top of page

Why Robots Still Struggle With Housework, and What China's "Robot Schools" Reveal

13 hours ago
5 min read

A robot that walks is impressive. A robot that reliably folds your washing, empties your dishwasher without smashing a plate, or safely helps you transfer from a bed to a wheelchair is a very different challenge. For disabled people hoping AI robots might one day help with household and caring tasks, this gap between walking and working is the whole story right now...


AI robot doing house chores

Why Household Tasks Are So Hard to Teach


Robots learn by copying huge amounts of data, in the same way a language AI learns from huge amounts of text. The trouble is that nobody has huge amounts of data showing exactly how a robot's hands should move to fold a fitted sheet, wipe a worktop, or help someone dress.


Goldman Sachs and the China Academy of Information and Communications Technology both estimate that a general purpose robot needs around 10 million hours of high quality, real world data before it can cope reliably with everyday tasks. Yet Goldman estimates that the leading humanoid programmes hold only around 500,000 hours, and Tesla is aiming for 1 million hours by the end of 2026. That is the gap the entire industry is racing to close.


Homes make this harder still. A factory floor is fixed and predictable. A kitchen is not. Every household has different cupboards, different crockery, different clutter and different routines, and a robot trained in one home may struggle in the next. For disabled people, whose homes are often adapted in highly individual ways, this variability matters even more than it does for the average household.


The Human Workforce Teaching Robots Their Trade


Because there isn't enough usable footage of ordinary chores already sitting on the internet, companies are now paying people to generate it deliberately. In China, this has become an industry in its own right.


Chinese firms have been paying residents and factory workers to film themselves doing everyday tasks such as folding laundry, wiping tables or cooking. This footage is known in the industry as "egocentric" data because it is shot from the worker's own point of view, and one home worker told Rest of World she earned around 20 yuan (about £2) an hour for it. Purpose-built training centres take a more hands-on approach. At China's largest, in Beijing's Shijingshan district, nearly 200 human trainers work with around 100 robots across more than 10,000 square metres, covering 16 full-scale scenarios from factory production lines to elder care.


One instructor, a former art teacher named Fudi Luo, described the job to CNBC. At first the robot has no awareness, so she steers its movements by hand, using cameras, controllers and motion capture, until it has gathered enough data to repeat the action alone. A robot hand maker working on the same campus told CNBC that learning a single new skill takes, on average, more than 10,000 repetitions. According to People's Daily, each robot at the centre, which was co-built with Leju Robotics, generates about four hours of training data a day, and 100 robots can complete at least 12,000 data collection tasks daily.


By December 2025, more than 40 state-owned robot data centres had been announced across China, according to research firm Interact Analysis, with around two dozen already running. Shanghai's Kylin training ground in the Zhangjiang district, which opened in 2025, covers more than 5,000 square metres and hosts over 100 types of robot from more than a dozen companies. This is now a formal part of China's industrial strategy. Its 15th Five Year Plan names embodied intelligence, meaning AI in physical machines such as robots, as one of six "future industries". Reuters has reported more than $20 billion in state subsidies for the humanoid sector in a single year, alongside a guidance fund expected to raise around 1 trillion yuan ($138 billion) over 20 years for AI, robotics and other technology.


What Robots Can Actually Do in the Home Today


This investment is starting to show results, though not fully independent ones. Several household robots are now close to market. 1X's Neo, priced at $20,000, sold out its first year of production within days of pre-orders opening in October 2025, although the company had not confirmed any customer deliveries by mid-October 2026. SwitchBot's Onero H1, unveiled at CES in January, is listed at $9,999 and is due to ship late in 2026. In China, UniX AI began deliveries of Panther in April. It is a wheeled humanoid that the company says can make a bed, clean a bathroom and cook a simple breakfast.


The catch is how many of these tasks still involve a hidden human. Bernt Børnich, chief executive of 1X Technologies, has openly said that early Neo owners must accept remote operators taking control of the robot for jobs it cannot yet manage on its own. The data from those sessions is then used to train the AI. For now, "autonomous" often means "mostly autonomous, with a person standing by."


Personal and caring tasks sit well beyond this. Dressing someone, helping with a transfer, or supervising medication all demand a level of reliability that a folded towel does not. None of these robots is designed or approved for personal care, and Neo is not due to reach markets outside the US until 2027 at the earliest.


Best Case, Worst Case: Two Timelines


In the best case scenario, China's enormous investment in training data, combined with falling component costs, drives a steady flow of cheaper, more capable household robots over the next three to five years. Basic chore assistance becomes genuinely affordable and reliable well before the more sensitive tasks around personal care do, freeing up time and energy for disabled people and unpaid carers alike.


In the worst case scenario, the data gap proves harder to close than expected. Robots continue to perform well in demonstrations and training centres but struggle with the sheer unpredictability of real homes, particularly homes adapted for a disability, where standard training data may not apply well at all. Costs stay too high for NHS or local authority funding, and progress on personal care robotics stalls for a decade or more while household robots remain a novelty for those who can afford one.


The honest answer sits somewhere between the two. Chores are likely to become genuinely automatable well before caring duties do, and price, not capability, may end up being the bigger barrier for disabled households.


What This Could Mean for Disabled People and Carers


Even partial success with everyday chores matters. A robot that reliably handles laundry, tidies clutter or loads a dishwasher would ease pressure on unpaid carers and free up energy for disabled people managing fatigue, pain or limited mobility, even if it never approaches anything more personal.


The people best placed to judge whether these tools are actually useful are disabled people themselves, and accessibility experts continue to argue they need a seat at the table now, while these systems are still being trained, rather than later, once the design choices have already been made.



Comments


bottom of page