A person walking into a friend's house sees a low bed, a sofa by the window, towels folded on a rack—and immediately knows what to do. No month-long homestay required. Robots have never had this ability: they learn the specific place they worked in and must relearn elsewhere. Helix 2.5, released by Figure on September 17, 2026, asks whether a humanoid robot can walk into a house it has never seen and start working with its whole body, entirely on its own [Figure official release, 2026-09-17].
🏠 The Experiment
Figure rented 30 homes in the San Francisco Bay Area. None of the houses, room layouts, or manipulated objects appeared in training data. The robot collected no data, did no fine-tuning, and performed no adaptation in these homes—it used the same fixed checkpoint for every task.
An important boundary: "zero-shot" here means the houses, layouts, and objects were all new, but the three tasks themselves were specified by fine-tuning data collected elsewhere. Figure states that the toys, towels, and bedding used in evaluation did not appear in the task-specification data; evaluation objects were isolated before the experiment and verified with an AI model plus human review [direct quote].
📏 Scoring: No Partial Credit
All three tasks used all-or-nothing pass criteria:
| 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 terminates the attempt | 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 | Two pillows and both duvet corners in the upper third of the bed; duvet smoothed | 1 minute per pillow and per duvet side | 67% |
Across 420 attempts, 237 completed: 56% overall. Any run requiring a human safety intervention was scored as a failure.
The living-room task is the obvious drag: 13–15 toys mean a dozen-plus independent grasp-and-place sequences, each with its own 1-minute timeout, and any single failure voids the whole run. Bed making is one of the few long continuous motions with large targets, so it tolerates errors better [inference].
🧪 The One Variable: Pretraining
The cleanest part of the experiment made pretraining the only variable. Two policies received identical task-specification data, with architecture, optimizer, hyperparameters, downstream training, and evaluation all held fixed. The only difference: one started from random weights, the other from Index pretraining weights.
Result: 9% vs. 56% zero-shot success—a more than sixfold gap.
Figure also reported that in the three evaluation tasks, no single task accounts for more than 1.90% of Index. Helix 2.5 itself was trained from random initialization on Index, whereas the previous Helix 02 started from a pretrained vision-language model [direct quote].
♻️ Self-Correction and Data Efficiency
Figure's press materials show the robot backing up to reposition, changing stance, and moving around the entire bed to correct a fold—capabilities attributed to Index pretraining, though no standalone error-recovery rate was given.
On data efficiency: Helix 2.5 used roughly half the task-specification data of a representative Helix 02 behavior while generalizing it to 30 unseen homes. No absolute data volume was disclosed, so "half" should only be read as a relative figure [inference].
📈 A Scaling Curve Predicted to Four Decimal Places
Figure describes this as "to our knowledge, the first measured human-to-robot transfer scaling law on humanoids." They trained 4 models on nested subsets of Index spanning an 8× data range, holding model scale and downstream training constant. Each doubling of pretraining data reduced held-out robot action-prediction loss.
The company 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—prediction error equivalent to 0.54% of the total loss swing across the 8× data range.
Watch what the curve measures: action-prediction loss, not household task success rates. Extending the same scaling relation to completion rates would require repetition across more behaviors and environments—a much harder problem [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, the company signed a multi-year compute agreement with AI infrastructure provider Nscale, locking in roughly 100,000 NVIDIA Vera Rubin-series GPUs with expansion headroom—potentially worth up to $6 billion total.
Timeline of the Helix series:
- 2024–2025: Helix 01 launches; Figure 02 pilots at a BMW factory
- 2025–2026: Helix 02 achieves long-horizon whole-body coordination; logistics tasks run autonomously for 200 hours
- 2026-09-17: Helix 2.5 released; zero-shot whole-body generalization in 30 unfamiliar homes
- 2026-09: Nscale agreement signed for ~100k Vera Rubin GPUs
⚖️ What to Watch
| Dimension | Boundary | |---|---| | Zero-shot scope | Environments and objects are new; the three tasks themselves were specified by fine-tuning data collected elsewhere | | Generalization metric | The scaling law is built on action-prediction loss, not task completion rate | | Error recovery | Qualitative examples only; no standalone recovery rate | | Data volume | Relative only—"half the task data" has no absolute value | | Completion rates | 56% overall; spread from 40% to 67% across tasks | | Interventions | Any safety intervention scores the run as a failure |
On the 56% overall figure, Figure itself states that "this does not mean the general-purpose humanoid problem is solved" [direct quote].
🧭 Where This Sits
Over the past year-plus, public humanoid results have split into two tracks: hardware and factory lines on one side, data scale on the other. Helix 2.5 sits at a third position: measuring the ability to carry what was learned in one environment into another, and trying to fit an extrapolatable curve to that transfer.
The hardest open question is not the model but the ledger. If action-prediction loss falls predictably with data, "seeing more human behavior" is usefully predictable. Whether completion rates scale the same way—or hit a wall at some data volume—nobody knows. The hardest task, tidying the living room, only reached 40%, and its failure mode (a dozen-plus independent grasps chained together) is precisely the shape of real housework.
Will more data close those 20 percentage points, or expose a different 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