A Ballot dataset is a set of episodes from one task: people piloting a simulated robot, recorded inside the simulation. You’ll be able to preview episodes before you buy, and choose the format you need.

What’s in a dataset

  • Episodes from one task and release, with the robot, controller and simulation version they came from.
  • Actions the pilot asked for and the actions the controller applied, at every step.
  • The full simulation state at every step, not just a camera feed.
  • Outcomes and events, so you can keep only successes, or study failures.
  • Credits for the simulation’s creators and the pilots.
This is simulation data. It isn’t recorded on real robots, and results on a real robot aren’t guaranteed.

Multi-robot data

Datasets from co-op scenarios hold several robots working together: a humanoid handing a crate to a rover, two arms passing a part. Every seat keeps its own tracks, labelled by robot and seat, on one shared timeline, so you can train on one robot, on the handoffs or on the whole team. Filter by robot type, seat or step when you choose what to include.

Choose at purchase, or request new data

Because episodes keep the full simulation state, several things can be chosen when you buy. Others were fixed when the pilots played, and changing them means collecting new episodes. If the dataset you need doesn’t exist, or it has the wrong robot or controller, fund a request for new episodes instead.

Licences

Each listing will show its licence, including any conditions carried over from the simulation’s assets. Some robot models and scene assets come with their own terms, such as non-commercial use only, and those terms can apply to the data made with them.

In the first release

The marketplace isn’t part of the first release. Developers get the episodes from their own simulations directly. See Episodes and data.