Figure, the humanoid builder now valued close to US$40 billion, has shipped three moves in quick succession that together look less like product updates and more like a bet on when embodied AI tips into the home.
The headline is Helix 2.5. The company rented 30 residences around the Bay Area and let its robot work directly in unfamiliar living spaces, publishing more than four hours of footage. With no on-site training, the robot tidies living rooms, folds towels and makes beds, adjusting its grip, stance and route as furniture, materials and layouts change.
The word zero-shot here means new rooms and new objects: no training data was collected for the test houses, and no on-site model fine-tuning was done. To measure what pretraining buys, Figure held task data, model architecture, training setup and evaluation conditions constant and changed only whether Index pretraining was used. In blinded cross-home tests, full-task success rose from 9 per cent to 56 per cent, a 47-point jump. Fifty-six per cent is not yet enough to hand over the daily chores, but the trajectory is the point.
Index crowdsources real-world operation at global scale
Helix 2.5 leans heavily on Index, a data platform Figure ran quietly for about four months before opening it on 25 August. Contributors worldwide record themselves performing real tasks at home or at work and get paid; cooking, laundry, restocking and clearing tables all count. The app has been downloaded more than 264,000 times across 108 countries, with over 44,000 weekly active users. Contributors have uploaded more than 16 million videos and earned US$15 million in rewards. By the Helix 2.5 launch, Figure said Index was adding roughly 35 minutes of new human-operation footage every second.
Figure has promised to invest more than US$1 billion in data and compute over the 12 months after Index opened, with a 100-fold scale-up as the stated direction. The hard part for any rival copying this path is the same: how to keep contributors coming, and how to turn a mountain of video into training data that actually moves robot performance.
Up to 100,000 GPUs booked ahead of commercial scale-up
On the compute side, Figure said in early September it would deploy up to 100,000 Nvidia Vera Rubin GPUs through AI cloud provider Nscale, with initial deployment in Bastrop, Texas from the second half of 2027. The initial compute commitment is US$3.5 billion, with both sides open to pushing past US$6 billion. A data-scaling experiment inside Helix 2.5 showed that lifting Index pretraining eightfold over the baseline produced a regular drop in action-prediction error, the evidence Figure will cite as it keeps expanding data.
A stranger’s living room is now a validated test bed, the global data tap is already running, and the next tranche of compute is reserved years ahead. Figure is checking the route while preparing the conditions to scale it.
More from the original report
Editor’s note: This is an adapted translation of the original OFweek report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://robot.ofweek.com/2026-09/ART-8321203-8110-30703772.html.