- Pudong is one of four Chinese regions selected to develop national data-labeling standards for embodied intelligence
- The pilot comes as robot makers worldwide race to build the vast real-world datasets needed to train machines to perform physical tasks
Shanghai’s Pudong New Area has been selected as one of China’s national pilot regions for data labeling in embodied intelligence, giving the district a role in setting standards for the data needed to train robots.
Pudong was named one of four regions selected for the embodied-intelligence track under the second batch of China’s national data-labeling pilot program at the China International Big Data Industry Expo on August 28-30 in Guiyang, capital of Guizhou Province, Shanghai Observer, a news portal, reported on September 3.
The designation puts Pudong in charge of exploring standards and practices that could eventually be adopted across the country.
For embodied AI, that is becoming an increasingly urgent task.
Robots need a different kind of data
Large language models can learn from trillions of Tokens collected from the internet. Robots have no equivalent reservoir of ready-made training data.
To learn how to tighten a screw on a factory line, fold clothes in a kitchen or restock a supermarket shelf, robots need first-person, real-world records of those actions.
Industry estimates suggest embodied AI could require at least 1,000 times more data than large language models. Yet the entire industry has only about 500,000 hours of high-quality embodied data today, leaving a huge gap between supply and demand.
That is why more companies are turning to real-world data collection. The challenge is no longer simply gathering data, but producing enough high-quality data in a consistent format that different robots and AI models can actually use.
Pudong bets on standardization
Pudong is positioning itself around four strengths: industry, data, real-world scenarios and research.
It has built what it describes as the world’s first training facility for heterogeneous humanoid robots, with more than 50,000 hours of multimodal data collected from physical machines.
A dataset released by AgiBot (智元机器人) has also been incorporated into international model-training systems, while its data operation produces 30,000 to 50,000 high-quality motion trajectories a day.
The district is home to more than 150 embodied-intelligence companies spanning chips, algorithms, components and complete robots.
Zhangjiang AI Innovation Town and Zhangjiang Robot Valley provide two major industrial bases for the sector.
Pudong has already released initial standards covering data formats, labeling practices and quality assessment, while developing tools for multimodal labeling and systems that allow data collected from different robot platforms to work together.
The national pilot will allow those efforts to be tested on a larger scale.
A global race for robot data
The race is not confined to China.
In the US, Figure has built a global data-collection network through an app that has reportedly reached 108 countries and regions and accumulated 16 million video clips in four months.
China is pursuing a more industrialized approach. Wuhan opened a high-quality data center in August designed to produce training data at scale through standardized, scenario-based collection, earning comparisons to a “vocational school” for robots.
Pudong’s national pilot adds another layer: turning data collection and labeling into a standardized part of the embodied-AI infrastructure.
As physical AI moves from demonstrations to real-world deployment, the bottleneck may increasingly shift from building robots to teaching them what to do.
Data is becoming the textbook for embodied intelligence. The countries and companies that can produce high-quality data at scale — and establish the standards for using it — may have a say in writing that textbook.
For the robotics industry, data labeling is emerging as a third competitive front alongside hardware and algorithms.
Header image credit: AgiBot’s official WeChat
