In the first half of the year, 93.5 billion yuan was raised, and the gap is still over 99%; people who teach robots to work are in kitchens in Jilin and rental houses in Shanxi.
In June this year, a young man named Xiaoyuan saw an online recruitment post and became a part-time data collector. The equipment not only has a heavier head-mounted camera, but also has mechanical grippers on both hands that simulate the grasping action of a robot. He collects for about 3 hours every day. He cannot shake his head greatly, otherwise the collection will be ineffective. His arms remain suspended in the air for a long time. After repeating the same action dozens or even hundreds of times, his arms become sore and numb, and his hands feel sore. “This hard work indeed exceeded my expectations.” Xiaoyuan said.
In Baicheng, Jilin, 62-year-old Wang Xiaohua signed up as a collector after retirement. A lightweight headband was handed to her hand, and cooking in the kitchen, wiping the table and folding clothes, washing pots and dishes by the sink, all became collection scenes. She joked with her old sisters that she was “paid to do housework.” Ms. Li, a mother of two children, collects about 6 hours of work on weekdays and rests on weekends. She can earn more than 4,000 yuan a month.
They have an official professional name – “Embodied Intelligent Data Collector”, or more accurately: people who serve as teachers for robots. In 2026, when Yushu Technology is listed on the Science and Technology Innovation Board and 93.5 billion yuan in 322 financings in the first half of the year has poured into embodied intelligence, the most scarce thing in this industry is not motors or joints, but the real action data “made” by people like Xiaoyuan and Wang Xiaohua frame by frame.
Conceptual diagram of data collector and robot shadow, image generated by AI
Hardware is no longer the bottleneck, the brain is
2026 is regarded by the industry as the “first year of mass production” of embodied intelligence. Yushu Technology was listed on the Shanghai Stock Exchange’s Science and Technology Innovation Board on August 19, becoming the “first A-share humanoid robot stock”; Zhiyuan Robot launched its listing in Hong Kong in July, but progress has slowed down significantly. Public reports indicate that the possibility of completing an IPO within the year is decreasing.
A consensus is forming: the “cerebellum” and hardware of robots have advanced extremely fast, and walking, running, grabbing, and carrying are already quite efficient, but the “brain” – the ability to understand the complex world and make decisions autonomously – is significantly slower.
Market money gives the answer. The China Academy of Information and Communications Technology’s “Embodied Intelligence Development Report (2026)” quoted IT Orange statistics: In the first half of 2026, there were 322 financings in the domestic embodied intelligence field, amounting to approximately 93.5 billion yuan, a year-on-year increase of 5 times. Looking at the segmented flow direction, money is clearly migrating to the “brain” and data infrastructure.
500,000 hours vs. 10 million hours
According to estimates by the China Academy of Information and Communications Technology and other institutions, as of the beginning of 2026, there will be approximately 500,000 hours of real physical interaction data that can be used for model training in China; training an embodied model with universal generalization capabilities is generally believed to require at least tens of millions of hours, and high-end intelligence will require hundreds of millions of hours, with a gap of more than 99%. The reference given by Gao Shaolong, chairman of Jiyuan Zhihang, is even more extreme: the industry consensus is that 10 billion hours are needed, but only 500,000 hours are available for domestic compliance.
There is a difference of three orders of magnitude between the two figures – the calculations of the required data magnitude by different institutions and practitioners are so different, which itself shows that the industry has not yet unified on the line of “enough”. But by any measure, the gap is too big to be filled with existing supplies.
Unlike large language models that “eat” ready-made text on the Internet, embodied intelligence “eats” actions in the physical world. The material, thread, friction, and force control behind the bottle cap have never been stored on the Internet. They must be “made” one by one in the real scene, recorded, and clearly marked. The market price of real-machine remote operation data is about 500 to 1,000 yuan/hour. Although simulated synthetic data can be sold for 20 to 50 yuan/hour, there is always a migration gap between “virtual and reality”.
500,000 hours did not come from the laboratory. Taking it apart, it can be seen that it is Wang Xiaohua who is cooking and wiping the table in Jilin’s kitchen, it is Ms. Li who is folding clothes and laying sheets at home, and it is Xiaoyuan who is wearing a gripper and repeating the same action dozens or hundreds of times until his arms are numb. In order for a robot to work like a human, someone must first do the work like a robot.
Yao Maoqing’s sentence is cruel in the second half
Yao Maoqing, partner and senior vice president of Zhiyuan, chairman and CEO of MiFeng Technology, said something at the WAIC “Zhiqi Embodied Forum” that is worth quoting: “The model determines the starting point, and the data defines the ending.”
The cruelty lies in the second half of the sentence. Models can rely on papers, open source and talent to catch up, but data cannot. Yao Maoqing has added a judgment many times: in order for physical AI to scale, it must break through three “walls”, the first of which is the data wall. MiFeng Technology completed three rounds of financing six months after its establishment and positioned itself as a “one-stop physical AI data service platform.” For a person with a background in modeling to bet on data infrastructure for the next stage is a signal in itself.
The path of Daimeng robot is more specific. In August 2026, this Shenzhen-based haptic company, incubated by a team at the Hong Kong University of Science and Technology, completed hundreds of millions of yuan in strategic financing, led by Ant Group. On April 15, Daimon teamed up with Google DeepMind, China Mobile, Tsinghua University and dozens of other institutions to release the Daimon-Infinity data set. The first batch of 1,000 hours of real-world data was launched on the Alibaba Magic Community. The open source scale will reach 10,000 hours, covering 16 major industries, more than 80 real-life scenarios, and more than 2,000 types of tasks. According to Daimon and media reports, the cumulative downloads since the launch have exceeded 8.5 million times. This number illustrates better than any funding news: the industry’s hunger for high-quality physical interaction data is real.
In Chenzhou, Hunan, the “Embodied Data Collection 5S Store” launched by Dai Meng and China Mobile launched 1,000 sets of equipment in the first phase. Citizens received the equipment at the mobile business hall, went home after training, and completed collection during daily housework. At full production, it can produce 1 million hours of real scene data every year.
Selling is not the same as actually working
The influx of capital does not mean a closed business loop.
Yushu Technology’s prospectus and reply to the Science and Technology Innovation Board’s inquiry letter make the gap clear: in the first three quarters of 2025, scientific research and education accounted for 73.60% of its humanoid robot revenue, commercial consumption accounted for 17.39%, and industry applications accounted for only 9.01%; among less than 10% of industry applications, the revenue scale that truly points to clear industrial scenarios such as smart manufacturing and smart inspections is even smaller. The income structure of “the number one humanoid robot stock” is still highly dependent on the scientific research and education market.
The judgment of the China Academy of Information and Communications Technology report is more direct than any media comment: “Shipment does not mean employment, a considerable proportion of equipment is powered on for less than 10 days on average per month, and the gold content of orders is low.” The report also pointed out that the price of the complete machine only accounts for 40% to 60% of the full life cycle cost, and the daily rent for performance rentals has dropped by 80% to 90%. The robot was “sold” but sat vacant between the showroom and the set.
Galaxy Universal is another example of observing this paradox. Since its establishment, it has raised a total of approximately 7 billion yuan, and its valuation has exceeded 20 billion yuan. It has been endorsed by major funds in three phases. Data from third-party organization Counterpoint shows that it shipped more than 1,100 units in the first half of 2026, ranking third in the world. But its customer structure highly overlaps with its shareholder list: CATL is both an investor and the largest deployer of Galbot S1; Meituan is a shareholder, and the robots have entered Meituan pharmacies; BAIC and SAIC are shareholders, and Galaxy robots are running on the production line.
Regarding this “shareholders are customers” structure, Gao Chengyuan, deputy secretary-general of the Guangdong Provincial Social Policy Research Association, said that these orders are real and the income has been audited and confirmed. Data shortage and difficulty in implementation are two sides of the same problem.
Who is “selling teaching materials”
Data gaps have given rise to new tracks. According to statistics from the third-party data platform Interact Analysis, as of the end of April 2026, at least 90 humanoid robot data collection and training centers are in use or planned and constructed in China, and at least 64 of them have been put into operation. However, there are still doubts whether “purely selling data” is a profitable business.
The pressure is transmitted layer by layer along the chain and finally falls on the collector. Gao Shaolong revealed that Jiyuan Zhihang’s effective hourly salary for collectors is 70 yuan, with special subsidies in some areas. According to a survey by the Beijing Evening News, the daily salary for directly recruited positions by human resources companies can reach 200 to 300 yuan, while practitioners who take orders through intermediaries have a significant discount. In July, Xiao Pan worked part-time through an agency and earned only 15 yuan an hour. After working 10 hours a day, he only got 147 yuan after deducting insurance. “For a period of time, my only task was to simulate the operation of beating eggs in the kitchen, using beans instead of eggs, and stirring the beans repeatedly with an egg beater. This recording lasted all day.”
Another force is open source. Zhiyuan has open sourced the AGIBOTWORLD data set in 2026, and the third phase was opened on August 31, covering 14 types of interactive tasks such as industrial and household, with a total of 11,430 trajectories, including 1,369 failure trajectories – allowing the model to learn from failures, risks and corrections, which is regarded by the industry as a key signal to reduce the industry’s trial and error costs, and is a catfish for commercial data service providers.
Infrastructure is also being laid. The China Academy of Information and Communications Technology’s “Research Report on Embodied Intelligent Training Fields (2026)” shows that as of the end of June 2026, more than 70 embodied intelligent training fields have been built and in use across the country, and more than 40 are under construction or planning. But compared with the demand of tens of millions of hours, 70 companies are far from enough.
A moat can only be considered after crossing three passes.
Data is not “harvested” but “raised”. From collection, quality inspection, annotation, to real feedback after deployment to form a rolling closed loop, it is a systematic project that requires the mobilization of the whole society. A coordinate that can be referenced is: In July 2026, the Ministry of Industry and Information Technology approved the release of the first quality standard for embodied intelligent data sets, “Quality Requirements and Evaluation Methods for Embodied Intelligent Data Sets of Key Basic Technologies of Artificial Intelligence”, which was drafted by the Academy of Information and Communications Technology in conjunction with more than 40 units in the industry and will be implemented on November 1 – the industry has finally begun to set standards for “teaching materials”.
For China, the richest application scenarios and the most complete manufacturing system are barriers that are more difficult to copy than any single technology. However, to transform “scenario advantage” into “data advantage” and then into “model advantage”, there are still three hurdles in between: industrial quality control, closed-loop flywheel and data standards. If you can’t cross it, it’s a moat. If you can’t cross it, it’s just “mining for many hours.”
Return to Xiaoyuan. In the chain of embodied intelligence, collectors are at the forefront and farthest from the final product. He didn’t know which robot the grabbing, placing, inserting, and stacking movements he recorded would eventually be fed to. But those repetitions of arms hanging in the air until they feel sore have become a small section of a certain model parameter.
The “textbook” of embodied intelligence is not written. It is made over and over again by someone at home in Jilin, in a rental house, or in front of the equipment they took home at the Chenzhou Mobile Business Hall. 500,000 hours sounds like an industrial figure, but if you break it down, it turns out to be the daily actions of countless ordinary people.
In order for a robot to work like a human, someone must first do the work like a robot. In this competition for “teaching materials”, whoever succeeds in commercializing the data flywheel first will have the pricing power for the next stage of embodied intelligence.
Disclaimer: This article is based on public reports, official reports and corporate disclosure materials. The data is based on the Institute of Information and Communications Technology reports, prospectuses, exchange documents and authoritative media reports. The companies involved are only examples of industrial phenomena and do not constitute any investment advice. The market is risky and investments must be made with caution.
This article is produced by New Yellow River Technology and Big Health Research Institute
Related Reading
- Nvidia live updates: Jensen Huang ousted, CFO Colette Kress to serve as interim CEO2026-11-13
- Are mangoes not afraid of red fruits?2026-11-12
- Original: After Huang Renxun, who is most likely to take over Nvidia?2026-11-12
- Samsung’s roll-screen mobile phone is exposed, creating a new form ahead of Huawei and Apple, and folding screens are expected to be eliminated2026-11-12
- Musk’s favorite car is going to be discontinued.2026-11-12
