A new preprint on arXiv (2610.22496) examines the limits of scaling in robot learning. The authors note that progress in the field has come from scaling demonstration datasets and model capacity, echoing a broader trend toward larger data and models in robotics.
The abstract then identifies a persistent bottleneck: deploying a new robot embodiment or task typically still requires a dedicated teleoperation corpus. This implies that even with larger models and more demonstrations, policies remain tied to the specific data collected for each new setup, rather than generalising freely.
The source abstract cuts off mid-sentence at "The resulting policies remain bounded by…", so the full argument and any proposed remedy are not available from the provided text. This article therefore covers only the stated portion of the paper.