The same rig in many places
DROID was not assembled from whatever each lab happened to own but collected on one standardised rig: 76,000 demonstrations, 350 hours, 564 scenes and 84 tasks, gathered by fifty people over 12 months across North America, Asia and Europe.
Why it matters
Open X-Embodiment proved transfer between different bodies; DROID asks the complementary question of what one identical body records in many different places. That is the only way to separate the diversity of scenes from the diversity of hardware.
The paper carries 101 authors. The dataset holds 76,000 demonstration trajectories and 350 hours of interaction, 564 scenes and 84 tasks; fifty people collected it over 12 months. Collection sites were in North America, Asia and Europe. The paper reports that training with DROID produces policies with higher performance and improved generalisation. Because the rig is identical everywhere, the dataset later became the shared platform on which distributed evaluation of policies across institutions became possible.