We study what makes robot training data actually work. We then produce it at a cost and scale no one else can compete with.

Right now, nobody knows which demonstrations matter, which tasks transfer, or why 1000 hours from one setup outperforms 5000 from another. The relationship between data and capability is still a black box.

Some questions only answer themselves at scale. You can't discover what makes data work by collecting a little of it carefully. You need volume to run the experiments — ablations across a wide variety of tasks, environments, operators, and embodiments. The study requires the infrastructure.

We have both. Now we're running the experiments nobody else can.

Gigascale embodied AI datasets shaped by research into what works.

Researchers and engineers from Nvidia, Anthropic, Mercor, Bridgewater, Draper, SpaceX, Toyota Research Institute, and growing.