--- title: "Robots Are Cheap Enough to Buy, but Still Cannot Do the Housework" date: 2026-11-04 category: Builders site: NeuroAI canonical: https://neuroai.site/a/na-wangxingxing-30-1104 language: en --- # Robots Are Cheap Enough to Buy, but Still Cannot Do the Housework > A humanoid robot that can tumble and do backflips appeared only in short videos from trade shows a few years ago; today it sits on a shelf with a price tag. The one who kept pushing that price down A humanoid robot that can tumble and do backflips appeared only in short videos from trade shows a few years ago; today it sits on a shelf with a price tag. The one who kept pushing that price down is Wang Xingxing, founder of Unitree, a company that began with four-legged robot dogs. But a robot that can run and jump is a fair distance from one that can come into your home and tidy up. The checkable numbers, meaning price, shipments and the listing, are one thing; the small capability an ordinary family is actually waiting for is another. ## The Buzz on Listing Day According to The Paper, Securities Times and other media, Unitree listed on the STAR Market on August 19, 2026, and was dubbed by the market the first humanoid-robot stock. The issue price was set at 150.8 yuan; the shares opened sharply higher on the first day and then went through marked swings. Behind the excitement are real books. According to China Daily and Shanghai Securities News, the first quarterly report the company released after listing showed revenue of 1.817 billion yuan for the first three quarters, up 55.71 percent year on year. The growth is still considerable, but where the revenue comes from deserves a closer look. In that same report, the rise in profit comes with an accounting wrinkle: a one-off share-based payment charge recognized in the same period a year earlier pushed down the comparison base, so the year-on-year figure looks especially high. Peel back that layer and you can see how far the business itself has actually run. According to Securities Times, the share price traced a sharply volatile curve after listing, falling noticeably at one point before a rebound. Whether money is buying expectations or the present, any small stir gets magnified into swings on the screen. ## Pushing the Price Within Reach Unitree turned robots from costly exhibits into goods that ship in volume, and that path has a clear public record. According to Tencent News and Sina Finance, the company shipped 5,500 humanoid robots in 2025, and set a target of 10,000 to 20,000 units for 2026. Those shipments did not come from nowhere. According to public records, the company started with four-legged robot dogs, launched its first commercial quadruped around 2017, and only later rolled out general-purpose humanoid models, steadily raising the share of core components it develops in-house. That supply chain is exactly what lets the price fall. Making joint motors and modules in-house lets the body, motion control and production improve together, so the cost curve slides down as volume rises, and entry-level models have been pushed into the tens-of-thousands-of-yuan range, nudging the threshold from institutions-only toward something an individual can reach. But cheap is only the ticket in. According to Guangming Online, the starting price of an entry-level humanoid has fallen to tens of thousands of yuan, yet buyers are still most often universities, research institutes and developers; scenarios where it is used as a long-term production tool are only slowly opening up. ## Confidence Built on Shipments According to Tencent News and Sina Finance, as of July 2026 the company's cumulative output of bipedal humanoid robots over the years had passed 10,000 units, a running total, not the same book as the single-year shipments above. The manufacturing base funded by the IPO is still under construction, and its planned annual capacity of tens of thousands of units remains, for now, on paper. Scale brings not only cost but data. The more robots are sold and the longer they run, the more real movement and failure samples the team collects, which in turn feed the control algorithms and models, forming an engineering loop that spins faster and faster. According to Securities Times, Shanghai Securities News and other media, the company keeps increasing research spending in embodied intelligence, especially on embodied large models and computing power. Being able to run and jump is the ticket in; what keeps customers renewing is whether the robot understands instructions and makes few mistakes. According to trade-press reports, research and teaching scenarios still account for a sizable share of the company's revenue, while real industrial and commercial deployment is still being cultivated. With volume up, how long and how steadily a robot gets used becomes a more worth-asking question than how many units were sold. ## Housework Still on the Calendar What ordinary people really care about is when a robot can come into the home. According to the South China Morning Post and other media, as models improve, Wang Xingxing judges that robots could enter homes and take on some housework as early as 2027, while also describing this as a forecast. In the same round of reports, he also noted that fine manipulation remains a weak point: success rates for grasping and carrying are already fairly high, but tasks like taking a phone apart or threading a needle are still clumsy. Between a model's output and the real physical world lie the last few millimeters. According to Securities Times and other media, he said the company is doing early research on letting models write their own code and train and iterate in simulation or real settings, with the goal of relying less on manual debugging. This is exploration in a research direction, not machines having taken over development outright. ## Models That Train Themselves To make that exploration clear, one has to separate what is done from what is under way. What several outlets describe is an internal loop in which models gather their own data, score results and retrain, while people mainly evaluate outcomes and set goals; the loop is still at an early validation stage. According to TMTPost, Wang Xingxing judges that 2027 may become a tipping point for the robotics industry, on the grounds that the understanding and planning abilities of top models are improving fast. A judgment like that belongs to expectations about the future, and is best read alongside the pile of unsolved problems. Another bottleneck he returns to is data efficiency: many robots are built and much data is collected, yet a great deal of it goes unused. Getting models to digest that data on their own is seen as the key to the next gains in efficiency, and it too remains a plan in progress. ## What Is Still Missing Beyond Hardware The company's approach has always been clear: first open the path to scale with high cost-performance, then turn the cost advantage into customer stickiness. That path worked once with quadruped robots; whether it can be repeated in the more complex humanoid setting is the real question ahead. Generalization is the harder gate. Succeeding in one prepared room is not the same as succeeding in a different room with a different cup; practicing a single motion at a fixed station until it is smooth is a different matter from hearing one instruction in an unfamiliar place and finishing a whole chain of tasks. Back to the tumbling, backflipping robot at the start. That it can run and jump is no longer a novelty; the novelty is when it can steadily finish a whole chain of work in an unfamiliar room. An engineer can drive the price down, but generalization has to be ground out slowly by the whole industry. **Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.** --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-wangxingxing-30-1104 Free to quote with attribution and a link to the original.