What's this about?
Scaling robot data is treated as the path to generalist vision-language-action models, but robot trajectories are far harder to scale than web-scale image-text data — embodied collection is costly and covers the physical world sparsely. The paper's argument: that makes representation quality, not data volume, the central bottleneck, and proposes representation-centric continued pre-training. 78 upvotes.
Why can't you just collect more robot data?
Because every trajectory requires a physical robot, in a real place, doing a real thing, in real time. You cannot scrape it, cannot parallelise it beyond how many robots you own, and each one covers a vanishingly small slice of physical situations. Text and images exist in the quantities they do because humans already produced them for other reasons. Nobody has produced a billion hours of labelled robot manipulation as a byproduct of daily life, and nobody will.
Why does this matter to you?
- It is the counter-argument to the scaling orthodoxy in the one domain where the data genuinely does not exist.
- Anima Anandkumar made the same argument for physics on Latent Space last week — structure over scale where data is scarce.
- "Representation quality is the bottleneck" is a claim you can test, which is more than most bottleneck claims.