On the afternoon of August 25, Figure founder Brett Adcock posted the numbers on a project the company had run in stealth for four months. It is called Index, it was previously known internally as Project Go-Big, and it is now a consumer app on iOS and Android.
Four months produced 16 million video uploads from 108 countries. 264,000 downloads. More than 44,000 weekly active users. Footage arrives at 30 minutes per second, which works out to roughly 4.9 years of human work uploaded per day.
Figure has paid $15M to contributors so far and committed more than $1B to data collection and compute over the next 12 months.
That is under a dollar a video.
What is being collected
Index opens on a task list. Set a table, empty trash, fold clothes, tidy a room. A user picks one, mounts a phone or wears a head rig, records the chore in first person, uploads it, collects a bounty. Figure calls these people Creators.
The platform has a second side. A user can book a gig worker through the app to come and do the chores on site, capturing first-person footage while they work. In that transaction the customer pays, the gig worker does the labour, and Figure keeps the recording.
Collection is not limited to homes. Restaurants, retail stockrooms, and logistics facilities all sit on the list.
Figure also published a diversity figure: per 1,000 hours logged, the dataset carries 373 unique tasks, 1,146 distinct manipulated objects, and 116 unique environments.
Uploads pass five stages before they reach the model. Automated vision filters screen for resolution, frame rate, lighting, and task relevance. Human auditing teams check accounts for people gaming the bounty. Vector embeddings identify and discard near-duplicate segments. A rebalancing step applies task quotas and embedding clusters so easy tasks stop crowding out rare ones. Accepted episodes then get hierarchical text captions that align the high-level task with the low-level manipulation. Downstream of all five sits Helix, Figure’s robot neural network.
Three ways to solve the same problem
The comparison is what makes Index worth logging.
China builds buildings. On July 17 we covered the 64 data-collection centres Beijing has open with 20 more under construction, each a 1:1 replica of a real environment with people hired to work in front of the cameras. On August 21 we covered the CAICT count of 70 embodied-AI training grounds, 86% of them laid out as industrial manufacturing. That route runs on rent, rigs, and wages.
The other American route is hiring. On July 30 we covered Agility opening a Fremont hub and hiring about 200 people to teach Digit warehouse work. That route runs on salaries and benefits.
Figure built an app. No buildings, no hires. It bought coverage no fixed set of sites can produce: 108 countries, 116 distinct environments per 1,000 hours. That route runs on per-clip bounties.
Only the third one creates no employment relationship.
A bounty sheet is not a payroll
This is the load-bearing difference.
The people on camera in China’s collection centres are hired. The 200 people at Agility’s Fremont hub are employees. The 44,000 weekly contributors on Index are independent contractors paid per submission, with no guaranteed hours, no benefits, no minimum volume, and no claim on what the footage is later used for. Between the $15M already paid and the $1B committed sits more than an order of magnitude, and most of that $1B will not be spent as labour.
The gig-booking side deserves its own line. When a customer books someone through Index to come and do their chores, that customer pays a market rate for household help, the worker earns a wage for the visit, and Figure acquires the recording as a by-product of a transaction it did not have to fund. Every other collection method on the board has to pay for the labour it films. This one gets the labour paid for by a third party.
Note also which job Index does create. Stage two of the pipeline is fraud review, and Figure staffs it with human auditing teams watching for people gaming the bounty. A data-collection engine running at 30 minutes of footage per second generates exactly one durable human role, and it is policing the contributors.
Now read the task list against the sales pipeline. Restaurants, retail stockrooms, and logistics facilities are exactly where Figure already sells. Figure 03 has run logistics sequencing at BMW’s Plant Spartanburg, the company signed a deployment with Catalyst Brands in Reno, and BotQ recently built its 1,000th unit. The tasks in the app are not a random sample of human activity. They are the product roadmap, written a second way. The laundry you fold on camera trains the machine being prepared for your industry.
Adcock has stated the thesis plainly: hardware scale is no longer the primary hurdle, onboard intelligence is. The $1B commitment to data and compute is what that sentence looks like as a bet.
Which jobs should file this away
Domestic and home services. The tasks Index leads with are folding clothes, making beds, cleaning, and setting tables. Household work has long been considered automation-resistant because no two homes are laid out alike. The 116-environments-per-1,000-hours figure exists to attack precisely that argument.
Restaurant back and front of house. Restaurants are named in the collection scope. Food service has short repetitive motion sequences and a wide error tolerance, which puts it near the front of the queue behind logistics.
Retail stockroom work. Figure’s Reno deployment is already retail-side, and “retail stockroom” is broken out as its own collection category. Product and data are pointed at the same target, which means this is not exploratory collection.
Warehouse and logistics. This one is underway, not pending. What to watch is whether Index closes the edge cases the BMW line left open. The 1,146-distinct-objects figure is aimed at the standard objection that SKU variety makes warehouse automation uneconomic.
None of this justifies panic about timing. First-person human video carries no force feedback and no tactile signal, and whether it can support fine manipulation is still an open research question. One thing is settled, though. Since August 25, the work above has a public, structured, annotated dataset behind it, running at 373 labelled tasks per 1,000 hours and growing at 4.9 years of human work per day.
Being recorded is the step before being automated. That step now pays, at just under a dollar a clip.
