物理 AI 的「GPT 时刻」，可能不远了。
10 月 22 日，杭州国际博览中心热闹非常，不仅人多，而且机器人也特别多。
这是 2026 年宇树科技举办的世界物理 AI 大会的现场，从四足机器人到人形机器人，你甚至可以说这里是全球「机器人密度」最高的地方。
2026 年，也是宇树科技创立的第十年。有意思的是，在今天的舞台上，宇树科技创始人王兴兴，在谈机器人之外，聊的最多的就是当下最火热的 AI。
他不仅劝所有人要「相信 AI」，而且还反思，觉得自己「相信 AI 还相信得不够」。
为什么这家中国、乃至全球机器人行业中最有实力的公司掌舵者，会这么说？物理 AI 的拐点，真的要到来了吗？
必须相信 AI
在王兴兴 36 分钟的演讲中，宇树自己的产品只占了开头和结尾。中间最长的一段，王兴兴在做一件事——劝人相信 AI。
从 AI 攻下了一批顶尖数学难题，「人类最智慧的金字塔的尖端」已经被击穿；到 AI 逆向工程让「传统软件公司的壁垒没了」；AI 视频里人物的表情和情绪「比真人演的还好很多」；顶尖大模型公司内部「已经实现了模型的自进化」。
「这不是闹着玩的」。他的结论是，以当下的模型能力，哪怕不再进步，信息世界也已经被「完全击穿」。
然后他说了一句不太像上市公司董事长会说的话，「我感觉我自己还不够相信，还不够坚定」。
这不是即兴发挥。十天前的上海工博会上，他已经当众「认错」过一次，「过去十几年我犯的最大的错误，就是对 AI 的相信还不太够」。他还给过一个倍数，5 年前或 10 年前你喜欢、相信 AI 的程度，「现在乘 10 倍都不为过」。
王兴兴在会上探讨物理 AI 的未来｜图片来源：宇树科技
更直观的变化在时间表上。今年 8 月的世界机器人大会，王兴兴对具身智能「ChatGPT 时刻」的预期还是快则 2 到 3 年、慢则 5 到 10 年；两个月后，这个窗口被他压缩成「1 到 2 年，慢的话 2 到 3 年」。至于随便把一台宇树机器人带进陌生的家里，一句话让它把房间收拾好，「快的话可能明年就可以实现」。工博会上他说得更直白，「明年 GPT-7 出来，可能直接把具身智能问题打穿」。
一个硬件出身的工程师，在两个月里成了全场最激进的 AI 布道者。这看上去像一次认知突变，但把宇树最近半年的处境摊开来看，他的「相信」几乎是一道必答题。
但是，宇树的强项——机器人本体，依然是王兴兴骄傲的事情。
王兴兴在台上最骄傲的是 G1。这款 2024 年 5 月发布的人形机器人，累计下线大约 1 到 2 万台，他称之为全球单一型号出货量最大的人形机器人。
10 月 21 日，宇树放出 H2-X「功夫传奇 守护」的视频，一脚踢瘪车门、徒手抬车、4 米高台空翻落地，续航从 H2 的约 3 小时拉到约 8 小时。可以看到，宇树团队依然在快速迭代公司的产品。
本体之外，一直在布道 AI 的王兴兴，推销的 AI，并不完全是宇树自己的 AI。
随后的论坛环节王兴兴和姚明等嘉宾一同探讨物理 AI 的未来｜图片来源：宇树科技
他在台上很坦白，「我们毕竟是一家新公司」，跟全球最头部的多模态大模型公司「做一模一样的模型，那是不太现实的」。宇树会自研，也会和国内外大模型公司合作。他展示的最亮眼的 demo，是把 GPT-6 接到机器人上，零训练就能自己洗笔、换颜料，在围棋盘上精准落子。他判断规划和导航这类问题「某种意义上已经解决掉了」，剩下的是多模态大模型的输入输出和真实机器人之间的对齐——传感器的误差、执行器的误差，「我觉得这个也是很快可以解决的问题」。
把这几句话连起来，就能看清这份「信仰」背后的商业逻辑。如果机器人的大脑最终由 OpenAI、谷歌这样的公司提供，而且像他说的那样越来越强、越来越便宜，那具身智能这场竞赛的胜负手，就会回到谁能最便宜、最大规模地造出可靠的身体。这恰恰是宇树最擅长的事。
所以在王兴兴列出的未来最重要的几件事里，排第一的不是模型，而是硬件的大规模制造能力。宇树现在一年生产几万台机器人、几十万个关节电机，在他看来「完全不够」，真正普及至少要「年产千万台」。他甚至谈到了算力集群、芯片产能和矿。至于宇树自己的模型能力，他押注的是一套「自进化」循环，让 AI 自己检索论文、写控制代码、在仿真里训练、到真机上测试，再把打分结果喂回去。这套方法的前提同样是基础模型每个月都在变强，用他的话说，「模型能力越提升」，这个循环「也会越来越强」。
世界物理 AI 大赛启动｜图片来源：宇树科技
换句话说，对一家机器人公司而言，「相信 AI」不是锦上添花的表态，而是整套战略成立的前提。AI 进步得越快，宇树的硬件规模优势就越早兑现。
演讲的最后，王兴兴讲了一条自己和 AI 的「交叉时间线」。2009 年上大一，他做的第一台机器人就是双足人形；大学一门计算机课要写小论文，他的题目是「随着神经网络规模的增加，能力会越来越强」，理由是人的大脑比鸡、狗、猪的大，所以人更聪明。然后是 2016 年的 AlphaGo，以及 2022 年底的 ChatGPT，后者直接促成了宇树 2023 年正式进军人形机器人。
这段回忆讲得很真诚，也很有选择性，它把一个硬件工程师的十几年，讲成了一条始终指向 AI 的直线。也许他确实很早就相信了，只是相信得不够；也许是这个时代逼着每一家机器人公司，都必须讲出同一个故事。
展台的热闹之外
在这届宇树举办的世界物理 AI 大会上，展台区域可以说是机器人比人多的经典场景。除了传统的擂台赛、打乒乓球等 showcase 之外，你甚至还能看到抖空竹、玩「剑玉」的机器人——是的，玩剑玉这个机器人的开发者说的是日语。
玩「剑玉」的机器人｜图片来源：极客公园
各种各样形态的机器狗｜图片来源：极客公园
在整个展区我们能看到，目前无论是四足还是人形机器人，最大的一个场景，可能还是导览和娱乐。例如你可以看到戴着牛仔帽的 G1 机器人，也能看到被打扮成熊猫、恐龙和舞狮的四足机器狗。一家叫逸璟科技的公司，还为机器人配上了能做出表情的「人脸」，团队负责人告诉极客公园，在一些康养、导览等场景，有一张「人脸」的机器人，还是会比普通的机器人，更让人感到亲近，虽然有一些些「恐怖谷」效应。
此外，安全、巡逻等场景，也在真正意义上的落地。一家改造机器狗做巡逻场景的负责人告诉极客公园，配上激光雷达和相关设备的四足机器狗，已经能代替人工在夜晚进行小区、园区的执勤工作，遇到危急情况，人类操作员还能进行远程操作。而如果背上一些灭火设备，这些四足机器人还能完成灭火等任务。
工作人员正在遥操操作｜图片来源：极客公园
展区中比较热的一个展台，是机器人教育和课程的展台。宇树的同学告诉极客公园，团队有成熟的方案，不仅能提供硬件，而且能提供从硬件、软件和人才的一套方案，为学校提供整套的机器人教育实验室，让更多的学生，能真正上手体验机器人编程的魅力。
学习投篮的机器人｜图片来源：极客公园
在大会当场，王兴兴和一众嘉宾，也开启了世界物理 AI 大赛，宇树提供高达数千万人民币的奖金，鼓励全球开发者进行机器人的开发，这也是自大疆曾经的「RoboMaster」大赛之后，国内少见的真正培养机器人开发者的大赛。
正如这几年新一波的 AI 浪潮，培养出了无数的 AI 创业者和超级个体一样，当物理 AI 开始进入临界状态时，这个行业同样需要无数的机器人人才，而世界物理 AI 大赛，无疑是个很好的起点。
头图来源：宇树科技
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Physical AI's GPT Time, probably not far.
On October 22nd, the Hangzhou International Expo Center was very popular, not only with a large number of people, but also with a particularly large number of robots.
This is the scene of the World Physics AI conference in 2026, from a four-foot robot to a human robot, and you can even say it's the highest robot density in the world.
In 2026, it was also the tenth anniversary of the creation of the technology of the Oki tree. Interestingly, on the stage today, Wang Hing-Hing, the founder of the Wow tree technology, is talking about the hottest AI in the day, beyond robots.
Not only did he persuade everyone to "believe" the AI, but he thought he "believe the AI and not enough."
Why would the most powerful company in the robot industry in China and around the world say that? Physical AI's turning point. Are you really coming?
Trust AI
In Wang Heung-Hing 36 minutes of speech, Woo-soo's own product was only the beginning and the end. The longest part in the middle, Wang Hing Hing is doing one thing -- convincing people to believe in AI.
From AI, which has taken down some top-of-the-art math puzzles, "the tip of the most intelligent pyramid of mankind" has been blown through; to AI, the reverse engineering has left "the barriers to traditional software companies" out; to AI, the face and emotions of the characters in the video are "better than real people"; and within the top LLM, "the model has evolved itself."
"This is not a joke." He concluded that the information world had been “completely pierced” with current modelling capabilities, if not progress.
Then he said something less like the Chairman of the Board of Directors of the listed company, "I don't feel convinced enough myself, not firm enough."
It's not improvising. Ten days ago, at the Shanghai Labor Fair, he made a public mistake, "the biggest mistake I've made in the last decade or so is not enough to believe in AI." He also gave you a multiple, five years ago or 10 years ago, the degree to which you liked and believed in AI, "not even ten times now."
Wang Hing Heung at the conference discusses the future of Al's photo source: Woo Tree Technology.
More intuitive changes are on the timetable. Last August, at the World Robotics Congress, Wang Xing Xing Xing expected the smart "ChatGPT Time" faster than 2 to 3 years or 5 to 10 years; two months later, the window was condensed into "1 to 2 years, in slow terms 2 to 3 years." As for bringing a wild tree robot into a strange home, let it pack up the room, "probably next year." At the fair, he said more clearly, "The GPT-7 will come out next year, and it's possible that the smart question will go right through."
A hardware engineer who became the most radical AI preacher in two months. It looks like a cognitive mutation, but to take the last half of his life apart, his "belief" is almost an answer.
But the power of the owl tree, the robotic body, is still the king's pride.
Wang Xing Xing is proud of G1. This man-form robot, publish, in May 2024, accumulated downlines of about 1 to 20,000, which he calls the singlest human-form robot ever delivered.
On October 21st, Woo Tree released a video of H2-X "Kung Fu Legend Guardian" with a kick in the door, a hand-lifting car, and a 4-metre-high platform, which went from about 3 hours of H2 to about 8 hours. As can be seen, the Woo Tree team is still working on the products of the Rapid Overlay Company.
Outside of the body, you've been preaching the king of AI, selling the al, not exactly the al of Oki.
Then Wang Hing Xing and Yao Min came together to discuss the future of Al's future photo source: Woo Tree Technology.
He was quite honest on the stage, "We're a new company after all," and "do the exact same model as the world's largest multi-modular LLM company, which is not realistic." Woo-ki will study on his own and will also work with LLM companies at home and abroad. The brightest demo he's ever shown is to put GPT-6 on a robot, so zero training can wash your pens, change your paint, and be precise on the board. He judged that the problems of planning and navigation were "resolved in a sense" and that the rest was the alignment between the input output of multimodular LLM and the real robot — the error of the sensor, the error of the implementer — “I think this is also a problem that can be solved soon”.
If we put these words together, we can see the business logic behind this "belief". If the robot's brain is eventually supplied by companies like Open AI, Google, and, as he says, is getting stronger and cheaper, then the winner of the smart race will go back to who can build a reliable body at the cheapest and largest possible scale. That's exactly what Woo-hee is good at.
So of the most important things in Wang Xing Xing's future, it's not a model, but a large-scale manufacturing capacity for hardware. In his view, the production of tens of thousands of robots and hundreds of thousands of joint power units a year is "never enough" for real coverage, at least "ten million units a year". He even talked about calculator clusters, chip capacity and mines. As for his own modelling capability, he was betting on a "self-evolutionary" cycle that allowed AI to retrieve his papers, write control codes, train in simulations, test on real machines and feed the scores back. It is also premised on the fact that the base model is getting stronger every month, and in his words, "the more the model is capable," the more the cycle becomes stronger.
The world's physics.
In other words, for a robot company, "Believe AI" is not a rhetorical statement, but a precondition for the establishment of a strategy. The faster the progress, the sooner the hardware-size advantage of the Woo-tree materializes.
At the end of the speech, Wang Hing Xing spoke of a "cross-the-clock line" between himself and AI. The first robot he made in 2009 was a double-footed man; a computer class at the university had to write a small paper entitled "As the nervous network increases in size, power will grow" on the grounds that people are smarter than chickens, dogs, pigs. Then AlphaGo in 2016, and ChatGPT at the end of 2022, which directly contributed to the official entry of Woo-tree into the military robot in 2023.
The memoir was sincere and selective, and it presented a hardware engineer for more than a decade as a straight line to AI. Maybe he did believe it early in the day, just not enough; maybe this is the time when every robot company has to tell the same story.
Outside the booth.
At the World Physics AI conference, the exhibition area was described as a classic scenario with more robots than people. In addition to traditional battles, ping-pong, etc. showcase, you can even see the robots that are shaking and playing "Sword Jade" -- yes, the robots who are developing them speak Japanese.
It's called the "Sword Jade" robot.
A variety of forms of machine dog and dog picture: the polar park.
Throughout the exhibition area, we can see that at the moment, whether it's a four-foot or human robot, one of the biggest scenes, perhaps a guide and entertainment. You can see, for example, G1 robots in cowboy hats and four-legged robot dogs dressed as pandas, dinosaurs and dancing lions. A company called Quake Tech, which also equipped robots with face-based faces, the team leader told them that there was a robot with face-to-faces that was closer to normal robots, even though there were some "terrorist valleys" effects.
In addition, security, patrolling and so on are actually landings. The director of the patrol scene, who modified the machine dog, told the polar park that the four-foot machine dog with laser radar and associated equipment had been able to perform night-time, night-time and night-time operations and had been able to operate remotely by human operators in critical situations. These four-foot robots, on the other hand, would have been able to perform tasks such as fire suppression if they had some fire suppression equipment on their backs.
The crew is operating remotely from the photo source of the magma: the Polar Park.
One of the more hot stands in the exhibition area is that of robotic education and courses. Woo-ki's classmates told us that teams have mature programs that not only provide hardware, but also a package of hardware, software and talent, provide schools with a full-fledged robotic education laboratory, and allow more students to truly experience robotic programming.
Learning to shoot from robotic pictures: Polar Park
On the spot, Wang Xing Xing and a group of guests also launched the World Physics AI competition, which provided a prize of up to tens of millions of yuan to encourage the development of robots by global developers, which was a rare match for real robotists in the country after the once-roboMaster competition.
Just like the new wave of AI in the last few years, which has produced countless AI entrepreneurs and super-individuals, when physical AI begins to be critical, the industry also needs countless robotics, and the World Physics AI competition is certainly a good starting point.
Source of headline: Woo Tree Technology
Auto-fetched and translated by AI Atlas. All rights belong to the original author.