Introduction
On August 25, 2026, humanoid robotics company Figure came out of stealth with Index, a smartphone app that pays ordinary people to film themselves doing everyday chores. The company says it has already paid $15 million to users after four months of quiet operation, and that it intends to spend more than $1 billion on data and compute over the next 12 months.
The pitch is a straightforward bet on scale: the physical data a general-purpose robot needs is not on the internet, so Figure is buying it directly from the people who generate it.
What Figure Has Collected
The numbers Figure published for its four-month stealth period:
- 264,000 app downloads across 108 countries
- 44,000+ weekly active users, whom Figure calls Creators
- 16 million videos uploaded
- 30 minutes of video ingested every second — what Figure describes as 4.9 years of human work uploaded per day
- $15 million paid out to Creators to date
Figure also reports a diversity measure rather than a raw volume one: per 1,000 hours collected, the dataset contains 373 unique tasks, 1,146 unique manipulated objects and 116 unique environments. That framing matters more than total hours, because a million clips of the same kitchen teaches a robot very little.
The tasks range from household work — cooking, cleaning, laundry — to commercial settings including logistics centers, restaurants, factories and offices. Figure cites cleaning kitty litter, changing oil and busing restaurant tables among the more obscure submissions. Creators can record in their own home or workplace, and the app also lets users book a Creator to come to their home or business and do chores while filming.
Why Buy Data This Way
Figure says it tried the conventional route first: "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."
The underlying claim is that generalization in robotics is a data problem before it is an architecture problem — the same scaling laws story that played out in language models, applied to physical tasks. Crowdsourced human video is one answer to the embodied AI data shortage; synthetic data from simulation is the other, and most labs use both.
The long tail is the argument for crowdsourcing specifically. As Figure puts it, every new Creator brings "an unseen environment, unfamiliar objects, and their own idiosyncratic way of completing a task" — variation that is nearly impossible to specify in advance in a lab or a simulator.
Ingesting consumer video at this rate forced Figure to rebuild its data infrastructure around consumer-app constraints. The processing pipeline has five stages: automated filtering for technical, visual and semantic quality; human fraud review auditing users for deliberate evasion; embedding-based deduplication that discards clips too similar to accepted data; rebalancing against task quotas and embedding clusters; and finally hierarchical text captioning of every episode.
What Is Not in the Announcement
Figure says the generalization results from training Helix on this data are "already validating this thesis" and that it will share more soon. That is the load-bearing claim of the whole post, and it arrives with no benchmarks, no evaluation protocol and no comparison against models trained without Index data. Until those appear, the verifiable part of the announcement is the collection operation, not its effect on robot performance.
Two other gaps are worth noting. First, human egocentric video is not robot demonstration data: it lacks the joint states, forces and action labels a policy ultimately needs, and bridging that embodiment gap is an open research problem rather than a solved preprocessing step. Second, a global network of paid contributors filming inside homes and workplaces raises consent and privacy questions — for bystanders and employers as much as for Creators — that the announcement does not address.
Figure frames Index as "laying the groundwork for ordering robots as a service," with today's human Creators as a placeholder for tomorrow's machines. That is a long way from 16 million phone videos, but the funding commitment behind it is specific: 100x the current collection rate, and over $1 billion in the coming year.
Conclusion
Index is the largest public attempt yet to solve robot training data by paying consumers for it, and the operational numbers — $15 million out the door, 30 minutes of video per second — are real and substantial. Whether the data translates into robots that generalize is the question Figure has promised to answer and has not yet answered. The next release, with actual Helix results attached, is the one that will settle it.
Sources
- Introducing Index: Building The World's Largest and Most Diverse Physical Dataset — Figure, August 25, 2026