Robotics remains early. Progress is unlikely to come from scaling data volume alone, particularly when the underlying signal is weak or poorly matched to the policy being trained.

Raw egocentric data is an important substrate, but it is not sufficient for every problem. The more important question is which data changes policy behavior: what improves control, what transfers across settings, what remains robust in deployment, and what meaningfully shifts learning curves.

We think data should be benchmarked against policies, evaluations, and deployment outcomes, rather than treated as a volume problem in isolation. Many settings also require custom modalities and hardware-in-the-loop work conducted in close collaboration with labs.

Our view is that data is not merely an input to scale. It is an experimental instrument that should be designed, measured, and revised with the same care as the policies it helps train.

A data mix is chosen and evaluated in service of a specific policy, not as a generic pile of examples.