Figure Tested Its New Robot in 30 Homes. Did It Work

We've all seen those videos of robots handling all kinds of chores around the house, cooking meals and some even driving a car. But none of us has actually had one of these robots at home to find out whether it's really that good. Until now that is. For the first time, Figured decided to take its robots out of the lab and test them not in controlled demos with engineers watching, but in the homes of regular people. 30 lucky people were given humanoid robots to see what they could do in real world conditions. And the results turned out to be surprising and pretty controversial. The head of Sunday Robotics, which is also building home robots, even called the whole experiment a flop. The trial started where it only made sense to start, the living room. This is the first of three tasks. Before each attempt, a person scatters up to 15 toys around the room, completely at random. The robot has to pick them all up without ever letting go of the basket. It reaches for each item in a way that keeps it from losing its balance or bumping into the furniture, then scans the room for the next one. Each toy gets one minute, and if the robot fails to clear even one of them in time, the whole attempt doesn't count. Engineers call this whole body control. The legs, torso and arms all have to work as one unit. For an industrial arm, that's not a problem. It's bolted in place, and everything it works with is within reach. A wheeled robot needs clear floor space, and enough room to drive up and turn around. And in an apartment, you can't always count on that. A humanoid has to walk up to the object on its own, position itself so it can reach it, and actually see it through its own cameras. Figure stresses that vision, walking and handwork can't be tuned separately here. The robot went through 30 living rooms like that. The test took place in homes around San Francisco, and everywhere it went, the robot had to work with whatever the owners had. Their couches, beds and tables. Furniture height, open space, lighting, all of it changed from house to house. And here's the key part. The robot wasn't prepped for these homes ahead of time. Not a single frame of footage was recorded inside them for training, and the machine wasn't fine tuned after arriving. The toys, towels and laundry for the test were set aside in advance, and the team checked that the robot had never seen them in training. It had been taught to fold and tidy, sure, but in other places and with other items. Basically, each of the 30 homes was to the robot what your apartment will be someday. But there are some caveats. Those 30 lucky people didn't exist. Figure rented the homes specifically for the tests, and there were people standing by next to the robot, ready to step in. So this wasn't a robot moving in with an ordinary family. It was the same old test, just moved out of the lab and into other people's apartments. That said, at least Figure didn't hide from the owners why they wanted the house, unlike another startup, the Bot Company. In May, it ended up in court over exactly that. An Airbnb host in San Francisco claims the company's employees rented the apartment posing as tourists so they could secretly test robots there, and left damage behind. And he's not alone. According to local press, at least 12 other hosts complained about the same guests. So far, these are only allegations. But the story shows how desperate companies are to find other people's homes for their robots. Back to the tasks. After the toys came the towels, and that's a whole different kind of challenge. A solid object is something you can just pick up and carry, but fabric changes shape with every move. Pull on one edge and the whole towel shifts. Make one bad fold and the corners aren't where they used to be. On top of that, every home has its own towels, different sizes, thicknesses, and materials. The robot had to fold all four and put them in a basket, with no more than 3 minutes for each one. Otherwise, that's a strike. Next up was the bed, the biggest surface of the three. The robot can't reach the far edge from where it stands, so it has to walk around the bed and work from different sides. Under the rules, it had to put both pillows at the head of the bed, pull the corners of the comforter up, and smooth it out. Each pillow and each side of the comforter got 1 minute. Here's an interesting detail. The footage shows pretty clearly how the new version of the robot differs from the older ones. If the comforter lands crooked, the robot doesn't just freeze. It steps back, shifts its stance, comes at it from the other side, and fixes the wrinkle. Figures engineers point to this ability to correct its own mistakes as one of the most noticeable improvements in the new generation. In total, the robot made 420 attempts, 140 for each task. It completed 237 of them. That's 56%. A little better than every other attempt. The bed performed best at 67%, and towels came in at 62%. The worst result went to what looks like the easiest task. The robot picked up all the toys in only 40% of attempts. Why toys of all things? Let's do some quick math. Say the robot handles each individual toy 94 times out of 100. That's nearly perfect. But it only gets credit if it doesn't miss once across a set of 15 items. One slip, and the whole attempt fails. At those odds, the whole room gets cleaned up about 40% of the time, which is exactly what the test showed. Figure didn't publish how the robot did on each individual toy, so that's not their data. Just simple arithmetic. But the point is clear. When a task is long, tiny slip ups at every step pile up and sink the whole attempt. That said, the success criteria are more lenient than you might think. A towel counts if it's folded and in the basket. The quality of the fold is scored separately. The best score means all four corners lined up to within an inch, and the worst means none of them did. A sloppy fold. But even that counts as a pass. The bed is similar. A pillow counts even if it's turned almost 45 degrees. And the corners of the comforter can be off by up to a foot. Figure doesn't say how many of those sloppy folds and messy beds went into the final percentages. And finally, time. The report doesn't say how long the robot actually took. But under the rules of the test, it was allowed a full 12 minutes to fold four towels, and up to a quarter of an hour to clean up the toys in the living room. For comparison, back in 2010, a robot at Berkeley took 25 minutes to fold a single towel. And at CES this January, an LG robot did it in about a minute, with a human handing it the towels. Not everyone thinks that's a good result though. Tony Zhao, the head of Rival Sunday Robotics, wrote that Figure's robot fails half the time. Meanwhile, his own wheeled robot, Memo, folded clothes in strangers' homes back in July with a 99.1% success rate. Sounds like a massacre. But there's a catch. At Sunday, one attempt means one folded item. At Figure, an attempt means the entire task from start to finish. And the lowest score is on the task that had 15 steps stacked into it. So the 56% and the 99% can both be true at the same time. What's more, if you scale 99.1% up to a basket of 15 items, you get a success rate of around 87%. Not 99 anymore. On the other hand, by our estimate, Figure's robot handles a single toy about 94% of the time. So on the reliability of a single action, Zhao has a point. His formula goes like this, useful work is the ability to operate in an unfamiliar place plus reliability. Figure has shown the first part. The second one, not so much. The industry has a concept called 9s. Andrej Karpathy, Tesla's former head of AI, explained it like this, if a demo works 90% of the time, that's just the first 9. Getting from 90% to 99% takes as much work as it took to get to 90% in the first place. And getting from 99% to 99.9% takes just as much again. Rodney Brooks, co-founder of iRobot, thinks it takes more than 10 years to go from a lab demo to 99.9%. And another 10 years for every 9 after that. Figure doesn't even have the first 9 yet. If the robot did these tasks every day, it would nail all 3 on only 1 day out of 6. Still, by industry standards, 56% isn't embarrassing. Take Unitree for example, it's one of China's best known robot makers. Its R1 humanoid costs less than $6,000, making it one of the cheapest models on the market. In August, Unitree's CEO, Wang Xingxing, laid out a goal, a robot that can handle about 80% of tasks in an unfamiliar home, and he added that we're at least 2-3 years away from that. Meanwhile, at a Stanford household tasks competition, the best team in the world got its success rate up to just under 12.5%. And that wasn't even in real homes, but in a computer simulation. By the way, Figure calls its experiment the first of its kind in the world. But that's only true with some qualifiers. In April 2025, Physical Intelligence sent its wheeled robot into 3 homes it had never seen. And back in 2024, New York University ran an experiment where a robot moved 10 items around in unfamiliar homes. The success rate then was 58.5%. But for a humanoid on legs, controlling its whole body, and at the scale of 30 homes at once, Figure really is the first. But the most interesting part of Figure's report isn't even these percentages. The company ran a control experiment. Two identical models were trained on the same three tasks and tested in the same homes. There was only one difference. The first had been pre-trained on a huge archive of human actions beforehand, and the second hadn't. Here's what came out. Without that kind of prior experience, the robot successfully finished only about 8% of attempts. With it? 56%. Toys, 5% vs. 40%. Towels, 9% vs. 62%. Bed, 11% vs. 67%. In other words, the same robot, shown the exact same specific tasks, did roughly 7 times better in an unfamiliar home if it had first learned from human experience. And this is where it's worth explaining what that archive actually is, and why it made such a difference. This spring, Figure launched an app. Anyone can film themselves cooking, doing laundry, cleaning, or working in a cafe, upload the video, and get paid for it. That's the archive, and it's called Index. Originally, the company tried to buy this kind of data from contractors, but they couldn't keep up with either the volume or the variety. The app, on the other hand, had just 99 users in its first week after launching in May. By the end of August, there were more than 44,000 a week. And by early September, according to the company's founder, Brett Adcock, almost 70,000. All told, people have uploaded more than 16 million videos to Index, and Figure has paid them $15 million. Sounds substantial, but that's less than a dollar per clip on average. Right now, the app takes in 35 minutes of footage every second. So while you watch one minute of this video, more than a full day of other people's lives gets uploaded to Index. Ego4D, the largest academic dataset of first-person video, put together by Meta and 13 universities, is 3,670 hours. Index racks up that much in under two hours. And people film absolutely everything, from scooping the cat litter box to changing the oil in their car. Even so, none of the test's three tasks makes up more than 2% of the archive. The robot didn't learn to fold towels by re-watching thousands of folded towels. It saw how people handle things in general. Grabbing, pulling, smoothing, bending down, walking around furniture. The specific task was shown to it afterward through separate examples. By the way, the app has a second, pretty practical side. Through it, you can invite one of the other users to your home for an actual cleaning. Figure explains the move by pointing to its main goal. Today, people come to your place to help you clean. Someday, a robot will do it all. And while the idea behind Index is pretty original, Figure isn't the first to try it. Tesla switched to the same approach back in the summer of 2025. The people collecting data for Optimus work 8-hour shifts wearing a helmet and a backpack with 5 cameras, which weighs somewhere around 30-40 pounds. Over a shift, they have to bring in at least 4 hours of usable video. One of the collectors described the job like this, a lab rat under a microscope. A whole market has already grown up around this kind of filming, and the price for the same work depends on the country. In the Indian state of Tamil Nadu, a worker folds towels in a mock-up bathroom, about 90 clips a day for 250 rupees an hour, which comes to roughly $2.60. In Los Angeles, they pay $80 for 2 hours of filming your own cleaning. And DoorDash, the biggest delivery service in the US, launched a separate app for its couriers, where they can make some extra money by filming themselves folding laundry, washing dishes, or making a bed. China is going at it on a national scale. The online retailer JD.com, together with the city government of Sichuan, plans to record 10 million hours of video over 2 years. A housewife there who films herself cooking and doing laundry earns about $3 an hour, and she says, And according to The Guardian, So the test in 30 homes isn't one company's discovery. The whole industry has bet on the same recipe at the same time. Teach the robot from videos of people, then give it as much data and computing power as you can afford. But why the race for millions of videos? Let's go back to Figure's report. They have an argument that seems to be the whole reason the document was written. They trained 4 versions of the model on different amounts of the archive. A portion, twice that, 4 times, and 8 times. With every doubling of data, the model's error dropped along the same straight line. So neatly, in fact, that they predicted the error of the biggest version to 4 decimal places before it was even trained. Predictability matters a lot to investors. It means you can work out ahead of time that if you put in twice the data, you'll get this much improvement. Without that, billion-dollar investments look like a lottery. And with a graph like this, they look like a calculation. Patterns like that are exactly what turned language models like ChatGPT into a race for billions. And Figure has put its money on it. In early September, it signed a $3.5 billion deal for access to 100,000 next-generation NVIDIA GPUs. For scale, that's about as many GPUs as were in Colossus, the supercomputer where Elon Musk's ex-AI trained Grock when it launched in the summer of 2024. The $3.5 billion itself is almost double what Figure has raised from investors in its entire history, around $1.9 billion. And the company plans to expand the deal to $6 billion. Meanwhile, Figure doesn't disclose its revenue. But the graph itself deserves a closer look. The scale on the left doesn't start at zero, which makes the line look steep. In reality, an eight-fold increase in data cut the error by only about 4%. And there's a catch. A drop in this error just means the model gets better at guessing the right movements on test videos. That doesn't necessarily mean the robot got better at cleaning rooms. Researchers at Tsinghua University, who have studied scaling laws for robots themselves, pointed out that this kind of error rate often doesn't match how well a robot actually performs in the real world. On top of that, NVIDIA already published a similar pattern for human video back in February. But this bet has some tough critics. Rodney Brooks wrote an essay a year ago on why humanoids won't learn dexterity from video. His argument? Our hands don't work by sight, they work by touch. The skin of the palm has around 17,000 receptors that sense pressure, slip, and vibration. None of that is in a video. Believing a robot will pick up dexterity from recordings like these any time in the next few decades, Brooks says, is pure fantasy. Even people who back learning from data have doubts. Sergei Levine, a Berkeley professor and co-founder of physical intelligence, warns that a model trained on humans will solve a problem the way a human would, not the way that's easiest for a robot. And Ken Goldberg, also at Berkeley, calculated that language models were trained on text that would take a person about 100,000 years to read. Robots have no way of getting a stockpile of data like that, and in his view, that gap won't close in five years or in 10. Still, Figures Test has something to say back. Human video took the success rate from 8% to 56%, and that's hard to chalk up to chance. But towels, toys, and a comforter are exactly the kind of tasks where delicate finger work barely matters. Zipping a zipper, tying a shoelace, or washing a fragile wine glass are things Figure hasn't tried in public. So this test didn't settle the argument between Brooks and Figure. And the last question about those 56%? How far can you actually trust Figure? The company already has a history with BMW. In early 2025, Adcock wrote about a whole fleet of robots working at the factory every day. Fortune reporters asked BMW itself, and the answer was, there's one robot, and it moves parts outside working hours. Adcock called the article outright lies and threatened to sue. Then, when the year-end numbers came in, it turned out that one Figure 02 really had been working. 1,250 hours, three parts per car. The home story is similar. In February 2025, Adcock announced he was moving up the home timeline by two years and starting alpha tests in homes that very year. No public results from those tests ever appeared. In January 2026, he predicted that this year, humanoids would be doing multi-day tasks unsupervised in homes they'd never been to. In September, what we got was three short tasks, rented houses, and a 56% success rate. And back in the spring, Adcock said you'd soon be able to rent a Figure 03 for your home for around $600 a month. To this day, there's no price, no pre-order, not even a wait list. There's a safety question, too. In November 2025, Robert Grundle, Figure's former head of product safety, sued the company, and his lawsuit contains some genuinely chilling details. He claims the robot's strike is more than twice as strong as what it takes to fracture a human skull. And according to the suit, during one malfunction, the robot left a dent in a steel refrigerator door right next to an employee. Figure denies all of it, calls it lies, and has filed a countersuit, so the court hasn't established anything yet. But note that children and pets weren't part of the 30-home test. That said, this time, Figure did more than is typical for the industry. The scoring criteria were locked in ahead of time, the grading was blind, and any human intervention for safety reasons counted as a fail. Adcock also posted several hours of raw footage, failed attempts included. Compared to polished clips where the robot nails it on the first try, that's rare. So what home robots exist today? The best known one you can buy right now in the US is the Neo Humanoid from One X. The company says it collected more than 10,000 pre-orders in five days. But at least through early September, there hadn't been a single confirmed delivery to a customer. And when a Wall Street Journal reporter went to see Neo in person, a human in a VR headset was doing most of the hard work for the robot. That's actually the One X model. Whatever the robot can't do on its own, a company employee does for it. The owner schedules a time, and an operator remotely controls the robot right there in their apartment. While a person is connected, the glowing rings on the robot's ears change color. In a July YouGov poll, almost half of Americans said they'd turn down even a free cleaning if the recordings went toward training robots. Tesla doesn't even have that much. In 2022, Musk promised Optimus in three to five years for under $20,000. And now he's talking about sales by the end of 2027, and only if the robot turns out to be reliable and safe. And the pioneer of home robots, iRobot, maker of the Roomba vacuum, went bankrupt last December and ended up owned by its own Chinese contractor. Even a robot vacuum turned out to be a tough business. On top of that, according to surveys, more than two-thirds of Americans aren't willing to pay more than $5,000 for a home robot. That's why the first robot to show up at your door probably won't be a humanoid. Sunday is giving away its wheeled memo to 50 families by Thanksgiving. The startup Weave is already selling a two-armed robot on wheels for $7,999. It folds laundry on its own, and a remote operator helps out with the tricky items. So the nearest thing to a home helper is a cabinet on wheels with arms that folds laundry slowly and sometimes calls a human for backup. In other words, nobody is selling a helper that doesn't need help yet. So did Figures Robot pass? As a household helper, no. In a stranger's home, it finished the job a bit more than half the time, and that's counting sloppy work as a pass. Getting all three tasks done in one go would happen on roughly one day out of six. It's still a long way from the 80% Unitree's CEO is talking about, and even further from the nines Karpathy and Brooks described. But as an experiment, the test worked. Without human video, the robot succeeded in about one attempt out of 12. With it, more than one out of two. And Tesla, Nvidia, and dozens of startups all arrived at the same conclusion at the same time. If they're right, the home robot stops being a question of invention and becomes a question of scale. How many more hours of their lives people will film and how many billions will go into computing? And if Brooks is right, as he himself warns, an awful lot of money is simply going to vanish. Meanwhile, thousands of people around the world, some for $3 an hour, some for 40, are filming themselves folding towels and washing dishes so that someday nobody will have to. And if you need a cleaning right now, Figure will gladly send a human over. So what about you? Would you get paid to film yourself cleaning your apartment? And would you let a robot into your home that learned from other people's homes? Let us know in the comments.