What's this about?
Foundation models learned to see and talk by consuming the internet, but embodied agents (robots) get no such free lunch — they need data linking observations to physical states and actions, which is scarce and costly. This paper organizes the available sources into a "pyramid" to make the most of each tier (15 upvotes).
What exactly is "embodied" data?
It's data from acting in the physical world — what a robot's camera saw, what state it was in, what action it took, what happened next. Unlike text, you can't scrape it; it has to be collected by real or simulated bodies doing things, which is why robotics is data-starved compared to language.
Why does it matter?
Data scarcity is the central bottleneck in robotics. A principled way to combine plentiful-but-weak sources (video) with scarce-but-precise ones (real robot logs) is how the field stretches limited data — the same "better data" theme as DataPrep-Bench, applied to atoms instead of text.