Signals Inbox·August 26, 2026·Humanoid Robotics

What is Figure’s new Index dataset, exactly?

Figure Index is a global human-video collection network built to give Helix far more physical-world experience than Figure’s own robot fleet could gather alone. With 16 million+ uploads already flowing in, the real question is no longer whether Figure can collect enough data, but whether it can turn that scale into better robot manipulation.

Q1What did Figure actually launch with Index?

Figure Index is currently a huge private training-data operation that pays people around the world to record themselves doing physical tasks so Figure can train its humanoid robots.

Figure had been running the system quietly for about four months before opening it more broadly through the Index app. By launch, Figure said the network had reached 264,000 app downloads, more than 44,000 weekly active users across 108 countries and over 16 million uploaded videos. Contributors had earned $15 million.

The tasks look deliberately ordinary: folding clothes, unloading dishwashers, cleaning rooms, stocking shelves, serving customers, working in restaurants and handling jobs inside warehouses and factories. Figure also says it has received much stranger examples, including changing oil and cleaning kitty litter.

Behind the app sits the more interesting part. Figure filters incoming footage, checks for fraud, removes videos that are too similar to material it already has, rebalances the remaining data and generates text descriptions for the episodes. The resulting material goes into Helix, the AI system controlling Figure's humanoid robots.

So Index gives Figure both a growing training corpus and a way to keep producing new data whenever Helix needs more experience with a particular part of the physical world.

Figure Index today

Q2Is Figure Index basically a gig-work app for robot training?

Figure Index currently looks a lot like a gig-work marketplace whose output happens to be training data for humanoid robots.

People can apply to become Index "Creators." Figure says accepted Creators receive a recording device, film themselves completing approved tasks and get paid according to the amount of footage they upload. The current App Store description gives examples such as folding clothes, unloading a dishwasher and tidying a room.

There is also another side to the marketplace. People can request help with chores through Index, while businesses can invite contributors into restaurants, retail locations, warehouses and other workplaces. For now, Figure is even advertising some of those services as free.

That setup gives Figure more control over what gets recorded. If Helix keeps struggling with messy kitchen counters, certain warehouse tasks or a particular kind of object, Figure can potentially push collection toward those situations. A normal internet-video dataset would leave Figure searching for whatever people happened to film.

There is also a longer-term commercial idea hiding inside the app. Figure openly says that people may use Index workers for services today and eventually order robots for the same work. The collection network could therefore become both a source of robot data and an early map of where customers actually want physical labor.

Q3What exactly does an Index Creator record for Figure?

An Index Creator records real human activity in homes and workplaces, although Figure still has not published enough technical detail for us to know the full contents of every accepted training episode.

The Index app says Creators receive a dedicated recording device and capture themselves performing everyday physical tasks. Figure's earlier Project Go-Big work focused specifically on egocentric human video, meaning footage recorded from a human-like first-person viewpoint. Index clearly extends that same human-data strategy, although Figure has not said that every Index submission uses exactly the same camera setup or viewpoint.

Once a video enters Figure's system, it carries more structure than a random clip scraped from YouTube. Figure segments the footage, creates embeddings that represent what is happening visually, associates it with a task and generates hierarchical text captions for each episode.

We still lack several important technical details. Figure has not disclosed the camera specifications, frame rates, whether depth or inertial measurements are collected, or the exact metadata attached to every episode. It also has not released a public schema showing what an Index training sample looks like internally.

So we know quite a lot about how Figure gathers and processes Index data, while the raw sensor package remains surprisingly opaque.

Q4How big is Figure Index really?

Figure Index is already collecting physical-world video at an extraordinary rate, although the usable training corpus is smaller than the headline upload numbers suggest.

Figure says more than 16 million videos have been uploaded. More revealingly, the system is currently receiving about 30 minutes of footage every second. At a constant rate, that works out to 1,800 minutes each minute, 43,200 hours every day, or almost five years of recorded human activity arriving within 24 hours.

That calculation tells us how fast the intake pipe is running today. It does not tell us how many accepted training hours Figure actually owns.

Figure screens submissions for technical quality, semantic usefulness, fraud and duplication before accepting them into the useful corpus. Yet the company has not disclosed the rejection rate, average video length or total number of hours remaining after those filters.

We also should not take today's upload rate and extrapolate it backward across the entire stealth period. A network like this almost certainly grew over time.

So the scale today is clear at the ingestion layer: Figure has built an enormous video-ingestion system. The size of the final, cleaned Index training corpus remains undisclosed.

Q5Are Figure’s 16 million Index videos actually 16 million robot demonstrations?

Figure's 16 million Index uploads should currently be treated as human-video submissions rather than 16 million executable robot demonstrations.

A conventional teleoperated robot trajectory can contain camera images, joint positions, gripper states, commanded movements and other synchronized measurements showing exactly what the robot did at every moment. DROID, for example, distributes video together with robot actions and proprioceptive trajectories.

A person folding a shirt on camera gives Figure a different kind of information. Helix can see the objects, sequence of actions, goal, hand movements and interaction with the environment, but the raw footage does not automatically provide the exact actuator commands that a Figure robot needs to reproduce those movements.

That difference is central to the Index bet. Figure is choosing access to far more people, tasks, objects and environments while accepting weaker direct supervision over robot actions.

Figure may eventually extract surprisingly rich physical information from those recordings. Recent work across robotics increasingly uses human video to teach perception, navigation and manipulation priors. But the conversion from watching a person work to controlling a humanoid's fingers, arms and balance remains a hard technical problem.

Calling every Index upload a robot demonstration would therefore exaggerate what Figure has collected.

Q6Why is Figure collecting human video instead of using more robots?

Figure is betting on human video because millions of people can generate varied physical experience far faster than Figure can manufacture robots and teleoperate them.

Traditional robot data collection becomes expensive very quickly. Every parallel stream needs hardware, operators, maintenance and access to an environment where the robot can safely work. Changing the kitchen, warehouse or object distribution also takes effort.

DROID shows the scale of that challenge. Fifty data collectors spent a year gathering about 76,000 robot trajectories, adding up to 350 hours across 564 scenes. AgiBot later pushed past one million robot trajectories, but it built a collection system involving roughly 100 physical robots.

Index uses humans as the collection fleet. Another accepted contributor can bring another home, another workplace, another set of objects and another way of completing a task without Figure first shipping another humanoid into that location.

The trade becomes especially attractive for home robotics. Warehouses contain variation, but people's homes contain an almost absurd long tail of cupboards, towels, tools, toys, dishes, furniture, clutter and personal routines.

Figure could spend years trying to recreate that diversity in a robotics lab. Index lets the company go directly to where the variation already exists.

Q7Has Figure proved that Helix can actually learn from human video?

Figure has already proved that human data can produce useful behavior on Helix, especially for navigation and whole-body movement.

Project Go-Big provided the clearest example. Figure trained Helix using only egocentric human video for a navigation experiment, then asked a real humanoid to move through cluttered spaces using commands such as going to a fridge. According to Figure's technical release, the robot performed the behavior without robot demonstrations being used for that training result.

Helix 02 gives us another form of human-to-robot transfer. Figure says its System 0 whole-body controller was trained with more than 1,000 hours of retargeted human motion data combined with reinforcement learning in simulation. The resulting controller now helps Figure 03 coordinate walking, balance and manipulation.

Those experiments make Figure's Index strategy much more credible than a purely theoretical proposal. The company already knows that human behavior can teach useful components of humanoid control.

The remaining challenge is dexterous manipulation. Learning where a person walks from video is different from learning exactly how much pressure a robotic finger should apply to a slippery glass. Figure has made real progress across the embodiment gap, but the hardest part of Index still needs stronger evidence.

Q8Has Figure shown that Index makes Helix better at manipulation?

Figure has not yet published the experiment that would prove Index itself meaningfully improves Helix's hardest manipulation tasks.

What Figure has shown very clearly is that Helix improves when it receives more high-quality robot demonstrations. In its logistics work, Figure trained otherwise comparable models with roughly 10, 20, 40 and 60 hours of demonstration trajectories. Moving from about 10 to 60 hours cut average package-processing time from 6.84 seconds to 4.31 seconds while barcode-orientation success rose from 88.2% to 94.4%.

That is a strong scaling result. The same company has also taught Helix new behaviors such as laundry folding and dishwasher loading by adding data to a common learning architecture.

The original Helix itself used about 500 hours of multi-robot, multi-operator teleoperated behavior. So Figure already had a valuable source of precisely labeled robot data before Index arrived.

The missing experiment now is obvious: hold the Helix architecture and conventional robot-training data roughly constant, add increasing amounts of Index footage, then measure how much better Figure 03 becomes on unfamiliar manipulation tasks.

Figure says its internal generalization results are already supporting the Index thesis and that more details are coming. Until those results appear, the strongest public evidence supports "more robot data improves Helix" and "human video can transfer to robots." The combination of those two claims at Index scale is still being tested.

Q9How does Figure Index compare with DROID, Open X-Embodiment and AgiBot World?

Figure Index is much bigger by raw video count, while DROID, Open X-Embodiment and AgiBot World currently give researchers much clearer robot-action data.

DROID contains roughly 76,000 trajectories and 350 hours of interaction across 564 scenes and 86 tasks. Open X-Embodiment pooled more than one million real robot trajectories from 22 robot embodiments and 34 research labs. AgiBot World crossed one million trajectories collected with about 100 real robots.

Those datasets are smaller than Figure's upload count, but each trajectory comes from a robot executing actions. Researchers can inspect the relationship between what the robot saw and what it physically commanded.

Index scales along another axis. Humans can cover far more environments and tasks without requiring another expensive robot for every collection stream.

The ranking changes depending on what we measure. AgiBot World and Open X-Embodiment remain easier to call large robot-action datasets. Index appears to be operating at a much larger scale as a purpose-built collection of human physical behavior for robot training.

For now, Figure has not disclosed accepted Index hours or an equivalent number of robot-ready trajectories, so there is no honest single-number ranking between them.

Figure Index compared with major robot datasets

Q10How does Figure stop Index from becoming millions of repetitive chore videos?

Figure has built Index around aggressive filtering and diversity controls because sheer upload volume would otherwise become almost meaningless.

Every incoming submission first goes through automated checks for technical, visual and semantic quality. Figure then uses human reviewers to audit contributor-level samples and catch people trying to game those filters.

The more interesting step comes afterward. Figure turns video segments into embeddings, which let its system estimate how similar one recording is to material already accepted. When a new video is too similar, Figure can discard it rather than keep accumulating near-duplicates.

The remaining data is rebalanced using both task quotas and clusters found automatically in embedding space. That second method can catch differences that broad labels such as "fold laundry" would miss.

Figure's own diversity statistics give us a sense of what this produces. For every 1,000 hours collected, the company says Index contains 373 unique tasks, 1,146 unique manipulated objects and 116 unique environments.

Those numbers come from Figure itself, so we cannot independently audit how "unique" is defined. Still, they show that Figure is actively measuring whether additional collection expands coverage instead of merely pushing the hour count higher.

Diversity inside every 1,000 Index hours

Q11Why does all this task diversity matter so much for Figure robots?

Index's diversity could be more valuable than its sheer size because Figure wants Helix to work inside places the model has never encountered before.

Industrial automation can often rely on tightly defined environments. A home robot gets no such luxury. Two people who both ask a robot to "unload the dishwasher" may have different dishwashers, racks, glasses, lighting, cupboard heights and starting arrangements. One kitchen may be spotless; another may be full of clutter.

Figure's recent Helix demonstrations increasingly target exactly this kind of variation. Helix 02 has performed long household sequences such as unloading and reloading a dishwasher, tidying a living room and working with another robot to reset a bedroom. Those demos involve navigation, deformable objects, doors, furniture, containers and two-handed manipulation within the same task.

Training the next jump in reliability requires seeing many versions of those situations. Collecting another thousand identical plate pickups in one laboratory eventually gives diminishing returns.

That is also why Figure removes near-duplicate Index submissions. The company is trying to increase the number of situations Helix understands, rather than maximize a raw storage statistic.

If Index works, its advantage will come from covering the weird edge cases that appear once robots leave carefully prepared demos and enter thousands of real homes and workplaces.

Q12Why is Figure willing to spend more than $1 billion on Index data and compute?

Figure is spending heavily on Index because the company increasingly believes that the next big improvement in humanoid intelligence will come from scaling training, not from manually programming thousands of new robot skills.

Figure says it plans to spend more than $1 billion on data and compute over the next 12 months. That figure combines several expenses, so we cannot treat it as a $1 billion budget for paying Index Creators. Still, it shows how central the data strategy has become.

The comparison with Index's early economics is striking. Figure says contributors have earned $15 million during roughly four months of stealth operation, equal to an average of about $3.75 million a month so far. Spreading the announced $1 billion data-and-compute commitment evenly across a year would imply more than $83 million of spending per month.

Figure can afford to make that bet after raising more than $1 billion at a $39 billion post-money valuation. That financing was explicitly intended in part to fund GPU infrastructure and large-scale human-video data collection.

We should be careful with Figure's claim that Index is now on a "path to 100x." The company has not specified whether that means users, data, compute, throughput or some combination. The spending commitment is concrete; the 100x target is still too vague to model.

Figure's current data push

Q13Could Figure Index become a bigger advantage than Figure’s robot hardware?

Figure Index could become one of Figure's hardest advantages to copy if the company can connect its growing data network directly to the mistakes Helix makes in the real world.

Robot hardware can eventually be studied, benchmarked and matched by competitors. A growing history of millions of varied human interactions is much slower to recreate, especially when the collection system is already running across more than 100 countries.

The deeper advantage would come from the feedback loop. Imagine Figure robots repeatedly failing on unusual drawer handles. Figure can identify the weakness, collect more examples of people opening different drawers through Index, retrain Helix and test again. The same process could run across hundreds of behaviors.

Figure is also combining data access with deployment access. Its Brookfield partnership gives the company potential exposure to more than 100,000 residential units as well as large commercial and logistics footprints. More recently, Figure has signed commercial deployments in logistics environments, giving Helix another source of real-world experience outside homes.

Competitors can build their own collection systems, of course. AgiBot has already shown how aggressively robot data itself can be scaled. Index becomes a true moat only if Figure converts its broader human dataset into better robots faster than rivals can obtain the same capabilities through robot demonstrations, simulation, internet video or their own contributor networks.

Right now, that outcome looks plausible rather than proven.

Q14So what is Figure’s Index dataset, exactly?

Figure Index is currently best understood as a global data factory for humanoid intelligence: Figure pays humans to expose Helix to far more physical tasks, objects and environments than its own robot fleet could realistically collect alone.

As we saw above, the 16 million uploads should not be confused with 16 million robot trajectories. Figure is collecting human demonstrations, cleaning and annotating them, then trying to transfer what people know about the physical world into Helix.

That strategy already has some technical backing. Figure has demonstrated human-video-to-robot transfer for navigation, trained Helix 02's whole-body controller with more than 1,000 hours of human motion, and separately shown that Helix manipulation continues improving as demonstration data increases.

What Figure has yet to publish is the result that would settle the debate: a clean scaling curve showing how much Index data improves a Figure 03 robot on difficult unseen manipulation tasks.

Our judgment today is fairly sharp. Index is already a serious data-collection breakthrough. Whether it becomes a robotics breakthrough depends on Figure's ability to turn abundant human video into precise robot behavior.

If Figure solves that conversion efficiently, Index gives it something far more scalable than a warehouse full of teleoperators: tens of thousands of people continuously showing Helix how the physical world works. If the conversion remains weak for dexterous manipulation, the giant collection network becomes much less valuable.

For now, Figure has convincingly built the data engine. The next thing we need to see is what happens to the robot when that engine really starts scaling.