One dataset from many robot bodies
Laboratories pooled their demonstrations into the shared Open X-Embodiment dataset — 22 robot types and over a million real trajectories — and a model trained on it beat the same methods trained each on its own data by 50%.
Why it matters
Until then data from one robot was assumed useless for another, so every laboratory started from zero. That 50% is the measurement that transfer between bodies exists, and therefore that a shared store of robot data is worth building.
The dataset covers 22 robot embodiments, over 1 million real trajectories, 527 skills and 160,266 tasks. RT-1-X, trained on it, outperformed RT-1 or the original methods trained each on their own individual dataset by 50% in the small-data regime. RT-2-X tripled RT-2’s score on emergent-skill evaluations. The number of contributing laboratories differs between the project’s own sources: the paper abstract says 21 institutions, the Google DeepMind post says 33 academic labs, and the project page says 60 existing datasets from 34 research labs. The range is given here rather than a single figure.