Figure Launches Crowdsourced Robot Data Platform Index, Pays Global Users to Collect First-Person Real-World Operation Videos
Figure has publicly launched the Index platform, which has been operating secretly for four months, paying global users to collect first-person real-world operation videos to train its humanoid robot vision-language-action model Helix.
Figure claims that Index has accumulated 264,000 downloads, covering 108 countries, with over 44,000 active contributors weekly, and has uploaded 16 million video segments; the platform currently receives about 30 minutes of video per second, which the company estimates equates to approximately 4.9 years of human activity recorded daily. These figures are disclosed by Figure and have not been independently audited.
Users can record tasks such as cooking, laundry, cleaning, moving, and organizing to earn rewards. Figure states that it has paid contributors $15 million and plans to invest over $1 billion in data and computing power over the next 12 months, aiming to scale data operations by 100 times.
The core of Index is not an ordinary video library, but rather converting human first-person actions into training data for robots. Figure claims that every 1,000 hours of collected video averages 373 tasks, 1,146 manipulated objects, and 116 environments; the data is then filtered, anti-fraud checked, deduplicated, rebalanced, and hierarchically annotated.
Figure previously demonstrated in Project Go-Big that Helix can learn end-to-end control solely from human videos and transfer to humanoid robots with zero-shot learning; the company claims this is the first time a humanoid robot has learned low-level SE(2) speed control directly from human videos, using images and language. This conclusion is part of the company's technical statement, and external parties have not yet obtained complete training data, evaluation protocols, or independent reproduction conditions.
Helix 02 has been used to control the full-body movements, operations, and balance of Figure 03. Figure disclosed that the model completed approximately 4 minutes of dishwasher loading and unloading tasks under conditions without resets or human intervention, involving 61 continuous movement-operation actions while maintaining task status and implicit error recovery.
Figure has been advancing data, models, and manufacturing in parallel. The BotQ factory has increased the production speed of Figure 03 from one unit per day to one unit per hour in less than 120 days, delivering over 350 units; the first line's designed annual capacity is 12,000 units, and the company aims to produce a total of 100,000 units within four years.
In terms of market mechanisms, Index uses cash incentives to transform ordinary users' homes and workplaces into data supply sources, with funds flowing to video contributors, annotation reviews, cloud storage, training computing power, and data governance; Figure aims to convert diverse human operation data into Helix capabilities and robotic deployment advantages. Beneficiaries include companies with data collection channels, GPU computing power, edge-side sensors, and robotic deployment scenarios; humanoid robot teams relying on a small amount of remote operation data, single-scene demonstrations, or expensive custom integrations will face pressure from data scale disparities.
Source: Public Information
ABAB AI Insight
Figure's path is to replace "remote operation collection" with "humans experiencing the world for robots." Traditional robot training requires engineers to control real machines to collect actions, and limitations in hardware quantity, space, maintenance, and safety make data expansion costly; Project Go-Big has moved the training source to human videos and claims to achieve zero-shot human-robot transfer. Index has productized this research direction: establishing a distributed collection network with 264,000 downloads and paid incentives, aiming not to record fixed actions but to capture the long-tail differences of real homes, object placements, lighting, clutter levels, and body movements.
The capital path presents a closed loop of "data procurement—model capabilities—robot deployment—operational data." Figure has committed to investing over $1 billion in data and computing power over the next 12 months, having already paid contributors $15 million; this amounts to initially purchasing human behavior samples with capital, then compressing those samples into reusable control strategies with Helix. Figure 03 has a 10Gbps millimeter-wave data offloading capability, allowing fleets to upload TB-level data for continuous learning; as the number of real deployments increases, data sources will expand from paid human videos to the robots' own failures, recoveries, and long task trajectories.
Historically, the closest comparison is the data flywheel of Tesla Autopilot and the data operations of Scale AI, but humanoid robot data is harder to standardize: the camera poses, vehicle dynamics, and traffic rules for road driving are relatively uniform, while home scenarios face varying cabinet heights, materials, occlusions, variable objects, and contact mechanics. Figure has allowed Figure 02 to operate for about 1,250 hours at the BMW Spartanburg factory, handling over 90,000 parts and supporting the production of over 30,000 X3s; Figure 03 subsequently entered the same factory to handle more complex logistics sorting. This indicates that its commercialization first validates through more constrained industrial tasks before expanding to high-variability home tasks.
This represents capital concentration. The true moat is not just the robotic arm or a single demo, but whether it can continuously cover the fixed costs of data procurement, computing power, manufacturing ramp-up, and on-site deployment. Figure completed over $1 billion in C-round financing commitments in 2025, with a post-investment valuation of $39 billion, with investors including NVIDIA, Intel Capital, Brookfield, LG, Salesforce, and Qualcomm Ventures; this type of capital and industrial connection can simultaneously lock in computing power, manufacturing resources, and real-world scene access. The mechanism is that the more real-world data there is, the more the model can cover long-tail tasks; capability improvements also enhance deployment value and data feedback speed, with laggards facing compounded scale disparities rather than just single algorithmic differences.
ABAB News · Cognitive Laws
The most scarce resource for robots is not motors, but reality.
Whoever first turns human behavior into data will gain the corpus of the physical world.
Hardware determines the starting line, while the data flywheel determines the outcome.