Watch almost any promotional clip for a humanoid robot and one shot keeps turning up: a neat stack of folded laundry.
LG Electronics showed off the trick early this year, and a run of home-humanoid makers soon followed — Figure AI, Sunday Robotics, 1X, Weave Robotics. Tesla got there first, releasing footage of its Optimus robot straightening a shirt back in January 2024.
It looks a little odd, spending tens of millions of dollars to build a robot and then showing it off by folding shirts and trousers. There's a reason, though. What comes easily to human hands is brutal for a machine. A piece of laundry is crumpled differently every time, its front and back jumbled, and its shape shifts the instant you touch it.
Here's the first hidden point. The companies aren't checking how neatly the robot folds. They want to know whether it can fold a garment it has never seen before. Folding laundry, in other words, is a way to measure a robot's ability to generalize — to cope with the unfamiliar.
Clothes are so awkward for robots because they are deformable. A cup stays a cup however you grip it; a chair stays a chair when you move it. A T-shirt becomes a different object depending on where you grab it. Lift one corner and a sleeve flops down; pull in one spot and some other part trails along.
So the robot can't just decide where to grip. It also has to predict how the rest of the fabric will move once it does. In robotics this counts as its own field of study — handling objects that change shape — and researchers have built benchmarks to score it.
The variables pile up from there. Clothes differ in type, size, material, how badly they're wrinkled, and where they happen to be lying. Laundry, then, squeezes the endless unpredictability of a home — an environment nobody controls — into a single chore. Less a test of nimble fingers, more a test of reading a reality that keeps shifting.
Ayanna Howard, a roboticist and president of Spelman College, told Business Insider that teaching a robot to fold clothes has long been the "holy grail" of the field. Figure AI called it "one of the most difficult dexterous manipulation tasks for a humanoid robot."
Folding robots have appeared before, but most never made it as products. Laundroid, a Japanese firm behind a $16,000 laundry-folding machine, filed for bankruptcy in 2019. FoldiMate drew crowds at CES several times in the late 2010s, yet never shipped.
High prices and shaky reliability sank those earlier tries. The ground has shifted since. AI has advanced fast, hardware has gotten cheaper, and talent and money have rushed toward so-called physical AI. That has raised hopes of doing far more than laundry. Where engineers once fitted the robot to the wash, they now fit the AI to it.
The center of gravity has moved, too. The question is no longer simply whether a robot can fold a shirt. It's whether the machine can adapt to a real home by failing and learning, over and over. When it slips, a person steps in to correct it, and the record of that moment feeds back into the model.
Data is the deciding factor here. The internet overflows with cat photos and documents, but there is almost nothing that teaches a robot how to fish a sock out from under the couch.
Some builders are already chasing that gap. Sunday Robotics unveiled a robot last year that gathers data and learns inside people's homes. The US startup Micro AGI goes further still: in New York it sends professional cleaners and a personal chef to applicants' houses free of charge, and in return it harvests the flood of data those jobs produce.
In the end, the real contest among robot companies may not be about dexterity at all. It's about how many of the odd, unscripted situations that pop up in actual homes a robot has lived through and absorbed.
There's a practical hurdle as well. Humanoid robots usually run well past a few thousand dollars. At that price, sending your laundry out is far cheaper. To earn a place inside the home, a robot has to do a great deal more than fold clothes before real demand shows up.
Which is why the test list is long. "Tidying toys" probes whether the robot can sort objects and grasp space. Setting out dishes checks force control and fine manipulation. Making coffee tests tool use and stringing actions together: the robot has to locate the cup, the coffee, the water and the machine, then chain several moves toward one goal. Sunday Robotics says it is stretching its ACT-2 model beyond laundry into cleaning, tidying toys, zipping zippers, turning inside-out clothes right-side out, and brewing coffee.
And a true final boss waits for the home humanoid: the command "clean the house."
Folding laundry at least follows a rough script — spread the garment, square it up, fold. Tell a robot to clean the whole house and the script vanishes. It doesn't know what's on the floor, every object has its own proper place, and even the "clean" the user has in mind is fuzzy. The robot has to size up the scene and set its own order of work.
That's why recent research has begun grading a separate skill: carrying out several chores across a long stretch of time rather than one quick motion. A 2026 benchmark called LongAct does exactly this, judging how well a robot plans and executes long-horizon housework described in plain language. The best current models finished the full task just 16 percent of the time.
Folding laundry is close to doing a fixed job well. Cleaning a house means deciding what the job even is. The last wall standing between home humanoids and the living room may not be manual skill but the judgment to read a situation and choose an action.
Seen this way, the laundry-folding shot that recurs in every robot video isn't a humble chore demo. It's the first exam of how well a machine can understand the real world and learn from it.
This article was rewritten from reporting by AI타임스. Source: AI타임스