Figure takes its robot into thirty homes it had never seen before: it finishes the job in just over one case in two
On September 17, 2026, Figure presented Helix 2.5 and tested it in 30 Bay Area homes where the robot had never been and of which it had no data. Tidying up, folding towels and making the bed succeed in 56% of the trials: it is the number the company itself publishes.
On September 17, 2026, Figure presented Helix 2.5, the model that drives its humanoid robot, and put it to the test in 30 Bay Area homes the robot had never entered and from which no data had been collected. On three household tasks — tidying the living room, folding towels, making the bed — the robot completed the job in 56% of the trials. It is the company itself that publishes that number, on its own website, together with the point of comparison: without that preliminary training the rate was 9%.
What "never seen before" means
The point of the test is not the three tasks in themselves: it is where they are carried out. The comparison Figure proposes is with one of its previous models, Helix 02, trained with data collected in the same environment in which it would then be evaluated — on which of the three tasks, the page does not say. Not here: Figure declares that no data was collected in the 30 homes, that there was no on-site fine-tuning, and that none of the objects used in the evaluation — the toys, the towels, the sheets — appeared in the data with which the tasks had been specified. The company also writes how it checked this: the objects were set aside before the experiments began, and an artificial intelligence model, followed by a human review, verified that they were not there. The robot used the sofas, the beds and the surfaces of those homes.
This is what in this field is called zero-shot: the system faces a new situation without having examples of that situation. Figure specifies what it means by that term: it concerns the environments and the objects of the evaluation, while the tasks had been taught with data collected elsewhere. There is only one model — what the company calls a foundation model, that is, a base trained broadly and then adapted to specific tasks — and all three skills were derived from that base.
The number that matters, and what it does not say
Figure declares that preliminary training on Index, its archive of human behavior, brings success from 9% to 56% in blind evaluations. That is more than six times as much, and it is the result the company lists among its own "Key Results".
It is worth, however, reading the 56% for what it is: it is the overall number for the three tests, and in just under half of the cases the task was not completed. And "completed" has a strict definition, set before the evaluation and published: all 13-15 toys scattered around the room inside the basket; all the towels folded and placed in the basket; both pillows and the corners of the duvet in the right place, with the duvet pulled up. Each toy has one minute: beyond that the trial is stopped.
The company adds two more figures. The first: compared with a Helix 02 model trained on the same task with data collected in the environment in which it was then evaluated, Helix 2.5 matches its success rate using half of the adaptation data, and does so in the 30 never-seen homes. The second: as the Index data doubles, the result improves regularly enough to be predicted — using only the smaller training runs, Figure declares, the error of the largest training run was predicted to four decimal places before launching it. It is a prediction about the model's error, not about the success rate in the homes.
What changes for those waiting for a robot at home
Two things, and neither of them is a date. The facts are in the paragraphs above: what follows is how we read them, and we present it as such.
The first: Figure presents the test as the first demonstration of zero-shot whole-body generalization with this scope on a humanoid. It is a statement by the company, which writes it with the reservation "to our knowledge", and not a verification of ours. What you read in the report is more concrete, and for us it is the part that matters: the homes are real and the robot had never seen them, and the success criteria were set before the evaluation and published.
The second: almost one task in two is not finished, and there are three tasks, chosen by the company. A robot that completes the job in just over one case in two is not a household appliance: it is a research system that has passed a difficult test. Figure itself writes it, in a line that is worth more than many headlines: the point is not that general humanoid robotics is solved.
For Italy nothing changes in practical terms: on its website Figure does not publish a price or dates of sale to private buyers (verified on September 25, 2026 on the figure.ai pages of the robot, of Helix and of the home page: no price, no pre-order, no date). What changes, as we see it, is the yardstick with which to read the next demonstrations: the question to ask is in how many never-seen homes, and with what success rate.
Figure's page does not say which version of the robot was used in the tests. The company's current generation is Figure 03, for which we keep a profile.
Sources
- Figure (official website) primary source
- Figure (official website) primary source
- Figure (official website) primary source
- Figure (official website) primary source
Next to each source we state what kind of source it is: “primary source” means the manufacturer, an official document or the author of the data; “press” a news report; “retailer” a distributor or a store, reliable about its own offer — price and availability in Italy — but not about the product’s features; “analysis” a professional firm or a research center that signs its own reading; “Chinese state media” an outlet controlled by the state, whose content should be weighed as such; “blog/forum” a non-editorial source; “unclassified” a source we have not assessed yet: the link stays, the judgment does not.