Better representations first: where does data scaling break down in robot learning?

KyleMorgenstein · x · 2026-11-17

In this thread, the author contrasts two approaches to robot contact-interaction problems: the mainstream route of synthetic data generation over a huge object distribution, hoping all real interactions are covered and the policy interpolates; versus work that finds a better representation first, reducing dependence on data volume. The open question: to what extent can robot learning collect its way out of these problems, and where does that break down?

Related event: Roboticists explore action representations as alternative to mass synthetic data(2 posts)→

Original post →

More from Embodied

Embodied channel →