Beijing Targets 200,000 Robots a Year as China Pushes Embodied AI Into Real Work

A manufacturing plan puts robot performance to the test

Beijing wants factories capable of producing 200,000 embodied intelligence robots a year, machines that can work for more than four hours on a battery charge, and a regional supply chain accounting for a quarter of China's production capacity. Its latest robotics plan sets concrete manufacturing and reliability targets as Chinese cities compete to turn advances in artificial intelligence into useful physical work.

The Beijing Municipal Bureau of Economy and Information Technology issued its Action Plan for High Quality Development of the Embodied AI Robotics Industry (2026–2028) on November 11, 2026. By 2028, it aims to cultivate at least five humanoid robot manufacturers considered internationally leading and 10 considered leading within China.

Shanghai, Shenzhen, Hangzhou and Hunan are pursuing related policies, combining subsidies, industrial supply chains, training sites and legislation. The national ambition is to make embodied intelligence a new source of economic growth under the 15th Five Year Plan.

Embodied intelligence means AI operating through a physical machine: sensing its surroundings, deciding what to do and acting. A humanoid robot is one possible form. Other forms include robotic arms, vehicles and machines with four legs. For a warehouse customer, the practical issue is whether a robot can reliably move goods, rather than whether it resembles a person.

That shift toward useful work runs through Beijing's plan. Alongside production capacity, it sets targets for loads, equipment life, shared manufacturing services and customer returns. The challenge is to connect those measures so that more machines leaving factories also means more machines doing jobs that customers will pay for.

What Beijing wants to build by 2028

Beijing's proposed annual capacity of 200,000 complete robots would combine manufacturing by contractors with production lines built by robot developers themselves. Across Beijing, Tianjin and Hebei, the target is 25% of national production capacity, with more than 60% of key components supplied through the regional industrial network.

The plan calls for three to five robot platforms suitable for deployment across different settings. These would use modular designs, allowing manufacturers to adapt common hardware and software rather than design every product from the beginning.

Its performance goals include a rated load exceeding 40% of the robot's own weight, an average operating time between failures exceeding 2,000 hours, a design life exceeding five years and battery endurance exceeding four hours during continuous work. A robot weighing 50 kilograms, for example, would need a rated load above 20 kilograms to meet the weight target.

Core moving components and end effectors, the hands or tools that perform tasks, would have annual failure rates below 10%. The plan also targets dexterous robotic hands costing less than 10,000 yuan each and batteries with an energy density of 350 watt hours per kilogram.

Beijing already has a substantial industrial base. Municipal economy and information technology bureau director Jiang Guangzhi said the city had more than 940 companies along the humanoid robot supply chain and had established the first humanoid robot innovation center jointly developed at national and local levels. More than 30 complete humanoid robot manufacturers are based in Beijing.

Production is growing faster than commercial certainty

China produced more than 40,000 complete humanoid robots in the first half of 2026, according to figures cited by Vice Minister of Industry and Information Technology Xiong Jijun. Full year production was expected to exceed 100,000.

Beijing's industrial operating figures put production by its key humanoid robot manufacturers at nearly 6,000 units during the same six months, an increase described as nearly tenfold from a year earlier. The city's broader robotics industry generated 51.8 billion yuan in revenue in 2025, up from 15 billion yuan in 2023, or about 3.5 times the earlier figure.

Market researcher IDC measured a different part of the business: shipments. Its figures put global humanoid robot shipments at nearly 25,000 in the first half, including more than 19,000 in China, representing about 77.9% of the global market. Production counts machines made; shipments count machines dispatched into the market.

Beijing's 200,000 figure describes future annual manufacturing capacity, while the national figures describe current output. How fully the planned factories will be used depends on orders, operating costs and the speed at which customers adopt the machines.

Regional cooperation aims to shorten the factory journey

Beijing plans at least five manufacturing service platforms and at least 10 shared processing platforms across Beijing, Tianjin and Hebei. Their services would include structural parts, surface and heat treatment, circuit board assembly and wiring harness production.

The target is to deliver new component samples within one to two days after a design change. Faster access to samples would allow developers to test revised parts sooner, identify manufacturing problems and repeat the process before committing to larger production runs.

Zhang Yufeng, founder and chief executive of Wujie Power, said his company had established a pilot production platform in Beijing's Haidian district for manufacturing and validating its K15 robot. He described closer regional coordination as a way to improve quality, reduce costs and identify production problems during design.

Beijing also wants suppliers from the intelligent connected vehicle industry to enter robotics. Yizhuang already hosts more than 300 robotics and embodied intelligence businesses, with a supply chain valued at more than 20 billion yuan. Precision machining, wiring and manufacturing at scale are capabilities that can serve both vehicles and robots.

The first jobs are in factories, shops and warehouses

Beijing's application targets cover at least 20 industrial settings, 20 service settings and 10 specialist settings, including emergency rescue and workplace safety. It also wants at least 15 solutions that can be deployed repeatedly across customers, rather than remaining individual demonstration projects.

Galbot robots have entered production lines at CATL, Bosch, Toyota and BAIC for inspection, loading and unloading, and sorting. In a Zhongguancun convenience store, the company's robot clerk has handled drinks, coffee and grilled sausages, including performing different tasks with its two arms at the same time.

Wujie Power's robots have also started work at Beijing Qixian coffee outlets and entered ZF LIFETEC and Aumovio factories for testing and improvement. Such deployments provide opportunities to train machines around changing objects, layouts and production requirements.

Experience in Shanghai illustrates how demanding a complete shift can be. During a 24 hour store trial on November 2–3, six Agibot robots completed 1,730 tasks, but only four passed their assessments. Shelf inspection and cleaning caused difficulties, and one cleaning robot left its route and went outside.

Galbot cofounder and head of large models Zhang Zhizheng identified several tasks still facing developers: helping models handle unfamiliar situations, improving the fit between software and hardware, and making deployment adapt to different working conditions. These capabilities determine how easily a successful installation can be repeated elsewhere.

Logistics operators are opening real workplaces for training

Cainiao announced in Hong Kong on November 5 that it would open its global warehouse network to industry partners as a training and testing environment for embodied intelligence. Cainiao International president Xiong Wei described access to millions of square meters of warehouses and operating experience across more than 25 countries.

The initiative offers more than 100 practical testing scenarios across dozens of industries. Xiong divided logistics robotics into three levels: short sequences such as placing parcels on conveyors and attaching labels; more precise picking of individual items; and longer planning tasks such as packing boxes or arranging complex stacks.

Each level requires different abilities. Moving a parcel involves recognizing its shape and gripping it securely. Picking an item from a crowded container requires finer judgments about overlapping objects. Packing a box requires planning how several objects will fit together and in what order they should be placed.

Xiong forecast that commercialization of the simpler tasks could accelerate in 2027. JD is pursuing a related approach, with plans to collect more than 10 million hours of video showing human activity in real settings and more than 1 million hours of robot operating data over two years. Its logistics network has more than 3,600 warehouses.

Mech-Mind founder Shao Tianlan proposed three practical tests of maturity: how long a robot works without intervention, how many people and how much money installation requires, and the customer's return on investment. His suggested benchmarks include two weeks of operation without human intervention and deployment that an ordinary person can complete in half a day.

Data and computing power are becoming shared infrastructure

Beijing plans to collect more than 1 million hours of high quality data from at least 100 real settings and add dedicated computing clusters totaling more than 50,000 petaflops. A petaflop measures a computer's ability to perform a million billion numerical operations per second.

The city supports several ways of collecting training material, including cameras showing a person's view and methods that gather demonstrations without requiring a robot at every recording session. It also proposes a permanent evaluation system where customers define tasks and models are tested on their ability to complete them.

Research support covers vision language action models, which connect images and instructions to movements; world models, which help predict how a physical environment changes; and reinforcement learning, which improves behavior through feedback on attempted actions.

Computing access remains a constraint. Carnegie Endowment researchers Pavlo Zvenyhorodskyi and Scott Singer identified advanced AI chips, computing resources and imported torque and force sensors as weaknesses in their November 24, 2025, embodied AI study. The pressure on computing access has continued, with H200 chips awaiting entry into mainland China while buyers seek import approvals.

Developers are also trying to use existing equipment more efficiently. Qujing Technology founder and chief executive Ai Zhiyuan said its system distributes model inference, the process of running a trained model, across graphics processors, conventional processors, memory and storage. Reusing intermediate calculations helps reduce computing demand.

Other cities are choosing different policy tools

Shanghai's July 2025 implementation plan targets at least 20 breakthroughs in core algorithms and technologies, at least four specialist incubators and a core industry exceeding 50 billion yuan by 2027. Its five proposed support platforms cover computing, practical training, pilot production, investment and leasing.

Shanghai offers support of up to 30% of approved investment, capped at 50 million yuan, for key technology projects. Innovation service platforms can receive up to 50%, capped at 20 million yuan. Eligible companies can receive computing vouchers worth up to 40 million yuan a year. The city is also exploring investment arrangements in which computing resources and training data contribute to company equity.

Shenzhen issued its first action plan in March 2025 and revised it in June 2026. Its priorities include AI chips, dexterous hands, robot control and domestic computing chips and modules installed inside robots. By 2027, it seeks more than 10 additional companies valued above 10 billion yuan and more than 20 with annual revenue above 1 billion yuan.

Hangzhou's regulation promoting embodied intelligence robotics took effect on May 1, 2026. It addresses application access, standards, supervision and coordination across suppliers. A separate industrial plan targets more than 20 billion yuan in output from complete robot manufacturers and more than 50 billion yuan across the supply chain by 2027.

Hunan's September policy draws on its construction machinery, rail equipment, aerospace, materials and electronics industries. It focuses on industrial robots, specialist operations and consumer services, linking component development with testing, production, deployment and feedback from operating data.

How the policy push has developed

Local plans have accumulated over nearly two years, moving from identifying an emerging industry to specifying manufacturing services, commercial tests and legal responsibilities.

• February 2025: Beijing released an earlier embodied intelligence action plan covering 2025–2027.
• March 2025: Embodied intelligence entered the national Government Work Report, and Shenzhen issued its initial robotics action plan.
• July 2025: Shanghai published its embodied intelligence implementation plan.
• May 1, 2026: Hangzhou's robotics industry regulation took effect.
• June 2026: Shenzhen revised its plan.
• September 2026: Hunan issued its intelligent robotics development policy.
• November 5, 2026: Cainiao announced the opening of its global warehouse network for training and validation.
• November 11, 2026: Beijing issued its latest plan, effective through December 31, 2028.

The next major target year is 2027 for Shanghai, Shenzhen and Hangzhou, followed by Beijing's 2028 deadline. Cainiao's forecast of faster adoption in simpler logistics tasks also centers on 2027.

Investment, safety and the risk of duplication

China's rapid expansion has drawn warnings from its economic planning authorities. National Development and Reform Commission spokesperson Li Chao said the country had more than 150 humanoid robot companies, with startups and businesses entering from other sectors making up more than half.

Li described the industry's commercial and technical position at a press conference:

"Meanwhile, humanoid robots have yet to fully mature in terms of technological pathways, business models and application scenarios," she said.

The commission is working on standards, evaluation systems and mechanisms governing entry into and exit from the industry. Li warned about similar products crowding the market and reducing room for research and development.

Investment remains substantial. Consultancy iiMedia Research put financing for China's embodied intelligence sector at 109.174 billion yuan in the first half of 2026. Carnegie's 2025 study described a Beijing AI and robotics investment fund of 100 billion yuan with a 15 year lifespan, alongside a Shanghai embodied AI fund with an initial close of 560 million yuan. Its authors warned that overlapping local investment could add duplication and debt.

Export figures show how much commercial development remains. A 2026 robotics white paper put overseas revenue above 80% for some leading commercial delivery and cleaning robot companies, and above 70% for warehouse robot company Geekplus for four consecutive years. Many embodied intelligence companies were in the 10%–25% range.

Beijing's response combines business support with safeguards. It proposes individual assistance for companies seeking stock market listings, support for at least 200 innovative businesses, 100 university and research institute robot teams, and participation in at least eight international standards and 30 national standards.

Safety measures include emergency braking, protection against malicious attacks, handling of sensitive personal data and a unique identification code for each robot throughout development, production, retirement and recycling. For workers sharing factory space with robots, and people receiving care or public services, these requirements are part of making deployment workable.

Gong Piming, an associate researcher at the NDRC's Academy of Macroeconomic Research, has called for coordination across regions and industrial alliances that connect software, hardware and actual workplaces. The central commercial task is now concrete: build machines that operate reliably, can be installed affordably and earn repeat orders.

Key Points

• Beijing targets annual manufacturing capacity of 200,000 embodied intelligence robots by 2028.
• The Beijing, Tianjin and Hebei region aims for 25% of national capacity and regional sourcing of more than 60% of key components.
• Robot performance targets include more than 2,000 hours between failures and more than four hours of continuous battery operation.
• Beijing plans more than 1 million hours of training data, shared manufacturing facilities and at least 50 benchmark application settings.
• Shanghai, Shenzhen, Hangzhou and Hunan are using different combinations of subsidies, technology programs, legislation and manufacturing strengths.
• Warehouse access and workplace trials are putting reliability, installation costs and customer returns at the center of commercialization.
• National authorities warn against duplicate products and are developing standards and evaluation systems.