Figure's Helix 2.5 cleans 30 homes it has never seen
Figure says its humanoid tidied rooms, folded towels and made beds in 30 unseen Bay Area homes with zero local data, backed by a $3.5B compute deal with Nscale.
In short: Figure's new Helix 2.5 model ran three household chores in 30 Bay Area homes with no data collected in any of them. Full-task success rose from 9% (trained from scratch) to 56% when pretrained on Figure's Index dataset of human behavior.
What Figure showed
- One foundation model, three behaviors: tidying living rooms, folding towels and making beds.
- Zero-shot: no data collection, fine-tuning or adaptation in the test homes. Success counted only if the whole task was finished.
- Figure says transfer improved smoothly as Index data grew, "enough to measure as a scaling law".
The compute behind it
On Sept 3 Nscale announced a multi-year deal to supply Figure with up to 100,000 NVIDIA Vera Rubin GPUs, starting in H2 2027 in Barstow, Texas, with an initial $3.5B compute commitment and intent to scale past $6B. Nscale also takes an equity stake. Separately, The Information reported on Oct 7 that Nvidia has discussed another $1B investment in Figure; no deal has been confirmed.
Why it matters
Homes are the hardest test for a humanoid robot: every layout and object is different. A 56% strict success rate is far from product-ready, but it is evidence that large-scale pretraining — the recipe behind embodied AI and VLA models — can transfer to places a robot has never seen.
Sources Sources verified
Related briefings
More in Humanoids →The humanoid factory race: Korea, Shanghai and China's 10,000-unit plant
In one week Korea cleared 20 humanoids to work beside people, Shanghai set a goal of 8,000 factory humanoids by 2028, and UBTECH's 10,000-a-year plant came into focus. Here is what each move means.
Galbot's PASSAGE teaches humanoids to duck, step over and squeeze through clutter
On a Unitree G1, Galbot's planner–tracker stack finished 50 unseen physical layouts without falls; 45 were contact-free. Destination success in sim was 98.67% — contact-free 70.3%.
Unitree's upgraded G1+ humanoid goes on pre-order in Japan
GMO AIR began Japanese pre-orders for the G1+ on Oct 13: a neck, stronger motors, 3D LiDAR and voice interaction. Founder Wang Xingxing now sees an embodied-AI tipping point in 1–2 years.
Newsletter
Robrief Weekly
One email a week with the robotics news that matters. Free, unsubscribe anytime.