On September 17, 2026, Figure published an announcement introducing Helix 2.5 and describing an evaluation in 30 Bay Area homes. Figure characterized the work as the first demonstration of zero-shot whole-body generalization at this scope on a humanoid. That priority claim has not been independently established.
Figure said Helix 2.5 used an Index-pretrained foundation model to perform three behaviors: living-room tidying, towel folding and bed making. The company said the system was tested in homes and with objects it had not encountered before, without data collection, fine-tuning or adaptation in those homes or on the evaluated objects. Those details describe Figure’s account of the system and evaluation; the underlying claims remain unverified in the reviewed evidence.
The term “zero-shot” requires particular care here. In the reported evaluation design, it referred to generalization across unfamiliar homes and manipulated objects, not to learning the chores from scratch during the test. The three behaviors had been specified using task-specific training data collected elsewhere. The result therefore concerns transfer into new physical settings, rather than the discovery of household tasks without prior task-specific preparation.
Figure reported 237 successful full-task trials out of 420, or about 56%, across the three behaviors. It reported task-level success rates of 67% for bed making, 62% for towel folding and 40% for tidying toys. The reported rubric awarded no partial credit and counted safety interventions and timeouts as failures.
Humanoids Daily and RobotTesters, two independent accounts reviewed for this article, repeated the aggregate and task-level figures. Their reporting independently confirms that these numbers were publicly reported; it does not independently validate the trials themselves. The evaluation was run by Figure rather than independently administered, and the reviewed evidence contains no independent reproduction of the test. The reported 56% result therefore remains a company-reported evaluation rather than an externally established measure of household performance.
Why this enters the Record
Embodied AI is evaluated not only by whether a system can perform a behavior, but by whether that behavior transfers when the physical context changes. Homes contain different layouts, surfaces, objects and arrangements. A system that succeeds only under the conditions represented in its training or testing environment would show a narrower capability than one that can adjust to unfamiliar surroundings. Figure’s evaluation made that transfer question its central object by moving three previously specified behaviors across 30 homes and using objects the system had not encountered before.
That makes the announcement significant as a reported evaluation design and capability claim, even though it does not settle the claim. It placed whole-body generalization in ordinary domestic environments, rather than describing performance only in a single controlled setting. For the historical record, the distinction matters: the event documents an attempt to measure how an embodied system handles environmental variation, while also preserving the limits of the evidence used to support it.
The result itself should not be read as evidence of dependable household service. A 56% full-task rate, if the company’s figures are accurate, means that 44% of the reported trials did not meet the stated success standard. The task-level spread also matters: the reported rate was highest for bed making and lowest for toy tidying. Because the rubric allowed no partial credit, the figures record completion under a strict all-or-nothing measure, not the quality of intermediate progress.
Figure’s announcement also included a scaling analysis. Across four nested Index-data runs spanning an eightfold increase in pretraining data, the analysis measured robot-action prediction loss. It did not directly measure household-task success. The analysis can therefore describe how that training objective changed across the reported runs, but it does not establish a corresponding improvement in real-world chore performance.
What the evidence establishes is narrower than the announcement’s broadest framing. Figure made the announcement and reported the evaluation, its setup and its results. Independent accounts repeated those reported numbers and clarified the meaning of zero-shot, but neither independently administered the test. The reviewed record consequently preserves both the claimed advance in unfamiliar-environment generalization and the unresolved question of how the result would withstand independent assessment.