Figure AI, led by CEO Brett Adcock, has released Helix 2.5, an AI model that enables its humanoid robots to perform household tasks like tidying living rooms, folding towels, and making beds in unfamiliar homes with significantly higher success rates than previous versions. The breakthrough highlights a critical shift in how robotics companies are approaching artificial intelligence: instead of relying on synthetic data or remote control, Figure is building a competitive moat by collecting real-world video footage from people performing everyday chores.

How Does Figure's Data Collection Strategy Work?

Figure operates Index, a mobile app where people film themselves doing household and work tasks. The company has turned this into a systematic data pipeline that now ingests 35 minutes of video footage every single second, according to reporting on the company's latest developments. By late August, Figure had paid contributors $15 million for their video submissions, and the app now has users across more than 100 countries. This approach mirrors how large language models like ChatGPT and Claude were trained on freely available text from the internet, but with a crucial difference: video data from inside people's homes cannot be scraped for free and must be purchased.

The robots trained on this footage showed dramatic improvements. In company-run tests that have not been independently verified, Helix 2.5 completed household tasks in 56% of trials, compared to just 9% success for the same model without Figure's human video training. The company also discovered something encouraging about scaling: each time it doubled the amount of training footage, the robot's performance improved by a predictable amount. This finding suggests that the same scaling laws that made chatbots more capable as they processed more text now apply to robotics, meaning Figure can forecast performance gains before running expensive training runs.

What Makes Figure's Approach Different From Competitors?

While other humanoid robotics companies like XPENG and Apptronik are racing to mass production, they face a different bottleneck: hardware supply chains. Apptronik CEO Jeff Cardenas recently warned that the United States lacks the domestic gear supply base needed for humanoid robotics, noting that actuators (the assemblies that make robots move) can account for up to 60% of a robot's bill of materials. Tesla has reportedly been auditing suppliers in China's Yangtze River Delta region, while European companies like Schaeffler are developing new manufacturing techniques to reduce gearbox costs.

Figure's competitive advantage sits elsewhere. The company has signed a $3.5 billion computing power agreement with Nscale, a London-based AI cloud company, to train Helix 2.5 and future models on its growing video library. This positions Figure to improve its robots faster than competitors who lack equivalent data collection infrastructure. As more households contribute footage through Index, the company's models should improve at a compounding rate, creating a widening gap between Figure's capabilities and those of rivals still relying on synthetic training data or manual programming.

Key Factors Driving Figure's Robotics Strategy

• Data Ownership: Figure controls a proprietary dataset of real-world household videos that competitors cannot easily replicate, giving it a defensible advantage in training more capable robots.
• Predictable Scaling: The company has demonstrated that robot performance improves by a measurable amount with each doubling of training data, allowing it to forecast capability gains before investing in expensive compute resources.
• Global Contributor Network: With users in over 100 countries and $15 million paid to contributors by late August, Figure is building a distributed workforce that continuously generates fresh training material.
• Compute Infrastructure: The $3.5 billion computing power agreement with Nscale ensures Figure has sufficient resources to train on its video library at scale, avoiding bottlenecks that could slow model development.

The timing of Figure's announcement matters. XPENG has launched mass production of its IRON humanoid and raised over $900 million at a $6.3 billion valuation, while Tesla is reportedly building hundreds of Optimus robots per week. Yet none of these competitors have publicly disclosed a data collection strategy as systematic as Figure's Index app. If the scaling laws Figure has observed hold true across different task domains, the company's ability to continuously improve its models through crowdsourced video could prove more valuable than any single manufacturing facility or funding round.

The robotics industry is still in its early stages, comparable to personal computing in the 1980s according to some analysts. Figure's bet on home video data suggests that in this emerging era, the companies that control high-quality training data may ultimately outpace those that focus solely on manufacturing efficiency or hardware innovation. As Helix 2.5 demonstrates household capabilities in 30 Bay Area homes, the real test will be whether Figure can scale this approach to thousands of homes and maintain the quality and diversity of training data needed to keep improving.

" }