Inspiration
Robot learning runs on demonstrations: a person remotely drives a robot through a task and the recording becomes training data. Labs now buy this data by the hour, and investors expect more than $3B of robot data to be bought over the next two years.
The problem is that nobody tests whether a recording works before it is sold. One audit of 1,000 purchased hours found 26% unusable. Buyers pay first and discover the bad data after weeks of training. We wanted to know if a demonstration could be tested like a part coming off a production line.
What it does
The Station is a data broker for robot training data that sells only recordings that pass a physics test.
- Source: data vendors consign recordings to us and are paid when the buyer pays.
- Screen: a learned quality model rejects obvious failures cheaply.
- Test: each recording is replayed in simulation ten times with small disturbances to the arm.
- Certify: a recording passes if it still finishes the task in at least 7 of 10 replays.
- Deliver: the buyer pays only for recordings that pass and receives the matching simulation environment.
On 300 public recordings, 90% from skilled operators passed, against 80% from middle operators and 33% from weak ones. The test never saw the operator ratings.
How we built it
- Test harness: MuJoCo replays of public robomimic recordings, with randomised disturbances applied to each replay.
- Interactive scorer: a demo where you pick a recording and watch it pass or fail.
- Policy experiment: we trained the same robot policy on three sets of data to see if the test improves training, not only sorting.
- Market research: Claude ran six sub-agents that operated Velocity AI in the browser, four at a time. They produced the competitor map, buyer personas, price book and launch plan. We then checked the key figures on primary sources.
Challenges we ran into
- Sizing a market with no audited figure. Published estimates for robot data range from about $450M to $1.5B for this year. We built our own estimate from vendor revenues and cross-checked it against an investor forecast.
- Our uniqueness claim was too broad. We first believed no seller tests its data. Research showed several grade or replay it. We narrowed the claim to what holds: none uses a disturbance test to decide what the buyer pays for.
- Proving the test matters. Separating good operators from weak ones is not the same as training a better robot. Our first policy experiment did not show a gain, and we had to rework it.
- Treating simulated interviews honestly. Six model-generated buyers all asked for a small audit first. That shaped our sales plan, but it is not customer evidence and we do not present it as such.
Accomplishments that we're proud of
- A working test, not a mock-up. The scorer runs on real public recordings.
- A first training result. In one task, 648 recordings kept by our test gave 64.2% robot success, against 55.8% for all 900 and 47.8% for a random 648. These are first readings and a larger re-test is running.
- A complete business case in two days. Market, competitors, pricing and go-to-market, with sources.
- Knowing what we have not proven. Every claim in the deck carries its caveat.
What we learned
- Buyers already pay for "accepted" data. The opening is in who defines acceptance and how.
- A broker earns trust by reporting the usable fraction, including for its own suppliers.
- AI research tools are fast at finding companies and slow at being right about numbers. Every figure needed a primary source.
- Buyers want to start small, so the first sale is an audit, not a contract.
What's next for The Station
- Finish the larger policy re-test across more tasks and repeated runs.
- Test on real-robot recordings. Everything so far uses simulated public data.
- Measure cost and time per recording, which our pricing depends on.
- Publish a free report certifying the commercially licensed open datasets and stating the usable fraction of each.
- Hold real customer conversations with robot learning labs, using the audit as the first offer.
- Sign the first five suppliers and first paid audits within six months.
