A humanoid robot cleaned a reporter's San Francisco apartment, with remote human pilots

A San Francisco startup called Tau Robotics is sending humanoid robots into people's homes to clean them. An MIT Technology Review reporter booked the service without saying they were a journalist and had two visits. In their firsthand account, the robots wiped counters, cleaned a kitchen and scrubbed a toilet. They were slow and clumsy, and for most of the work a human was controlling them from somewhere else.

The service matters less as a cleaning product than as a way to collect data. Tau's stated goal is autonomous cleaning robots. Each paid visit records how a cleaning task is done in a real home, and that material is used to train the models that will drive later versions of the robot.

How the service works

The robots are four and a half feet tall. Each has a 3D-printed head with antennae for communicating with Tau's headquarters, and two 3D-printed claw hands. According to the report, hands that look more like human hands were not durable enough, and claws handle most cleaning tasks well. The robot that visited wore a company shirt and grippy socks so it would not scratch hardwood floors.

For now, human pilots at Tau's headquarters operate the robots remotely. They watch through cameras mounted in the robot's head and wrists. Many tasks, including making beds and folding laundry, use what the company calls kinematic guidance: the pilot performs the motions, sensors capture them and the robot copies them. The report says these guided tasks produce especially valuable training data. Tau cofounder Cornelia Weinzierl told the reporter that the robot can already vacuum and mop without human help.

Because the data is worth something to Tau, the price is lower than a new technology might suggest. When the reporter booked, an hour with one robot cost $30. The service now costs $60 for a weekly cleaning with two robots, or $100 for a single visit. Tau has completed only a couple hundred cleanings so far.

What happened in the apartment

The first visit went badly. The robot came in a minivan with Weinzierl and another employee. It climbed the eight steps to the building's front door unsteadily, often staggering back down after a step, though it never fell. It could not manage the carpeted indoor staircase, and Weinzierl carried it up the last few steps. Weinzierl said the odd gait comes partly from remote operation, because a short transmission delay affects the robot's coordination. "The amount of work and time that we spend just optimizing, for example, timings and connectivity is crazy," she said.

During that visit the robot kept stalling. Weinzierl first thought the robot was overheating in hot weather. A colleague then told her the heat had shut down the equipment the remote pilot was using to watch the robot. Without its pilot the robot could not clean, and the visit ended early. Tau did not charge for it and offered a second visit.

The second visit, about a month later, went better. The robot got up the stairs with little help. In an hour it cleaned the kitchen, though not the stove, and it cleaned the toilet. The head came off the toilet brush three times, and three times the robot reattached it by probing with the handle until it fit. It also left taps dripping, knocked over a closed bottle of olive oil, and toppled an air purifier it could not set upright again. The reporter concluded that the robots are thorough but slow. At $30 an hour, someone who stands and watches would save time by cleaning themselves.

In September, a Tau representative stayed with the robot for the whole visit. Weinzierl says that is no longer necessary.

The company's case for robot cleaners

Weinzierl argues that cleaning is work that people leave. "Cleaning is not a job that a lot of people really actively want to be doing," she said. "There's high turnover, there's a high risk of injury." She added: "We need people, and we don't have them, and no one wants to do it." She said Tau is also starting to think about using the robots for other dangerous or physically hard work, such as mold remediation and home health care.

The report also notes the other side. If Tau or its competitors manage to train robots that clean on their own and cheaply, people who clean homes for a living could lose their income, and the United States is a long way from policies that would support workers displaced by AI. The amount of data needed to teach a robot the varied tasks of cleaning an apartment is probably enormous, and a couple hundred visits is a small start.

What it means for work

Right now Tau's robots do not replace human labour. They move it. Someone still does the cleaning, but the work runs through a remote pilot at headquarters instead of a person in the home, and the motions are recorded so a model can learn them. Vacuuming and mopping are the first tasks the company says the robot does alone. Folding laundry and making beds still depend on human demonstration. For domestic cleaners, the immediate effect is close to nothing, given the speed, the errors and the small number of visits. The open questions are how much data autonomy actually needs, whether connectivity and hardware can be made reliable in ordinary homes, and whether the people who now pilot the robots are training the systems meant to take over their work.