Topics / Embodied AI
Embodied AI
Embodied AI is where China’s freshly available self-driving talent and its most aggressive capital are meeting a physical problem neither can shortcut. The rounds run ahead of shipments, and the smart money has moved upstream to the scarce input, the training data a robot has to collect one motion at a time. The state has moved there too, with a named MIIT and SASAC campaign that orders central enterprises and ten provinces to open real work stations to the machines rather than paying for the machines directly. We track the bodies, the brains, the policy instrument and the shovel sellers who supply the data, with dates and sources.
The state of play
- Where this stands, as of the most recent development on this page, 20 August 2026. The two newest facts in Chinese embodied AI point opposite ways and both are load-bearing. On 19 August forty-eight central state-owned enterprises appeared together for the first time at the World Robot Conference in Beijing, showing twelve application scenarios between them under the joint MIIT and SASAC training campaign; the next morning, on the same stage, Unitree founder Wang Xingxing said robots have not been rolled out into factories at scale because efficiency and general-purpose capability are still not high enough, and put embodied AI’s ChatGPT moment two to three years away if things move quickly and five to ten if they do not. Unitree fell 16.05 percent to 709.4 yuan that day, its second session as a listed company. Behind those two days sit the year’s three measured quantities, and they disagree with each other in the same way: about 19,100 humanoid units shipped worldwide in the first half of 2026 against about 5,100 a year earlier, more than 97 percent of them from Chinese manufacturers; about 93.5 billion yuan, roughly 13 billion dollars, invested across 322 disclosed deals in the same half; and a best score of 13.93 percent on RoboDojo, the generalization benchmark from Tsinghua, the University of Hong Kong and Berkeley, with every other model under 10. The money and the units are both up several fold. The capability that would justify either is not. Read the shipment and funding totals as private estimates and reported figures rather than audited counts (中华网, citing 中评网, 19 Aug 2026; Caixin, 20 Aug 2026; Smart Analytics Global via 南方都市报, Aug 2026; 36Kr; Huxiu; via CAD, “The Internship” and “The Down Payment”).
- The state demand behind the sector now has a document number, and it is not buying robot labour so much as the data that would make robot labour possible. On 3 June 2026 the general offices of the Ministry of Industry and Information Technology and of SASAC, the regulator that owns China’s central state enterprises, jointly issued 工信厅联科函〔2026〕256号, launching a 2026 real-scene training campaign for humanoid robots and embodied intelligence. It is addressed to the industry departments of ten named regions, Beijing, Tianjin, Shanghai, Jiangsu, Zhejiang, Shandong, Hubei, Hunan, Guangdong and Sichuan, and to designated central enterprises. Each region must nominate at least twenty real work-scene units spanning at least two of the industrial, service and special-operations domains; each central enterprise at least ten drawn from its own industry. The scenario list runs through production, testing and analysis, maintenance, warehousing and logistics, food service and retail, medical and elder care, workplace safety, emergency rescue and disaster mitigation, and the notice says to pick units with clear working conditions and a high degree of standardisation, which is what you select for if the output you want is clean training data rather than a demo. The stated end-2026 goal is more than one hundred high-value scenarios and 万台级规模落地能力, deployment capability at the ten-thousand-unit scale. Work plans were due 30 June 2026 and results summaries are due 30 November 2026 (MIIT and SASAC, 工信厅联科函〔2026〕256号, read in full; via CAD, “The Internship”).
- The first look at what the campaign produced arrived in August, and it arrived alongside its own rebuttal. On 19 August, forty-eight central state-owned enterprises appeared together at the World Robot Conference in Beijing for the first time, showing twelve application scenarios between them, at an event whose exhibitor count was up about 36 percent on the year. The next morning, on the same stage, Unitree founder Wang Xingxing gave his first public remarks since his company listed and said robots have not been rolled out into factories at scale because efficiency and general-purpose capability are still not high enough, putting embodied AI’s ChatGPT moment at two to three years away if things move quickly and five to ten if they do not. Unitree fell 16.05 percent to 709.4 yuan that day. Both statements can be true at once, and the second is a large part of why the first is happening: a technology no one wants to buy yet is being given somewhere to practise (中华网, citing 中评网, 19 Aug 2026; Caixin, 20 Aug 2026; via CAD, “The Internship”).
- The largest single identified buyer is the power grid, and the composition of what it is buying is the tell. State Grid’s internal 2026 embodied-intelligence plan, reported exclusively by Jiemian and carried by 21st Century Business Herald on 29 April 2026, covers roughly 8,500 units for about 6.8 billion yuan of total investment, of which about 5.8 billion is the equipment itself and 1.2 billion is R&D and training: 5,000 quadruped inspection robots, 3,000 dual-arm inspection robots and 500 humanoid live-line-work robots. Inspection is the category even the sceptics concede robots can already do, and live-line work on an energised system is the category where a machine failing is cheaper than a person failing. State Grid’s own arithmetic puts an inspector at about 300,000 yuan a year against a robot dog at 500,000 to 1,000,000 yuan, for a payback of two to three years. Treat this as a reported internal plan rather than a published tender; the figures are Jiemian’s reporting of a document State Grid has not released (界面新闻 via 21世纪经济报道, Apr 2026; via CAD, “The Internship”).
- China placed a record down payment on embodied AI in the first half of 2026. Investors put about 93.5 billion yuan, roughly 13 billion dollars, into the sector across 322 disclosed deals, a total Chinese coverage reports as five times the year-earlier figure, with more than 800 institutions taking part, from marquee venture funds to state capital to the strategic arms of the big manufacturers. The money clustered at the top: the twenty largest raises took close to 60 percent of it, fifteen startups crossed a roughly 1.4-billion-dollar valuation inside those six months, and by layer the robot bodies pulled the most at about 31.5 billion yuan, the embodied brains 14.5 billion, and the chips and actuators underneath the rest. Hold the pace and full-year funding clears 150 billion yuan. Read the 93.5-billion and five-times figures as reported by 36Kr and Chinese finance channels, directional rather than audited (36Kr; Caixin; via CAD, “The Down Payment”).
- The demand behind the money is still mostly the state, and the robot brain is the limit. The buyers are overwhelmingly the national team, provinces and state firms following a policy to gather robot-interaction data, so a maker sells robots to a government-backed data center that then pays again to have them collect the training data the whole field is short of; UBTech, the best-selling maker, booked about 1.4 billion yuan in contracts last year, and a Bilibili teardown put 59 percent of that at local-government data centers. On RoboDojo, a generalization benchmark from Tsinghua, the University of Hong Kong and Berkeley, the best humanoid scored 13.93 percent and every other model came in under 10, the measured distance between the capital and a robot that can cope with a world it was not drilled for (Huxiu; 36Kr; a Bilibili teardown, via CAD, “The Down Payment”).
- Training a general-purpose robot brain is estimated to need on the order of 10 million hours of real-world interaction data, against a current high-quality supply near 500,000 hours; AgiBot’s CTO Peng Zhihui puts the gap at three to five orders of magnitude (36Kr; Sina Finance, Jul 2026).
- Capital has moved upstream to whoever can manufacture that data. LightWheel (光轮智能) raised around 1 billion yuan to become what Chinese coverage calls the world’s first embodied-data unicorn, and PokeBot (破壳机器人), founded in April 2026, raised a Pre-A near 100 million dollars on 3 August 2026 led by Shunwei and Matrix Partners (Pandaily; Caixin).
- Touch is the binding capability gap. At WAIC 2026, 242 embodied-AI firms exhibited, but demos still failed an edge-of-bin pick after about ten tries; NeoteAI and Fudan released a 30,000-hour paired vision-touch dataset and Yimu Tech (一目科技) raised over 1 billion yuan to take a tactile sensor to mass production (甲子光年; 量子位; InfoQ, Jul 2026).
- The rounds run ahead of shipments, and much of the money and talent comes from China’s self-driving shakeout. Moqi Intelligence (墨奇智能), six months old and without a product, disclosed a valuation near 1 billion dollars, and AI2Robotics (智平方) raised roughly 5 billion yuan (36Kr, Jul 2026).
- The physics still caps the machine. UBTech has conceded that full-size humanoids today run only two to four hours on a charge, and as the bodies converge the value is shifting to the brain (Ant’s Lingbo, adapted to more than 20 robot configurations) and the data suppliers; Unitree, which makes bodies, priced its Shanghai STAR Market IPO at 150.80 yuan a share for a post-issue value near 61 billion yuan (about 8.5 billion dollars), with DeepSeek taking a strategic stake, then listed on 19 August 2026 (688836.SH), opening at 1,100 yuan for a market value above 444 billion yuan and closing its first day at 845 yuan, above 340 billion. It gave back 16.05 percent the next day, to 709.4 yuan, on its own founder’s remarks about how far the technology still has to go (TechNode; Caixin, 20 Aug 2026; CAD, “The Cash-Out” and “The Internship”, Aug 2026).
- Set the money against the units and the sector is still small, concentrated and largely selling to itself. Global humanoid shipments in the first half of 2026 came to about 19,100 units, against about 5,100 in the same half of 2025, a rise of roughly 272 percent. Chinese manufacturers accounted for more than 97 percent of those shipments and the Chinese market for more than 85 percent of global demand, so China is overwhelmingly buying these machines from itself. Two makers took three quarters of the volume: AgiBot about 8,400 units for a 44 percent global share, Unitree about 5,900 for 31 percent, and then a cliff, with Galbot near 900, UBTech near 700 and Leju near 600. Industrial and commercial deployments were more than 70 percent of shipments, up from about half a year earlier. These are estimates from Smart Analytics Global, a San Francisco research firm, published 9 August 2026 and reported by Southern Metropolis Daily, which says AgiBot confirmed its own figure; they are a private firm’s modelling rather than audited shipments, and the same firm’s full-year projection of about 60,000 units and roughly 1.6 billion dollars of industry revenue is a forecast, not a result (Smart Analytics Global via 南方都市报, Aug 2026; via CAD, “The Internship”).
Best issues on this topic
- Beijing’s First Robot Data Factory Went Quiet After 17 Months
硬氪, the hardware desk at 36Kr, walked through Beijing’s first humanoid-robot data training centre in Shijingshan and found a 3,000 square metre hall with the zone signs still hanging, the display stands still holding their plants, and the robots gone. It stopped seventeen months after it opened, and the third phase that was to run 108 robots across a thousand real tasks in 10,000 square metres was never built. The structure behind it is the story. A robot maker and a local government co-found a data-collection company, usually at a 1:9 or 2:8 split with the maker putting in the small share; the joint venture buys robot bodies, arms and supporting hardware from the maker; a third-party operator sells the data on to companies training models. So the manufacturer contributes one or two yuan in ten, books an order for all ten paid mostly with public money, and the buyer is a company it partly owns. 硬氪 reports more than a hundred such centres went up across 23 provinces in a year, that one robot company heading for a ChiNext listing had as much as 80 percent of its orders from the model, and that another valued above 20 billion yuan shipped 500 robots this year into data factories it had co-funded, against total expected shipments of 1,000 to 2,000. Figures from Interact Analysis cited in the piece put state-capital participation in these centres at 53.8 percent against 88.4 percent a year earlier, and an official at a provincial capital investment-promotion office told 硬氪 they had just turned down two data-collection proposals and that internally some of these projects look like a new kind of Ponzi scheme. The economics were thin even while the money flowed: an operator told 硬氪 that a collector wearing a rig for a full eight-hour shift yields roughly one hour of usable data once idling, out-of-spec motions and broken arms are subtracted, and collection has since gone bodiless to head-mounted egocentric video and UMI-style hand-worn grippers that need no robot present, yielding five to six usable hours out of eight at a crowdsourced rate that has fallen from 45 yuan an hour to 35. Real-robot data still matters, and a recent Tsinghua experiment found that stripping it out drops task success from 81 percent to 40 percent, but it is the seasoning now rather than the main ingredient. The listed-market version surfaced the same day: ASD, a Zhejiang cookware and small-appliance maker, filed an abnormal-fluctuation notice disclosing that its embodied-intelligence robot business has booked 137,200 yuan of sales to date, about 20,500 dollars, against 1.351 billion yuan of first-half revenue and a 93.4 million yuan net loss, and Sinovatio, filing after two straight limit-ups on AI-security enthusiasm, said its AI application products come to under 2 percent of first-half revenue.
- Spain and Croatia Opened Public Roads to Chinese Robotaxis
Spain’s Dirección General de Tráfico issued the country’s first operating licence for a Level 4 autonomous passenger vehicle on 10 September, and it went jointly to three companies: WeRide, Uber, and AVOMO, the self-driving unit of the Spanish rental group Moove Cars, with the Madrid city government supporting the application. The permit was granted under Spain’s ES-AV framework and clears the three of them to put vehicles on public roads, with commercial service to the public expected before the end of 2026 once the operating and regulatory conditions are met. Madrid is the fourth city in a WeRide and Uber plan the two say will reach 15 cities and tens of thousands of vehicles before 2030, and WeRide says it is now the only company whose products hold autonomous-driving permits in nine countries, naming China, Spain, Switzerland, Belgium, France, Singapore, the United Arab Emirates, Saudi Arabia and the United States. The same morning Pony.ai and the mobility operator Verne said they had begun fully driverless robotaxi passenger testing on public roads in Zagreb, which they call the first anywhere in Europe. Set against both, Anfia chairman Roberto Vavassori asked the EU for an 80 percent tariff on Chinese vehicles and parts above a quota of 8 percent of annual new-car registrations, per Reuters, on top of the combined 18 to 45 percent the 2024 measure already charges Chinese electric vehicles. The proposal covers vehicles and parts and says nothing about software, fleet operations or a permit: a tariff is charged on an object crossing a border, and an operating licence is granted for a service performed inside one. In both deals that day the Chinese company supplied the driving stack while a European operator supplied the cars, the depot, the insurance and the relationship with the regulator, which is a template that travels without a vehicle being imported. Nine permits is nine permissions and no commercial service, Madrid is an expectation for later this year, and Zagreb is a test with passengers in it.
- Unitree Earned Half as Much in Q1 on 68% More Revenue
Unitree booked 50,013,826.62 yuan of net profit attributable to shareholders in the first quarter of 2026, against 95,602,511.45 yuan a year earlier, on 68.49 percent more revenue. The half-year headline runs the other way, a 274 million yuan profit against a 32 million yuan loss, but Chinese listing rules make a company print earnings both with and without non-recurring items, and the 0.67 against 0.75 yuan ratio puts underlying first-half profit down 19.34 percent while revenue grew 48.54 percent. The gap is a 349 million yuan share-based payment charge that hit the first half of 2025 and did not recur. Research spending went from 6.9 to 11.8 percent of revenue and selling expense from 6.0 to 14.2 percent.
- Everyone Else’s Ceiling
Unitree listed on Shanghai’s STAR Market on 19 August at 150.80 yuan a share and opened at 1,100. Four sessions later it closed at 602.80, taking about 200 billion yuan off the peak valuation. Only 7.44 percent of its shares were free to trade on listing day, and roughly 85 percent of that float changed hands in a single session. The mark matters most to companies with no ticker: eight domestic embodied-AI firms were valued at 20 billion yuan or more as of June, and those private marks were set against Unitree’s screen price. Its own filed half-year numbers are good, with revenue up 48.5 percent and a 274.0 million yuan profit against a loss.
- The Internship
Forty-eight central state-owned enterprises appeared together at the World Robot Conference in Beijing on 19 August, showing twelve application scenarios under a joint MIIT and SASAC notice that orders each central enterprise to open at least ten real work stations and each designated province at least twenty, targeting more than a hundred high-value scenarios by the end of 2026. The next morning on the same stage Unitree founder Wang Xingxing said robots are not yet efficient or general enough to be rolled out into factories at scale, and put embodied AI’s ChatGPT moment two to three years away at best. His stock fell 16.05 percent that day.
- The Down Payment
Chinese investors put about 93.5 billion yuan (roughly 13 billion dollars) into embodied AI in the first half of 2026, a figure Chinese coverage reports as five times the year earlier, across 322 deals with more than 800 institutions taking part; the twenty largest raises took close to 60 percent of it and fifteen startups crossed a 1.4-billion-dollar valuation in the half. The asterisk is the demand: the buyers are overwhelmingly the state gathering robot-interaction data, and on the RoboDojo generalization benchmark the best humanoid still scored under 14 percent.
- The Cash-Out
Unitree priced the year’s most anticipated hardware IPO at 150.80 yuan a share, a listing value near 61 billion yuan (about 8.5 billion dollars) and well above the 40 billion the market expected, with DeepSeek taking a strategic stake. That is about 103 times trailing profit for a company that shipped roughly 28,000 robots in 2025, most to research and education, so the price assumes it grows into 100,000 or more a year of paying industrial work.
- The Shovel Sellers
A four-month-old robot startup, PokeBot, raised close to 100 million dollars, but the more telling money went a layer up the supply chain to the data sellers. Training a general-purpose robot brain needs an estimated 10 million hours of interaction data against a supply near 500,000, and LightWheel became what Chinese coverage calls the world’s first embodied-data unicorn.
- The Last Centimeter
Chinese robots can see and reason; the sense they still lack is touch. At WAIC 2026, 242 embodied-AI firms exhibited, yet demos still failed an edge-of-bin pick after about ten tries, even as NeoteAI and Fudan released a 30,000-hour paired vision-touch dataset and the tactile-sensor maker Yimu Tech raised over 1 billion yuan to reach mass production.
- The Transplant
A product-less humanoid startup hit a billion-dollar valuation, built by Huawei’s former self-driving lead. China’s autonomous-driving talent is transplanting into robots.
These link to the full issues on the newsletter. New pieces on this topic go out in the daily first.
Common questions
What is the latest news in China’s embodied AI?
The most recent development covered on this page is dated 20 August 2026. On 19 August forty-eight central state-owned enterprises appeared together for the first time at the World Robot Conference in Beijing, showing twelve application scenarios under the joint MIIT and SASAC real-scene training campaign, at an event whose exhibitor count was up about 36 percent on the year. The next morning, on the same stage, Unitree founder Wang Xingxing gave his first public remarks since his company listed and said robots have not been rolled out into factories at scale because efficiency and general-purpose capability are still not high enough, putting embodied AI’s ChatGPT moment two to three years away if things move quickly and five to ten if they do not; Unitree fell 16.05 percent to 709.4 yuan that day. The two figures that set the year so far are about 19,100 humanoid units shipped worldwide in the first half of 2026, more than 97 percent of them from Chinese manufacturers, and about 93.5 billion yuan of disclosed investment across 322 deals in the same half. Sources are 中华网 citing 中评网 and Caixin for the conference and the remarks, Smart Analytics Global via 南方都市报 for the shipments, and 36Kr and Chinese finance channels for the funding total; the shipment and funding figures are a private firm’s estimates and reported totals, not audited counts.
Is the Chinese government requiring state-owned companies to deploy humanoid robots?
Yes, and there is a specific document. On 3 June 2026 the general offices of MIIT and SASAC jointly issued notice 工信厅联科函〔2026〕256号, launching a 2026 real-scene training campaign for humanoid robots and embodied intelligence. It requires each designated central state enterprise to nominate at least ten real work scenes from its own industry, and each of ten named provinces and municipalities, Beijing, Tianjin, Shanghai, Jiangsu, Zhejiang, Shandong, Hubei, Hunan, Guangdong and Sichuan, to nominate at least twenty. The target is more than one hundred high-value scenarios and deployment capability at the ten-thousand-unit scale by the end of 2026, with work plans due 30 June and results summaries due 30 November. The point of the campaign is training data rather than robot labour: the notice asks for work stations with clear conditions and a high degree of standardisation, which is what produces clean data. The first public showing was on 19 August 2026, when forty-eight central enterprises appeared together at the World Robot Conference with twelve application scenarios.
How many humanoid robots did China ship in 2026?
About 19,100 humanoid robots shipped worldwide in the first half of 2026, up from about 5,100 a year earlier, and Chinese manufacturers accounted for more than 97 percent of them. China was also more than 85 percent of global demand, so the country is mostly buying these machines from itself. AgiBot led with roughly 8,400 units and Unitree followed with roughly 5,900; every other maker was in the hundreds. The figures are estimates from Smart Analytics Global published in August 2026, not audited shipment counts, and the same firm projects about 60,000 units for the full year, which is a forecast.
How much money is going into embodied AI and humanoid robots in China?
A record amount, and it is running ahead of the shipments. In the first half of 2026 Chinese investors put about 93.5 billion yuan, roughly 13 billion dollars, into embodied intelligence across 322 disclosed deals, which Chinese coverage reports as about five times the year-earlier total, with more than 800 institutions taking part. Fifteen startups crossed a roughly 1.4-billion-dollar valuation inside those six months, and at the current pace full-year funding would clear 150 billion yuan. Read the totals as reported by 36Kr and Chinese finance channels, directional rather than audited. The caution is that most of the robots are still bought by the state to gather training data, not by companies with a job for them to do.
Why is data the bottleneck for humanoid robots in China?
Training a general-purpose robot brain is estimated to take on the order of 10 million hours of real-world interaction data, while the high-quality supply today is near 500,000 hours. Unlike text, that data has to be physically collected one motion at a time, so the companies that manufacture it, such as LightWheel, have become some of the most valuable in the sector.
Which Chinese companies lead embodied AI?
The field splits by layer. Unitree, UBTech, AgiBot and Galbot are the most visible body makers; Ant’s Lingbo builds the robot brain, a model adapted to more than 20 robot configurations; and a data-supply layer led by LightWheel and Delta Intelligence sells the training data. Much of the founding talent came from China’s self-driving industry.
Are Chinese humanoid robots ready for real work?
Not yet, on the evidence of 2026. At the World AI Conference, demos ran longer than a year earlier but still failed tasks like picking a package wedged against a bin edge, and full-size humanoids run only two to four hours on a charge. The financing is running ahead of deployment.
What is the “shovel sellers” thesis in embodied AI?
It is the Chinese investing phrase 卖铲人, the people who sell shovels in a gold rush. In embodied AI the scarce shovel is training data, so capital has moved from the robots themselves to the asset-light companies that collect and sell the data, which can resell the same recorded scene to several buyers.