Startups & Funding

A Humanoid Robot Cleaned a San Francisco Apartment for $30 an Hour

Tau Robotics charges $30–$100 for humanoid robot cleanings in San Francisco. The robots are teleoperated, collecting data to train fully autonomous descendants — and they still struggle with stairs.

I watched a humanoid robot clean my apartment. It felt like a taste of the future.
I watched a humanoid robot clean my apartment. It felt like a taste of the future.karishea / pixabay
By Sophie Lindqvist5 min read

Why it matters

  • Tau Robotics charged $30 for an hourlong single-robot cleaning; current pricing is $60 weekly or $100 one-time.
  • The robots are remotely piloted by humans via head- and wrist-mounted cameras, and collect training data for future autonomous models.
  • Tau has completed only a couple hundred cleanings so far.
  • Cofounder Cornelia Weinzierl says the robots can already vacuum and mop without human help.
  • The four-and-a-half-foot robot uses 3D-printed claws instead of humanlike hands for durability.

A San Francisco startup called Tau Robotics sent a four-and-a-half-foot humanoid robot to clean a customer's apartment for $30 an hour last month — and the machine, piloted remotely by a human, mopped floors, wiped countertops and even scrubbed the toilet. The service is one of the first real-world deployments of humanoid robots in ordinary homes, and the trial, described in a detailed first-person account, shows both how far the technology has come and how much human labor still sits behind the curtain.

The customer, who joined the waitlist after seeing an X post about the service and did not tell Tau she was a journalist, booked an hourlong cleaning with a single robot at the $30 introductory rate. Tau's pricing has since changed: the company now charges $60 per weekly cleaning with two robots, or $100 for a one-time visit.

Those prices are low for a reason. The robots are not autonomous. Human pilots watch through cameras mounted in the robots' heads and wrists, and every counter wiped and T-shirt folded becomes training data for the AI models Tau hopes will eventually run the robots on their own. In effect, customers are paying a discounted rate to generate the dataset that could put the robots' human operators out of a job.

How much of the robot is actually a robot?

Right now, less than you'd think. Many tasks, including making the bed and folding laundry, require what Tau calls kinematic guidance: a human pilot performs the motions, sensors transmit them, and the robot copies the movements. These guided tasks are especially valuable as training data, because the models learn from real-world examples of completed work.

Tau cofounder Cornelia Weinzierl said the robots can already handle some tasks — vacuuming and mopping among them — without any human help. The laundry problem remains unsolved, but learning from robots that have physically folded clothes in real apartments is, in Tau's view, how the models will get there.

The hardware itself is deliberately unglamorous. Each robot has a 3D-printed head with antennae that communicate with Tau's headquarters, and two 3D-printed claw hands, because more humanlike hands apparently aren't durable enough. Claws, it turns out, do surprisingly well at most cleaning tasks. The visitor wore a Tau Robotics shirt and grippy socks so it wouldn't scratch hardwood floors.

What happened when the robot met the stairs?

The arrival was already a spectacle. The robot showed up in a blue minivan, helped out by Weinzierl and another Tau employee, and immediately faced the eight stairs to the building's front door. It climbed, in the customer's description, as if it were drunk — each successful step was as likely as not to be followed by a stumbling backward stagger that erased its progress, though it never fully lost its footing.

The indoor carpeted staircase to the apartment defeated it entirely, and Weinzierl had to lug the robot up the last few steps. The explanation for its strange gait, she later said, is the signal transmission delay between Tau's remote headquarters and the machine. "The amount of work and time that we spend just optimizing, for example, timings and connectivity is crazy," she said.

The visit then hit a stranger snag. The robot repeatedly stalled mid-clean, which Weinzierl first attributed to overheating in unseasonably hot weather. The real culprit turned out to be the heat shutting down the equipment the human controller was using to watch the robot's progress. Without remote assistance, the robot could not clean, and it left before the booked hour was up.

Tau's customer service, though, drew praise. Before learning the customer was a journalist, Weinzierl emailed an apology, comped the visit and offered another one free. That's the takeaway: the hardware is fragile, but the company behind it is responsive.

How did the second cleaning go?

A month later, the return visit went much more smoothly. The robot climbed the outdoor stairs largely unassisted, still with the drunken stagger. In its hour on site, it did what the customer called a creditable job on the kitchen — though it steered clear of the stove — and took on the toilet, the most dreaded task on the list.

The toilet performance was its finest moment. Three times the head fell off the cleaning wand, and three times the robot reattached it by poking around with the handle until it made contact at exactly the right angle. For a beta product, that kind of recovery matters as much as the cleaning itself.

The robot remains rough around the edges. It repeatedly left the taps dripping, knocked over a closed bottle of olive oil, and toppled an air purifier it couldn't right. The customer forgave all of it, describing the same fond gratitude she feels when a Waymo delivers her home safely. There's a pattern here: imperfect autonomy, when it's visibly trying, earns patience that a broken appliance never would.

Why does this matter beyond one apartment?

The stakes are larger than novelty. If Tau or its competitors succeed in training robots to clean autonomously and cheaply, people could lose their livelihoods, and the United States is far from implementing policies to support workers displaced by AI. That is the real bet underneath the charming claw hands and grippy socks.

Weinzierl argues the labor market case is on her side. "Cleaning is not a job that a lot of people really actively want to be doing—there's high turnover, there's a high risk of injury," she said. "We need people, and we don't have them, and no one wants to do it." She said Tau is also considering eventually deploying the robots for other dangerous or physically demanding work, including mold remediation and home health care.

The scale remains tiny. Tau has completed only a couple hundred cleanings so far, and the volume of data needed to train robots for the varied, complicated tasks of apartment cleaning is probably enormous. Back in September, a Tau representative stayed with the robot throughout each visit; Weinzierl says that's no longer necessary.

Still, the direction is clear. Humanoid robots that clean toilets and mop floors now exist in the commercial wild, even with generous human help, and every paid cleaning session makes the next generation a little more independent. The Jetsons, as the customer put it, no longer feels quite so far away.

Original: nytimes.com

Share this article:

More from Sophie Lindqvist

Sophie Lindqvist

Show full bio

Staff writer covering regulation, policy and their impact on artificial intelligence and machine learning at AI In Context.

371 articles

Related articles

  1. Danu Robotics raises $5M to sell cheaper AI recycling robot H.E.R.O.
  2. OpenAI confirms it will build a humanoid robot; Altman says the brain is the hard part
  3. Eka Robotics Raises $150M at $1.45B Valuation on Force-First Robot AI
  4. 4.5% of US AI Users Pay — and the Top 1% Spend $900 a Month
  5. Toyota to Deploy 400,000 Robots, Spending $6.42 Billion a Year

« Previous article