Coldstack for Humanoid & embodied AI
Training data search for humanoid and embodied-AI robots
Your data engine is only as good as the episodes you can find. Coldstack turns raw teleop and autonomy logs into curated training sets.
Your data
Humanoid and embodied-AI teams collect enormous volumes of teleoperation and autonomous rollout data — multiple cameras, joint states, and force/torque, recorded continuously. The whole business is the data engine: finding the demonstrations and failures worth training on, and doing it fast enough to keep the model improving.
The problem
Finding good episodes is the bottleneck, not collecting them
You have petabytes of rollouts. The scarce resource is finding the clean demonstrations and the informative failures inside them — the episodes that actually move the model.
Curation is manual and does not scale with the fleet
As you add operators and robots, hand-labeling and scrubbing recordings stops keeping up. The data grows faster than your ability to mine it.
Building the search pipeline in-house is a project of its own
A Spark or Iceberg pipeline over petabytes is real infrastructure to build and run — time your ML team would rather spend on models.
What you can ask
Search your whole fleet in one query — by signal, image, and metadata.
- episodes where the gripper slipped on a transparent object
- rollouts with force spikes above threshold during a grasp
- clean demonstrations of the pour task from operator 4
- every autonomy run that ended in a recovery behavior
Matched moments export straight to a LeRobot-compatible dataset. Raw MCAP stays in your own bucket the whole time.
Questions
Does Coldstack export to LeRobot?
Yes. Matched episodes export directly to a LeRobot-compatible dataset, so the moments you find flow straight into training. RLDS-style exports are on the roadmap.
Can we search by both signal and vision?
Yes, in one query. You can combine a time-series signal (a force spike, a joint limit) with a semantic image or text match (what the scene looked like) and metadata (which robot, operator, time) — agreement across them ranks the best episodes highest.
Where does our data live?
Your raw logs stay in your own bucket. Coldstack custodies only the derived index, so nothing raw is copied out of your account.