Spill recovery · open now

Train the next
generation of
robots.

Household robots are not held back by their hands. They are held back by never having watched anyone do the work. Every clip of an ordinary cleanup is a lesson they cannot get from the internet, from a simulator, or from each other.

A humanoid household robot, arms raised

Drag to turn it. This is the thing your twenty seconds is teaching.

You already make a mess. You already clean it up. The same two minutes, filmed, is what teaches a robot how.

Start training Why this matters No account · one take · judged in under a minute
What you are teaching(01)

It can already move. It cannot yet decide.

The hardware problem is largely solved — joints, grip, balance, battery. What a household robot still lacks is the ordinary judgement a person spends no thought on at all. That part is not engineered. It is learned, from examples, one at a time.

Sequencing

What to gather before what. Sweep the rice up first or the water spreads it further — nobody ever writes this down.

Order of operations

Object sense

Which thing is a tool, which is rubbish, which goes back in the cupboard.

Semantic placement

Recovery

What to do when the first attempt smears it instead of lifting it.

Failure data

Judgement

When a surface is actually clean, rather than merely wiped once.

Terminal state
Contributed clips(02)

Every lesson ends with a real person doing it — on camera.

Four clips from our own capture set, with the decision the agent published for each. Two were kept. Two were not, and the rules that failed are printed on them.

Read the full ledger ↗
accepted 53.8s

Wiping down bottles

contributor capture · 1080p source

BOTH HANDSONE TAKEWRIST-TRACKEDCLOTHEXCERPT
surface cleaningaccepted
capture ok 11.4s

Egg and fork, fixed table

contributor capture · 1080p source

BOTH HANDSONE TAKEWRIST-TRACKED
kitchen manipulationqueued
refused 25.5s

Components, patterned cloth

contributor capture · 2160p source

CAMERA MOTIONMOTION BLURNO CLEAR TASKEXCERPT
three rules failednot accepted
refused 27.8s

Walkthrough, interior

contributor capture · 2160p source

CAMERA MOTIONSUBJECT CHANGESNO BEFORE STATEEXCERPT
four rules failednot accepted

One in two of these did not make it, which is close to the discard rate the whole industry runs at. The difference is that the two refusals are on this page rather than deleted from a bucket, each one naming the rule that stopped it. A contributor can see what to change. A lab can see what it is not being sold. The duration on each card is the clip as submitted; where that runs long, the player shows a trimmed excerpt of it.

Where we stand(03)

Real people. Real tasks. Real records.

Every figure below is something you can check on this page today. We are not going to print a total we cannot evidence — the whole point of the ledger is that the numbers hold up when somebody looks.

Rules enforced per clip 8

Named checks, each answered separately in writing by the agent before any verdict is computed.

Time to a decision <60sec

From upload to a published verdict with the reasoning attached. No human reviewer in the loop.

Decisions published 100%

Accepted or refused, every decision goes on the record with the rule that produced it.

Tasks written 6

One open for contributions now, five more specified and waiting in the queue behind it.

When there are ten thousand hours in the set, that number goes here and the ledger will back it. Until then this row stays small and true.

Written rules

Eight named checks. Your clip is judged against words you can read before you film.

Rubric v1

SERV Reasoning

The agent answers every rule in writing. It never gets to announce the verdict.

BRAID

Published decisions

Accepted or refused, the reasoning goes on the record where anyone can read it.

Open ledger

Your name on it

Every clip keeps its contributor record. The lab that trains on it knows where it came from.

Provenance
Why it matters(04)

Robots can read every book. They have never watched anyone tidy up.

A language model learned from text a person would need a hundred thousand years to read. A robot arm has nothing comparable to learn from. Controlling joints is harder than producing sentences, and no simulator gets the physics of a wet cloth on a counter right. The only source is a person doing the thing, on camera, once.

That footage barely exists — so the largest robotics companies on earth are now paying to create it from scratch.

Figure AI $1B

Committed over twelve months to data and compute, after launching a platform that pays people to film household work.

Forbes, August 2026
Already collected 16M

Videos uploaded to that one platform so far, at a rate of thirty minutes of footage every second, from 108 countries.

Forbes, August 2026
Bought, per year $100M+

What robotics companies already spend annually buying real-world manipulation data from third parties.

MIT Technology Review, 2026
One vendor alone 100k hrs

Footage gathered by a single data contractor. Others run thousands of workers across more than fifty countries.

MIT Technology Review, 2026
A woman folding towels at a table beside a humanoid robot
A robot learns to fold by watching a person fold. There is no shortcut around this step.
A woman on a step ladder changing a bulb while a robot holds out a replacement
Knowing which object to hand over, and when, is judgement. It is learned from examples or not at all.

Filming is not training. Footage becomes training only when something can tell a good demonstration from a bad one — and say why.

The part every closed collection app leaves out

A fifth of it is wasted

Between twenty and thirty per cent of collected footage never becomes usable training data. Hands drift out of frame, the task is half finished, the light goes. Someone pays for those hours anyway.

Sorting beats collecting

One folding policy went from eight per cent success to eighty-three on the same two hundred hours, simply reweighted by a model that could judge quality. The bottleneck is not volume. It is judgement.

The failures are the lesson

Grading failed attempts instead of deleting them took real-world success from thirty per cent to ninety in two hundred episodes. Most pipelines throw that data away by default.

How it works(05)

Your normal cleanup, with the camera running.

Nothing here asks you to stage anything. You make the mess you were going to make, you clean it the way you always clean it, and the recording is the only thing you do differently.

See the full brief ↗
01

Set the phone down

On a stand, or propped against a jar. It must not move while you film, and both your hands need to stay inside the frame.

02

Make the mess

The one you were going to make anyway — the dal that goes over, the flour, the tea. Then clear it exactly the way you always do. No demonstrating, no slowing down. Twenty seconds is usually the whole thing.

03

Upload it

Send the clip straight from your camera roll. The agent samples frames, reads it against all eight rules, and comes back in under a minute.

04

Get paid, or get told why not

Accepted, and your cleanup joins the training set with your name on it. Refused, and you get the exact rules that failed and the sentence that decided each one. Nothing is left unexplained.

Take part · spill recovery(06)

Spill recovery is open now.

It is the first task because it is the one a robot fails hardest at: a mess has no fixed shape, no fixed place, and no instruction manual. Watching a person decide what to gather first is the whole lesson.

Spilled
Contained
Clear

Spilled, contained, clear. Three states the agent has to tell apart in your clip — and so will the robot.

Filming20S
TakeONE
CameraFIXED
ChecksEIGHT
Answer<60S
RecordPUBLIC

What you need

  • A phone, and anything to prop it against so it does not move.
  • A flat surface — a kitchen counter, a table, a tray.
  • Something to spill. Rice, dal, flour, sugar, water, tea.
  • Whatever you would normally clear it with. A cloth, a hand, a brush, a sheet of paper.
  • Daylight, or one lamp that does not move.

What makes a clip useful

  • Let the mess be fully visible before you touch anything. No before, nothing to learn from.
  • Move at your normal speed. Careful demonstration speed teaches the wrong thing.
  • Gather before you wipe, if that is how you would do it. Sequence is the signal.
  • Finish properly. Clear surface, tools set down, hands lifted away.
  • Vary it. A different surface, a different substance, a different tool each time.

What gets it refused

  • Hands leaving the frame, even briefly.
  • Cuts, speed-ups, or anything trimmed out of the middle.
  • A moving camera. Held in your other hand does not count as fixed.
  • Other people, other faces, readable screens or documents in shot.
  • A near-repeat of a clip already in the set. Same spill, same surface, same motion.

The eight rules, in full

Every one of these is answered separately, in writing, before anything is decided. Seven of them can refuse a clip. The eighth only ever raises its grade.

  • FramingBoth hands and the whole affected area stay in frame for the entire clip.
  • Mess visibleThe spill is clearly visible before any cleaning starts.
  • Containment firstThe mess is gathered or contained before it is absorbed or wiped.Graded, not required
  • CompletenessThe area is visibly clear by the final frame and the hands have released their tools.
  • ContinuityA single continuous shot. No cuts, no speed changes, nothing removed.
  • DurationLong enough to show the whole task, measured from the file rather than guessed.
  • VisibilityHand position and the state of the mess followable throughout.
  • One subjectOne person. No other people, no faces but yours, no readable screens.

Containment first is scored rather than enforced. Cleaning it the other way round is still useful — how people actually sequence a mess is a large part of what we are here to collect, and a rule that punished the unusual order would quietly delete it.

On payment. Accepted clips are paid in USDC on Base, between $0.30 and $1.20 depending on how demanding the task is. It is deliberately not the reason to be here: nobody gets rich filming spills, and a contributor chasing volume is exactly who the repeat check is built to catch.
Open the portal and submit a clip Read the eight rules first · answer in under a minute
What counts(07)

Four kinds of mess worth teaching.

All four sit on a fixed surface inside one frame. Mopping and vacuuming are deliberately excluded — they break the framing rule, and a clip that breaks framing teaches a robot nothing.

See the setup ↗
01

Spilled dry goods

02

Liquid spill

03

Crumbs and debris

04

Cluttered surface

Every decision, published(08)

Refusals go on the record too.

The middle one is why this exists. A closed collection app tells you your clip was rejected. This one tells you which rule, in whose words, and what measurement overruled the model.

Acceptedgrade 1.0

clip a41c8e07 · 38.0s

All seven required checks cleared. Rice was gathered toward the centre before the cloth came out, so containment scored too.

→ added to the training set

Refused

clip 7b20d914 · 11.0s

Framing
The hands leave the bottom edge for roughly a third of the clip.
Duration
Measured from the file: 11s, too short to show the whole task. This overruled the model, which had passed it.
Completeness
Liquid is still visible on the surface in the final frame.

→ not accepted · film it again

Already held

clip c93f1a55 · 41.0s

Every rule cleared, and it was still refused. Too close to clip a41c8e07 — distance 0.11 against a threshold of 0.40.

→ not accepted · try a different mess

  1. The model answers, the code decides

    SERV Reasoning answers each rule separately with a written reason. Code computes accept or refuse from those answers. If the model volunteers an opinion that contradicts its own answers, the code wins and the disagreement is recorded.

  2. Some rules score instead of gate

    Containment never causes a refusal — it sets a grade. Reweighting an identical dataset by a quality model once took a folding policy from 8% to 83% success, so a graded set is worth more to a lab than a merely admitted one.

  3. Measured facts beat the model

    Duration comes from the file, not from an opinion about the file. When the two disagree, the file wins and the override appears on the record.

  4. A broken agent never costs you the work

    If the reasoning pass fails to return a usable verdict, your clip is queued for another look. Nobody is refused because our software misbehaved.

For robotics teams(09)

Every sample arrives with its paperwork.

Recent copyright litigation has separated the training activity, often defensible, from how the data was acquired — which is where the liability actually sits. Buyers now ask for a rights record per asset rather than a catalogue-level claim. Ours is built in.

What a camera captures

What gets picked up, in what order, where each thing belongs, whether a mess is contained before it is absorbed, when the tool changes. Sequencing and semantic placement — the part long-horizon household policies are weakest at, and the part no force sensor records.

What it does not

Not force data. Surface-wiping datasets are specified around 6-axis force/torque at 500Hz, collected by physically guiding a robot arm, because wiping means regulating normal force in a 2–15 newton band. A phone cannot record newtons and we do not pretend otherwise. This is pre-training data, not a teleoperation substitute.

Opening next(10)

Five more tasks.

Each opens as the one before it fills. They get harder, longer and less scriptable in that order — which is exactly the direction household data is shortest in.

What this does not do yet(11)

Stated plainly.

A dataset nobody can audit is a dataset nobody can buy. So here are the gaps, before anyone has to ask.

Somebody has to show them how.

The robots arriving in kitchens over the next decade will have learned from ordinary people doing ordinary things on camera. The only question is whether anyone kept a record of who taught them, and whether the lesson was any good. The agent is open source, the rules are written down, and every decision it has made is public.

Start training For robotics teams