Figure's Helix 2.5 cleans 30 homes it has never seen, and human video pretraining lifts success from 9% to 56%

Figure AI said in a research post on Sept 17 that its new Helix 2.5 model ran three whole-body household tasks, tidying living rooms, folding towels and making beds, in 30 Bay Area homes where no data had been collected and with objects absent from its training set.

In blind trials scored pass or fail on the whole task, a policy pretrained on Figure's Index dataset of human behaviour succeeded 56% of the time, against 9% for an identical policy trained from scratch, and it matched a Helix 02 policy's success rate with half the task-specific data.

Figure also reported what it calls the first human-to-humanoid transfer scaling law: across an 8x range of Index data, test loss fell predictably enough to forecast its largest run to four decimal places.

The company said Index now adds about 35 minutes of human experience every second and that it has committed $3.5 billion of compute to training Helix. Figure's last disclosed round in MEKAWIRE's records is a $1 billion Series C at a $39 billion post-money valuation in September 2025, nearly three times Skild AI's $14 billion in January 2026.