The humanoid race is industrialising first-person human video. Figure Index is paying creators by the minute to record real tasks in homes and workplaces because Helix still has to learn those rooms.
Figure brought Index out of stealth in August. The company is still scaling it in October. The pitch is not a research dump and it is not a hall demo. It is a consumer-facing pipeline: people record first-person video of real work in the places robots are supposed to enter, then those files train Helix. Figure has talked about tens of millions of videos. Creators are paid by the minute. Some are sent recording devices. The other side of the app books a creator into a home or a business so the task happens on someone else’s floor, under someone else’s lighting, with someone else’s objects. Before that, Figure says it tried buying vendor data and could not get the throughput, the diversity or the quality Helix needed, so it built the pipeline itself.
That is the market post. Not “what is egocentric.” The bottleneck moved from a model announcement to an industrial capture loop, and a well-funded lab decided the loop was too important to rent. A data factory is still a written skill, a rig that actually sees the grasp, reject codes, a freeze for eval, and people whose job is to say no. Headcount and an app store listing without that sentence is a film shoot with a robotics logo. You can spend a quarter uploading takes and still have lifestyle footage.
Why they stopped shopping for a zip
Third-person beauty shots do not show occlusion the way a chest camera does. Simulation does not give you a damp sponge or a drawer that sticks. Teleoperation is useful and it is a different object: robot cameras, robot dynamics, a different rights conversation. Found first-person video on the open web is first-person-shaped. It is rarely a spec, and it is rarely a chain that covers commercial robot training. Figure’s own account of the vendor market is the blunt version of a conversation every humanoid team is having. Buying “humanoid data” often means buying the cheapest of those four objects. The policy then fails on the mug that was never in the files.
Index is an attempt to industrialise the second object: a person doing the task, from a wearable, in a real room. Cooking, cleaning, laundry, a cafe, a warehouse aisle, a factory bench. The company has described creators in more than a hundred countries, a consumer-app ingest rate, and a five-stage pipeline after the upload – filtering, fraud review, deduplication, rebalancing, annotation. Helix 2.5 was pretrained on Index and then walked into thirty Bay Area homes it had never seen. Holding the rest of the recipe fixed, Figure said Index pretraining moved zero-shot success from 9 percent to 56 percent on three long-horizon behaviours. That is an existence proof that first-person human video, at this scale, is now a training input with a product number attached. It is not a licence for your kitchen, and it is not a reason to skip the spec.
Imagine the skill the app is asking a stranger to perform. Fold a towel on a bed that is not the lab’s bed. Unload a dishwasher that is not the lab’s dishwasher. Pick a mug off a cluttered counter in a house the policy has never entered. A camera in the corner of the room shows a body moving through space. You can label the scene. You still do not see the grasp, the occlusion when the hand covers the mug, or the moment they change their mind and use a different cup. A chest-mounted or glasses-mounted camera sees the cupboard handle, the mug, the other objects in the way. For a policy that has to reach and place, that difference is the dataset. Index is betting that if you pay enough people, in enough rooms, the long tail of objects and grips shows up. The bet only works if the take is actually the skill.
Crowdsourcing does not retire the reject codes
Scale without a reject code is how you pretrain on the wrong thing at great expense. Figure knows that – the pipeline talk is about fraud, near-duplicates, and throwing away clips that do not match the task. Every other lab that copies the shape of Index without copying the no still has to write the codes. Task not completed. Camera on the ceiling. Beauty filters. Digital zoom that destroys the hands. A take that is walking through a house while the dashboard counts a minute. Automation can flag black frames. A person still has to apply the codes that mean “this is not the job.” If gold items do not exist before production, the first month is the network learning your kitchen on your budget.
Faces are not a sidebar. First-person cameras see kitchens, children, post on the fridge, laptops on the counter, a visitor walking through. If faces are not required for the skill, keep them out. Blurring later is a processing choice. It is not retroactive consent, and it often destroys the hands you needed. Booking a creator into someone else’s home or business adds a second consent conversation: the people who live or work there, the screens in frame, the customer in the cafe. A pilot in a lab is one pack. A wearable in a real home is another. Found “POV” video is a third, and usually a dead end for commercial training. Internal capture does not make the legal problem disappear. It makes it your problem. A factory without a pack is a factory that cannot sell the model, or cannot survive diligence.
Eval-by-person still applies when the footage comes from an app. A random split of clips is a test of whether the policy remembers those hands, those rooms, that way of folding a towel. Split by person so the same creators are not in train and test. Public robot boards will not save you. They have the same contamination story the coding labs just admitted, with a mug. If the same three people filmed the SFT-like demonstrations and the quarterly eval, you have style leakage even when the kitchens look different. Freeze a set that looks like the rooms you will actually enter. Hold those people out. Re-run when you claim Helix-class transfer.
Four objects still pretending to be one zip
Crowdsourced first-person video. Useful. Different files, different QA, different consent from teleop.
Teleoperation. A person driving a robot. Useful. Robot cameras, robot dynamics, a different rights conversation.
Third-person hall video. Useful for checking whether the task finished. Almost none of the viewpoint the policy will have once it is on a head, a chest, or a wrist.
Found POV. First-person-shaped. Rarely a spec, rarely a chain that covers commercial robot training. Europe is tightening the web-copyright side this autumn. Homes and faces were never covered by “it was online.”
If the brief says “humanoid data,” procurement will buy the cheapest of those four and the policy will fail on the mug that was never in the files. Mix them if you must. Label them. When the mug fails, you need to know which object to collect more of. Index is Figure choosing to own the first object at industrial scale. That does not collapse the other three into it. A lab that quotes Index in a slide and then fine-tunes on hall video plus a YouTube dump has not copied Index. It has copied the hour-count.
The papers this year already said hours of aligned human video move robot performance. DreamDojo-scale pretraining, EgoScale’s log-linear story, the retargeting work that tries to turn a human grasp into a robot view. Figure’s contribution is operational: pay people, send devices, book the house, run fraud review, keep going in October after the launch blog. That is industrialisation. It is also a reminder that the hard parts were never the model card. They were whether the skill was written down, whether the camera saw the hands, whether a person could reject the take, and whether anyone in the room had said yes.
What this means if you are not Figure
You probably should not copy their app. You should copy the sequence. Write the skill in one sentence. “Pick a mug off a cluttered counter, chest camera, indoor daylight, faces out, start to finish including the failures.” Decide teleop versus human demonstration versus first-person video. Lock the rig. Lock faces. Look at a catalogue match with the actual trainer before you stand up a consumer ingest. Custom when the object set, the kitchen, or the faces decision is yours.
If you are standing up even a small capture loop this quarter, treat it as an operation with gold and reject codes, not as a film shoot and not as a gig marketplace with a robotics logo. Headcount comes after the sentence. Starting with a warehouse of cameras, or with an open call for “everyday tasks,” is how briefs balloon and recapture becomes a line item. Everyday is not a skill. Making a bed in this lighting, with this linen, to this finished state, with the failures kept, is a skill. A creator network will give you diversity you cannot script. It will also give you a long tail of takes that are not the job. The factory is the sentence that lets you throw those away.
Rights travel with the camera. Training a manipulation policy is not a marketing still. If marketing stills or a launch video might appear later, say so now. You will not get a clean expansion of rights after the clips are famous inside a weight. A booked home is not a demo hall. The people who did not put the camera on still have to be in the pack, or they have to be out of frame on purpose.
Physical AI is not short of model announcements. It is short of specified, consented behaviour in the rooms the robots will enter. That is why Index exists. That is why Figure says it stopped buying zips and built a pipeline. That is why a generic “egocentric for robotics” brief still produces lifestyle footage. The industrial fact this autumn is that one lab is paying the world by the minute to film the work. The operational fact is unchanged. Spec the skill. Reject the rest. Hold people out of eval. Put faces and homes on a yes or no line before anyone records. Scale on the wrong object is not a smaller version of Index. It is a different dataset with a bigger bill.
People4.ai would rather kill a bad spec on the first call than recapture in month two. hello@people4.ai · Contact
FAQ
Is crowdsourced first-person video the same as teleoperation?
No. Crowdsourced video is a person doing the task from a wearable. Teleop is a person driving a robot. Both can train policies. They are different files, different rights, different QA. Mixing them in one zip called “humanoid data” is how you cannot debug the next failed mug.
Does paying creators by the minute replace reject codes?
No. Payment buys volume. Reject codes decide whether the volume is the skill. Task not completed, camera on the ceiling, non-consented faces, beauty filters – those still need a person who can say no. Scale without that person is lifestyle footage.
Can we skip a rights pack if the footage is “just chores in a home”?
Homes, faces, children, screens and a booked workplace are the pack. “It was a household task” is not consent. If faces are not required, keep them out before the upload, not with a blur after the fact.
Related reading
Physical AI is suddenly hungry for first-person human video · Humanoid labs are building data factories, not just models