Figure's Helix 2.5 hits 56% zero-shot success across 30 unseen homes, vs 9% baseline

APPSO · wechat · 2026-09-18

Figure released Helix 2.5 and ran a zero-shot, blind evaluation in 30 real Bay Area homes it had never seen: 237 of 420 trials succeeded (56%), versus 9% for the same policy without Index pretraining. Robots folded towels, made beds and picked up objects with no on-site data collection or fine-tuning, with self-correction and retries on long tasks.

Key findings

With half the adaptation data of its predecessor Helix 02, Helix 2.5 matched Helix 02's success rate — which required site-specific data collection.
Performance scales predictably with pretraining data, like an LLM-style scaling law: doubling Index data lowers action-prediction error, with loss predictable from small-scale runs to within 0.54%.

Index dataset

The Index data app launched August 25: 264K downloads, 108 countries, 44K weekly active creators, 16M+ videos uploaded (peaking at 35 minutes of new video per second, 4.9 years of human work daily). Figure has paid creators $15M and plans $1B+ over the next 12 months to scale data 100x, with five processing stages (screening, anti-cheating, dedup, rebalancing, captioning).

Context and limits

1X's Neo ($20K) still relies on VR teleoperators; Tesla's Optimus uses large-scale human teleoperation data.
Only 3 task types were tested; 10-12 tasks at 80-90% zero-shot success would make home deployment realistic.
Nscale committed $3.5B (potentially $6B, up to 100K Nvidia Vera Rubin chips by H2 2027) in compute to Figure and took equity.

CEO Brett Adcock called it "the most important project we've ever done," with data and compute as the biggest bottlenecks.

Original post →

Figure's Helix 2.5 hits 56% zero-shot success across 30 unseen homes, vs 9% baseline

More from Embodied