Index is a proprietary data-collection program and mobile application operated by Figure AI. It recruits people to record first-person video of everyday activities and uses the resulting material in Figure's robot-learning pipeline. Figure introduced the Index name publicly on August 25, 2026, after describing the app as having operated in stealth for about four months. The company said the material was intended to train Helix, its humanoid-control system.[1][3]

Index is best understood as a private collection and processing system, not as a downloadable research dataset. At launch, Figure reported 264,000 app downloads, more than 44,000 weekly active users, availability or activity across 108 countries, and more than 16 million uploaded videos. Those figures were company-reported product metrics. Figure did not release the underlying records, a dataset card, or an independent audit that would allow the counts or its claims about scale, quality, and diversity to be reproduced.[1][2]

Launch and program design

Figure said it built Index after purchased data from vendors did not meet its throughput, diversity, or quality requirements. The launch article described a Figure-exclusive pipeline for sourcing physical-world recordings from people in homes and workplaces. It also framed the project as a response to the relative shortage of real-world training material for general-purpose humanoid robots.[1]

The public app combines data collection with a service marketplace. Its current materials describe three related modes:[2][4][5]

A prospective creator can apply, receive a recording device if accepted, perform tasks in a home or workplace, and seek payment based on uploaded recording time.
A household user can request a creator to visit and help with chores while the work is recorded.
A business can arrange collection in settings such as logistics facilities, restaurants, factories, offices, retail locations, warehouses, or skilled-trade environments.

The United States App Store listing describes an application and review process with a waiting list, so an app download does not establish that a person became an approved or paid creator. It lists Figure AI Inc. as the seller, classifies the app under Business, limits the listed iOS version to iPhone with iOS 18 or later, and displays an 18+ age rating. Google Play lists the Android package as ai.figure.index. These storefront facts confirm that the apps were publicly listed, but their descriptions and data-safety answers were supplied by Figure.[4][5][6]

Reported scale and metric boundaries

Figure's launch materials present several large numbers. Each measures a different stage or aspect of the program, and none should be substituted for another.[1][2]

Figure-reported measure at launch	What the measure describes	What it does not establish
264,000 app downloads	Downloads of the application	Unique people, approved creators, active contributors, or paid contributors
More than 44,000 weekly active users	Activity within a company-defined weekly window	The counting method, contribution volume, creator approval, or payment status
More than 16 million uploaded videos	Videos submitted to the app before the full disclosed pipeline is accounted for	Retained or accepted training data, usable episodes, or robot trajectories
30 minutes of video uploads per second	Aggregate duration entering the app at the reported moment	Post-filter processing speed, accepted-data throughput, or training throughput
4.9 human-years per day	Figure's time-equivalence conversion from the upload rate	Years of distinct labor, annotations, or robot operating experience
$15 million paid to creators	Aggregate payout reported by Figure	A per-minute rate, per-video rate, distribution, eligibility rule, or future guarantee
More than $1 billion over the next 12 months	A planned commitment for data and compute	Money already spent or allocated only to Index collection

Figure also reported that each 1,000 hours collected contained 373 unique tasks, 1,146 unique manipulated objects, and 116 unique environments. The company did not publish the taxonomies, counting procedure, sampling window, confidence intervals, or overlap rules behind those diversity measures. It is therefore possible to report the numbers as Figure's definitions, but not to compare them directly with independently documented corpora.[1]

Video duration is also absent from the upload count. Without a public duration distribution, 16 million videos cannot be converted reliably into hours. Conversely, the 30-minutes-per-second inflow implies an aggregate duration rate but does not reveal how many uploads passed filtering or how much material remained after deduplication and rebalancing. The product page's 108-country counter and the store description's broader wording about more than 105 countries are compatible at a coarse level, but only Figure's launch article gives the exact 108-country claim used here.[1][2][4]

Processing pipeline

Figure described five stages between upload and annotated episodes: automated filtering, human fraud review, deduplication, rebalancing, and annotation. Automated systems first screen submissions for technical, visual, and semantic quality. Human analysts then audit samples at the user level for deliberate attempts to evade those filters. Figure said it embeds video segments and discards segments above a similarity threshold relative to previously accepted data.[1]

The remaining material is then rebalanced using task quotas and embedding clusters. Figure said the quotas consider how closely a submission matches its selected task, while the clusters are intended to capture variation that a task label may miss. The final disclosed step generates hierarchical text captions associated with every episode.[1]

This description establishes that uploaded videos can be filtered, segmented, rejected as similar, regrouped, and converted into episodes. It does not disclose the acceptance rate, number of retained segments, number or duration of episodes, similarity model or threshold, captioning model, review protocol, or error rate. It also does not show that every paid minute was retained, that every upload became one episode, or that an episode contains robot control actions.

That last distinction is important in robot learning. A human first-person video records appearance and activity from a person's body, while a robot trajectory normally includes actions, poses, joint states, or another representation tied to a robot embodiment. Figure publicly described filtering and captioning, but did not disclose whether Index also performs pose estimation, action reconstruction, retargeting, tactile alignment, or visual embodiment conversion. The public record therefore supports calling the launch corpus human activity video intended for robot training, but not counting its uploads as robot trajectories.[1]

Relationship to Helix and Project Go-Big

Figure says Index supplies human data for Helix, a vision-language-action model intended to control its robots. The Index launch article said internal generalization results supported Figure's data-scaling thesis, but it did not publish the evaluated checkpoint, data subset, benchmark, task definitions, baselines, or failure cases. Its statement about generalization is therefore a company-reported internal result, not a public experiment.[1]

Index follows an earlier Figure initiative called Project Go-Big. Announced in September 2025, that program involved egocentric human recordings in properties connected with Brookfield and a company-reported demonstration of human-video-to-robot navigation transfer. Figure said that particular navigation system was trained on human video without robot demonstrations.[10] TIME separately observed headset-wearing human pilots collecting task video at Figure's campus in August 2025 and reported mixed robot demonstrations, including repeated failures during towel and laundry tasks.[11]

The two initiatives share a focus on egocentric vision, human demonstrations, and Helix. However, Figure's Index announcement did not state that Index replaced Project Go-Big, that the two names cover the same corpus, or that all earlier Go-Big recordings are included in the Index counts. Their precise organizational and data relationship was not publicly documented by the cutoff.[1][10]

Privacy, consent, and contributor terms

Index-specific privacy policy

Figure published a separate Index Privacy Policy, last updated August 11, 2026. It says it applies only to people collecting data through the app, whom it calls Data Creators. People using other Figure services are directed to Figure's general privacy policy. The Index policy says that using or accessing the app acknowledges and consents to the practices it describes, and it references separate Application Terms of Use.[7][8]

The Index policy lists profile and contact information, device and network information, age or date of birth, geolocation associated with content, and voluntary identifying information. Its sensory-data category includes photos, videos, audio, or other recordings of a creator or the creator's environment, together with characteristics derived from those recordings. That scope can include more personal information than is visible in a narrow description of task video alone.[7]

Figure says it may use creator data to operate and improve the application, improve Figure's and its affiliates' artificial intelligence models, conduct testing and research, develop products, prevent fraud, enter contracts, manage and pay creators, and communicate with them. The policy lists service providers, affiliates, creator-management partners, employers, and commercial purchasers as potential recipients. It specifically describes selling or disclosing personal data collected through the app to commercial purchasers for their own marketing, analytics, research, product development, or other purchaser-determined purposes.[7]

Rights depend on a creator's jurisdiction. The policy lists access, rectification, erasure, objection, restriction of processing, and an opt-out from sale to commercial purchasers, with requests directed to Figure's privacy email address. It bars people under 18 from registering or using the application themselves. The policy provides no specific retention period for creator data. Figure's separate general privacy policy contains generic retention language, but that document is not the creator-specific policy and should not be used to fill the Index policy's omission.[7][8]

App store disclosures

Apple's privacy label says Figure may collect coarse location, email address, name, user ID, and other data linked to a user. It separately lists device identifiers, product interaction, crash information, and diagnostics as data that may be collected without being linked to identity. Apple explicitly says the declaration came from the developer and was not verified by Apple.[4]

Google Play's data-safety page says Figure may share approximate location and personal information. It lists approximate location, personal information, app interactions, device identifiers, crash logs, and diagnostics among data that may be collected. The page also says data is encrypted in transit and that users can request deletion. Google presents those statements as information supplied by the developer, not as a Google audit.[6]

Neither captured store summary separately names photos or videos among its displayed privacy categories, while the Index policy expressly covers sensory recordings. This difference should not be read by itself as proof of a violation because platform taxonomies and summaries differ. It does mean that a prospective creator needs the full Index policy, rather than only a store label, to see Figure's broader description of recording data.[4][6][7]

Public documentation gaps

The creator policy refers to Application Terms of Use, but a public Figure-specific copy was not linked from that policy, the Figure navigation, or the captured store listings as of August 25, 2026. Figure's publicly linked Terms and Conditions were last updated in 2023 and identify themselves as terms for the www.figure.ai website. They do not establish the app's content license or creator-payment rules.[7][9]

As a result, the public materials do not fully document who owns uploaded content, what license a creator grants, the rate and currency of payment, how accepted minutes are measured, when submissions can be rejected, or how taxes, disputes, and account termination affect payment. They also do not spell out a bystander-consent workflow, workplace authorization procedure, on-device redaction system, face or license-plate treatment, or rules for confidential material. This is a boundary of the public record, not proof that the app contains no controls or that no additional terms appear during onboarding.

These questions are material because wearable computer vision can capture people and spaces beyond the camera wearer. Research on egocentric privacy has demonstrated that first-person video may reveal attributes and routines of the wearer, while separate wearable-camera work has focused on the privacy of bystanders who enter the field of view.[16][17]

Public evidence and reproducibility

Index was not publicly reproducible at launch. Figure did not release a dataset download, public sample, dataset card, schema, license, API, manifest, checksums, task taxonomy, annotation guide, sensor specification, retained-hour count, episode count, acceptance rate, train-test split, or evaluation protocol. No public artifact connected a named Index snapshot to a named Helix checkpoint and a repeatable benchmark. App store listings independently confirm distribution of the app, but they do not validate Figure's dataset metrics or quality claims.[1][4][5]

Academic datasets illustrate what more detailed documentation can make possible. Ego4D reported 3,670 hours from 931 camera wearers across 74 locations in nine countries, described consent and de-identification procedures, and released data for research benchmarks.[12] EgoVerse reported 1,362 hours, about 80,000 episodes, 1,965 tasks, 240 scenes, and 2,087 demonstrators, along with standardized formats and downstream tooling.[13] Those projects have different goals and collection rules from Index, so their counts are context, not a ranking.

Academic robot-learning systems also show why raw video and robot-ready data are separate layers. The 2026 Ego2Robot preprint converts human egocentric video through action retargeting, robot-arm visual synthesis, and multi-level quality curation before treating it as robot training data.[14] UniDex, a peer-reviewed CVPR 2026 system, used human-in-the-loop retargeting and explicit three-dimensional representations to derive more than 50,000 robot-centric trajectories from egocentric sources.[15] These studies do not test Index. They demonstrate that imitation learning across human and robot bodies is a technical transformation problem, not merely an upload-count problem.

Figure may have proprietary processing beyond what it disclosed, and private data can still be useful for embodied AI. The evidence available at launch, however, supports only a narrower conclusion: Index is a large app-mediated human-video collection effort by Figure, with a described filtering and annotation pipeline and substantial company-reported participation. Its retained scale, training contribution, comparative diversity, privacy implementation, and effect on robot performance remained unavailable for independent verification.[1]

References

Figure AI, "Introducing Index: Building the World's Largest and Most Diverse Physical Dataset," August 25, 2026. https://www.figure.ai/news/introducing-index
Figure AI, "Index" product page, accessed August 25, 2026. https://www.figure.ai/index-app
Figure AI official X account, "Introducing Index," August 25, 2026. https://x.com/Figure_robot/status/2092303621392376314
Apple App Store, "Figure INDEX," accessed August 25, 2026. https://apps.apple.com/us/app/figure-index/id6800436796
Google Play, "Figure INDEX," accessed August 25, 2026. https://play.google.com/store/apps/details?id=ai.figure.index&hl=en_US&gl=US
Google Play, "Figure INDEX Data safety," accessed August 25, 2026. https://play.google.com/store/apps/datasafety?id=ai.figure.index&hl=en_US&gl=US
Figure AI, "Index Privacy Policy," last updated August 11, 2026. https://www.figure.ai/index-privacy-policy
Figure AI, "Privacy Policy," last updated January 21, 2026. https://www.figure.ai/privacy-policy
Figure AI, "Terms and Conditions," last updated February 1, 2023. https://www.figure.ai/terms-and-conditions
Figure AI, "Project Go-Big: Internet-Scale Humanoid Pretraining and Direct Human-to-Robot Transfer," September 18, 2025. https://www.figure.ai/news/project-go-big
Billy Perrigo, "The Robot in Your Kitchen," TIME, October 9, 2025. https://time.com/7324233/figure-03-robot-humanoid-reveal/
Kristen Grauman et al., "Ego4D: Around the World in 3,000 Hours of Egocentric Video," CVPR 2022. https://openaccess.thecvf.com/content/CVPR2022/papers/Grauman_Ego4D_Around_the_World_in_3000_Hours_of_Egocentric_Video_CVPR_2022_paper.pdf
Ryan Punamiya et al., "EgoVerse: An Egocentric Human Dataset for Robot Learning from Around the World," arXiv, 2026. https://arxiv.org/abs/2604.07607
Ye Wang et al., "Ego2Robot: Scalable Robot Data Synthesis from Egocentric Human Data," arXiv, 2026. https://arxiv.org/abs/2608.02580
Gu Zhang et al., "UniDex: A Robot Foundation Suite for Universal Dexterous Hand Control from Egocentric Human Videos," CVPR 2026. https://openaccess.thecvf.com/content/CVPR2026/html/Zhang_UniDex_A_Robot_Foundation_Suite_for_Universal_Dexterous_Hand_Control_CVPR_2026_paper.html
Yijiang Li et al., "EgoPrivacy: What Your First-Person Camera Says About You?," ICML 2025. https://proceedings.mlr.press/v267/li25ds.html
Haotian Li et al., "Extra-Lightweight AI-Based Privacy Preserving Framework for Egocentric Wearable Cameras," CVPR Workshop 2025. https://openaccess.thecvf.com/content/CVPR2025W/MTF/papers/Li_Extra-Lightweight_AI-Based_Privacy_Preserving_Framework_for_Egocentric_Wearable_Cameras_CVPRW_2025_paper.pdf

Editing: Index (Figure AI)