China's Embodied AI Sector Confronts Over 99% Physical Data Deficit

China's embodied artificial intelligence sector is confronting a severe data deficit that hampers real-world robot deployment, according to industry executives at the 2026 World Robot Conference reported by People's Daily. Current domestic compliant physical interaction data stands at approximately 500,000 hours, falling short of the tens of millions of hours required for commercialization by more than 99%. High collection costs, heterogeneous formats, and low generalization have prompted calls for universal standards, simulation technologies, and large-scale data collection initiatives from supply chain giants such as JD.com.

Embodied artificial intelligence represents the frontier of integrating AI models into physical manufacturing, logistics, and household robotics. However, unlike large language models trained on massive internet text, physical robots require high-fidelity interaction data that cannot be easily scraped. Resolving the high costs of real-machine data collection and establishing cross-industry formatting standards will determine whether the robotics industry can successfully scale into commercially viable autonomous labor.

China's embodied artificial intelligence industry is facing a severe bottleneck in physical interaction data, posing a critical hurdle to deploying robots across manufacturing and service sectors, according to industry figures speaking at the 2026 World Robot Conference reported by People's Daily. While hardware showcases highlighted dexterous robotic hands and heavy-lifting units, industry participants warned of an acute data deficit. Data cited at the conference indicates that compliant real-world physical interaction data in China totals roughly 500,000 hours, whereas commercial deployment requires tens of millions of hours—leaving a shortfall exceeding 99 percent. Cao Peng, president of JD Cloud, noted that robotic brains lack generalization capability primarily because real-world operational datasets remain profoundly scarce. Industry founders stressed that physical interaction datasets cannot be substituted by conventional internet text or video feeds. Wang Qian, founder of Variable Robot, observed that physical properties require modalities distinct from digital web media. However, capturing real-machine interaction is both time-consuming and expensive. Traditional teleoperation methods cost approximately 10 to 50 yuan per valid entry, while training an operational model typically demands tens to hundreds of thousands of entries. Compounding the expense is format heterogeneity, which prevents cross-platform data reuse. Xie Shaofeng, chairman of the OpenAtom Open Source Foundation, urged the establishment of universal industry standards covering data collection, labeling, formatting, and quality benchmarks to resolve pervasive data silos. Huang Qingqiu, co-founder and chief technology officer of Mage Intelligence, emphasized that data quality in real-world scenarios matters far more than gross volume, cautioning that much of the sector's current data remains substandard. To bridge the gap, enterprises are turning to hybrid approaches combining real-world collection, simulation, and hardware-agnostic sensor systems. Fan Haoqiang, co-founder of ForceLink, highlighted a framework combining real-machine data loops, native embodied architectures, and synthetic physics simulation to handle edge cases and long-tail scenarios. JD.com announced plans to collect over 10 million hours of real human scenario data and more than 1 million hours of robotic data within two years, leveraging its logistics network. Concurrently, companies such as Seer Robotics and Orbbec are introducing dedicated data subsidiaries and modular capture systems, including first-person egocentric and wrist-mounted cameras, to standardize and accelerate collection.