This week brought one of the more transparent robotics disclosures of 2026 — and, by contrast, it exposed how little the home-robot industry's front-runners are willing to say about the same question. On September 17, Figure AI published results from Helix 2.5, a new control model it tested across 30 rented homes in the San Francisco Bay Area that the robot had never entered before. With no on-site data collection and no fine-tuning, zero-shot success on three household tasks — tidying a living room, folding towels, making a bed — rose from 9% with a randomly initialised policy to 56% with Figure's pretrained one. Figure gave no partial credit for incomplete tasks and published the comparison itself.
9%
Figure, random-init baseline
56%
Figure, pretrained (Sept 17, 2026)
Undisclosed
NEO · Isaac 1 · Memo
That 56% is not a finished product. But the way Figure reported it — a defined test set of unseen homes, a named baseline, a stated scoring rule — is closer to the kind of evidence DOMULU looks for than almost anything else published by a home-robot company this year.
Compare that to 1X, which plans to deliver its NEO robot to US homes by the end of 2026 at $20,000 or $499 a month. NEO leans on what 1X calls "Expert Mode": when the robot can't complete a task on its own, a remote human operator takes over through a live camera feed. A figure of "60-70% autonomous" circulates widely in coverage of NEO — but at least one detailed review of the product traced that number and found it has no identified primary source. 1X itself has not published what share of tasks in a typical home actually require a human operator.
It's the same open question we flagged about Isaac 1 and Memo in our previous piece: both also route unresolved tasks to a remote human, and neither has published how often that happens in practice. Three different companies, three different robots, the same undisclosed number.
The pattern isn't limited to homes. Skild AI, whose software runs industrial and warehouse robots rather than home ones, told Bloomberg on September 10 that its S1 model reaches 66% success on entirely new tasks learned from a single video — up from 9% for a language-only baseline. Unlike Figure's home test, that 66% comes from an internal benchmark Skild has not released, with no disclosed trial count and no external baseline. It is a different product in a different setting, but it is the same habit: a single headline percentage standing in for evidence a reader can't actually check.
DOMULU isn't asking any of these companies to be fully autonomous today — nobody credible claims that. We're asking for the number they already have internally: what share of tasks in a real home needs a human, this month, for this product. Figure showed this week that publishing a falsifiable test is possible. Until the others do the same, 'autonomous' and 'assisted' remain whatever each company's marketing team wants them to mean.
This article is a research preview. Unless explicitly stated, it does not report a completed independent product test by DOMULU.