A Humanoid Walked Into 30 Homes It Had Never Seen. It Finished the Chore 56% of the Time.
The company Figure rented 30 homes around the Bay Area. It walked a robot shaped like a person through each front door. The robot stands 5 feet 8 inches tall. The team told it to clean up. Nobody gathered data inside those houses first. Nobody adjusted the robot after it arrived.
The robot finished the whole job 237 times out of 420 tries. That works out to 56%. That one number tells the whole story, good and bad.
What the video shows
Figure put the clips online on September 17, 2026. They show the Figure 03 robot doing three chores in normal rooms. It picks toys off a living-room floor and puts them in a basket. It folds towels. It makes beds it has never seen before.
Founder and chief executive Brett Adcock speaks in the video. He calls the robot’s new brain, Helix 2.5, “the most important project we’ve ever taken on at Figure.” He and artificial intelligence director Corey Lynch say the robot succeeded at least once in every one of the 30 houses. That tells you it can cope anywhere. It does not tell you how often it works.
Figure’s own captions point out the moments worth watching. The robot steps back to reposition itself. It shifts its stance. It walks around the far side of a bed to fix a fold it botched. The company calls this whole-body self-correction. It says this is the clearest sign that its new training method works.
Adcock also posted almost four hours of extra footage from the homes. The site Humanoids Daily watched it. They note that it does not show all 420 graded tries.
The 44% that went wrong
Figure marked itself strictly. That makes the failures more honest than most demo videos. A tidying run only counted as a win if all 13 to 15 toys landed in the basket. The robot had one minute per toy.
A towel run needed every towel folded and put in the basket, within three minutes each. Bed making meant both pillows and both duvet corners at the top of the bed, with the duvet pulled smooth. If a person stepped in for safety, the try counted as a failure.
Here are the results by chore. Each chore got 140 tries. Bed making: 94 wins out of 140, or 67%. Towel folding: 87 out of 140, or 62%. Toy tidying: 56 out of 140, or 40%. There was no half credit. So a run where the robot got 14 of 15 toys in the basket counts as a failure. Keep that in mind when you read the 56%.
Tony Zhao works at a rival firm, Sunday Robotics. He pushed back in public. He said the robot fails about half the time. He argued that useful work needs two things: coping with new places and being reliable.
His company says its own system folded 778 times successfully out of 785 tries. You cannot compare those numbers directly. Sunday counts one piece of clothing at a time. Figure counts a whole chore with many objects. But Zhao’s main point holds. A robot that wrecks your bed twice a shift is not one you leave alone in the house.
The real news is where the skill came from
This part matters more than the video. Figure trained two versions of the robot’s brain on the same robot data. It changed one thing only. One version first learned from Index. Index is Figure’s crowd-gathered collection of ordinary people filming themselves doing everyday tasks.
The version trained from scratch finished 9% of tries. The version that first watched human videos finished 56%. Same chore data. Same design. Same marking.
Figure also says the new brain needed half as much robot-specific training data as an older version. And that older version had been trained inside the very room where it was tested. Figure says it doubled the amount of human video and the model’s accuracy improved so predictably that it could forecast the result of its biggest training run before running it.
Index now takes in about 35 minutes of human footage every second. More than 90,000 people contribute each week. The company has committed $3.5 billion worth of computing power to the training.
What this means for you right now
Nothing yet. Figure gave no price for shoppers and no delivery date alongside Helix 2.5. The company ran all these tests itself. Nobody outside has repeated them. So treat the percentages as Figure’s own report card.
What has changed is the money side of the idea. Until now, a home robot mostly had to be taught your particular house. Now chores learned once may carry over into a stranger’s bedroom. If so, the costly part of a home robot is no longer setting it up for each customer. That matters more than any single demo.
So watch for one thing next. Not a prettier video. Watch whether someone outside the company can put one of these robots in an unfamiliar home and get numbers close to these.
