Figure 让机器人第一次「空手」进陌生人家：不遥操，也不提前预习，进了家门就干活

Figure AI's Helix 2.5 model achieved 56% zero-shot success (237/420 trials) across 30 unseen Bay Area homes, versus 9% without Index pretraining - a 522% overall improvement - using no site-specific data collection or fine-tuning. The Index dataset (264K downloads, 16M videos, 44K weekly creators, $15M paid out, with $1B+ planned for data/compute) yields a predictable scaling law: quadrupling pretraining data reduces action-prediction error smoothly, and Helix 2.5 matches the prior Helix 02's site-trained performance with half the adaptation data. Competitors like 1X Neo (Turing teleoperators) and Tesla Optimus still rely on remote human operation, while Figure's approach learns general behavior from human video first, then adapts with minimal task data - backed by Nscale's $3.5B+ compute deal (up to 100K Nvidia Vera Rubin chips, 2027 H2).

why now

Helix 2.5发布：机器人零样本进陌生家，成功率56%。

topic

AI Tech & New Models

source

爱范儿