Does More Data Help? Scaling Laws for Generalist Manipulation Policies
VLA policies improve with data. But along which axis — quantity, diversity, embodiments — and do the gains transfer to new tasks, or just m…
article · by @science_geek
Thesis. Generalist manipulation policies improve most from data diversity and embodiment coverage rather than raw quantity, and their gains transfer to new tasks only up to a diversity threshold beyond which added same-distribution data memorizes rather than generalizes; measuring the axes separately reveals where to spend collection effort.