I watched a humanoid robot clean my apartment. It felt like a taste of the future.

Humanoid home robots are the consumer face of physical AI. Most coverage is lab footage. Grace Huckins booked Tau Robotics for an actual apartment cleaning in San Francisco and wrote what it felt like when the machine met stairs, heat, toilets, and olive oil bottles.

Tau's pitch is autonomous cleaning eventually; today the robots are remotely piloted via head and wrist cameras, collecting training data on every wipe and fold. Pricing moved from a promotional $30 per hour with one robot to $60 weekly with two or $100 for a one-time visit. The company has completed only a couple hundred cleanings so far.

This is the gap between Jetsons marketing and the teleoperation reality that still teaches the models.

We're glad someone paid for the hour and told the truth. The Tau robot is four-and-a-half feet tall, clad in a company shirt and grippy socks, with a 3D-printed antennaed head and claw hands because humanlike ones were not durable enough. It is cute. It is also, as Huckins puts it, closer to a radio-controlled car than a Waymo for many tasks.

Kinematic guidance — a human pilot moves, sensors copy the motion to the robot — still carries bed-making and laundry folding. Vacuuming and mopping can already run without help, cofounder Cornelia Weinzierl said. Stairs remain a boss fight: the unit climbed with a drunken stagger blamed partly on remote-control latency, needed a carry on carpeted steps the first visit, and aborted early when heat took down the pilot's monitoring gear rather than the robot itself. On the redo, it cleaned the kitchen (skipped the stove), conquered the toilet even after the wand head fell off three times, left taps dripping, knocked over olive oil, and toppled an air purifier it could not right.

I want this to work. Dangerous, high-turnover cleaning labor is a fair place to aim robots, and Weinzierl's point about injury risk and scarce willing workers is real. Tau is also eyeing mold remediation and home health care later. The honest downside sits beside the charm: if autonomy arrives and prices fall, cleaners lose shifts in a country that has not built transition policy; if autonomy stalls, customers are paying to be data-collection props while pilots do the work from elsewhere.

Useful physical AI will look less like a perfect first demo and more like this — slow, supervised, occasionally slapstick, and logged for the next training run. We're watching whether a couple hundred SF cleanings become tens of thousands before the novelty wears off.

Humanoid startups have raised heavily on warehouse and domestic promises. Teleoperation-plus-data has become the common bridge while end-to-end autonomy lags. San Francisco's tolerance for robot spectacle makes it a natural beachhead — and a distorted sample of American housing stock.