Figure's Helix 2.5 Completes 56% of Chores Zero-Shot in 30 Unseen Homes

In late September, Figure released Helix 2.5, its most capable neural network to date, and validated it in an unusually rigorous way: the team rented 30 real homes across the San Francisco Bay Area and had humanoid robots enter entirely unseen households to do chores with no additional training (zero-shot). The Index-pretrained model completed 56% of tasks end-to-end, versus 9% for a from-scratch baseline, while using 50% less data. Figure also published roughly 4 hours of full test footage.

Confirmed

• Helix 2.5 achieved a 56% zero-shot success rate across 30 unseen homes (@TansuYegen, @FinanceYF5)
• Test environments were genuinely messy rentals—scattered toys, untidy towels, unmade beds—with no extra training (@lukasmziegler, @FinanceYF5)
• The team released about 4 hours of complete test video (@FinanceYF5)
• The from-scratch baseline hit only 9%, and the pretrained path used 50% less data (@TansuYegen, @lukasmziegler)
• Per QbitAI's account, the two key experiments compared random initialization versus Index pretraining under identical architecture and data; "zero-shot" means no retraining for new environments, though the tasks themselves were previously adapted using data from other environments
• @emmanuelvivier noted the jump from roughly 9% marks substantive progress in home-generalization, coming alongside Google DeepMind's new research institute announcement signaling intensified industry investment

Why it matters

• This is a landmark validation of humanoid robots moving from staged demos to real work in unfamiliar environments, with real rentals rather than labs lending strong credibility
• The 56%-vs-9% gap with half the data suggests that scaling via pretraining on human behavior data holds in robotics too—viewed as evidence of robot scaling laws (QbitAI)
• @thealexbanks argues the humanoid form is the best vessel for a "general worker"—carrying, climbing stairs, even driving forklifts—because the world is built to human scale, closing with "which side are you betting on?"

2026-09-21 ~ 2026-09-23 · 9 related posts

• Figure's Helix 2.5 completes 56% of chores end-to-end across 30 unseen homes — thealexbanks · 2026-09-21
• Figure releases Helix 2.5, robots do chores zero-shot in 30 unseen homes — FinanceYF5 · 2026-09-22
• Figure's Helix 2.5 achieves zero-shot whole-body autonomy across 30 unseen homes — FinanceYF5 · 2026-09-22
• [source] Figure Helix 2.5 hits 56% zero-shot success in 30 unseen homes vs 9% trained from scratch — TansuYegen · 2026-09-22
• Figure's Helix 2.5 robots complete 56% of tasks in 30 unseen Bay Area homes — lukas_m_ziegler · 2026-09-22
• Figure's Helix 2.5 lifts zero-shot tidying success in unseen homes from 9% to 56% — emmanuelvivier · 2026-09-22
• [source] Figure shows pretraining scaling works for robots: 9% to 56% success in unseen homes — 量子位 · 2026-09-23
• Figure Ships Helix 2.5, New Humanoid Robot Demo Goes Viral — adcock_brett · 2026-09-23

1 near-duplicate retellings: FinanceYF5