The Data Bottleneck

Ask why humanoid robots still fumble simple chores and the answer is not the model or the motors. It is data. A large language model trains on tens of trillions of words scraped from the internet; the largest open datasets of real robot manipulation hold only a few million recorded demonstrations.^[1] Robots have no internet to learn from; every example must be physically performed.

China’s response is to industrialise the demonstration. Shanghai startup AgiBot open-sourced AgiBot World Colosseo, a dataset of more than one million real teleoperated trajectories across 217 tasks, gathered on a fleet of humanoids, the raw material for its GO-1 “vision-language-action” foundation model, which reported a roughly 30% performance gain over prior methods.^[1]

A language model reads tens of trillions of words. The biggest open robot dataset holds a few million clumsy demonstrations. Whoever closes that gap owns embodied AI.

Data factories

The state is scaling collection like manufacturing. By late 2025, China had announced more than 40 robot data-collection centres, around two dozen already operating, staffed by human “robot trainers” who teleoperate humanoids through repetitive tasks.^[2] One centre in Sichuan, opened in early 2026, is designed to produce up to three million manipulation samples a year, trainers in VR headsets driving robots through chores, all day, and a single collection session can cost over 1,000 yuan.^[3]

The reason data is the binding constraint, and not compute or model design, is that robot data cannot be scraped. Every manipulation example must be physically performed, which makes it slow and expensive to accumulate, the opposite of the near-free text and images that trained the language models. China’s response is to industrialise the one input that resists industrialisation, treating teleoperated demonstrations like a manufacturing output with daily quotas. Whether that brute-force approach or a simulation-led one wins is genuinely unsettled; what is not in doubt is that a country deploying the most robots will capture the most operational data, and that data compounds.

How we read it

This reframes the embodied-AI race. Models are converging and hardware is commoditising; the scarce, compounding asset is real-world interaction data — and China is building an industrial pipeline to accumulate it, pulled by state mandates to deploy 10,000 humanoids by the end of 2026.^[4] With Chinese makers already around 90% of humanoid shipments, more robots deployed means more data captured, which trains better robots, a flywheel.^[5] We watch the data layer most closely: the datasets, the collection infrastructure, and the foundation-model teams positioned to compound on top of it.