Turn the city into a "testing ground" for robots, the embodied intelligence experiment in Changsha
At 0:00 a.m., Changsha Zoomlion Smart Industrial City. On the excavator final assembly line, dozens of self-developed humanoid robots shuttle between the stations of logistics, assembly, machining and quality inspection, and this production line rolls out one excavator every 6 minutes on average.
At 6:00 a.m., No. 18 Factory of Sany Group. AGVs and robotic arms circulate on the unattended production line, a steel plate cutting, blanking and sorting line runs automatically, hundreds of industrial cameras complete the detection of thousands of points in milliseconds with an accuracy of 0.15 mm.
At 10:00 a.m., the open square in front of IFS of Wuyi Commercial Circle in Changsha. In the 16-square-meter intelligent service cabin, the humanoid robot that once appeared in the Spring Festival Gala is busy picking up and delivering goods. Citizens scan the code to place an order, and the robot completes the whole process of identification, grasping and delivery within 48 seconds.
Late at night, Lens Intelligent Yong'an Park in Liuyang. Quadruped robotic dogs and wheeled robots roll off the production line, waiting to be delivered to the next city. The designed annual production capacity here is 500,000 units of embodied intelligent robots.
These scenarios are not rare when viewed separately.
Humanoid robots are in trial operation in factories in Shenzhen, Shanghai and Hangzhou; unmanned delivery vehicles travel through streets in Beijing and Suzhou; service cabins delivering beverages are also common in new retail trials in Chengdu, but all of them appear in the same city — Changsha.
At present, the whole country is competing for algorithm talents, financing quotas and large model bases. What is Changsha competing for?
The answer is hidden in a document recently issued by the General Office of Hunan Provincial People's Government: to cultivate intelligent robots as an emerging pillar industry during the "15th Five-Year Plan" period, strive to make the industrial chain revenue exceed 100 billion yuan by 2028 and 200 billion yuan by 2030, and Changsha is clearly defined as "the provincial technology source and innovation leading pole".
What this article aims to analyze is Changsha's layout logic of turning the whole city into a "test field", a unique route of "exchanging scenarios for data, nurturing intelligence with data, and feeding manufacturing back with intelligence".
01
Five Rings Interlock, Changsha Turns Heavy Working Conditions into a "Driving School" for Robots
To see Changsha's layout clearly, first look at an "interlocking" industrial map: complete machine manufacturing, core components, scenario application, scientific research talents, policy carriers, the five rings are interlocked, and each ring is closely connected with another.
The representative of complete machine manufacturing is Lens Technology. This giant in the "Apple supply chain" defined robots as its core track in 2025, and it does not manufacture self-owned brands, focusing on mass production and delivery.
In 2025, the shipment of Lens humanoid complete machines exceeded 3,000 units, and the quadruped robotic dogs exceeded 10,000 units. Lens Intelligent's revenue in that year exceeded 1 billion yuan; in November 2025, the Yong'an Park was put into operation, with a designed annual production capacity of 500,000 units.
This link gives Changsha a rare label: one of the few cities in the whole industry that has both "complete machine production capacity and consumer electronics-level quality control".
In terms of core components, according to the caliber of Hunan Provincial Department of Industry and Information Technology, upstream robot enterprises account for about 59% of the province's industry. Changsha is not an "assembly economy", but a complete chain with flesh and blood — the harmonic reducer of Wanxin Precision (Ningxiang) breaks the overseas monopoly and passes the continuous fatigue test of more than 10,000 hours; Jingjiawei is developing the first domestic physical AI chip for robots; Goke Microelectronics' intelligent vision chips have shipped more than 300 million units, ranking first in market share; Inovance's base in Hunan (Yueyang) produces 7 million servo motors per year, accounting for nearly 30% of the national total.
Scenario application is the most important link in Changsha's layout. Zoomlion has independently developed 8 humanoid robots, dozens of which have entered the trial operation of production line logistics, assembly and quality inspection links, and has built a 120-station training ground. The humanoid robot factory will be put into operation at the end of 2026, with a five-year plan to deploy 1000-2000 units; Sany Group opens scenarios to technology companies, and the steel plate sorting line built by SuperVision for it has increased production capacity by about 3 times after being put into operation in 2020; Xingshen Intelligent's unmanned delivery vehicles operate normally, and Super Intelligent Robot is deeply engaged in the health care service scenarios.
In the field of scientific research and policy, the National Engineering Research Center for Robot Visual Perception and Control Technology led by Academician Wang Yaonan, together with the talent supply from universities such as National University of Defense Technology, Central South University and Hunan University, forms a smooth talent chain.
On the policy side, the provincial implementation opinions deploy "10 key projects", carry out the "Intelligent Robot +" action on a regular basis, and set up a future industry sub-fund with a target size of 10 billion yuan; in terms of core carriers, Hunan Embodied Intelligence Innovation Center (Changsha Economic Development Zone, jointly operated by Lens and Agibot) has gathered nearly 30 enterprises, and Xiangjiang New Area has gathered 532 AI and sensor enterprises with an output value of over 30 billion yuan.
The interlocking of the five rings lies in that each ring creates value for other rings: complete machine factories provide orders, scenario enterprises supply data, universities supply talents, and policies provide support. This is exactly the difference between Changsha and "single-point breakthrough" cities. Interlocking requires the whole industrial chain to operate in the same city.
Why can Changsha build this chain? It relies on a "heavy working condition heritage". The training data most needed by embodied intelligence comes from real working conditions, and Changsha is one of the cities with the most dense heavy working condition scenarios in China.
Changsha's construction machinery industry cluster is one of the first national advanced manufacturing clusters, and its scale has ranked first in the country for 16 consecutive years; 5 enterprises including Sany, Zoomlion, CRCC Heavy Industry, Sunward Intelligent and Xingbang Intelligent are shortlisted in the 2026 Top 50 Global Construction Machinery Manufacturers, accounting for 9.5% of the global share and 45% of the national share.
This means that mines, construction sites, ports and production lines are densely distributed around Changsha, which are the training grounds most needed but hardest to obtain for embodied intelligence. The common feature of these scenarios is "dirty, tiring, dangerous and non-standard", which are exactly the places that traditional industrial robots cannot handle and general large models cannot reach.
This is also the raw material of Changsha's layout moat: the simulation environment can be built with capital, but "the data of the real physical world" cannot be accumulated faster by capital. It takes decades to complete the "dirty, tiring and dangerous" work step by step. While others are "feeding" models in the computing power center, Changsha turns the excavator production lines and mine sites into "driving schools" for robots. The first product produced by this school every day is not equipment, but data.
02
Data is Obtained through Practical Work, and Orders are Selected through Screening
Every time the production line runs one cycle, Changsha accumulates more data. Every time the humanoid robot in Zoomlion's factory completes a handling cycle, it leaves a complete data chain including trajectory, force sense and vision; the daily sorting records of special-shaped parts on Sany Heavy Industry's steel plate sorting line are tens of thousands, which feeds back SuperVision's "Kunwu" platform to be reused across industries.
However, data cannot turn into intelligence on its own. What makes this raw material flow, circulate and appreciate is a "data flywheel", whose technical core is a pain point recognized by the industry: sim-to-real gap.
Physical details such as friction, tolerance, light and sudden failures cannot be learned in simulation, so there is a saying circulating in the industry: It is better for a robot to fall once on a real production line than to fall ten thousand times in simulation.
Changsha's solution is to turn production activities themselves into data production, and precipitate physical details such as tolerance, light, stress and failures into training samples one by one, realizing that "working means generating data".
Here are two typical cases.
Zoomlion's self-developed humanoid robot starts from "auxiliary stations" such as pre-assembly and quality inspection, and does not replace workers all at once, but realizes progressive human-robot collaboration. It is supported by the self-developed RobotOps operating system and the "Cloud Valley" embodied intelligence large model. The 120-station training ground collects thousands of real machine data every day, and has completed verification of nearly 20 scenarios.
The more stations the robot works, the more batches of "data obtained through practical work" it accumulates.
The second case is SuperVision "staying in the factory for 10 months".
In 2019, the start-up company SuperVision took over the steel plate sorting problem of Sany Heavy Industry: there were tens of thousands of special-shaped parts that traditional industrial robots could not grasp, and the team stayed in the lighthouse factory for 10 months to debug closely.
In August 2020, the first full-automatic sorting line was put into operation, which increased the production capacity by about 3 times and reduced the sorting labor by more than 60%. Then this capability was replicated across industries. In the first quarter of 2026, it won nearly 200 million yuan of intelligent ship manufacturing orders, completed a 100 million-yuan level B++ round of financing, and the "Kunwu" platform reduced the project debugging cycle from 6 months to less than 2 weeks.
The common point of the two cases is that data is not collected, but "obtained through practical work".
Scenario usage generates data, data trains models, and the improved capability undertakes more complex tasks, which in turn generates more scenario data. The production line is the only place that can form a self-reinforcing closed loop, which is also the difference between "the city as a test field" and "gathering enterprises in the park".
Whether this chain can continue to operate also depends on a set of "self-rotation mechanism", and the most convincing evidence is not the investment promotion PPT, but the choice of enterprises. A large number of non-local enterprises have placed their orders and layouts in Changsha.
One question is, why did Agibot give the order to Lens? Agibot's headquarters is in Shanghai, but its cooperation with Lens is increasing, and they jointly operate the Hunan Embodied Intelligence Innovation Center; in July, the two sides signed a cooperation agreement on the intelligent production line project in Huizhou, entrusting Lens to carry out production.
The production capacity and quality control of Changsha's enterprises have become the choice of leading players, which is the first layer of the "self-rotation mechanism".
The same is true for component enterprises. As mentioned earlier, Changsha already has a complete interlocking of complete machines, components and scenarios. Being close to complete machine manufacturers and scenario parties shortens the supply chain radius, which is the most simple logic of manufacturing.
By disassembling the development logic of Changsha's robot industry, we can find that Changsha attracts not "pure algorithm teams", but "teams that need scenarios and manufacturing". On the contrary, scenario parties such as Zoomlion and Sany Heavy Industry need external intelligent capabilities, and this two-way demand constitutes the "self-rotation".
At the same time, the amplifying effect of policies also takes effect simultaneously. The "industrial chain closed loop" and "advantageous manufacturing capability transformation" projects in the provincial implementation opinions guide enterprises in construction machinery and other fields to extend to core components, and the 10 billion yuan sub-fund invests in the industrial chain according to the principle of "investing in early-stage, small-sized, long-term and hard technology projects".
In summary, the "self-rotation" of Changsha's layout does not rely on feelings, but on the interlocking of interests. Each link can make money and obtain data from other links.
03
"Test Field" or "OEM Factory"? Misalignment of Production Capacity, Data and Scenarios
Before evaluating this set of layouts, first put Changsha into the national coordinate system. Objectively speaking, Changsha has scenarios and manufacturing, but the density of algorithm talents, capital activity and data reserve of service scenarios still cannot be compared with first-tier cities.
Advantages and shortcomings are two sides of the same coin. After seeing the front, we have to look at the back, and on the back of this coin, there are three structural cracks. The metaphor of "the city as a test field" magnifies them to a certain extent.
The first crack: Is it a test field or an "OEM factory"? Do enterprises come to Changsha to "use scenarios" or to "build production capacity"?
The most direct test took place in July. Agibot signed the intelligent production line project with Lens, entrusting Changsha-based enterprise Lens to produce, but the planned location of the project is Huiyang Embodied Intelligence Industrial Park in Huizhou. In addition to the Southwest Base that Agibot put into operation in Pixian, Chengdu in May, leading players' production capacity maps are spreading across multiple points in the country, and Changsha is only one of the options.
The gap between "attracting enterprises" and "retaining enterprises" is real. Whether enterprises are "passing by" or "taking root" is the criterion to test the quality of the test field.
The second crack: Who owns the data?
The premise of the five-ring interlocking is data flow, but Zoomlion's production line data belongs to Zoomlion, and Lens's manufacturing data belongs to Lens. Will the "islandization" of enterprise data make the "city-level data flywheel" only a vision? Who will build the data middle platform? At present, there is no clear production line data sharing mechanism at Changsha's municipal level.
If data only flows within enterprises, "the city as a test field" is only a collection of "enterprises as test fields", and the five-ring interlocking lacks the most critical link.
The third crack: The "unbalanced development" of heavy working conditions.
Changsha's scenario spectrum is concentrated in the heavy industry, such as excavators, cranes, mines, ports... This is both an advantage and a limitation.
When the main battlefield of the industry turns to family, commercial and service scenarios, what is the universality of robots that "can only do heavy work"? Xiangjiang New Area focuses its breakthrough on "Changsha-made parts for robots" such as dexterous hands and bionic joints. Zhongke Huisi settled in the new area and released three dexterous hands. Academician Wang Yaonan said that "dexterous operation is the last centimeter for robots to enter the real world", which is a correction to the "unbalanced development", but the data density of service scenarios is still Changsha's short board in the short term.
It needs to be clarified that the three cracks are not independent, but point to the same root: The logic of "test field" is public goods, while the logic of enterprises is private property rights.
The government hopes that data will flow into a flywheel at the city level, but enterprises regard production line data as core assets; the government hopes that enterprises will take root and build production capacity, but enterprises naturally compare prices and choose locations across the country. When public logic meets private logic, the quality of the test field depends on whether Changsha can bridge the cracks through governance design, such as building a data sharing mechanism to give certainty for enterprises to take root.
There is no ready-made answer to this question at present, but it is the biggest variable that can determine whether Changsha's 100-billion-yuan target in 2028 can be fulfilled. Whether Changsha can upgrade the "test field" to a "mass production base + data hub", make enterprises willing to stay and build production capacity, make data willing to flow into a flywheel, and cultivate a more diverse scenario spectrum beyond heavy working conditions, the answer will not be written in policy documents, but in every game between enterprises and the city in the next stage.
04
Conclusion
China's urban industrial competition for embodied intelligence is emerging with two layout logics.
One is the "talent + capital" agglomeration model of Beijing, Shanghai, Shenzhen and Hangzhou, which uses the highest valuation and the densest algorithm talents to bet on the first breakthrough of general intelligence; the other is the "scenario + manufacturing" interlocking model of Changsha, which uses the most solid manufacturing base and the most dense