Recorded
Real people, real homes, real factories. Egocentric and teleoperated footage captured to spec, with hand pose, depth and multi-view on request.
- Egocentric
- Teleop
- Multi-view
Robotics training data
Robots don't get a 90% chance.
Ninescale AI records, simulates and labels the data that takes robots from 90% to 99.9999% reliable.
01 / The gapReliability
90%
At 90% reliable,
10 in 100 fail.
Good enough for a demo. Not good enough for a robot working next to people.
In a home, a hospital,
on a road — one in ten
costs lives.
Trust is earned
one nine at a time.
Every extra nine needs more data — and rarer data. That is what Ninescale makes.
● 10 / 100 failed
Today · clinical lab intake
Every sample is still sorted by hand.
Rack after rack, one tube at a time. It's exactly the work robots should take over — and the data they need to learn it.
02 / What we sell
The data robots learn trust from.
Simulated
Scenario-matched simulation that fills the long tail — the rare drops, slips and edge cases you can't wait to film in the real world.
- Long-tail
- Domain-randomised
- Paired with real
Labelled
Every frame passes through LabelNow: segment boundaries, success / failure, rewards and QA by trained reviewers before it ships.
- Segments
- Rewards
- Human QA
03 / From LabelNowReal and simulated.
Real and simulated.
Side by side.
Same episode. Two worlds.
A real robot episode rebuilt end to end — Gaussian-splat background, generated object meshes, MuJoCo physics. Drag the seam.
● Real robotSimulation ○
⟷
AUTOLab · Block into cupSource: public DROID episodes
04 / Real2Sim
Captured once. Rebuilt in physics.
Every capture is recorded so it can become a simulation environment — the same scene, the same camera, real contact physics.
0.6 mm
Median depth error on the best rebuilt object
5 / 5
±2 mm perturbation trials passed
0
Unintended contacts while lifting by physics alone
Sim → RealTrained in simulation.
Trained in simulation.
Works on the real robot.
0
Real-world training data · the policy learned only in simulation
05 / PipelineFrom capture
Grasp events labelled automatically from jaw width + camera pose
From capture
to checkpoint.
One pipeline, built on LabelNow, that turns raw human activity into training-ready robot data.
- 01 — CAPTURE
Capture
A distributed contributor network records tasks to your spec on phones and rigs.
- 02 — SIMULATE
Simulate
We mirror each real task in simulation and generate the variations reality can't.
- 03 — LABEL
Label
LabelNow annotates segments, outcomes and rewards — frame by frame.
- 04 — VERIFY
Verify
Multi-pass human QA with agreement scoring. Bad frames never ship.
- 05 — DELIVER
Deliver
Training-ready datasets in your format: RLDS, LeRobot, HDF5 or raw.
99.9%
Expert policy success in simulation
96%
Sim success after scaling up
21,504
Simulated robots practising at once
173M
Practice steps per experiment
06 / Contact
Add a nine.
Tell us the task. We'll send back a sample dataset.