China advances standards for embodied AI data

China’s National Data Administration is set to expedite standards for embodied artificial intelligence, aiming to alleviate one of the major hurdles in the burgeoning robotics industry: securing high-quality, diverse, and extensive training data. Focusing on establishing clearer rules for datasets utilized by embodied AI systems, which integrate artificial intelligence with physical machines such as humanoid robots, the regulator seeks to promote more uniform data collection, processing, evaluation, and application in the face of accelerating competition.

China has already established a comprehensive standards framework encompassing broader aspects of the sector. Published on August 27, these national standards dictate the quality of real-world embodied-intelligence data and the technical prerequisites for data-generation platforms, while several additional specifications are still under development.

Projects overseen by the National Data Administration include proposals for standards on the origins and components of high-quality embodied-intelligence datasets, the generation and processing of synthetic data, and guidelines for data collection and model training at training facilities. These initiatives are being crafted under the guidance of the National Data Standardisation Technical Committee.

A draft program, registered in April, stipulates a 12-month schedule for a standard concerning data collection and model training at embodied-intelligence training bases. Its drafting committee comprises the China Electronics Standardization Institute, Beijing Institute of Technology, the Institute of Software at the Chinese Academy of Sciences, and prominent robotics firms.

Another draft focuses on synthetic data, which has gained prominence due to the expense, slowness, and reproducibility challenges associated with gathering sufficient physical-world robot interactions. Encouraging simulation and synthetic techniques, China’s policy framework aims to augment the supply of data where real-world interactions are scarce or prohibitively costly.

As part of its broader artificial-intelligence strategy, the National Data Administration has emphasized high-quality datasets. A June implementation plan calls for accelerated dataset development in strategic and emerging fields such as embodied intelligence, intelligent driving, and the low-altitude economy. It also urges authorities and industry stakeholders to enhance data cleaning, enhancement, labeling, alignment, and quality inspection while developing national standards for formats, categories, annotation, and quality assessment.

Such efforts should yield structurally comprehensive, varied, accurately labeled, and model-friendly datasets.

Reflecting the unique data requirements of robots operating in the physical world, China’s advances in standards for embodied AI data target the complex demands of robots that must perceive, reason, and act in real-world environments. Unlike text-based systems primarily trained on digital information, embodied AI necessitates vast quantities of multimodal data, including video, images, sensor readings, point clouds, movement trajectories, and feedback from physical interactions.

Consequently, data supply has become a significant competitive constraint for developers striving to train machines to perform reliable sequences of actions across changing environments. With over 126,000 high-quality datasets already established in China, totaling more than 1,815 petabytes, the government has been actively expanding the underlying repository of AI-ready datasets through a nationwide dataset management platform, open-source community, data-labelling network, and sector-specific pilot programs.

Embodied intelligence has been a focal point in these pilots, alongside other emerging fields. Cooperation among robotics companies, universities, research institutes, and government bodies aims to augment the quantity and practicality of machine-training data while testing standards prior to wider deployment.