Home · September 17, 2026 · 2 min
Figure’s Helix 2.5 does chores in 30 homes it had never seen
Pretrained on Figure’s Index dataset of human behavior, the model tidied rooms, folded towels and made beds with no data collected in the test homes.
Figure has introduced Helix 2.5, which it calls its most advanced neural network yet. The company tested it in 30 Bay Area homes where no data had been collected, and on objects the robot had never trained on.
Video from Figure on YouTube · Helix 2.5 30-Home Generalization
A single base model, pretrained on Index, Figure’s large dataset of human behavior, was adapted into three whole-body tasks: tidying a living room, folding towels and making a bed. A trial counted only if the whole job was finished, with no partial credit.
In blind evaluations, a policy trained from scratch on the same task data succeeded 9% of the time. The Index-pretrained version succeeded 56% of the time. Figure said Helix 2.5 matched an earlier Helix 02 behavior while using half as much task-specific data. It also reported a scaling law: each doubling of Index pretraining data improved robot action prediction in a predictable way.
Figure said the result does not mean general humanoid robotics is solved. Index now adds roughly 35 minutes of new human experience every second, and the company has committed $3.5 billion of compute to training Helix.