Figure AI has discovered a data advantage that could define the next phase of humanoid robotics: access to millions of hours of real household video that competitors cannot easily replicate. The company's Helix 2.5 AI model, trained on footage collected through its Index app, enabled robots to successfully complete tasks like tidying living rooms, folding towels, and making beds in 30 Bay Area homes they had never encountered before, according to company-run tests.

How Does Figure's Robot Training Approach Differ From Traditional Methods?

Figure's strategy mirrors the scaling laws that made large language models like ChatGPT and Claude powerful, but applies them to physical robotics for the first time. Where text-based AI models trained on freely available internet data, humanoid robots need video of real people performing household tasks. Figure's Index app, which pays contributors to film themselves doing chores and jobs, has become the company's primary data engine.

The results suggest this approach works. In company-run tests that have not been independently verified, robots using Helix 2.5 completed full tasks in 56% of trials, compared to just 9% for the same model without Figure's human video training. The improvement represents a six-fold increase in task completion rates.

What Makes Figure's Data Collection Strategy Strategically Important?

The bottleneck in robotics has shifted from computing power to data ownership. Index now operates in over 100 countries and had paid contributors $15 million by late August. The app ingests 35 minutes of video footage every single second, creating a moat that would take competitors years to replicate.

Figure has also secured $3.5 billion in computing power from Nscale, a London-based AI cloud company, to train on this footage. This combination of proprietary video data and dedicated compute resources creates a compounding advantage: more footage enables better models, which attract more users to Index, which generates more footage.

Steps to Understanding Figure's Scaling Advantage in Robotics

• Predictable Improvement Curve: Each doubling of training footage improved the robot by a steady, measurable amount, allowing Figure to forecast the results of its largest training run before it began.
• Global Data Collection: Index operates across 100+ countries with paid contributors, creating a distributed network that captures household tasks across different cultures and home layouts.
• Dedicated Compute Infrastructure: The $3.5 billion computing power agreement with Nscale ensures Figure has the resources to train on massive video datasets without competing for cloud capacity.
• Real-World Task Validation: Testing in 30 previously unseen Bay Area homes demonstrates that models trained on Index footage generalize beyond their training environments.

The significance of Figure's approach lies in what it reveals about the future of physical AI. Chatbots became powerful because text was free to copy from the internet. Video from inside people's homes must be paid for, creating a new economic model where data collection becomes a core business function rather than a byproduct.

Brett Adcock, CEO of Figure, has positioned the company to own this data advantage at a critical moment. While competitors like Tesla, Boston Dynamics, and Chinese robotics firms are building humanoid robots, few have established systematic ways to collect and monetize training video at scale. Figure's Index app represents the first major attempt to industrialize this process.

The implications extend beyond Figure itself. If scaling laws hold in robotics as they have in language models, the companies that control the most relevant training data will likely dominate the market. Figure's early lead in video collection, combined with its ability to pay contributors globally, suggests the company has identified a defensible competitive advantage that capital and engineering talent alone cannot quickly overcome.