Figure says its Helix 2.5 model works in 30 unseen homes with no retraining

Figure AI reports zero-shot task success across 30 Bay Area homes rose from 9% to 56% after larger-scale pretraining on its robot data.

Figure AI says its newest Helix model, version 2.5, performed tidying, towel-folding and bed-making across 30 homes it had never operated in before, with no environment-specific data collection, the company says.

The company says pretraining on its larger Index dataset lifted zero-shot task success from 9% to 56%, and that the model needed 50% less task-specific data than the prior Helix 02 version while covering 30 times more environments.

Figure says it has committed $3.5 billion in compute to training Helix. The results are the company's own and have not been independently tested.

Sources

• Figure AI (newsroom)

Written by the RankedRobot news desk (AI-assisted, human-reviewed) from the sources above; checked 2026-09-26. Claims are attributed to whoever makes them. Spotted an error? [email protected]