Staff at China Xiongan Group Digital City Technology Co., Ltd. are training robots at an embodied intelligence training ground. The scene resembles a specialized technical school established for robots.

In the industrial goods handling area, robots repeatedly grasp and stack items, their movements appearing somewhat clumsy yet meticulous. In the home living area, a robot practices placing slices of bread into a toaster's slot. The commercial retail and logistics equipment zones are equally busy, with robots of various brands and designs learning at their respective workstations.

Behind each of these robots stands a fully focused data collector. Wearing VR headsets or exoskeleton suits, they gather data from multiple sensors capturing different perspectives from the robot's head, eyes, hands, and forearms.

Why is it necessary to teach robots in this hands-on manner?

Li Guoliang, Assistant General Manager of China Xiongan Group Digital City Technology Co., Ltd., provides the answer: robots are not born knowing how to perform tasks. The embodied intelligence training ground uses physical machines and wearable devices for data collection, annotation, and processing. It transforms the vast range of human actions across various scenarios into high-quality datasets that robots can "understand." These datasets are then used to train large models, and the refined capabilities are ultimately transferred to more embodied intelligence entities, enabling them to execute the actions humans desire.

Intelligent robots require high-quality "data textbooks" and patient "human teachers."

In the home living area, data collector Liang Lishan grips the control handles tightly, her gaze fixed on the robot's gripper. Following her movements, the robot slowly extends its mechanical arm, the gripper gently closing as it attempts to place a slice of bread into the toaster slot.

"Oops, dropped it," she says, deliberately giving a slight shake. The robot fails to control the bread. "Your hands need to be steady to capture effective data."

Liang Lishan explains that for every seemingly simple action—toasting bread, stacking plates, stacking bowls, sorting candies—thousands of instances of effective data must be collected. The actual number of training repetitions far exceeds the number of successful data captures.

In the adjacent logistics assembly area, data collector Guo Chenyang, wearing an exoskeleton suit, resembles a "mech warrior." He exerts force with his arms, operating corresponding buttons to drive the mechanical arm in picking up goods, turning, walking, and stacking them neatly in one fluid motion.

Guo Chenyang notes that the exoskeleton mechanical arm can synchronously record multimodal motion data from the robot, including joint torque and posture across seven degrees of freedom. Subtle changes in each joint—shoulder, elbow, wrist—are captured with precision. Actions like grasping, holding, placing, taking, shaking, and transporting all become learnable data through repeated training.

In 2025, China Xiongan Group Digital City Technology Co., Ltd. was approved for the "Innovative Development Pilot for Urban Trusted Data Space (Xiongan New Area)," accelerating the deep integration of "Cloud Xiongan" with the physical city. Prior to the establishment of the embodied intelligence training ground, several leading model companies had expressed procurement needs for tens of millions of hours of high-quality datasets, but the industry reality was challenging.

Requirements for robot strength, precision, and motion logic vary completely across different scenarios such as home living, catering, sanitation, industry, and warehousing. The number of细分场景 is vast, and the cost of full-dimensional data collection is extremely high, leading to a widespread "data drought" in the industry.

From Li Guoliang's perspective, the development prospects for the embodied intelligence training ground in Xiongan New Area are broad. On one hand, the embodied intelligence industry is entering a period of rapid growth, with market demand for physical machine training datasets continuously rising. On the other hand, Xiongan New Area offers a rich variety of scenarios, providing the training ground with numerous real-world operational environments, which helps in accumulating diverse training materials.

Driven by both supply and demand, establishing the embodied intelligence training ground became imperative. However, the absence of a unified national standard for embodied intelligence data and the differing data formats among major robot manufacturers also presented challenges for data universalization.

To address this, China Xiongan Group Digital City Technology Co., Ltd. collaborated with partners to develop an embodied intelligence corpus processing platform. This platform enables the unified collection, annotation, and processing of data in different formats from over 150 domestic robot brands.

In December 2025, the embodied intelligence training ground, constructed and operated by China Xiongan Group Digital City Technology Co., Ltd., officially commenced operations. The training ground introduced robots from multiple manufacturers, achieving full-form coverage including legged, wheeled, and collaborative robots. In terms of场景构建, the training ground adopted a 1:1 real-scene restoration method, creating five practical application scenarios: reception and guidance, industrial goods handling, home living, commercial retail, and logistics assembly. It pioneered a complete capability chain for multi-vendor, heterogeneous robots, from unified device management and unified data collection to unified data annotation and model training.

The widespread application of BIM (Building Information Modeling) technology in Xiongan New Area's construction sector also facilitates embodied intelligence training. They utilize architectural virtual spaces for simulation training and data synthesis of grasping, holding, placing, and taking actions, significantly reducing data collection costs and improving efficiency.

Currently, the embodied intelligence training ground has entered a phase of routine operation for physical machine data collection and annotation. High-quality data is simultaneously listed on the Xiongan New Area Urban Trusted Data Space, providing continuous "fuel" for rapid algorithm iteration and the practical application of large models. This effectively enables the scaled application and commercial deployment of embodied intelligence large models in complex, city-level scenarios.

Li Guoliang stated that the operational launch of the embodied intelligence training ground marks Xiongan New Area's entry into a new phase of "collective training and data-driven" development in the embodied intelligence field, planting the seeds for future industry development in this city of the future.