Tau Robotics, a San Francisco startup, is offering apartment cleanings with humanoid robots while harvesting data to train more autonomous machines. The company’s robots are not fully independent yet: humans remotely pilot the four-and-a-half-foot-tall machines by watching video feeds from cameras mounted in the robots’ heads and wrists.
Service and pricing
Because the data Tau collects is valuable for development, the company subsidizes its service. When the author joined the waitlist, an hour with a single robot cost $30; later referenced pricing in the article is $60 for a weekly cleaning with two robots, or $100 for a one-time visit.
How the system works and why Tau collects data
Many tasks—such as making a bed or folding laundry—are performed via kinematic guidance: a human operator demonstrates motions that the robot copies, and those demonstrations are especially useful training examples for machine-learning models. Tau’s long-term goal is to use these real-world examples to improve robot autonomy, but today the machines still depend heavily on human support.
The author’s experience: spectacle and effectiveness
The robot sent to the author’s apartment arrived in a blue minivan. Its first test was eight steps up to the building entrance; it climbed awkwardly and often stumbled backward, though it never fell completely. Tau cofounder Cornelia Weinzierl said part of that odd gait is caused by remote operation and short signal delays affecting coordination: “The amount of work and time that we spend just optimizing, for example, timings and connectivity is crazy.”
On a later, comped visit the robot performed better: it managed the stairs with a continued unsteady walk and, over an hour, cleaned the kitchen credibly (it avoided the stove) and handled the toilet—the latter task proving most notable as the toilet-cleaning wand head detached three times and each time the robot succeeded in reattaching it by probing with the handle until the contact was made.
Technical weaknesses observed
The robot stalled repeatedly during visits. On one occasion Weinzierl initially suspected the robot had overheated in unseasonably hot weather, but then learned that the remote operator’s monitoring equipment had shut down from heat; without that remote assistance the robot could not continue. The robot also could not ascend a carpeted interior staircase and had to be carried up several steps.
Physically, the robot has a 3D-printed “head” with antennae for communication and 3D-printed claw-like hands, chosen because more humanlike hands proved insufficiently durable. It wore a Tau Robotics shirt and grippy socks to avoid scratching hardwood floors. During cleaning it left taps dripping, knocked over a closed bottle of olive oil, and could not right an air purifier it toppled—yet the author found the machine’s charm and utility forgiving.
Economic and societal considerations
If Tau or competitors eventually train inexpensive, fully autonomous cleaning robots, there could be effects on employment. Weinzierl argued that some cleaning work is undesirable, has high turnover and injury risk, and that robots could be appropriate for hazardous or physically demanding jobs: “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.” Tau is considering future deployments in tasks such as mold remediation and home health care.
Outlook
Tau has completed only a couple hundred cleanings so far, which is a modest dataset relative to the enormous variety of tasks involved in apartment cleaning. The company still has significant development ahead to reach true autonomy, but the existence of humanoid robots that can mop floors and clean toilets—with substantial human assistance—is already notable. The Jetsons-feel of robot helpers is closer than it used to be, even if the technology remains in beta.
