This is Figure AI's own announcement, published August 25, 2026, of Index — a crowdsourced data-collection programme in which members of the public record video of themselves performing everyday physical tasks and are paid for accepted submissions. The data trains Helix, Figure's vision-language-action stack. The post is the sole published account of the programme's scale and pipeline, and every quantity in it is Figure's own; no third party has audited the dataset, the payouts, or the diversity counts.
The argument
Figure frames the release around a single claim: that generalization in robotics is a data problem, and that the data required does not exist on the internet. "The data needed to scale a truly general purpose robot doesn't exist on the internet — it has to come from the real world: a global sampling of physics captured across every environment on earth." The company presents this as an application to robotics of what it calls "the central finding of the last decade of AI research," and says internal generalization results are "already validating this thesis," with detail promised later.
The post also records a strategic reversal worth noting because it is a claim about the market rather than about Figure: "Prior to this we tried buying data. Vendors couldn't hit the throughput, diversity, or quality bar Helix requires, so we built the pipeline ourselves."
Reported figures
Figure states that Index ran in stealth for four months before the August 25 launch on Google Play and the App Store.
| Measure | Reported value |
|---|---|
| App downloads | 264,000 |
| Countries | 108 (the post's summary bullets say "100+") |
| Weekly active users | over 44,000 |
| Videos uploaded | over 16 million |
| Ingest rate | 30 minutes of video per second |
| Paid to contributors to date | $15 million |
| Forward commitment | over $1 billion on data and compute over 12 months |
Three of these carry qualifications the company's own text supplies. The "4.9 years of human work every day" figure is Figure's restatement of the 30-minutes-per-second ingest rate, not a separately measured quantity. The $1 billion is a forward commitment covering data and compute together, against $15 million in contributor payouts to date. And the post's summary bullets and body give different country counts, 100+ and 108; both are the company's text.
The diversity figures are stated per 1,000 hours collected, not as dataset totals: 373 unique tasks, 1,146 unique manipulated objects, and 116 unique environments. Figure attributes this variation to the contributors rather than to a collection design: "Every new Creator brings an unseen environment, unfamiliar objects, and their own idiosyncratic way of completing a task, the kind of long-tail variation that's nearly impossible to define upfront." Named task examples run from cooking, cleaning and laundry at home to logistics centres, restaurants, factories and offices, and include "cleaning kitty litter, changing oil, busing restaurant tables."
The Creator model
Figure calls its contributors Creators, and the arrangement runs in two directions that most coverage carries only in one. A contributor can record tasks in their own home or workplace — "making beds, folding laundry, serving guests at a local cafe, stocking retail shelves." Alternatively, a customer can book a Creator through the app to come to their home or business and perform tasks while recording: "you can book a Creator through the app who comes to you to help with daily tasks and chores. We'll even send one to your business."
The programme is a paid gig-work arrangement whose output is training data; see Data-labeling and data-collection labor. The post gives a cumulative payout figure but no rate, no acceptance rate, and no terms.
Processing pipeline
Figure describes rebuilding its data infrastructure around "constraints more typical of a consumer app: 24/7 availability, continuous large-scale compute for processing, and real-time feedback to users that helps improve future collections." Submissions pass five stages:
- Filtering — automated screening for technical, visual, and semantic quality.
- Fraud review — human analysts audit samples at the user level for deliberate attempts to evade the automated filters.
- Deduplication — each video segment is embedded and discarded if above a similarity threshold with previously accepted data.
- Rebalancing — by task quota, based on how well a submission matches its selected task, and by embedding-based clusters capturing variation beyond task labels.
- Annotation — hierarchical text captions generated for every episode.
The presence of a dedicated human fraud-review stage, and its stated framing around "deliberate attempts to evade" the automated filters, is a consequence of paying per submission that the post states directly.
Stated destination
"Index is laying the groundwork for ordering robots as a service. Today, you have people coming to help clean your house; eventually, a robot will do everything for you." Figure says it is "now on a path to 100x."
Comparison with the alternative approach
Figure's account is one of two vendor statements of what robot training data should look like published within months of each other, and they describe opposite strategies. Skild AI sets out four collection modes scaled in house, traded off against hardware proximity, diversity, and scalability, with roughly three dollars of quality control spent per dollar of collection (Introducing S1: In-Context Learning for Robotics). Figure reports having tried the vendor market, found it inadequate, and built a crowdsourced consumer pipeline instead. Both are the developers' own accounts of their own methods; neither has been independently evaluated, and no published result compares the resulting datasets.
Provenance
Published on figure.ai on August 25, 2026, authored to the company rather than a named individual. Pulled and verified August 26, 2026 from the developer's own domain with cache bypassed; OpenGraph metadata carries the title, URL and date, and glitchwire.com covered the announcement independently the same day.
Relationships
- related: Figure AI — the publisher; the page carries these figures.
- related: Helix (Figure AI Vision-Language-Action model) — the model the dataset trains.
- contradicts: Introducing S1: In-Context Learning for Robotics — sets out the opposite data-acquisition strategy, scaling collection in house rather than crowdsourcing it.
- related: Skild AI — the competing approach.
- supports: AI Robotics — a vendor account of the data constraint on general-purpose robotics.
- related: Data-labeling and data-collection labor — the Creator programme is a paid data-collection workforce.
- related: Training Data Walls — advances the claim that internet-scale data is unavailable for physical tasks.