Роботам в Китае не хватает «мозга» и данных: главный вызов embodied AI на WAIC 2026

What Chinese robots lack

Chinese robots lack two things at once: quality data about the physical world and a strong foundation model that links perception to action. Industry representatives spoke about this at WAIC, and according to South China Morning Post, it is precisely this dual deficit that is holding back embodied AI's transition from demonstrations to real products.

Unlike language models, which learn from texts on the internet, robots need data on movement, grasping objects, and reacting to an unpredictable environment — this data has to be collected by hand, which is expensive and slow. Without a large volume of such data, a robot's "brain" cannot learn to act confidently outside the lab.

• Event — the WAIC conference, ended July 20, 2026 in Shanghai
• Who's speaking — Wang Xiaogang, co-founder of SenseTime and chairman of its robotics division
• Barrier 1 — lack of data on interaction with the physical world
• Barrier 2 — absence of a strong "brain," a unified foundation model
• Task — link hardware, data, models, and scenarios into a closed loop

What the "closed loop" is

Participants at WAIC named building a closed, iterative system — in which hardware, data, models, and real-world scenarios reinforce each other — as the most critical challenge for the embodied AI industry. The idea is that a robot collects data while working, that data improves the model, the improved model makes the robot more effective — and the cycle repeats.

"Link hardware, data, models, and real-world scenarios into a closed, iterative system," —

Wang Xiaogang, co-founder of SenseTime and chairman of its robotics division, on the industry's main task.

Right now this cycle is broken: most Chinese companies have hardware platforms, but neither a sufficient flow of operational data nor a model capable of fully digesting that data. As long as the links aren't connected, every prototype remains a one-off demonstration rather than a scalable product.

Why data is the bottleneck

Data for robots is the main bottleneck because it can't simply be downloaded from the internet like text or images. Every action in the physical world — a step, a turn, grasping a cup — has to be recorded via a real device's sensors or through expensive simulation, and the volumes of such information are incomparable to the trillions of tokens that language models train on.

That's exactly why WAIC participants emphasized the "data + brain" link: a powerful model without data won't learn to act, and a huge trove of data without a unified foundation model won't turn into meaningful behavior. According to conference insiders, the industry is still looking for a way to put data collection on a production line so that it pays back the cost of the hardware.

What this means

Chinese robotics isn't stuck on hardware but on intelligence and data: the country is already strong on the hardware side, but without a "brain" and a flow of real-world data, robots don't leave demo mode. Closing the "hardware — data — model — scenario" loop is becoming the main race of embodied AI for the coming years.