Lessons from Unitree’s Ten-Year Journey to a STAR Market Listing: Build One Complete Thing First

Verification notes (2026-10-29): 1. In his freshman year in 2009, Wang Xingxing built a small bipedal robot for a little over RMB200. At the 2025 World Robot Conference, his exact wording was “approximately a little over two hundred renminbi”; the post says “reportedly…about 200 yuan,” which is consistent in order of magnitude. He once “firmly opposed” humanoid robots. A customer was willing to place an order with a deposit in 2022, and Unitree released H1 in August 2023. His own statements and reporting also describe XDog, the quadruped robot he built during his master's studies, costing less than RMB20,000 in total; his application to delay graduation by more than six months to finish it; and his view that a half-finished project has no value. Sources: https://www.robothot.com/2080.html https://www.yicai.com/news/102765655.html https://news.qq.com/rain/a/20240705A0488700 https://www.unitree.com/operate/h1/ 2. The “two 80%s” are Wang Xingxing's formulation for robotics' “ChatGPT moment.” Source: https://www.thepaper.cn/newsDetail_forward_33816672 ---- Lessons from Unitree's ten-year journey to a STAR Market listing When you do not know whether something will work, rather than keep thinking, build something first. Isn't that simply what we often call “low-cost trial and error”? What matters, though, is not actually “something,” but the word “one” before it. Build “one” thing first—not half a thing or 0.8 of a thing, but one complete thing. It can be small, rough, and cheap. But it must be complete. At nineteen, Unitree founder Wang Xingxing wanted to build a bipedal robot himself. He had little equipment or money. If components were too expensive, he found cheaper alternatives. Without machining equipment, he used small drill bits, files, and scissors to make parts himself. Reportedly, for about RMB200 in the end, he really did build a small bipedal robot. Later, in graduate school, he began building a quadruped robot. This was much more complex, but according to his own recollection, the entire early development effort cost only RMB10,000–20,000. He reportedly applied to delay graduation by six months to finish it properly. Explaining that decision later, he said a project only half completed has no value. Of course, this should not be interpreted mechanically to mean that anything half finished is worthless. Intermediate research results can certainly matter. In an engineering project, however, it points to a very specific issue. Making a leg means you have made a leg. Getting a motor turning means you have a turning motor. Getting a control algorithm running means the algorithm works. Each demonstrates that one local problem has been solved. Only when everything is assembled and a robot actually walks do you first discover whether these parts can become a product together. The real lesson is therefore not “be willing to try when problems arise,” but to shrink a large goal first, then complete that smaller goal in full. A complete thing, however small, provides different information from half of a large thing. First, problems between systems emerge only once the whole thing is complete. A motor testing well on its own does not mean it will not overheat in a robot. An algorithm running independently does not mean it will remain stable when connected to sensors and a mechanical structure. Every component can look fine separately, yet the assembled robot cannot stand. That is extremely common in engineering. Many problems exist not within any individual part, but only after everything is put together. Second, only a complete thing can truly enter the outside world. It can be evaluated, compared, shared, and eventually seen by potential customers and investors. From this perspective, “complete” marks the point at which a product stops being merely “my progress” and becomes “an object in the outside world.” Second: when conditions change, reassess. Unitree moved from quadruped to humanoid robots. Today, mentioning Unitree probably brings humanoids to mind. Earlier, however, Wang Xingxing “firmly opposed” them. The reasons are understandable. Humanoids were extremely expensive, engineering was difficult, and AI was far less capable than it is today. Unitree had therefore not formally begun making humanoids through 2022. By 2023, circumstances had changed. Unitree formally launched its humanoid project and released H1 that August. Why the change? Wang has described several specific developments. First, demand emerged. In 2022, while Unitree was still making quadrupeds, customers approached it wanting humanoids. Some were even willing to pay deposits before a product had taken shape. The second change was technological. Around 2021–2022, AI capabilities began accelerating noticeably. With the emergence of general-purpose large models in particular, whether robots could have a more general-purpose “brain” became a question worth revisiting. The third change: Unitree had by then spent many years making quadrupeds. It had motors, joints, and electronic controls. Making a humanoid is certainly not as simple as standing a robot dog upright, but neither does it mean starting from a blank sheet. The same question can have entirely different answers at different times. We like praising people for “steadfastness” and “long-term thinking”: they recognized a direction ten years ago and never changed. But there is another kind of steadfastness: consistency not in your conclusions, but in your criteria for judgment. Often, what we should hold to is not yesterday's conclusion, but the criteria that originally produced it. Third: before predicting the future, define how the prediction will be tested. Unitree still faces this issue today. One of the most popular questions in humanoid robotics is: when will its ChatGPT moment arrive? Roughly, it asks when robots will suddenly cross a threshold, as ChatGPT did in late 2022, turning from a technology watched by a few into a major change everyone can feel. But how exactly do we define robotics' ChatGPT moment? Does running count? Doing flips? Or must robots enter ordinary homes, tidying rooms, cooking, and caring for older people before it counts? If everyone uses different standards, expectations for the industry's prospects are hard to align. Wang Xingxing did something interesting. He proposed the “two 80%s.” Suppose a robot enters an unfamiliar everyday setting, understands about 80% of spoken or written instructions, and can actually complete about 80% of those tasks. Such an embodied AI model could then be considered to have reached an important breakthrough. Notice that 80% is not necessarily the right figure. What matters is not the number, but the act of specifying it. Previously, people asked, “When will robots become truly intelligent?” That is highly abstract. To discuss it, we first need to quantify the abstract standard. How much does instruction comprehension improve each year? What figure counts as a major breakthrough? The discussion can then move from everyone expressing opinions to everyone conducting tests together. The same method applies elsewhere. A boss says a project will definitely take off next year. What does “take off” mean? How much revenue? How many users? What repeat-purchase rate? Or you say you will master English this year. What does “master” mean—understanding the news, participating in English-language meetings, or scoring 7 on IELTS? A prediction without acceptance criteria can almost never be wrong. A judgment that can never be wrong, however, conveys little information. The important point about Wang's “two 80%s” is not the numbers, but his attempt to give his judgment conditions against which it can later be tested. What makes these three things interesting is not that Wang has some brilliant “method for success.” Quite the opposite: all three assume that human judgment can be wrong at any time. Will this thing actually work? We do not know. Is a conclusion from three years ago still right today? We do not know. When will robots truly cross the threshold? We do not know either. Given that uncertainty, do not rush to turn judgments into beliefs. Build something complete and let it be tested. Let answers change when conditions change. When talking about the future, specify a standard you can return to and test later. Good judgment need not be right from the outset. But it should at least leave reality a chance to correct it. #得到頭條20260820 ---- Robin: “Wang Xingxing was born in 1990 and is thirty-six this year. Unitree's market capitalization has surged since its recent listing. Wang's current shareholdings plus employee options amount to nearly 30% of Unitree, worth RMB100 billion, or about US$14.8 billion. Before forty, his wealth is roughly on a par with Barry Lam, ranked third in Taiwan. XDDD” Editorial note: Unitree listed on the STAR Market on August 19, 2026. Based on the first day's closing price, Wang's direct and indirect holdings were worth approximately RMB102.6 billion, as cited by China News Service. In Forbes' 2026 list of Taiwan's fifty richest people, published in March 2026, Terry Gou ranked first (US$15.3 billion) and Barry Lam second (US$14.2 billion). The post's exchange-rate conversion to “about US$14.8 billion” has not been verified and is not adopted here. Sources: http://www.chinanews.com.cn/cj/2026/08-19/10680543.shtml https://news.cnyes.com/news/id/6377369