The verified data layer for physical AI
We record the work robots need to learn from.
People in shops, homes, restaurants, factories and other workplaces book one of our kits, pick it up nearby and record their own work. On a head camera, that's a normal working day. On the gripper, it's short tasks the app shows them one at a time. We check everything that comes back and deliver it to the spec you gave us, with a receipt for every episode.
Almost all of it comes from a handful of buildings.
Manipulation data gets collected wherever someone can stand next to the rig and watch it. In practice, that's a lab or somewhere close to one.
| DATASET | VOLUME | WHERE IT WAS RECORDED | |
|---|---|---|---|
| 01 | AgiBot World | 1,001,552 trajectories | One facility of over 4,000 m² |
| 02 | DROID | 76,000 trajectories · 564 scenes | 52 buildings, by 18 research labs |
| 03 | BridgeData V2 | 60,096 trajectories | 24 environments, seven of them toy kitchens |
Tracking gets harder to keep once you leave the lab.EVIDENCE→
Ego-Exo4D recorded people doing skilled activities in 13 cities, in scenarios its research team set up. Tracking held for the whole recording 95.9% of the time, and 3.5% of recordings had gaps where it failed partway. Our contributors don't have a research team around them.
We put the supervision in the app.
Embodiment, action space, loader, and the tasks and rooms you're missing.
In the app, in English or Hindi.
From a partner location.
The app teaches the kit in about ten minutes.
Every minute is checked to decide the contributor's pay, and every episode is checked against your contract to decide what's delivered.
That's ₹100 an hour on the head camera and ₹300 on the gripper, worked out to the minute and paid by UPI within 24 hours.
Which variety you add matters more than how many demos you collect.
| WHAT VARIED | SUCCESS RATE | SETUP | EFFECT |
|---|---|---|---|
| Spatial arrangement | 30% → 60% | Real robot, 20 rollouts | Large |
| Camera pose | 43% → 90% | Simulation, 30 rollouts | Large |
| Spatial arrangement | 27% → 57% | Simulation, 30 rollouts | Large, in line with the real robot |
| Object texture | 93%, with textures that didn't match | Simulation, 30 rollouts | Close to none |
These come from What Matters in Learning from Large-Scale Datasets. We've rounded the simulation rows, and the paper doesn't report seeds for them, so treat them as support for the hardware result. On their own they don't prove much.
Data Scaling Laws in Imitation Learning gets there from another direction. Across more than 15,000 real-robot rollouts, about 50 demonstrations per environment was enough. After that, more demos in the same place barely helped, and new environments and objects did.
It isn't settled for every use. Dyna's report on Dyna-2 shows mean success across 14 tasks climbing with more egocentric video in pre-training, from 20% at 1,000 hours to 53% at a million, though some tasks fall. So “50 and stop” doesn't carry over to pre-training. Those extra hours probably bring new rooms too, but the report gives no environment counts, so the two effects can't be separated.
We don't build programs around exotic objects.WHAT'S SCARCE→
The scarce things are ordinary ones, like a counter that's too low, work done sitting on the floor, a galley kitchen where the wall decides how far you can reach, or a camera angle set by where there's room to stand.
What Matters in Learning from Large-Scale Datasets · Data Scaling Laws in Imitation Learning
What we don't claim.
“Nobody has filmed India.”CORRECTION→
People have. Ego4D has 3,670 hours from 74 locations in nine countries, and IIIT Hyderabad alone recorded 138 participants across five states. What's missing is manipulation data with calibrated poses, gripper state, action timing you can trust and a check on every episode.
“The biggest action datasets are American.”CORRECTION→
Mostly they aren't. AgiBot World, from China, collected over a million trajectories itself, in one facility. ARIO, also China-led, lists over three million episodes, though most of those are converted from open datasets. The largest of those is Open X-Embodiment, a global collaboration that includes Stanford, Berkeley and Google. We used to pitch on geography until the numbers talked us out of it. What we care about now is whether the room existed before the camera arrived.
“Head-camera data replaces robot data.”CORRECTION→
Not for most work, and not yet. Handheld gripper data is the most proven route to a policy that runs on a robot. Head-camera data helps with pre-training and co-training. EgoScale pre-trained on over 20,000 hours and still got its best results after adding aligned robot data. A few small studies have run policies on a robot with no robot data at all, like EgoZero at 70% across seven tasks and HumanEgo at 92.5% across four. Both hit limits at around a centimeter of precision, and EgoZero's authors call theirs a proof of concept. Teleoperation is still the last mile on a specific robot, so we expect most labs to buy both kinds of data.
“Our data improves your policy.”CORRECTION→
We can't show that yet, because we don't have a policy result of our own. Published curation work supports the idea, and the downstream checks are how we'll test it.
“This is easier than a lab.”CORRECTION→
It's harder. In a lab you fix a bad take while the person is still standing there. Our coach catches framing, tracking and drift live on the kits that have it, and everything else we only learn when the recording comes back. So the checks have to carry a lot of weight. Read them on the verification page and decide for yourself.
Email outreach@fourvectors.ai.