Figure’s $1 Billion Robot-Data Bet Turns Household Chores Into Gig Work

A new marketplace for robot training

Figure AI’s latest move is not another glossy robot demonstration. It is an attempt to industrialize the collection of human behavior. On August 25, 2026, the humanoid robot company brought Index out of stealth, describing it as a Figure-exclusive pipeline for collecting the real-world physical data needed to train Helix, its robot AI stack [1]. The premise is direct: people do ordinary work, cameras capture the work, software filters and labels the recordings, and the resulting data helps robots learn how physical tasks are actually done [1].

The scale Figure is claiming is unusually large for embodied AI. The company says Index has crossed 264,000 app downloads across 108 countries, has more than 44,000 weekly active users, and has already received more than 16 million videos from contributors it calls “Creators” [1]. It says the app is processing 30 minutes of video uploads every second, equivalent to 4.9 years of human work uploaded each day [1]. The company also says it has paid $15 million to creators so far and is committed to spending more than $1 billion over the next 12 months on data and compute [1].

That makes Index more than a side project. It is Figure’s answer to one of the hardest problems in robotics: how to give a machine enough examples of kitchens, bedrooms, shelves, boxes, plants, laundry, tools and human hands to act safely and usefully outside a tightly controlled laboratory. Figure says the internet does not contain the right data for general-purpose robots, because the missing ingredient is not text or still images but a global sample of physics across real environments [1].

How Index works

Index operates as both a data-collection app and a labor marketplace. Figure says anyone can become a creator and record real tasks in a home or workplace, while customers can also book a creator through the app to come perform chores or business tasks [1]. The work examples named by Figure include making beds, folding laundry, serving guests at a café, stocking retail shelves, cleaning, cooking and other household or commercial tasks [1].

Reporting by Spatial Insider described the app as a way for approved contributors to record everyday work for Figure’s humanoid robots, with recordings covering cooking, cleaning, shelf-stocking, equipment repair and other physical tasks in real homes and workplaces [3]. Gate’s report summarized the launch as an iOS and Android effort to collect real-world task data from global users for humanoid robot training [4]. In practical terms, Figure is trying to convert dispersed human activity into a structured training resource.

The company’s own description of the pipeline is important because raw video alone is not automatically useful robot data. Figure says each submitted recording passes through five stages: filtering, fraud review, deduplication, rebalancing and annotation [1]. Automated filters screen for technical, visual and semantic quality; human analysts audit user-level samples for deliberate attempts to evade the rules; similar clips are discarded to maintain diversity; the remaining material is rebalanced by task quotas and embedding clusters; and hierarchical captions are generated for every episode [1].

Figure says that for every 1,000 hours collected, Index contains 373 unique tasks, 1,146 unique manipulated objects and 116 unique environments [1]. Those numbers matter because robotics failure often emerges from the long tail: the odd drawer handle, the soft towel, the differently shaped mug, the cluttered counter, the poorly lit garage, the restaurant table after a rush. A humanoid robot that only sees a model kitchen may perform impressively in a video demo; one trained on thousands of idiosyncratic homes and businesses has a better chance of recognizing that the world rarely looks like a demo set.

The billion-dollar thesis

Forbes framed the announcement as a three-part bet: that volume and diversity are decisive, that passive or egocentric video can provide enough signal for transfer to robots, and that synthetic data alone cannot solve the problem [2]. Figure’s public language strongly supports that reading. The company says vendors could not meet the throughput, diversity or quality bar Helix required, so it built the collection system itself [1]. It also says the data needed for a truly general-purpose robot “has to come from the real world” rather than from the existing internet [1].

This is a strategic line in the sand. In language AI, scaling laws made it plausible that more data and more compute would reliably improve performance. Robotics is messier. A robot has to infer not just what an object is, but how heavy it might be, whether it will deform, how much force to apply, what the human demonstrator intended and what motion will be safe around people. The video of a person folding a shirt does not directly contain robot joint commands, gripper pressure or force-torque readings. The value of Index depends on whether Figure can transform human video into representations Helix can use to control machines.

Forbes contrasted Figure’s real-world push with approaches that lean heavily on synthetic data and simulation, citing Nvidia’s physical-AI strategy as a cheaper and faster counterexample [2]. The tension is central to the industry. Simulation can create endless variations, but simulators may miss the strange physical edge cases that make homes and businesses difficult. Real video captures those edge cases, but it is expensive, noisy and only partially labeled unless the pipeline can extract the right structure.

Figure is choosing to buy reality at scale. Its $1 billion commitment is not only a compute budget; it is also an admission that humanoid robotics may need a new data supply chain. If large language models were trained on the web, general-purpose robots may be trained on paid recordings of people washing dishes, moving boxes, making beds and tidying rooms.

What this signals about the home market

The task mix gives away part of the roadmap. Figure mentions logistics centers, restaurants, factories and offices, but the examples it repeatedly highlights include household chores such as cooking, cleaning, laundry, making beds, watering plants, loading groceries and taking out trash [1][2]. Spatial Insider also noted that the recordings are aimed at home and work tasks and that Figure’s larger goal is a network that can send people to do tasks now and robots later [3].

That matters because the home is one of robotics’ most difficult markets. Factories are structured. Warehouses can standardize bins, routes and lighting. Restaurants and retail spaces are variable but still partly designed for repeatable work. Homes are different: every home has its own layout, clutter, surfaces, pets, appliances, fragile objects and social rules. The reason a household humanoid is so valuable is also the reason it is so hard to build.

Index is therefore not only a training dataset. It is a map of intended deployment. By collecting video in kitchens, bedrooms, living rooms, cafés, retail aisles and business backrooms, Figure is trying to teach its AI the places where a general-purpose humanoid would have to work. The company’s own “next steps” language is explicit: Index is laying the groundwork for ordering robots as a service, with people helping clean houses today and robots eventually doing the work [1].

The labor paradox

Index also introduces a sharper social question than most robot demos. The people paid to create the dataset may be recording the very tasks Figure hopes robots will later perform. Figure calls them creators, and the word is not accidental: it puts the work inside the broader creator economy rather than the older language of data labeling or gig labor [1]. But the economic structure is familiar. People perform small units of work, the platform coordinates demand, and the output becomes an asset owned by the company.

The immediate benefit is real for participants who are accepted and paid. Figure says creators have already earned $15 million [1]. The broader question is how value is divided if the data they produce helps automate future service work. If a creator films laundry, dishwashing and shelf-stocking today, and those recordings help train a robot that competes for those tasks tomorrow, Index becomes both a wage opportunity and a displacement mechanism.

Privacy is another friction point. Spatial Insider reported that Figure’s Index privacy policy covers location data and photos, video, audio or other recordings of a creator and their surroundings, and that it allows personal data to be disclosed or sold to commercial purchasers for purposes including marketing, analytics, research and product development, while offering an opt-out process [3]. That matters because Index is not collecting abstract data. It is collecting images and sounds from homes and workplaces, often the most private physical environments people occupy.

The open questions

Figure’s numbers are impressive, but the key performance evidence remains undisclosed. Spatial Insider noted that Figure has not yet shown how much Index improves Helix or released performance results tied to the new dataset [3]. Figure says internal generalization results are validating its thesis and that more detail will be shared soon [1]. Until then, the industry has scale metrics, not proof of capability.

The unanswered questions are technical, commercial and ethical. Technically, can Figure convert enormous volumes of human video into reliable robot behavior? Commercially, can a billion-dollar data operation produce robots useful enough to justify the spend? Ethically, can a system that records private spaces and pays people to train their potential replacements earn durable trust?

The significance of Index is that Figure is no longer merely presenting humanoid robots as hardware. It is building a human data refinery around them. If the bet works, household chores and gig tasks recorded in 2026 could become the training substrate for a new class of general-purpose robots. If it fails, Index may become a cautionary example of how difficult it is to translate human motion into machine competence. Either way, Figure has made the robot race more concrete: before humanoids can work like people, companies may first need millions of people to show them how.

Sources from the last 72 hours

1. [1]Introducing Index: Building The World’s Largest and Most Diverse Physical DatasetAug 25, 2026, 12:00 AM UTC
2. [2]Figure’s Billion-Dollar Bet: Gig Platform For Humans To Generate Robot Training DataAug 26, 2026, 5:40 PM UTC
3. [3]Figure Launches Index to Pay People for Videos That Train RobotsAug 25, 2026, 12:00 AM UTC
4. [4]Figure Launches Index Project to Crowdsource Robot Training Data, Pays $15M to ContributorsAug 26, 2026, 5:58 AM UTC

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.