Humanoid robots face the unfamiliar home: why it matters beyond housework
Figure’s latest experiment tests whether learned physical skills can travel between homes. The result raises questions about useful automation, deployment costs and how much human attention a robot should need.
On this page
Making a bed looks like a modest ambition for artificial intelligence. Until the bed is a different height, the bedding behaves differently and there is barely enough space to reach the other side.
Domestic work contains countless small variations that people accommodate without much conscious thought. A robot must turn those variations into decisions about where to stand, what to grasp and how to continue when the first attempt fails. An unfamiliar home makes those demands difficult to avoid.
What Figure’s experiment shows
On 17 September, Figure reported testing its Helix 2.5 system on three behaviours across 30 previously unseen homes in the San Francisco Bay Area: tidying living rooms, folding towels and making beds. In a controlled comparison, pretraining on its Index dataset of human behaviour raised complete-task success from 9% to 56%. Figure’s research report
Figure describes the evaluation as “zero-shot”. Here, that means the homes and manipulated objects were unfamiliar; the three behaviours had been trained using data collected elsewhere. Each task used a fixed version of the model across all 30 homes, without adapting it to individual properties. Evaluation methodology
The experiment tests whether prior learning transfers to new circumstances. It does not establish that a robot can arrive and undertake whatever job somebody happens to request.
The wider significance concerns whether a learned physical skill can remain useful when its surroundings change. If that becomes reliable, it could alter the effort required to introduce robots into everyday environments.
Why adaptability changes the cost of deployment
Consider a hotel with bedrooms across several buildings. Furniture, floor space and bed heights may differ. A system requiring extensive demonstrations in every room could create a substantial installation burden. A robot able to carry its skills between rooms could reduce that burden, making deployment easier to repeat.
That is an illustrative possibility, rather than a hotel application demonstrated by Figure. It shows why adaptability has economic value. Every additional location that requires engineering time adds cost before useful work begins.
There is already a separate commercial story developing around more bounded tasks. Agility Robotics has reported that its Digit humanoid moved more than 100,000 totes at GXO’s Flowery Branch facility. That provides evidence of repeated work in a commercial setting, according to the supplier. Its task and measurement differ from Figure’s household evaluation, so the results cannot be treated as a shared league table. Agility’s deployment report
Together, these examples suggest two useful dimensions for assessing progress: how consistently a machine performs a defined job, and how well its abilities survive a change of setting. A prospective buyer needs evidence relevant to the combination their operation requires.
The practical calculation extends beyond the purchase price. Setup, charging, maintenance, supervision and unfinished work all affect the cost of a satisfactorily completed task. A machine can move impressively while still leaving its owner with an unattractive calculation.
Human attention is part of the equation
Imagine a robot that folds towels but regularly needs someone to untangle an awkward item. Each intervention may be brief. Across a working day, those interruptions could make staffing difficult to organise. The relevant measure would include how much human time the whole process consumes, alongside the work the robot completes.
The same issue becomes personal at home. Making a bed or moving household objects could offer considerable value to someone who finds those movements painful or exhausting. Suitability for such use would need its own evidence. A person should not have to be physically capable of rescuing a struggling machine to benefit from its help.
Useful assistance therefore includes knowing when to stop, how to request help and how to leave a task in a manageable state. The quality of failure can matter almost as much as the quality of success.
It also matters who carries the additional responsibility. If a robot reduces physical effort but leaves someone constantly watching for mistakes, the benefit is different from being able to leave a task in its care. Both may have value, but they answer different needs.
Where commercial opportunities could emerge
For new ventures, this creates several possibilities worth testing. Installation, maintenance, integration with existing operations and carefully defined services could become businesses around the robot itself. Their prospects would depend on customer demand and dependable operating performance. A technically impressive machine does not automatically create an affordable service.
An early commercial proposition might therefore be quite specific: a recurring physical task, in a defined environment, with an agreed standard of completion. That would give customers something concrete to evaluate and founders a clearer way to test whether the economics work.
The comparison should include the alternatives already available. A different tool, a simpler machine or a change in the way work is organised may solve the same problem. The strongest opportunity will be one where the robot’s adaptability earns its additional complexity.
The next test is repeated usefulness
Figure’s results remain company-reported research covering three behaviours. The distance to dependable everyday use is substantial. Independent evaluation and sustained operation would help establish how broadly the reported improvement holds.
For SingulX, the development deserves attention because adaptability could change where automation becomes practical. Skills that travel could reduce the work of installing a robot, extend the places it can serve and make assistance available in more varied circumstances.
The next meaningful test is repeated usefulness: work completed to an acceptable standard, across changing conditions, with a manageable demand on human attention. That is where an unfamiliar home becomes a test of something much larger than housework.