A humanoid robot cleaned my apartment: a glimpse of the future, still with human help (Tau Robotics)

MIT Technology Review publishes a first-person account by Grace Huckins, who hired a humanoid robot cleaning service in San Francisco. The service is offered by the startup Tau Robotics. The author, who doesn't normally pay anyone to clean her home, joined the waitlist after seeing a post on X. She didn't tell the company she was a journalist, though she used her real name. She disclosed her profession weeks after the cleaning.

The key point is that the robots are not yet autonomous. They are remotely controlled by human operators, who watch what is happening through cameras mounted on the robot's head and wrists. As they clean, the robots collect data that Tau will use to train more independent models. Every countertop wiped and every T-shirt folded helps the models improve next time. Because that data is so valuable to the company, the price is low. When the author booked, an hour of cleaning with one robot cost $30. It now costs $60 for a weekly cleaning with two robots, or $100 for a one-off visit.

Even at the $30 rate, the author doesn't consider it a great deal. The robots are thorough but slow. If you end up watching them for the entire visit, as she did, you'd be better off cleaning yourself.

Each robot is about four and a half feet tall (around 1.4 meters). It has a 3D-printed “head” with antennas to communicate with Tau's headquarters and two claw-shaped “hands,” also 3D-printed. According to the article, more humanlike hands aren't durable enough, and the claws work surprisingly well for most cleaning tasks. The author's robot wore a Tau Robotics T-shirt and non-slip socks to avoid scratching the hardwood floors.

According to the author, the system is more like a remote-control car than a Waymo. Tasks such as making the bed or folding laundry require “kinematic guidance”: the human operator performs the movements and sensors relay them to the robot, which copies them. These tasks generate especially valuable training data. Tau's models may not have solved laundry folding yet, but learning from real examples is the way to get there. According to co-founder Cornelia Weinzierl, the robot can already do some other tasks without human help, such as vacuuming and mopping the floor.

The robot's arrival was a spectacle in itself. It showed up in a blue minivan, and Weinzierl and another Tau employee helped it out. Its first obstacle was the eight steps at the building's entrance. The author says it climbed them “as if it were drunk.” Each successful step could end in a backward stumble that erased its progress, though it never fully lost its balance. Weinzierl explained that the odd gait is partly because the robot is operated from Tau's headquarters. The slight lag in signal transmission affects its coordination. She said the work and time they spend just on optimizing timing and connectivity is “insane.”

The carpeted staircase leading up to the apartment defeated the robot entirely, and Weinzierl had to carry it up the last few steps. During the first cleaning, the robot stopped several times. Weinzierl initially blamed overheating due to the unusual heat. Later, a colleague told her that what had shut down from the heat was the equipment the human operator used to monitor progress in the apartment. Without remote assistance the robot couldn't clean, so it left before completing the booked hour. In September, Weinzierl or another Tau representative stayed with the robot for the entire visit. She says this is no longer necessary.

On customer service, Tau delivered. Weinzierl emailed an apology, didn't charge for that visit and offered another one for free, “minus whatever price I put on the privacy of my data.” All of this happened before the author revealed she was a journalist.

The second cleaning, about a month later, went much better. The robot climbed the stairs without much help, though with the same wobble. In an hour it competently cleaned the kitchen, though it didn't attempt the stovetop. It also tackled the toilet, the task the author dreaded most. It was its most impressive performance: three times the brush head fell off, and three times it managed to reattach it, feeling around with the handle until it found the right angle.

The robot is still imperfect, as you'd expect from a beta product. It left the faucets dripping, knocked over a closed bottle of olive oil and couldn't pick up an air purifier it had toppled. The author forgave the mistakes because she found it so endearing. Watching it clean the countertops, she felt the same gratitude she feels when a Waymo gets her home safely after a night out.

The article also points to the less pleasant side. If Tau Robotics or its competitors manage to train robots to clean autonomously and cheaply, some people could lose their livelihoods. According to the author, the United States is far from adopting policies to protect those who lose their jobs to AI. Weinzierl argues that, in the long run, handing cleaning over to robots could make sense. She says it is work many people don't want to do, with high turnover and a high risk of injury: “We need people and we don't have them, and nobody wants to do it.” She added that the company is starting to consider using the robots for other dangerous or physically demanding jobs, such as mold remediation and home healthcare.

Tau still has a long way to go. The amount of data needed to train robots on the varied and complex tasks of cleaning an apartment is probably enormous, and the company has done only a couple of hundred cleanings. Even so, the author finds it remarkable that humanoids capable of cleaning toilets and mopping floors already exist, even if they need a good deal of human help. She concludes that the future of The Jetsons no longer feels so far off.

Editor's note: the material received corresponds to the text of the article. This summary is limited to what that text says and adds no outside information.