Early this month at a home in San Francisco. A Humanoid Robot with arms and legs pushes a vacuum cleaner across the floor. It wipes the countertop, picks up items dropped on the floor, and even empties the trash can. U.S. robot startup Tau Robotics began an in-home cleaning service using humanoids in San Francisco on the 28th of last month. The expense for a robot to clean the house for an hour is $30 (about 42,000 won).
However, this robot does not decide on its own how to clean based on the situation in the home. An employee trained at Tau Robotics' own facility in San Francisco remotely controls the robot in real time while watching the robot's camera feed, and an artificial intelligence (AI) system assists with motion control. Because of this, local reactions include, "It's cheaper than calling a person, so it's worth trying once," alongside, "If a person is controlling it anyway, wouldn't it be better to just call a human cleaner?"
The real purpose of Tau Robotics offering a relatively low-priced robot cleaning service, even assigning people to it, is to secure real-world data. While the robot cleans, cameras mounted on its head and wrists continuously film the inside of the home. It also records the robot's movement information such as joint positions, force, and balance; the control commands issued by the human; and 3D spatial information representing the room layout and object locations. This data is used to train a VLA (vision-language-action) model that links visual and language information to the actual robot's movements.
Tau Robotics said, "Data collection is not a secondary outcome of the service but the 'core purpose,' which is why we set this price." Unless a user requests deletion, the collected video and robot data are retained.
This is also why Tau Robotics chose homes as the first stage for its service. Each home has different layouts and furniture arrangements, and even when handling the same cup, its shape, location, and surrounding obstacles change every time. The force needed to wipe a countertop or fold laundry also varies. When a person directly controls the robot in these situations, action data showing "how to move in which situations" accumulates with every task.
Alexander Koch, Tau Robotics chief executive officer (CEO), said, "Home cleaning has clear consumer demand, and it allows robots to experience diverse, unstructured real-world environments, making it a good area for humanoids to start."
Robot companies in urgent need of training data are even directly hiring people to produce "ground-truth actions." Tesla is hiring "Optimus data collection operators" in Palo Alto. Employees wear motion-capture gear and move objects or walk along set routes, and those movements are turned into training data for the humanoid "Optimus." A condition was added that they must be able to walk for more than seven hours a day while wearing the gear.
U.S. humanoid startup Figure AI also trained its AI model "Helix 02," which controls full-body humanoid motion, on more than 1,000 hours of human movement data and repeated training in over 200,000 virtual environments.
China has gone a step further, operating "data factories" in various places that mass-produce data by deploying large numbers of people and robots. At the data collection facility of Shanghai robot startup AgiBot, about 100 robots are controlled by 200 people. They build data by repeating tasks such as folding T-shirts, making sandwiches, and opening doors for 17 hours a day.
The city of Shanghai provided space for this facility free of charge. At a training ground of more than 10,000 square meters in Shijingshan, Beijing, 100 humanoids are also collecting data by repeatedly folding clothes, sorting parcels, or opening doors with keys in spaces that recreate logistics and retail environments.
The same competition has begun in Korea. LG Electronics is building a large-scale robot training data factory at the Yangjae R&D Campus in Seoul, aiming to start operations within the year. There, the home robot "LG CLOiD" will immediately begin producing and learning motion data by repeatedly grasping and moving objects. Samsung also plans to establish a humanoid mass-production system and a robot data factory in Gumi, North Gyeongsang.
ROBOTIS has deployed 50 "AI Walker" humanoids and control personnel at its headquarters in Magok, Seoul, to gather action data. When a person moves while wearing an exoskeleton-style "leader arm," the humanoid mirrors the same motions as it records video, joint movements, and more. In Uzbekistan, the company is building a "physical AI data factory" that will deploy thousands of humanoids to collect and process action data at scale.
An industry official said, "For real-world data, it's important to capture not only successful actions but also where failures occurred and how a person adjusted force and posture to keep working, in meticulous detail," adding, "Once enough of this data accumulates, AI will go beyond merely imitating ground-truth actions and be able to judge and adjust its movements on its own even in first-encounter situations, so the competition to secure data in the robotics industry will become even fiercer."