A creator app that pays people by the minute to film chores is a capture funnel, not a written skill – hour-counts without reject codes, task-not-completed, camera-on-ceiling and an eval split by person are not a brief.
Figure’s Index is the existence proof that the funnel can be industrial. Creators get a device, film everyday work in homes and businesses, and are paid by the minute. The public launch talked about hundreds of thousands of downloads, millions of videos, a pipeline ingesting thirty minutes of footage every second, and a spend path measured in billions. Per thousand hours they quoted hundreds of unique tasks, more than a thousand manipulated objects, more than a hundred environments. That is a data factory with a consumer front end. It is not a sentence you can paste into your own collection and call a spec. If you copy the app and skip the skill, you will get lifestyle footage at app scale: busy dashboards, incomplete tasks, cameras on ceilings, the same household in train and test, and a policy that still cannot put the mug in the sink.
The pattern is bigger than one brand. Physical-AI teams are hungry for first-person human video. Vendors could not, in Figure’s telling, hit the throughput they wanted, so the lab built a consumer capture loop. Other labs will try the same shortcut with a white-label app, a gig marketplace, or a “film your chores” brief that is really a store listing. An app can recruit. It cannot tell you whether the take was the job. That is what a brief is for.
What the app actually is
A funnel. Apply, waitlist, device, task list, upload, pay by the minute. Diversity arrives because every new person brings an unseen kitchen and an idiosyncratic way of doing the work. Figure said that long-tail variation is nearly impossible to define up front, which is true as a pretraining story and dangerous as a product story. If you do not define the skill, you cannot reject the take that is not the skill. You can only count hours.
Index’s own pipeline, as published, has five stages after ingest: filtering, fraud review, deduplication, rebalancing, annotation. Automated filters for technical, visual and semantic quality. Human analysts on samples for people trying to game the filters. Embedding-based near-duplicate drop. Task quotas and clusters. Hierarchical captions. That is a factory. Notice what it is not. It is not your reject code for “task not completed.” It is not your rule for “camera on the ceiling.” It is not your faces-out decision. It is not your eval split by person. It is not your rights pack. Figure can afford to throw away a large share of a consumer firehose because the firehose is the point. A product team collecting for one manipulation policy cannot. Every hour you keep that is not the job is a prior you will then have to unlearn.
Hour-counts without those codes are how you pretrain on the wrong object at great expense. People walk through houses. They start the dishwasher and stop to answer a message. They film the ceiling because the mount slipped. They complete a similar chore in a way that never matches the object set you actually need. The dashboard looks like EgoScale. The policy still fails on the mug, because nobody labelled the grasp, nobody rejected the incomplete takes, and half the people in eval were also in train.
We already wrote how to brief a human data collection when the failure mode was a vague email. The news this season is that the vague email now looks like a funded app. The failure is the same. Most failed collections do not fail in the field. They fail in the paragraph nobody wrote.
The sentence the app will not write for you
Not the vibe. The sentence.
Bad: “Egocentric chores, lots of homes, paid by the minute.” Better: “Pick a mug from a cluttered kitchen counter and place it in the sink, chest-mounted camera, hands in frame, no faces, indoor daylight, start to finish including the failures.”
Bad: “Index-like scale for our humanoid.” Better: “Unload a dishwasher, glasses-mounted camera, British kitchens, laterality noted, other people are a reject, eval held out by person.”
If you cannot write the sentence, you are shopping for hours, not collecting a skill. Shopping is what a catalogue is for. Collecting without the sentence is how you get ten thousand minutes of lifestyle footage and no trainable behaviour. Humanoid labs standing up data factories are hiring pilots, teleoperators and data-quality people for this reason. Headcount without the sentence is a film shoot with a robotics logo. An app without the sentence is the same shoot with a waitlist.
The fields that change the files are the ordinary ones. Device and mount: chest, head, glasses, third-person, phone in the hand. Those are different datasets. “Any wearable” means you will never know why the metric moved. Environment: kitchen, warehouse, cafe, parked car. Locale is not a language name. Who may appear: age gate, faces in or out, whether a second person can walk through. Rights for the real product: training a policy is not a marketing still, and a creator tapping accept in an app is not automatically shipping, eval, or a third-party licence. If faces are not required, keep them out. First-person cameras see children, screens, post on the fridge. Paying by the minute does not make those into the skill.
Teleoperation is not that dataset. A person driving a robot is useful. It is a different object: robot cameras, robot dynamics, a different rights conversation. Third-person demonstration is useful for checking whether the task finished. It does not show occlusion the way a chest camera does. Found YouTube is first-person-shaped and almost never labelled for the skill you need. An app that pays for chores can produce the right viewpoint and still mix all of the failure modes of found video: incomplete tasks, the wrong mount, the same person everywhere, no pack. Mix sources if you must. Label them. When the mug fails, you need to know which object to collect more of.
Reject codes, or you are scaling noise
Write what a bad item is while everyone is still calm, before the app starts paying.
Task not completed. The mug never reaches the sink. The laundry is filmed until it is interesting and then abandoned. Camera on the ceiling. Camera on the creator’s face when you asked for the hands. Non-consented faces. Children. Screens. Beauty filters. Digital zoom that destroys the grasp. Obvious cheating: looping the same thirty seconds, filming a screen of someone else doing the chore, uploading a take from a different household than the enrolled ID. Scripted demonstration when you asked for ordinary behaviour. A suspiciously clean counter when you asked for clutter.
Automation can flag black frames, duplicate hashes, and some of the mount failures. A person still applies the codes that mean “this is not the job.” Figure’s fraud-review step exists because consumer funnels get gamed. Your collection will too, the moment payment is by the minute. Gold items before production, or the first month is the network learning your kitchen on your budget. Annotation quality is not a percentage of uploads accepted. It is whether the accepted takes are the sentence.
Split eval by person, not by random clip. A creator who is good at the app will dominate the hour-count. If their hands sit in train and test, the policy has seen that household. The graph will look like transfer. The next kitchen will not. Hold those people out. Rights to keep the eval set and not train on it are a separate line from “we paid you to film.” If you cannot answer “are eval people excluded from training?” in one sentence, assume no.
Duration belongs in the brief so QA has a reject for “too short to be the task” instead of an argument in week three. A manipulation skill needs start to finish, including the failures. Paying by the minute without a minimum that matches the skill is how you buy fragments.
What to copy from Index, and what not to
Copy the seriousness. Humanoids need real work in real rooms, at a volume vendors did not historically deliver. Copy the idea that quality is a pipeline with more than one stage, not a zip file. Copy the willingness to throw takes away.
Do not copy the belief that the app is the specification. Do not copy hour-counts as the success metric. Do not copy “diversity we could not define up front” as an excuse to skip the sentence. Long-tail variation is valuable after you can reject the takes that are not the job. Before that, it is noise with extra countries. Do not copy a consumer privacy notice and call it a rights pack for a commercial policy you might ship or sell. Do not stand up a marketplace that sends a stranger with a camera into a house unless you have written who else may appear.
If a catalogue item already matches the skill and the rig, look at it with the actual trainer before you build an app. Custom when the object set, the kitchen, or the faces decision is yours. Starting with a consumer funnel because it looks like Figure is how briefs balloon and recapture becomes a line item.
Physical AI is not short of capture software. It is short of specified, consented behaviour in the rooms the robots will enter. An app is how you recruit people into that capture. The brief is still the dataset: the sentence, the mount, the reject codes, the split, the pack. Pay by the minute if that is how you staff the loop. Do not confuse the pay mechanism with the job. Hour-counts will always look like progress. Task-not-completed and camera-on-ceiling are how you find out it was not.
If you are collecting first-person chores this quarter, write the skill before you ship the APK. Gold before the first payout. Split by person. Faces out unless the policy needs them. An app can waitlist ten thousand people by Friday. It still cannot tell you whether the mug reached the sink.
People4.ai would rather kill a bad spec on the first call than recapture in month two. hello@people4.ai · Contact
FAQ
Can we start with an app and write reject codes once we see the footage?
You will be paying for the learning curve, and you will keep takes you cannot use. Write the skill, the mount, the faces rule and the reject codes before anyone records. Gold items first. Then recruit.
Is paying by the minute a quality problem?
It is a metric problem. Minutes reward duration. The job is a completed skill from the right camera. Put a reject for task-not-completed and too-short, and split eval by person so prolific creators do not sit in train and test.
Do we need our own app to get Index-like diversity?
You need the sentence and a way to recruit people who can do that sentence. An app is one funnel. A catalogue match or a specified collection can be another. Diversity without reject codes is lifestyle footage.
Related reading
How to brief a human data collection so you do not recapture · Humanoid labs are building data factories, not just models