Figure paid 44,000 people $15 million to film themselves doing chores. That footage just taught a robot to make beds in 30 houses it had never entered. The dataset is called Index. People in 108 countries upload first-person video of everyday tasks, 30 minutes of footage arriving every second, and Figure pays them for it. Helix 2.5 was pretrained on all of it before it ever touched a task. The controlled experiment is the whole story. Figure trained two identical policies on the same task data. The one starting from random weights succeeded 9% of the time in unseen homes. The one pretrained on Index hit 56%. Same robot, same tasks, same fine-tuning. The only variable was watching humans. Bed making landed at 67%. Towel folding 62%. Tidying toys, the messiest job, 40%. All in rented homes the robot had zero data from, using whatever couches and bedding were already there. Figure also says transfer improved so smoothly with data that they forecast their largest run's final loss to four decimal places before training started. That's the same predictability that made GPT models scale. It now applies to physical movement. Which explains the $3.5 billion of compute Figure just committed. Helix 2.5 matched its previous model with half the task-specific data across 30x more environments. The bottleneck shifted from robot hours to human video. Every clip of someone loading a dishwasher is training data now. Figure just proved it's worth paying for.