A new China Academy of Information and Communications Technology count shows a national embodied-AI network that has passed the 70-site mark. TechNode reports that the count was current at the end of June, while another 46 facilities were either being built or planned. The significance is not the number alone. It is the effort to make data collection, model training, and robot testing part of a dedicated physical infrastructure network rather than isolated company experiments.
An embodied-AI training ground is not simply a room full of robots. TechNode describes these sites as physical environments in which organizations collect real-world data, train models, and test robotic systems. That framework puts China’s robotics push on a different footing from pure software development. A model meant to act in factories, warehouses, homes, or public spaces needs data about objects, movement, human interaction, and failure conditions that cannot be collected from text alone.
The report follows a series of commercial moves into the same data problem. EastFrontier’s coverage of 51World’s embodied-AI data platform showed one company building a product around training material for physical systems. CAICT’s count suggests that the broader infrastructure is becoming geographically distributed, with the state-linked institute treating training grounds as a national industrial capability rather than a niche research asset.
CAICT’s Count Maps a National Network of Physical AI Sites
TechNode says more than 70 sites were in service by the end of June, and they were present in a majority of China’s provincial-level regions. It identifies three principal clusters: the Yangtze River Delta, the Beijing-Tianjin-Hebei region, and the Pearl River Delta. Those regions already have deep manufacturing, logistics, research, and technology ecosystems. Locating training grounds there makes practical sense because the infrastructure can connect robot developers with potential industrial users and the real settings from which data must be collected.
The China Academy of Information and Communications Technology released its “Research Report on Tailored Intelligent Training Fields (2026)” on August 18, according to Seetao. The report’s language of “tailored intelligent” training fields refers to embodied AI, systems that learn from and act in physical environments. Seetao describes the full workflow as data collection, annotation, model training, and deployment on real machines.
That sequence matters because hardware alone does not produce a capable robot. A mechanical arm or humanoid platform needs examples of tasks, labels that make those examples useful for a model, and repeated trials in which the model’s decisions are tested against the world. A training ground can provide the controlled but physically grounded setting for those iterations. It is a bridge between a lab model and a factory floor.
Industrial Manufacturing Leads the Early Use Cases
CAICT’s summary classifies 86% of the counted locations as having industrial-manufacturing applications, TechNode reported. That is the clearest sign of where the first large-scale use cases are expected. Manufacturing offers repeated tasks, existing automation needs, and environments where success can be measured in concrete terms such as placement, inspection, handling, or assembly. It also provides the potential for continuous data generation as equipment performs similar activities many times.
The emphasis does not mean every training ground is the same. TechNode says the sites are physical environments for collecting data and testing systems, while Seetao identifies industrial manufacturing and material handling as early areas of application. Their common advantage is that they combine repetition with variation. A robot can encounter many examples of a task while still needing to respond when the object, position, or sequence changes.
That data-centered approach matches the strategy in EastFrontier’s report on OneRobotics using the 58.com network to train embodied AI. The companies and infrastructures differ, but the premise is the same: physical AI needs a supply of real interactions and feedback. Training grounds formalize that premise in facilities that can be shared by developers, manufacturers, and researchers.
From Training Grounds to a Competitive Data Infrastructure
CAICT’s report does not prove that China has solved the embodied-AI data problem. It does show that the country is building facilities intended to address it at scale. The additional 46 projects under construction or planning suggest the network is still expanding. The regional clustering also indicates that early growth may follow established industrial geography rather than spreading evenly across the country.
The strategic value lies in iteration. A training site can enable a loop in which data is collected from a task, used to improve a model, tested on a machine, and then augmented by new data from the next round of performance. That is different from treating a robot demonstration as the end product. The useful question is whether the training system can produce reliable behavior after many encounters with changing objects and settings.
China’s buildout will also be judged by the quality, not only the quantity, of the data it produces. EastFrontier’s account of SCALEFORCE’s funding for embodied-AI data infrastructure illustrates why startups see value in this layer of the robotics stack. CAICT’s 70-plus-site figure gives that business activity a national physical context.
The next milestone is not a larger site count by itself. It is evidence that the training grounds help a robot learn a task faster, transfer it to a production setting, and perform it with fewer failures. China now has a broad platform for building that evidence. The harder work is converting an expanding map of facilities into models and machines that operate dependably when the training environment gives way to the real world.