whether world-model training shows up as real dexterity: the leap from simulation to physical tasks that systems like Cosmos 3 and Genie Operator-2 are supposed to deliver. Watch how Unitree’s STAR Market debut trades, because it will set the risk appetite for the entire supply chain behind it. Moreover, watch the price floor, which is falling toward family-car territory years before the software is ready — a gap that will produce both delighted early adopters and a wave of disappointment stories. And when a vendor demos a humanoid for your board, open with one question: Was that autonomous or teleoperated? The answer sorts this industry faster than any pitch deck. For retail brands, the practical read is narrower. Robots as traffic drivers already work; the Jing’an store proves the novelty draw is real, and renting that spectacle is cheap relative to the foot traffic it generates. Robots as labor — fitting rooms, shelf work, genuine customer service — remain a promise, and any 2026 deployment should be budgeted as marketing with a data upside, not as headcount replacement. The smartphone era taught Chinese manufacturers a hard lesson: hardware leadership can evaporate when someone else owns the operating system. The robot makers know that history. It explains today’s spending pattern better than any shipment chart. The country that already builds the world’s robot bodies is racing to publish, open-source, and fund embodied AI research because it sees the brain as the only prize left. On that one point, Shanghai and Silicon Valley agree completely. Ashley Dudarenok is a Chinese entrepreneur, author, and media personality, recognized by Thinkers50 as a “guru on fast‑evolving trends in China.” She is the founder of ChoZan, a China‑focused research and digital transformation consultancy.

China solved the robot body. The brain is still up for grabs

How Bottega Veneta weaved China’s forever it bag

Upgrade to Pro

Sectors

Insights

Company

China solved the robot body. The brain is still up for grabs

How Bottega Veneta weaved China’s forever it bag