Figure's Index App Pays 44,000 People to Train Its Humanoid Robots

Figure has pulled back the curtain on what it calls the world's largest and most diverse physical dataset for training humanoid robots. The company, best known for its Figure 03 humanoid and its Helix vision-language-action model, has quietly run a crowdsourced data collection app for four months. Now branded as Index, the app pays regular people to record themselves doing everyday tasks so those videos can teach robots how to move through the physical world.

Four months ago, Figure launched the app in stealth to test whether it could gather the physical data Helix needs directly from humans at scale. Today the company is rebranding it as Index and shipping it on Google Play and the App Store. In that short window, it has built a global data flywheel that dwarfs anything publicly reported in robotics.

Numbers behind the flywheel

Figure disclosed a set of metrics that read more like a viral consumer app than a robotics research program:

• 264,000 app downloads across 108 countries
• Over 44,000 weekly active contributors, called Creators
• 16 million video uploads to date
• 30 minutes of video ingested every second, which the company estimates equals 4.9 years of human work uploaded per day
• $15 million paid out to Creators so far
• A commitment to spend over $1 billion in the next 12 months on data and compute, aiming to 100x the dataset

Per 1,000 hours of collected footage, Figure says Index contains 373 unique tasks, 1,146 unique manipulated objects, and 116 unique environments. That density of variation is the whole point. Every new Creator drops in an unfamiliar kitchen, an unseen object, or an idiosyncratic way of folding a shirt, feeding the long-tail diversity that supervised robotics datasets have historically lacked.

Why the internet isn't enough

Language models had it easy. The web is essentially a giant pretraining corpus for text. Robotics has no such shortcut. There is no equivalent of Common Crawl for how a human hand rotates a spatula, adjusts grip when a coffee cup is heavier than expected, or navigates a cluttered laundry room. Figure's argument is blunt: the data needed to scale general robotics doesn't exist online and has to come from the physical world.

The company says it initially tried buying data from vendors and found the results wanting. Suppliers couldn't hit the throughput, diversity, or quality bar Helix requires, so Figure built the pipeline itself as an in-house system for sourcing real-world physical data at scale. The crowdsourced approach also solves an economic problem. Teleoperation rigs and lab-based demonstrations are expensive and slow, while a smartphone in someone's kitchen is essentially free capture hardware.

Inside the data pipeline

Ingesting 30 minutes of video every second forced Figure to rebuild its infrastructure around consumer-app constraints: always-on availability, continuous large-scale processing, and real-time feedback loops to Creators. The processing stack has five stages:

1. Filtering: automated screens for technical, visual, and semantic quality
2. Fraud review: human analysts audit at the user level for people trying to game the payout system
3. Deduplication: each video segment is embedded, and clips above a similarity threshold to previously accepted data are discarded
4. Rebalancing: task quotas and embedding-based clusters keep the mix diverse rather than skewing toward whatever tasks Creators find easiest to record
5. Annotation: hierarchical text captions are generated for every episode

That deduplication step is worth pausing on. Paying strangers by the video creates an obvious incentive to spam repeated clips. By embedding every submission and rejecting near-duplicates, Figure ties compensation to novelty rather than volume, which happens to be the exact signal that makes data useful for generalization.

The competitive picture

Figure is not alone in racing to solve the robot data problem, though its approach differs from rivals. A Brookfield partnership announced in 2025 gave Figure access to residential and commercial properties for egocentric video collection under Project Go-Big, which aims to build the world's largest humanoid pretraining dataset by collecting first-person human video in real environments. The deal covers over 100,000 residential units, 500 million square feet of commercial office space, and 160 million square feet of logistics space. Index extends that logic to anyone with a phone.

The humanoid foundation model space has become fiercely capitalized. Figure AI, Skild AI, and Physical Intelligence have all raised large rounds around general-purpose robot foundation models, and Travis Kalanick's robotics company Atoms closed a $1.7 billion round led by Andreessen Horowitz. Data moats are increasingly seen as the durable differentiator once model architectures converge.

What it means in practice

For anyone building or researching robotic manipulation, three things stand out. First, Figure is betting that human egocentric video, without teleoperation or robot embodiment, is a sufficient pretraining signal for a VLA model like Helix. If the internal generalization results the company teased hold up, that changes the economics of every robotics lab investing in expensive teleop farms.

Second, the pipeline design itself is a useful reference. The combination of embedding-based deduplication, task-quota rebalancing, and hierarchical captioning is a template other groups will likely copy for their own crowdsourced efforts.

Third, there is a business model tucked inside the technical announcement. Figure frames Index as laying the groundwork for ordering robots as a service, where the humans cleaning your house today are eventually replaced by a machine that does everything for you. The same app that pays humans to do chores today becomes the marketplace that dispatches robots to do them tomorrow. The data pipeline and the go-to-market are the same product.