The Signal · Issue 01 · Physical Intelligence
Thirty homes, fifty-six percent, no partial credit
Figure's Helix 2.5 took a humanoid into 30 homes it had never seen. The most impressive thing about the announcement is how it measured itself.
Facts are attributed and linked to their sources. The analysis and opinions are the author's own.
What it is
On 17 September Figure introduced Helix 2.5, which it calls its most advanced neural network. Figure pretrained the model on Index, its large dataset of human behaviour, then adapted it to three whole-body tasks: tidying living rooms, folding towels and making beds. It evaluated those behaviours in 30 Bay Area homes where no data had been collected, using objects that did not appear in the task data.
Figure reports that, holding task data, architecture, training and evaluation fixed, Index pretraining raised zero-shot success from 9% (a model trained from scratch) to 56%. Success meant completing the entire task, with no partial credit, against criteria fixed before evaluation; any human intervention for safety counted as a failure. Figure also says Helix 2.5 needed half as much task-specific data as a comparable earlier behaviour, and that it observed a scaling law: doubling pretraining data improved robot action prediction predictably.
Figure is careful to add: "The point is not that general humanoid robotics is solved."
Why it matters
Robotics demos are notoriously easy to stage. A single good video says little about reliability. What Figure published is closer to an evaluation protocol: a fixed rubric, held-out environments, a controlled ablation, failure counted strictly. That is how claims should be made, and it lets outsiders reason about the result.
The most impressive thing about the announcement is how it measured itself.
It also lets us read the number honestly. Fifty-six percent full-task success in unseen homes is a large step from nine, and it also means the robot did not finish the job in a large share of trials.
The ablation is the part I would show any analytics team. By holding everything fixed except pretraining, Figure isolates the contribution of one factor, the same logic as a well-designed A/B test or a champion-challenger model comparison. Most product claims skip this step and attribute improvement to whatever was changed last.
The strict scoring choice also matters. Partial credit would have produced a much more flattering number. Choosing all-or-nothing success, and counting safety interventions as failure, is choosing the metric that matches what a user actually experiences. That is the opposite of Goodhart's trap, and it is worth praising when companies do it.
The scaling-law claim is the strategically important one. If robot performance improves predictably with more human-behaviour data, then progress becomes something you can plan and budget, much like language models. It also makes the data itself the core asset, which raises the familiar questions: how was it collected, with what consent, and how is it governed? Those questions will follow robotics into homes.
What I'd watch
- Independent replication in homes Figure did not choose.
- Failure modes, not just rates. A towel left unfolded and a toy thrown are different kinds of failure.
- How success rates move from 56% toward the reliability a household would accept, and what supervision looks like until then.
Sources
- Helix 2.5: Zero-Shot 30-Home Generalization. Figure, 17 Sep 2026.