We record your robot demonstrations,you get training-ready episodes.
We hire and train the operators, run the collection stations, and do the segmentation, calibration, proprioceptive sync and format conversion ourselves. What lands in your bucket is training-ready data, not footage somebody on your team still has to clean up.
Somebody absorbs the rejects.
It shouldn't be you.
Most vendors quote per hour, which quietly moves the cost of rejected episodes onto you. Acceptance on complex tasks runs 70 to 85 percent, so a real share of every hourly invoice buys data your training code throws away.
At a 70% acceptance rate, roughly 900 demonstrations on the invoice never reach your training loop.
Rejected episodes still get recorded, reviewed and re-run. They just never reach your invoice.
Five ways a collection
program goes wrong.
Each of these has cost somebody a quarter of collection. What follows is what we do about them, not a promise that they cannot happen.
The corpus doesn't transfer to your robot
Everything is recorded on your embodiment, against your acceptance criteria. Data captured on somebody else's arm is useful for pre-training and not much beyond that, and we'd rather say so than sell you one.
The corpus turns out too narrow to train on
You agree a task diversity matrix with us before collection starts, and every weekly acceptance report breaks deliveries down against it. If a cell is thin you hear about it that week, not at the end of the program.
Delivery slips and your training schedule slips with it
Episodes ship weekly from the first week, not in one drop at the end. If operator capacity or rig calibration puts a week at risk, you get told before the week starts.
A cheaper hourly quote beats us on paper
Quotes below $15 an hour usually exclude QA rejection and replacement, so you pay twice for every episode that fails review. Divide by usable demonstrations instead of hours and the cheap quote is normally the expensive one.
Your data ends up somewhere you didn't agree to
Data rights are written into the contract per program, and by default you own the delivered corpus outright. We don't resell collected episodes and we don't fold them into anything we license.
Have you started
collecting yet?
The answer decides the track. One starts from nothing but a task description, the other picks up where your own rig runs out of throughput.
Egocentric video
Operators wear head-mounted cameras and perform your tasks in real environments, across whatever range of scenes and lighting you ask for.
- Priced
- Per hour accepted
- Best for
- Pre-training scale
No action labels, so it bootstraps manipulation but won't teach precise control on its own.
Teleoperation
Operators drive your robot through a leader-follower or VR rig, capturing joint positions, torques and camera streams on the platform you ship.
- Priced
- Per accepted episode
- Best for
- Control policies
Rig build and calibration come first. What the wait buys is no embodiment gap.
Six answers is all a real quote needs.
You own the delivered corpus by default. We never resell collected episodes.