Embodied AI now has the attention of the agency that runs China’s data policy. On Thursday 10 September 2026, Liu Liehong, director of the National Data Administration (NDA), chaired a symposium on embodied intelligence with a state research institute, robot makers, data companies and a robot data training centre. The agency published its account on Sunday 13 September. It said it will promote data standards for embodied intelligence “at an appropriate time”, guide local data bureaus to carry out related work “in an orderly manner”, and actively support companies that invest more in data. Bloomberg reported the move as a standards push.
The qualifier is the story. The NDA’s Chinese text says it will push the standards 适时, which means “at an appropriate time” or “in due course”. The statement gives no date, names no document and opens no consultation. The standards work is also further along than a “plans” headline suggests. A national guidance document on real-world embodied AI data, GB/Z 218.1-2026, is listed as published. An industry standard on dataset quality takes effect on 1 November, and a benchmark-testing standard has applied since 1 June. What is new is that the data regulator, not only the industry ministry, has claimed a role.
This article sets out what the NDA said and who attended, lists the embodied AI standards already on the books, and explains why robot data has become the bottleneck. It then counts the training grounds and datasets behind the policy, compares China’s approach with Europe’s, and lists the signals companies should watch. For background, see our coverage of the startup selling egocentric robot training data, the humanoid robot that learned spin kicks from human motion data, and the AI agents that build virtual playgrounds for robots.
Table of contents
- What China’s Data Regulator Said About Embodied AI
- The Embodied AI Standards China Already Has
- Why Embodied AI Needs a Different Kind of Data
- Training Grounds: Where China Collects Embodied AI Data
- The NDA’s Wider Dataset Programme for Embodied AI
- What Industry Asked for Before the Embodied AI Symposium
- How China’s Embodied AI Data Policy Compares With Europe
- What Embodied AI Companies Should Watch Next
- Embodied AI Data Standards FAQ
- References
What China's Data Regulator Said About Embodied AI
The NDA’s account runs to three short paragraphs. Liu Liehong, who also leads the agency’s Communist Party group, chaired the meeting. Xia Bing, a deputy director and member of the party group, attended, as did the principal officials of the NDA’s Department of Digital Technology and Infrastructure Construction. Participants gave views and suggestions on a single theme: “data empowering the development of embodied intelligence”.
Chinese state media and the NDA use 具身智能, “embodied intelligence”. English-language coverage, including Bloomberg’s, mostly says embodied AI. Both describe the same thing: artificial intelligence that perceives, decides and acts through a physical body, such as a humanoid robot, a robot arm or an autonomous machine on a factory floor.
An industry the NDA calls “data-driven”
The statement opens with a diagnosis rather than a policy. “The embodied intelligence industry is booming and shows data-driven characteristics,” it says, “with demand for high-quality, diverse and large-scale data continuing to grow.” That framing places embodied AI inside the NDA’s remit. The agency does not make robots or approve them, but it does run China’s programme for building datasets, and it now describes robot development as a data problem.
Six commitments, and only one is about standards
The policy sentence that follows contains six separate commitments. The table lists them in the order the NDA wrote them, with the original wording and what each one does and does not bind the agency to.
| Commitment | NDA wording | What it commits to |
|---|---|---|
| Stay problem-oriented and track industry trends | 坚持问题导向、把握产业发展趋势 | A working method, with no deliverable |
| Strengthen planning and development guidance | 加强科学布局和发展引导 | Guidance on where capacity is built; no document named |
| Promote embodied intelligence data standards | 适时推动具身智能数据标准建设 | Standards “at an appropriate time”; no date or title |
| Guide local data bureaus | 指导地方数据系统有序开展相关工作 | Local work “in an orderly manner” |
| Promote healthy industry development | 促进行业健康发展 | A policy aim, with no measure attached |
| Support companies that invest more in data | 积极支持企业加大数据投入 | “Active” support; no instrument or budget named |
Read that way, the headline commitment is the most hedged of the six. The strongest verb in the sentence, “actively support”, attaches to company spending on data, not to the standards that made the headlines about embodied AI.
Who was in the room
The NDA named nine participating organisations. Two of them already appear on the drafting list of GB/Z 218.1-2026, the national guidance document on real-world embodied intelligence data: the Beijing Humanoid Robot Innovation Center and Guanglun Intelligence, according to the national standards information platform run by China’s market regulator. The regulator was hearing from some of the organisations that had already written the first embodied AI data rules.
| Participant, as named by the NDA | Chinese name | Type | GB/Z 218.1-2026 drafter? |
|---|---|---|---|
| Institute of Automation, Chinese Academy of Sciences | 中国科学院自动化研究所 | State research institute | Not listed |
| Beijing Institute for General Artificial Intelligence | 北京通用人工智能研究院 | Research institute | Not listed |
| Guanglun Intelligence | 光轮智能 | Company | Yes |
| JD Group | 京东集团 | Company | Not listed |
| Shendu Jizhi | 深度机智 | Company | Not listed |
| Shijingshan Humanoid Robot Data Training Centre | 石景山人形机器人数据训练中心 | Data training centre | Not listed |
| Xinghaitu | 星海图 | Company | Not listed |
| Zibianliang | 自变量 | Company | Not listed |
| Beijing Humanoid Robot Innovation Center | 北京人形机器人创新中心 | Innovation centre | Yes |
The Beijing Humanoid Robot Innovation Center, also known as X-Humanoid, built the Tiangong Ultra robot that won the 100-metre large-group final in 8.64 seconds at the second World Humanoid Robot Games, Global Times reported. Its presence alongside a research institute, a data training centre and robot companies shows the NDA consulting across the embodied AI data chain, from collection to use.
Why the data regulator, and why now
China’s humanoid robot and embodied intelligence standards have so far come mainly from the Ministry of Industry and Information Technology (MIIT), which set up a dedicated standardisation technical committee in December 2025. The NDA, which is administered by the National Development and Reform Commission (NDRC), controls different levers: the national programme for high-quality datasets, its labelling push, and data standardisation. Its statement adds data governance to a field the industry ministry had mostly treated as a question of hardware and systems.
The timing follows a busy fortnight. On 28 August the NDRC promised a “high-quality real-machine data collection system” for embodied intelligence. About ten days before the NDA’s statement, according to MLex as reported by The Next Web, seven companies met the agency and asked for public data infrastructure and common data standards. The symposium reads as the regulator’s formal answer to that request.
The Embodied AI Standards China Already Has
“Plans standards push” can read as if China were starting from nothing. It is not. At least four published documents already govern parts of embodied AI in China, and a longer list of projects sits in the national standards pipeline.
February: a standard system for the whole industry
On 28 February 2026, the MIIT technical committee for humanoid robots and embodied intelligence released the “Humanoid Robot and Embodied Intelligence Standard System (2026 Edition)” at its first annual meeting in Beijing, Xinhua reported. More than 120 research institutes, companies and user organisations drafted it. It has six parts: basic commonality, brain-like and intelligent computing, limbs and components, complete machines and systems, applications, and safety and ethics.
Data sits in the second part. Xinhua said the brain-like and intelligent computing standards regulate the full data lifecycle and the full chain of model training, inference and deployment. The system is a map of standards still to be written rather than a rulebook, which is why the documents that followed matter more to engineers building embodied AI products.
June: a benchmark method already in force
YD/T 6770-2026, an MIIT-approved industry standard on benchmark testing methods for embodied intelligence, took effect on 1 June 2026, China Daily reported. It sets a common framework for testing in simulated and real environments, covering environment set-up, task libraries, the test process and how metrics are calculated. Wei Kai, head of the AI institute at the China Academy of Information and Communications Technology (CAICT), called a credible evaluation system “the bridge and yardstick” between research and large-scale industrial use.
A separate humanoid robot lifecycle management standard was released around the same time. It requires every humanoid robot to carry a unique, unchangeable identity code from the factory to scrapping, which China Daily summarised as “one machine, one code”.
July: a quality standard for embodied AI datasets
On 24 July, the MIIT approved YD/T 6771-2026, which sets quality requirements and evaluation methods for embodied intelligence datasets, according to the Digital China Summit’s news site. It takes effect on 1 November 2026. CAICT drafted it with more than 40 organisations. It covers data production procedures, organisational safeguards and quality indicators across eight dimensions: completeness, consistency, diversity, authenticity, usability, practicality, scalability and security.
The same notice explains the problem it targets. Companies are releasing open datasets and building training grounds, but whether that data is usable, and good to use, is judged differently across the industry. The result, it says, is that dataset quality is hard to control and different datasets are hard to combine, a problem any embodied AI team that has merged two public datasets will recognise.
A national guidance document on real-world data
The most directly relevant document is GB/Z 218.1-2026, “Artificial intelligence — Specifications for embodied intelligence data quality — Part 1: Real data”. The national standards information platform lists it as published. It is a guidance technical document, which is what the “Z” in its code signals, registered on 20 June 2025 with a 12-month project cycle and set to take effect on publication.
Its drafters include the China Electronics Standardization Institute, Huawei Cloud, UBTech, AgiBot, Xiaomi’s robotics company, Galbot, Ant Group and the Beijing Academy of Artificial Intelligence. Arabian Post, in an article syndicated on MSN, dated national standards on the quality of real-world embodied intelligence data and on data-generation platforms to 27 August. The platform page does not show a publication date, so we have not confirmed that day.
Projects still in the pipeline
The same platform page lists related standards projects that show where the embodied AI rules are heading. Two carry a different committee code, 907, from the rest. Their titles match the NDA-supervised projects that Arabian Post attributed to the National Data Standardisation Technical Committee: data sources for high-quality embodied intelligence datasets, and simulated synthetic data.
| Document or project | Issuer or committee | Status | Scope |
|---|---|---|---|
| Standard System (2026 Edition) | MIIT technical committee | Released 28 February 2026 | Six-part framework, including the data lifecycle |
| YD/T 6770-2026 | MIIT industry standard | In force since 1 June 2026 | Benchmark testing in simulated and real environments |
| YD/T 6771-2026 | MIIT industry standard | Approved 24 July; in force 1 November 2026 | Dataset quality across eight dimensions |
| GB/Z 218.1-2026 | TC28/SC42, under the Standardization Administration | Listed as published | Quality of real-world embodied intelligence data |
| 20263053-Z-469 | Code 469 project | Listed as a related plan | Data generation technical requirements |
| 20255546-Z-469 | Code 469 project | Listed as a related plan | Construction guide for data training grounds |
| 20262582-Z-907 | Code 907 project | Listed as a related plan | Data sources and constituent elements of high-quality datasets |
| 20262583-Z-907 | Code 907 project | Listed as a related plan | Simulated synthetic data generation and processing |
| 20255547-Z-469 | Code 469 project | Listed as a related plan | Evaluation guide |
China also plans more than 30 national standards in the data field in 2026, including a batch in frontier areas such as intelligent agents and embodied intelligence, CCTV reported from the NDA, as carried by IT Home. An official of the national data standardisation committee said China developed 48 national standards and technical documents in the data field in 2025.
The NDA statement did not say whether its promised embodied AI data standards would extend these projects, replace them or sit alongside them. That is the first question for anyone who has already built a data pipeline around the telecom industry standards.
Why Embodied AI Needs a Different Kind of Data
A chatbot can learn language from text that already exists online. A robot that sorts parcels or loads a machine cannot learn the task that way, because the training data it needs has to be recorded from a body doing the work. That is why the NDA’s statement stresses “high-quality, diverse and large-scale” data rather than more data of any kind.
What a robot’s training data contains
In June, Liu Liehong said embodied intelligence depends on “visual, tactile, audio and other high-quality, multimodal training data” to adapt and carry out tasks in real environments, according to a China Securities Journal article carried by Xinhua. He called high-quality datasets the foundation of the “perception, decision and execution” loop that every embodied AI system runs.
Taiwan’s United Daily News, citing Zhejiang Daily’s Chao News, put the contrast plainly. A large language model can draw on text and images accumulated across the internet. Humanoid robots need motion, vision and force information captured during real operation, and collecting that on real machines costs more.
The hours gap, in three sources’ numbers
Every estimate of the shortfall is large, but the sources do not agree on the figures. Gasgoo, reporting from the World Artificial Intelligence Conference in July, said global high-quality embodied interaction data totalled about 500,000 hours at the start of 2026, against an entry threshold of 10 million hours for training general embodied models. The Next Web cited a CAICT estimate that embodied AI foundation models need roughly 10 million hours, while between 100,000 and 1 million hours of high-quality data exist worldwide.
The notice on the CAICT dataset quality standard set a lower bar for individual companies: in 2026, it said, a production capacity of one million hours of data is seen as the hard threshold for developing a model. Gasgoo described its own figures as “a gap exceeding 99%”. On its numbers, 500,000 hours is 5% of 10 million, which makes the gap 95%. That is still enormous, but it is not 99%.
The chart puts the published figures on one scale, where a full bar equals 10 million hours of data.
Four ways to collect embodied AI data
Companies are attacking the shortage by several routes at once. Chao News listed four: teleoperation of real machines, “body-free” collection, motion capture and simulation. Gasgoo counted more than 20 companies showing data-collection kits at the World Artificial Intelligence Conference, from master arms with force feedback to sensor gloves and camera headbands. Each route has a trade-off that a common embodied AI data standard would have to handle.
| Route | How the data is made | Strength | Weakness |
|---|---|---|---|
| Teleoperation of real robots | An operator drives a robot in real time, often through a master arm with force feedback | The robot’s own body records the task | Real-machine collection is costly and needs a human for every recorded hour |
| “Body-free” collection | People wear gloves, headbands or other sensors while doing the task themselves | Scales with people rather than robots | Human hands and bodies differ from the robot’s |
| Motion capture | Cameras or suits record human movement for robots to imitate | Captures whole-body motion precisely | Movement must be mapped onto the robot’s joints |
| Simulation and synthesis | Virtual physics environments generate robot operation data in batches | Cheap and fast to scale | Gaps remain between simulated and real environments |
Simulation is also where much reinforcement learning for robots takes place, because a virtual robot can repeat a task millions of times without breaking anything. The NDA’s June dataset plan endorses that route, telling the industry to “actively apply simulation and synthesis technologies to expand data supply”. The NDRC’s August measures lean the other way, towards data recorded on real machines.
Training Grounds: Where China Collects Embodied AI Data
China’s main answer to the hours gap has been physical: purpose-built sites where robots and operators generate data on real tasks. The NDA’s promise to guide local data bureaus matters because these sites are, for the most part, local projects.
More than 70 sites, with 46 more coming
More than 70 embodied AI training grounds had been built and put into operation across China by the end of June, according to a 2026 CAICT report cited by TechNode and IT Home. Another 46 were under construction or being planned. The sites span more than half of China’s provincial-level regions, and industrial manufacturing appears as an application at 86% of them. The Yangtze River Delta, the Beijing-Tianjin-Hebei region and the Pearl River Delta are the main clusters.
On those figures the pipeline would take China to at least 116 sites, and more than four in five of the existing grounds already serve factories. The chart uses a scale where a full bar equals 120 sites.
The NDRC’s August line: real machines, and no “blind rush”
At a press conference on Friday 28 August, NDRC spokesperson Li Chao said robotics development “must be tailored to local conditions and avoid blind imitation and a rushed, herd-like rush”, Global Times reported. The NDRC would next “build a high-quality real-machine data collection system” to raise both the quality and the scale of embodied intelligence data, and address the “data hunger” that constrains training.
Li added that the commission would expand the library of embodied training scenarios, encourage companies in relevant industries to open real-world environments, and support reliability and safety testing. The sectors he named for faster deployment were manufacturing, healthcare, consumer markets, services and public safety.
Why “in an orderly manner” points at local bureaus
Read next to the NDRC warning, the NDA’s phrase about guiding local data systems “in an orderly manner” looks like the same concern from the data side. Seventy-plus local sites recording in their own formats would produce exactly the problem the July dataset standard describes: datasets that are hard to judge and hard to combine. A shared embodied AI data standard is how a national regulator turns dozens of local collections into one usable national resource.
That is our reading, not the NDA’s explanation. The statement does not mention training grounds by name, and it does not say which local bureaus will receive guidance first.
The NDA's Wider Dataset Programme for Embodied AI
The symposium sits inside a larger data programme that the NDA has run all year. Its numbers show how quickly China’s stock of training datasets is growing, and why the agency now wants a say in embodied AI.
The June plan names embodied intelligence
On 3 June the NDA issued its implementation plan for building high-quality industry datasets, document 国数科基〔2026〕25号, after publishing a draft for comment on 15 April. Liu Liehong had described it as six actions: strengthening the base and expanding capacity, a labelling push, raising quality and efficiency, application, management services, and releasing value.
The plan lists embodied intelligence among its innovation fields, alongside the low-altitude economy, intelligent driving, the smart ocean and biomanufacturing. For embodied intelligence it calls for faster construction of real-machine interaction datasets covering “physical interaction, environmental perception and motion control” in key scenarios. It also calls for national standards on dataset formats, types, annotation and quality evaluation, and sets the end of 2028 as the date for a batch of industry datasets validated in real applications.
Counting China’s datasets
The NDA said more than 126,000 high-quality datasets had been built nationwide by August, with a combined volume above 1,815 petabytes, IT Home reported from CCTV coverage of the data industry expo in Guiyang. It said volume had grown more than 89% since the end of the first quarter, when the count stood above 116,000 datasets and 960 petabytes, according to Xinhua.
The count and the volume tell different stories. The number of datasets rose by about 8.6%, while total volume nearly doubled. Dividing volume by count gives an average of about 8.5 terabytes per dataset at the end of March and about 14.8 terabytes in August, using 1 petabyte = 1,024 terabytes. The roughly 10,000 datasets added in between averaged about 88 terabytes each. Both official figures are floors (“more than”), so treat the averages as indicative.
| Measure | End of Q1 2026 | August 2026 | Change |
|---|---|---|---|
| High-quality datasets built | 116,000+ | 126,000+ | About +8.6% |
| Total volume | 960+ PB | 1,815+ PB | About +89% |
| Average size per dataset | About 8.5 TB | About 14.8 TB | About +74% |
| Average size of datasets added since Q1 | Not applicable | About 88 TB | Derived from the two totals |
Larger datasets fit the direction the NDA is pushing. Arabian Post quoted China’s National Data Development Research Institute as saying embodied intelligence and world models are driving demand for three-dimensional, video and other multimodal data. Robot recordings, with synchronised cameras, joint positions and force readings, are large by nature.
A national platform and a dataset register
The NDA’s national dataset management service platform began trial operation on 29 April. By 31 May it had certified 516 organisations and published 1,350 datasets, Xinhua reported, and by late August it had published more than 1,700, according to IT Home. A register of that kind is where common embodied AI data standards would show up first, because a shared format is what lets one company’s robot recordings be listed, found and reused by another.
What Industry Asked for Before the Embodied AI Symposium
The NDA did not arrive at embodied AI unprompted. Companies and officials had been arguing for months that data, not models, was the constraint.
Seven companies and a request for shared infrastructure
About ten days before the NDA’s statement, the agency met seven companies that asked for public data infrastructure for embodied AI and common data standards, according to MLex as reported by The Next Web. The symposium’s promise to promote standards answers the second request. It says nothing concrete about the first. Public data infrastructure, meaning shared collection sites or shared datasets, does not appear in the NDA’s text.
Liu Liehong’s “data engineering” argument
The regulator’s position had been building since spring. At the 2026 World Intelligence Industry Expo, Liu said the country should “drive the development of embodied intelligence with complete data engineering” and carry out systematic practice in depth, according to the China Securities Journal. He made the point in the same speech that previewed the June dataset plan.
A year earlier, in June 2025, the NDA published an account of Liu researching high-quality dataset construction for the embodied intelligence industry. The agency has since featured embodied datasets among its high-quality dataset case studies, including a humanoid robot manipulation dataset and a real-machine dataset described as million-scale.
The quieter worry: quantity without quality
Not every signal from Beijing encourages more building. The NDRC’s warning against a “herd-like rush” came two weeks before the symposium, and the dataset quality standard exists because quantity has outrun quality. The notice on that standard described embodied AI data as facing a structural conflict of being “small in quantity and low in quality”. A data standard is one way to slow the rush without stopping it.
How China's Embodied AI Data Policy Compares With Europe
Europe regulates machine data too, but from the other end. Its rules decide who may access the data a machine generates. China’s programme is about producing enough of that data to train embodied AI at all.
The EU Data Act and connected products
The EU Data Act has applied since 12 September 2025. It gives users of connected products, including industrial machinery, rights to access the data those products generate, and it limits the use of that data to build competing products, The Next Web noted. Its design obligation, which requires connected products to make their data accessible, applies to products placed on the EU market after 12 September 2026. That obligation began two days after the NDA’s symposium.
A strategy without data labs
The European Commission’s Data Union Strategy, published on 19 November 2025, lists scaling up access to data for AI as its first priority. Ten months on, most of its actions, including the promised data labs, have not been carried out, The Next Web reported. The nearest European equivalent to a Chinese training ground is a company: NEURA Robotics is building robot “gyms”, ten planned with five meant to run by the end of this year, split between Europe, the United States and China.
| Dimension | China | European Union |
|---|---|---|
| Bodies involved in robot data | NDA for datasets, MIIT for industry standards, NDRC for data collection | European Commission, through the Data Act |
| Main policy question | How to produce enough high-quality data | Who may access data a product generates |
| Standards position | Industry standards and a national guidance document published; NDA data standards promised “in due course” | Data Act access rules in force; data supply largely left to companies |
| Collection infrastructure | 70+ training grounds operating, 46 more planned or under way | Company-led, such as NEURA’s planned robot gyms |
| Next dated milestone | YD/T 6771-2026 takes effect on 1 November 2026 | Design obligation for connected products from 12 September 2026 |
Neither model is complete. China is building the supply faster than it is settling the rules for sharing it, while Europe has settled access rules for data it is not yet producing at the scale embodied AI needs.
What Embodied AI Companies Should Watch Next
The NDA statement is a signal, not a rule. These are the developments that would turn it into something companies must comply with, or can build on.
A document with a number
Watch for a draft for public comment, a formal document number like the June plan’s, or a new embodied intelligence project on the national standards platform under the data committee’s code. Until one appears, “at an appropriate time” leaves the timetable open.
How the new rules fit the telecom standards
YD/T 6771-2026 takes effect on 1 November. If the NDA’s data standards cover the same ground, such as formats, annotation and quality evaluation, companies could face two overlapping sets of expectations. If they build on it, CAICT’s eight quality dimensions become the baseline for every embodied AI dataset in China.
Who writes the rules
The drafting list for GB/Z 218.1-2026 includes Huawei Cloud, UBTech, AgiBot, Xiaomi’s robotics company and Galbot. Companies that draft a data standard shape which fields, sensors and formats count as “quality”. Smaller firms, and foreign firms selling into China, will want to see whether the NDA’s process is equally open to them.
Whether demand keeps pace with supply
Data standards do not create buyers. China Daily reported about 17,000 humanoid robot shipments in 2025, from more than 140 makers and over 330 products, an average of about 52 units per product. The Next Web cited a survey this year in which only 23% of Chinese enterprises were satisfied with the robots on offer. A training ground collects hours of data; it does not collect customers.
| Signal to watch | Where it would appear | Why it matters |
|---|---|---|
| Draft NDA data standard for comment | NDA website | Starts the clock on “at an appropriate time” |
| New embodied intelligence project under code 907 | National standards information platform | Shows the data committee’s pipeline in use |
| YD/T 6771-2026 takes effect | 1 November 2026 | First industry-wide quality yardstick for embodied AI datasets |
| Guidance to local data bureaus | Provincial and city data authorities | Tests whether “orderly” means shared formats across training grounds |
| Details of the real-machine collection system | NDRC | Decides how much data comes from real robots rather than simulation |
Embodied AI Data Standards FAQ
What did China’s data regulator announce about embodied AI?
After a symposium on 10 September 2026 chaired by its director, Liu Liehong, the National Data Administration said it will promote data standards for embodied intelligence “at an appropriate time”. It also said it will strengthen planning guidance, guide local data bureaus, and actively support companies that invest more in data. The statement was published on 13 September and named no date or document.
Does China already have embodied AI standards?
Yes. The MIIT released a six-part standard system in February 2026. Industry standard YD/T 6770-2026 on benchmark testing took effect on 1 June, and YD/T 6771-2026 on dataset quality takes effect on 1 November. National guidance document GB/Z 218.1-2026 on real-world embodied intelligence data is listed as published, and more projects are in the pipeline.
Why is data the bottleneck for embodied AI?
Robots need multimodal recordings of real tasks, including vision, touch, force and motion, which cannot be scraped from the internet. Published estimates put the need for general embodied models at around 10 million hours, against 100,000 to 1 million hours of high-quality data worldwide.
How many embodied AI training grounds does China have?
More than 70 were operating by the end of June 2026, with 46 more under construction or planned, according to a CAICT report. Industrial manufacturing appears as an application at 86% of them.
When will the NDA publish its embodied AI data standards?
It has not said. The statement uses the phrase 适时, “at an appropriate time”. The first concrete sign would be a draft for public comment or a new project on the national standards information platform.
How does China’s approach to embodied AI data differ from the EU’s?
China is building data supply through training grounds, a national dataset programme and standards on quality. The EU’s Data Act focuses on who may access data that connected products generate, with a design obligation for new connected products from 12 September 2026.
References
国家数据局局长刘烈宏主持召开具身智能座谈会 (Jiemian)
加速發展機器人 陸提出適時推動具身智能數據標準建設 (United Daily News)
China’s data regulator plans standards for embodied AI, ten days after industry asked (The Next Web)
人工智能 具身智能数据质量规范 第1部分：真实数据 (National Standards Information Public Service Platform)
关于推进行业高质量数据集建设行动的实施方案 (National Data Administration)
首个具身智能数据集质量标准发布 (Digital China Summit)
加速产业规模化 具身智能标准持续完善 (China Daily)
China has more than 70 operational embodied-AI training grounds, report says (TechNode)
国家数据局：全国已建成高质量数据集超12.6万个，总体量超1815PB (IT Home)
国家数据局：以完善的数据工程驱动具身智能发展 (Xinhua)
Embodied Intelligence Data, How Long Until Supply Is Sufficient? (Gasgoo)
China advances standards for embodied AI data (Arabian Post)
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.