Shanghai Turns Factory Workstations Into Embodied AI Training Grounds

The gap is stark. A 2026 report by the China Academy of Information and Communications Technology, or CAICT, estimated that training a general embodied model requires at least tens of millions of hours of real interaction data, while the world's available high-quality real-machine embodied corpus is only in the million-hour range, about one order of magnitude short. Collecting data mainly through expert teleoperation of real robots often yields only dozens of hours a day, according to the report cited by Leiphone. The missing piece is not another exhibition hall but a data pipeline that can collect, label, train, deploy and feed failure samples back into the model.

Training venues themselves are no longer scarce. CAICT data showed that by the end of June 2026, China had built and put into use more than 70 embodied AI training venues, with another 46 under construction or planned. The harder problem lies between training and use: tasks are homogenized, data output is limited, circulation is weak and a sustainable business model has yet to form. In June 2026, the Ministry of Industry and Information Technology and the State-owned Assets Supervision and Administration Commission launched a special action for real-scene training of humanoid robots and embodied intelligence, using production workstations and service operation points as training carriers to accumulate high-quality real-machine data.

In a training laboratory run by Mifeng Technology in Zhangjiang, workers are teaching robots simple actions hundreds of times. Zhu Mianrui, a data collection team leader born after 2000, said even picking up a bottle or placing an object requires repeated instruction before the robot becomes proficient, according to Leiphone. The laboratory, dubbed a robot school in the report, reflects the industry's data hunger: unlike large language models that can learn from massive text on the internet, robots must grind out action data in the physical world. At Zhiyuan's data collection factory in Pudong, the AgiBotWorld project covers 217 complex manipulation tasks, 87 atomic skills and more than 3,000 objects across 106 real scenes, the report said.

Pudong is also building public capacity. During the 2026 World Artificial Intelligence Conference, the national-local humanoid robot innovation center and Huawei released China's first embodied intelligence training venue sample, located in Pudong. Shanghai Electric's humanoid robot pilot service platform began operating in Pudong Robot Valley, covering about 6,800 square meters and open to the industry for prototyping, testing and scenario validation. It has an annual prototype capacity of about 2,000 units and can complete more than 400 component and whole-machine performance validations. In August 2026, Shanghai Instrument, a municipal state-owned AI-focused industrial group, signed an agreement with the Shanghai Institute of Physical Intelligence and Robotics to explore the conversion of source technology into scenarios and engineering applications.

On the production side, humanoid robots have entered real assembly lines. At SAIC-GM's Ultium plant in Jinqiao, Pudong, Nengzai No.1, developed by SAIC-GM and Zhiyuan, handles cell grabbing and loading on a Buick E7 battery production line. The companies described it as the first humanoid robot application on a mass-production line for SAIC and China's auto industry. The robot has completed nearly 80,000 workpiece operations, with an operating rate stable at 98%, according to Leiphone. The national-local center's Linglong robots have also entered Yanfeng automotive production lines for real-scene data collection and autonomous operation validation.

Similar tests are under way at global automakers. In 2025, at BMW's Spartanburg plant in South Carolina, Figure 02 ran for about 1,250 hours over a roughly 10-month project, moving more than 90,000 sheet-metal parts and participating in the production of more than 30,000 BMW X3 vehicles. In Berlin, Mercedes-Benz brought Apptronik's Apollo humanoid robot to its Marienfelde digital factory campus for tests around production logistics, parts handling and initial quality checks.

A key link is feeding deployment data back into training. Shanghai Innovation Institute and Zhiyuan jointly released research on an LWD, or Learning While Deploying, framework in which real-machine interaction data from deployment is fed back in real time to update policies. Online LWD raised success rates on long-horizon tasks such as making kung fu tea, juicing, mixing cocktails and shoe-box storage by up to 17%, with an average multi-task success rate of 95%, according to the report. On Sept. 9, 2026, the Ministry of Human Resources and Social Security and other departments released the eighth batch of new occupations, including embodied intelligent robot application technician, with data collector and trainer roles.

The next public test is scheduled for Oct. 23-25, when the 2026 Global Developer Pioneer Conference and International Embodied Intelligence Skills Competition will be held at Zhangjiang Science Hall. The event will convert 2,000 square meters into a permanent embodied intelligence scenario training space. The Shanghai Artificial Intelligence Industry Association, a co-organizer and operator of the training platform, will evaluate real task success rates, efficiency gains, safety and reliability, and economic feasibility through formats including work skill packages. Shanghai is seeking to build a closed loop of training, competition, production and finance, the report said.

Editor's Summary

Shanghai is linking real-scene robot training, factory deployment and data feedback in Pudong as China tries to close a shortage of high-quality real-machine data. The model is moving from laboratories into automotive production lines and public training venues, with a permanent space set to open during an Oct. 23-25 competition. The effort tests whether embodied AI can move from demonstrations to verifiable, repeatable industrial work.