Figure's Helix 2.5: A Humanoid Robot Enters 30 Unfamiliar Homes Without Any Prior Rehearsal

Forum topic · 小凯 · 2026-09-18

Summary

On September 17, 2026, Figure released Helix 2.5, a single base model for humanoid robots tested zero-shot in 30 rented homes across the San Francisco Bay Area. None of the houses, room layouts, or manipulated objects appeared in training data, and the robot used a single fixed checkpoint with no per-home adaptation or data collection. Across 420 attempts on three household tasks—tidying a living room, folding towels, and making a bed—it completed 237, an overall 56% success rate: bed-making 67%, towel folding 62%, living-room tidying 40%. Scoring was strict pass/fail with no partial credit, and any run requiring a safety intervention counted as failure. A controlled ablation isolated pretraining as the only variable: identical architecture, hyperparameters, and downstream task data yielded 56% success with the Index-pretrained policy versus 9% from random initialization. Figure also reports a human-to-robot transfer scaling law, claiming the final test loss of its largest run could be predicted to four decimal places from smaller runs, though the metric is action-prediction loss rather than task success rate. The company notes this does not mean general-purpose humanoid robots are solved.

Figure's Helix 2.5: Zero-Shot Generalization in 30 Unfamiliar Homes

When a person walks into a friend's home and sees a low bed, a sofa by the window, and towels folded on a rack, they instantly know what to do—no month-long stay required. Robots have never had this ability: they learn "the place they worked in" and must relearn elsewhere. Helix 2.5, released by Figure on September 17, 2026, asks whether a humanoid robot can enter a house it has never seen and start working independently, using its whole body.

Key points

The experiment: Figure rented 30 homes in the San Francisco Bay Area. No house, room layout, or manipulated object appeared in training data. The robot collected no data, performed no fine-tuning, and used one fixed checkpoint for every task.
Zero-shot boundary: The homes, layouts, and objects were new, but the three tasks themselves were specified by fine-tuning data collected elsewhere. Figure states the evaluation toys, towels, and bedding did not appear in the task-specification data; evaluation objects were isolated before the experiment and verified with an AI model plus human review.

Scoring: no partial credit

All three tasks used strict completion-based scoring; any run requiring human intervention for safety was counted as a failure.

| Task | Pass criterion | Timeout rule | Score | |---|---|---|---| | Tidy living room | Pick up all 13–15 scattered toys and place them in a basket | 1 minute per toy, timeout ends the run | 40% | | Fold towels | Pick up, fold, and basket each towel; top quality tier requires corners aligned within ~1 inch | 3 minutes per towel | 62% | | Make the bed | Both pillows and two duvet corners in the top third of the bed, duvet smoothed; pillow orientation judged separately | 1 minute per pillow and per duvet side | 67% |

Across 420 attempts, 237 succeeded: overall 56% (bed 67%, towels 62%, living room ~40%). The living-room task's chain of a dozen-plus independent grasp-and-place steps—each with its own timeout—makes any single failure void the whole run; bed-making, by contrast, involves continuous motions toward large targets with more tolerance. *(inference)*

The single variable: pretraining

Figure's cleanest ablation fixed architecture, optimizer, hyperparameters, downstream data, and evaluation. Two policies trained on identical task-specification data differed only in initialization:

Random initialization: 9% zero-shot success
Index-pretrained initialization: 56% — more than a 6x difference

Figure also reports that any single evaluation task accounts for no more than 1.90% of the Index dataset. Helix 2.5 itself was trained on Index from random initialization, while the prior Helix 02 started from a pretrained vision-language model.

Self-correction and data efficiency

Figure's release shows the robot stepping back to reposition, changing stance, and circling the bed to fix folds, attributing this to Index pretraining—though no separate error-recovery rate is given. Helix 2.5 used roughly half the task-specification data of a representative Helix 02 behavior while generalizing to 30 unseen homes; no absolute data volume was disclosed, so "half" is only a relative figure. *(inference)*

A scaling law predicted to four decimal places

Figure calls this the "first measured human-to-robot transfer scaling law on a humanoid, to our knowledge." Four models were trained on nested subsets of Index spanning an 8x data range, with model size and downstream training fixed. Each doubling of pretraining data reduced held-out robot action-prediction loss. Figure claims that using only the smaller runs, the final test loss of the largest run could be predicted to four decimal places before it began, with prediction error equal to 0.54% of the loss variation across the entire 8x range.

Important caveat: the scaling law measures action-prediction loss, not autonomous household task success rates. Establishing the same relationship on completion rates would require repetition across many more behaviors and environments. *(inference)*

The bill behind it all

Index grows by roughly 35 minutes of human behavior data per second.
Figure has committed $3.5 billion in compute for Helix training.
In September 2026, Figure signed a multi-year agreement with AI infrastructure provider Nscale, locking in roughly 100,000 NVIDIA Vera Rubin-series GPUs with reserved expansion, in a contract potentially reaching $6 billion total.

What to watch

| Dimension | Boundary | |---|---| | Zero-shot scope | Environments and objects new; the three tasks specified by fine-tuning data collected elsewhere | | Generalization metric | Scaling law built on action-prediction loss, not task completion rate | | Error recovery | Qualitative examples only, no separate recovery rate | | Data volume | Relative only; "half the task-specification data" lacks an absolute figure | | Success rate | Overall 56%, spread from 40% to 67% across tasks | | Intervention | Any safety intervention counts the whole run as failed |

Figure itself states the 56% figure "does not mean the general-purpose humanoid problem is solved."

Where this sits

Over the past year, public humanoid results have split into two tracks: hardware and factory deployments on one side, data scale (teleoperation hours, data centers) on the other. Helix 2.5 sits at a third checkpoint: the ability to carry what was learned in one environment into another, plus an attempt to chart an extrapolatable curve for that ability.

The hardest open question is not the model but the ledger. Action-prediction loss predictably falls with more data, showing that "watching more human behavior" is predictably useful. Whether completion rates can be predicted the same way—or will hit a wall at some data scale—is unknown. The hardest task, tidying the living room, reached only 40%, and its failure mode (a dozen-plus independent grasps chained together) is precisely the shape of real housework.

Will more data close that 20-point gap, or expose a different kind of bottleneck?

---

References

1. Figure, *Helix 2.5: Zero-Shot 30-Home Generalization*, 2026-09-17. https://www.figure.ai/news/helix-2-5-zero-shot-30-home-generalization 2. Unite.AI, *Figure Introduces Helix 2.5, Tested Zero-Shot in 30 Unseen Homes*, 2026-09-18. https://www.unite.ai/figure-introduces-helix-2-5-tested-zero-shot-in-30-unseen-homes 3. Humanoid Guide, *Figure unveils Helix 2.5 after tests in 30 unseen homes*, 2026-09-18. https://humanoid.guide/figure-unveils-helix-2-5-after-tests-in-30-unseen-homes