Community Discussion · Policy

Embodied AI Data Infrastructure: A Golden Age for Selling Shovels, But the Tools May Be Fragile

Tian JiJul 162026/07/16 90 views

Data is the fuel for artificial intelligence—a truth validated countless times in the LLM era. When it comes to Embodied AI, the problem becomes: fuel is extremely scarce, and extracting it is harder than mining Bitcoin. LLM pre-training has tens of trillions of tokens, autonomous driving has billions of hours of data, while publicly available operational data for Embodied AI is only on the order of hundreds of thousands of hours. This magnitude gap is like trying to water the Sahara Desert with a single cup of water.

Policy

1 replies

Ctrl + Enter to reply

Demo Still FarJul 26(edited)

[quote="tianji, post:1, topic:764"]

Data is the fuel for artificial intelligence; this truth has been validated countless times in the LLM era. When it comes to Embodied AI, the problem becomes: fuel is extremely scarce, and mining it is harder than mining Bitcoin. LLM pre-training has tens of trillions of tokens, autonomous driving has billions of hours of data, yet publicly available operational data for Embodied AI is only in the hundreds of thousands of hours range. This order-of-magnitude gap is like trying to water the Sahara Desert with a cup of water.

So, some people saw a business opportunity: helping robots get data. There is indeed a lot of money, financing news is everywhere, but the bubble is also palpable...

[/quote]

If 80% of the million-level dataset is simulation, this number can support a round of funding in a pitch deck, but during deployment, the sim-to-real gap knocks you back to square one. Do companies dare to disclose their simulation data ratio and real-world generalization test results?