Add physical supervision to the robot data you own.
Episode-aligned labels for archived or live robot video, after setup on your data. Same two outputs as a public pack.
Send us your worst hour.
An hour of your hardest footage: occlusion, clutter, the failures in your success rates that your data can't explain. We return a coverage report scoring each span readable, unsure, or needs-an-instrument, abstentions shown.
Ask for on-prem instructions if your data can't leave your network. By sending footage you confirm you hold the rights to share it.
The pilot: one task, one frozen baseline, one pre-agreed decision.
Agreed in writing before labeling: task and dataset, signals, frozen baseline, held-out split, success criterion, and the comparison that decides continuation.
- 1
ID and clock audit
- 2
Validation path defined up front
- 3
Labels, generated and reviewed
- 4
Overlay and integration recipe
- 5
Matched evaluation on your robot, your instruments as ground truth
The ID-and-clock audit comes back within days, before any labels exist.
The pilot brief
Fill it before the first call and the call becomes a review instead of a pitch.
Good fit
- Contact-rich manipulation
- Video aligned to actions or proprioception
- Repeated task families
- Enough demonstrations for grouped splits
- Optional physical truth for validation
If your task is visually easy for a large policy, our labels may add nothing; we measured that case on ourselves and will say so.