In a modest third-floor apartment in San Francisco, two humanoid machines from Tau Robotics performed a full suite of household chores while a human operator watched from a distant control room. The service, billed at $30 per robot per hour, is marketed as a practical cleaning solution, but its deeper purpose is to feed real-world observations into the company’s artificial intelligence pipelines. Founder and engineering lead Ted Pinkerton was on-site, listening to the operators through an earpiece and stepping in when the robots needed a nudge.
Pricing model and operational scale
The fee structure is straightforward: one robot costs $30 an hour, so a pair costs $60 per hour. Tau Robotics currently operates a fleet of eight modified Unitree G1 units, and the company reports roughly 160 paid robot hours spread across almost 100 cleaning sessions. Each Unitree G1 platform carries a base price of about $55,000 before Tau’s custom modifications, turning every deployment into a high-value data-gathering opportunity rather than a low-margin service.
How the robots work and what limits them
While the machines navigate stairs and vacuum rooms, they are not autonomous. They are jointly controlled by a human operator and AI a hybrid model that blends remote decision-making with on-board perception. Cameras mounted on the robots’ heads and wrists stream live video to the headquarters, allowing operators to direct fine-grained actions such as moving a book, clearing a charging cord, or repositioning a spoon rest before wiping a countertop.
Hardware constraints become evident during more delicate tasks. The robots wield pincer-style hands that struggle with smooth glassware; one unit finally managed to grip a cup by its rim after several attempts. Retrieving a small knife from a utensil basket proved similarly tricky, and a broken dish from a previous session was cited as evidence of the mechanical limits. Pinkerton noted that five-finger hands exist but are currently too fragile for the wear and tear of everyday cleaning, so the company sticks with the more robust, albeit less dexterous, pincers.
Speed is another variable. Folding a towel took about four minutes, and denim pants required just over a minute—metrics that highlight the gap between the appearance of a humanoid worker and the practical tempo needed for routine household chores.
Data collection, privacy, and future ambitions
Every cleaning session is also a structured data-collection exercise. The unblurred video feed is stored indefinitely to refine the general-purpose robotics model that runs on custom third-party hardware. Customers may request deletion of their footage after the job, and they can also block specific rooms from camera view; reporter Christine Ji exercised this option for her bedroom and bathroom.
CEO Alexander Koch explains that simple tasks—like vacuuming an empty hallway—provide limited training value, whereas unexpected scenarios, such as a vacuum becoming entangled with a mouse trap, generate rich data for the AI. The company aims to scale the operation to 1,000 cleanings per week in San Francisco by and to expand to other cities and international markets shortly thereafter.
Until the next generation of more dexterous manipulators arrives, the service remains a hybrid of human guidance and machine endurance, offering a glimpse of how everyday environments might serve as laboratories for the next wave of household robots.
