Figure Unveils Helix 2.5: Zero-Shot Household Work Across 30 Unseen Homes

On Sept 18, Figure launched Helix 2.5, its strongest humanoid foundation model to date, alongside a roughly 4-hour raw recording (later condensed to 30 minutes) of robots working zero-shot in 30 never-before-seen rental homes across the Bay Area. Without any additional training or on-site data collection, the robots performed chores such as tidying living rooms, folding towels, making beds, and picking up toys. The headline quantitative result: Index pretraining lifted zero-shot whole-body success from 8% to 56%.

Confirmed

• Helix 2.5 is Figure's most powerful neural network and the first model trained on the Index platform, powering the Figure 03 robot; Index produced training data at 35 minutes of data per second of shooting, per Figure engineer coreylynch.
• The model executes three whole-body behaviors zero-shot across 30 real Bay Area homes, with no training data collected in those homes; tasks include laundry, bed-making, towel folding, living-room tidying, and toy pickup.
• Deployment is zero-shot into unfamiliar homes: no on-site fine-tuning. CEO Brett Adcock called generalization to unseen environments the "holy grail of robotics," framing Helix 2.5 as answering whether humanoids can enter never-before-seen homes and immediately start autonomous whole-body work.
• The key ablation (which coreylynch repeatedly called the most important result): with identical model, task data, and evaluation, zero-shot whole-body success was 8% without Index pretraining versus 56% with it; under the strict criterion of fully completing a task, the gain exceeds 6x per rohanpaulai's summary.
• Index shows predictable scaling: doubling data yields gains in next-action prediction predictable enough to forecast the final run's loss to decimal precision before training begins.
• Watching the long-horizon live demos, coreylynch noted the experience changes qualitatively when perception and self-correction are strong enough—the model proactively finishes the task.
• The raw video honestly shows which parts of zero-shot technology work in real homes and which still struggle; authors such as mhdfaran praised this transparency as worth emulating.

Why it matters

• Zero-shot deployment validated across homes with no environment data is a key step from lab to general-purpose home robots; the 8%→56% ablation and predictable scaling curve provide rare quantitative evidence for scaling generalization via data volume.
• Cautionary voices exist: Waymo research lead mwulfmeier praised the work highly while noting such systems are not yet deployable, with discussion emphasizing long-tail reliability as the real challenge. Optimists like Linus Ekenstam predict home robotics could produce a ten-trillion-dollar company within a decade—a personal opinion.

2026-09-18 ~ 2026-09-19 · 27 related posts

• [source] Figure posts 4 hours of its humanoid doing zero-shot work in 30 rental homes — adcock_brett · 2026-09-18
• Figure rented 30 Bay Area homes to test robot generalization with no new training — Rewkang · 2026-09-18
• Figure's Helix 2.5 runs zero-shot whole-body tasks across 30 unseen homes — adcock_brett · 2026-09-18
• Figure ships Helix 2.5: zero-shot full-body robot behaviors across 30 real homes with zero home training data — coreylynch · 2026-09-18
• Figure ships Helix 2.5: zero-shot full-body behaviors across 30 real homes — coreylynch · 2026-09-18
• Index pretraining lifts Helix 2.5 zero-shot success from 8% to 56%, generating 50 min of data per second — coreylynch · 2026-09-18
• Figure's Brett Adcock shows robot doing zero-shot work in 30 rental homes over 4 hours — XxSpookxX · 2026-09-18
• Index pretraining lifts humanoid zero-shot success from 8% to 56% — coreylynch · 2026-09-18
• [source] Figure ablation: Index pretraining lifts robot zero-shot success from 8% to 56% — coreylynch · 2026-09-18
• Watch Helix 2.5 run zero-shot for 4 straight hours across 30 homes — coreylynch · 2026-09-18
• Figure Shows 4 Hours of Its Robot Working Zero-Shot Across 30 Unfamiliar Homes — mhdfaran · 2026-09-18
• Figure's humanoid robot completes household tasks across 30 unseen homes with no retraining — Polymarket · 2026-09-18
• Figure ships Helix 2.5, tested on 30 never-seen homes with zero prior data — xiaohu · 2026-09-18
• Figure rents 30 Bay Area homes to test zero-shot robot tasks; experts say long-tail reliability is the real bottleneck — chris_j_paxton · 2026-09-18
• Figure's Helix 2.5 hit 56% success in 30 unseen homes, vs 9% from scratch — CyberRobooo · 2026-09-18
• Figure's Helix 2.5 hits 56% zero-shot success across 30 unseen homes, vs 9% baseline — APPSO · 2026-09-18
• Figure runs humanoid robots in 30 rented Bay Area homes with no new training; predicted a $10T company — LinusEkenstam · 2026-09-18
• Figure demos zero-shot robot tasks in 30 Bay Area homes; expert says not deployable yet — m_wulfmeier · 2026-09-18
• Figure sends Helix 2.5 humanoid robots into 30 unseen homes with no retraining — minchoi · 2026-09-18
• Figure Releases 4-Hour Video of Humanoid Working Across 30 New Homes With Zero Extra Training — Distinct-Question-16 · 2026-09-19
• Figure runs zero-new-training robot tasks across 30 rented Bay Area homes, hinting at 'baby RGI' — DeryaTR_ · 2026-09-19

6 near-duplicate retellings: inductionheads · coreylynch · xiaohu · curiousinquirer007 · rohanpaul_ai · adcock_brett