Data Scarcity

The core bottleneck in embodied AI: usable real robot-interaction data falls far short of what's needed.

Data scarcity refers to embodied AI's lack of enough high-quality training data. Large language models can draw on the entire text internet, but what a robot needs is data about “how to actually move” — joint angles, force, touch — which is almost nowhere on the internet and can only be gathered one demonstration at a time through teleoperation or wearable devices, which is expensive, slow, and doesn't transfer easily between different robots. UC Berkeley professor Ken Goldberg wrote in Science Robotics in August 2025 that the text used to train large models is equivalent to what a person would need 100,000 years to read, and argued that robotics correspondingly faces a “100,000-year data gap.” Responses include collecting more real-robot data, generating data with simulation and generative models, learning from human video, sharing data across robot embodiments, and letting robots accumulate data during deployment.

ExampleAccording to a July 2026 report from tech outlet TMTPost, the CEO of Mifeng Technology estimated embodied AI would need about 100 million hours of training data to reach GPT-3.5-level capability, while as of early 2026 the world's available high-quality real physical-interaction data totaled only a few hundred thousand hours.

Also called

Data Bottleneck, Data Gap

Related

Data Pyramid · Simulation Data · Synthetic Data · Human Video Data · Data Flywheel · Teleoperation

Sources

UC Berkeley CDSS: Humanoid robots face challenges in gaining real-world skills
Rockingrobots: Humanoid robots are advancing but face a massive data gap
钛媒体：机器人还没学会做家务，卖数据的已经先赚到了钱 (Chinese)

As of

2026-07