Why Embodied Data Is the Biggest Gold Mine: Inside the World’s First Hundred‑Billion‑Scale Multimodal Data Cloud Mall

Paxini announced the global debut of a data cloud mall dedicated to embodied intelligence, built in partnership with JD Cloud, Tencent Cloud, and Baidu Intelligent Cloud. The platform aggregates over a hundred billion high‑quality multimodal samples collected by five proprietary data‑capture factories.

The core "evolution service engine" combines massive multimodal real‑world data with top‑tier cloud compute and storage. It defines four service dimensions to accelerate embodied AI research, model iteration, and industrial deployment, positioning the mall as a "data oasis" that addresses the chronic data scarcity in the field.

Key data features include:

Full‑modal high‑dimensional interaction data: the highest‑degree‑of‑freedom (82) hand‑motion dataset, integrating 30 six‑dimensional tactile modules to form a closed‑loop of vision‑touch‑language‑action.

Industrial‑grade millisecond precision with strict spatiotemporal alignment, enabling direct training of high‑success‑rate, highly generalizable control models.

Real‑machine closed‑loop validation that bridges the Sim‑to‑Real gap across multiple robot platforms.

A sustainable supply chain powered by five self‑built data‑capture factories, delivering standardized, scalable data for scaling‑law research.

The marketplace offers a transparent, tiered pricing model where users can select and combine data packages according to application scenarios and precision requirements. A fully self‑service workflow lets users sign contracts, pay, and receive data within minutes, dramatically shortening the traditional weeks‑long acquisition cycle.

Data can be downloaded or migrated directly to a user's cloud account, where cloud compute resources from the three partners enable end‑to‑end preprocessing, model training, and deployment without local hardware constraints.

Technical validation includes the "RoboPaint" study (Paxini et al., 2026, arXiv:2602.05325), which demonstrates a tactile‑guided redirection method achieving 84% success across ten manipulation tasks, and Isaac Sim simulations where a VLA policy trained solely on Paxini data reaches 80% average success on three benchmark tasks. Real‑world tests on the TORA humanoid robot show industry‑leading operation success rates and cross‑scene adaptability.

Through a "data flywheel"—hardware capture → data growth → model optimization → scenario deployment—the platform continuously feeds newly collected data back into model improvement, progressively raising the capabilities of the OmniVTLA large model for physical interaction, fine‑grained control, and robust generalization.

By turning previously scarce embodied data into an industrial resource comparable to electricity, the data cloud mall aims to accelerate the emergence of intelligent agents that can reliably operate in complex, unstructured physical environments across a wide range of industries.