Data

For VLA and world-model training

Robot data, verified before it ships.

EGO is our system for deploying robots on real jobs, and every deployment also produces verified training data. Data is the layer available today. Each episode comes from a skill proven on situations it never saw: behavior known to work, with clean labels on every frame.

60 tasks. One standard of proof.

Six of the 1,020 episodes. Every episode looks different, so models learn the task, not the scene.
1,020 Successful episodes, about one hour at 30 Hz
60 Tasks across place, fill and peg insertion, one verified policy each
80%+ Success on unseen conditions for every policy behind the data
129 Labeled failure episodes, for world models and value learning

Simulated in NVIDIA Isaac Sim on a Robotiq Hand-E parallel gripper. Cubes, cans, boxes and pegs of varied size, weight and grip, placed into bins, trays and shelves, and inserted into tight sockets.

Ready for your VLA today, your world model tomorrow.

  1. Three camera views

    A scene camera, a wrist camera and a camera that follows the gripper, so models learn from the viewpoints real cells use.

    Images
  2. Robot state and actions

    Every frame paired with the robot’s pose and the action it took, in the LeRobot format that GR00T and LeRobot policies train on.

    LeRobot
  3. Language

    The task in plain English, alternative phrasings, and a caption for each step of the motion.

    Text
  4. World-model ready

    Object positions, camera data and motion phases on every frame, for models that predict video as well as actions.

    World models
  5. Proven source

    Each episode is traced to the verified policy that produced it, with that policy’s success rate.

    Traceable
  6. Searchable

    An index of every episode by part, container, colors and outcome, so you can pick exactly the clips you need.

    Index
  7. Quality checked

    Every episode passes an automated check before it ships, with a quality report in the pack.

    Checked
  8. Photoreal on request

    Depth and segmentation for every view, ready for NVIDIA Cosmos to re-render the footage photoreal.

    Optional

Buy it, subscribe to it, or order your own.

Test it on your own model.

Request sample pack v1 for evaluation. We send a private download link and the datasheet after a short evaluation agreement.

Request the sample pack

Tell us what you are training, so we can point you at the right skills.