Figure AI announced Helix 2.5, a new version of the model that controls its humanoid robots, on September 17, 2026. In its announcement, Figure says Helix 2.5 achieved “zero-shot whole-body autonomy across 30 real homes,” with “no data collection, fine-tuning, or adaptation in those environments or manipulated objects.”
On this page
What Figure tested
According to Figure, the robots ran three household behaviors in 30 unseen homes in the San Francisco Bay Area: tidying living rooms, folding towels and making beds. In Figure’s framing, “zero-shot” means the robot worked in those homes, and with the objects in them, without any data gathered there or any tuning for them. The claim covers both parts of the problem: unfamiliar rooms and unfamiliar objects.
Results Figure reports
All of the following figures come from Figure’s own experiments:
| Measure (Figure’s experiment) | Result |
|---|---|
| Zero-shot success without Index pretraining | 9% |
| Zero-shot success with Index pretraining | 56% |
| Task-specific data used vs. a representative Helix 02 behavior | Half as much |
| Scaling-law forecasting error | 0.54% of variation across an 8× data range |
Figure says “Index pretraining alone increased zero-shot success from 9% to 56%,” with all other variables held constant. On Figure’s measure, that is a little over a sixfold increase: from roughly one success in eleven attempts to more than one in two. Index is the human-experience data Figure pretrains on; the company says it “is now generating roughly 35 minutes of new human experience every second.”
On data efficiency, Figure writes that “Helix 2.5 used half as much task-specific data as a representative Helix 02 behavior, then generalized that behavior across 30 unseen homes.” Taken together with the Index result, Figure’s numbers attribute most of the gain to pretraining on human data rather than to more robot data for each task.
A transfer scaling law
Figure also claims “the first human-to-robot transfer scaling law measured on a humanoid.” The company reports a forecasting error of 0.54% of variation across an 8× data range for that law.
Compute
Figure says it has “committed $3.5B of compute to training Helix.”
Helix 2.5 follows Helix 02 in Figure’s model line. For other model releases from labs and robotics companies, see the AI model release timeline.
