Robotics and embodied-AI labs are hitting the same wall LLM builders hit years ago: there isn't enough labeled data, and the labeled data that exists is expensive to make. Unlike text, physical-world data needs sensor fusion, spatial annotation, and real-world capture—work that annotation shops can't just crowdsource cheaply, which pushes per-unit pricing well above typical text/image labeling rates.
Expect frontier robotics players to start writing the same kind of premium licensing and annotation checks OpenAI and Anthropic wrote for text data two years ago, with specialized data-labeling vendors positioned as the winners.
Physical AI Hits A Data Labeling Wall That Only Cash Can Fix
— Forbes
