智东西（公众号：zhidxcom）
作者 | 李水青
编辑 | 心缘
智东西10月22日杭州报道，今日，在2026年世界物理AI大会暨宇树合作伙伴生态大会上，宇树科技董事长、创始人兼CEO王兴兴发表约30分钟演讲，谈及AI信仰、物理AI临界点、具身智能ChatGPT时刻、自进化V1.0与量产生态等重要话题。
“全世界几乎所有人对AI的重视和信仰都还不够，包括我自己。”王兴兴说，物理AI已看到被击穿的苗头，很可能在最近1到2年发生；产业正迈向“具身智能的ChatGPT时刻”，“在80%的陌生场景中，机器人按语音或文字指令可完成约80%任务”。
之江实验室主任、阿里云创始人王坚也作为重磅嘉宾登台分享，他提出，AI概念有些被滥用了，实际上汇聚了很多不同技术。相信AI和机器人很可能汇聚变成一个新东西，“灵魂”和“身体”真的有机会真正合二为一。
在他看来，大语言模型仍留有巨大想象空间。语言离不开阅读、打字、发声这类肌肉动作，把这些动作信息纳入模型，这也是创业公司的机会。他坦言早年波士顿动力高管曾称赞宇树，自己却没能更早留意这家公司，并称“期待迎来AI的宇树时代”。
宇树昨日线上发布的新款人形机器人“H2-X功夫传奇 守护”今日首次亮相，官方称最长续航约8小时，可徒手抬汽车、从4米高空平稳落地、一脚踢凹车门。
现场，王兴兴还解读预研的物理AI机器人自进化V1.0，它依靠更强基础模型、多源数据、更多真机部署和技能生态扩展，形成感知、决策、执行、评估、学习、进化循环。
他畅想，希望今年年底到明年，用户告诉机器人整理房间，它就能做到。
同时，大会现场大量合作伙伴方案集中展示，王兴兴看后感慨：“连我自己都没有看过这么全面的解决方案。”宇树的平台型生态意图，也随之清晰。
一、物理AI要被击穿：今年底到明年，机器人自己帮你整理房间？
王兴兴的演讲从AI信仰切入，他提到，全世界几乎所有人，对AI的重视程度和信仰程度都还不够，包括他自己。
他在现场讲到一段插曲，四五年前曾让公司同事都去关注下头部AI公司的股票，如果大家相信，现在财务情况可能已经涨了数十倍。现在，真正看到顶级AI能力效果的人依然很少。
近来，AI已经实现更多里程碑，包括攻克前沿数学研究问题、软件逆向及漏洞发现和利用、高保真视频生成、自动驾驶出租车等；顶尖大模型公司内部已基本实现模型自进化，未来一两年模型能力的迭代还会持续加速。
王兴兴感慨，尖端AI模型已击穿数学世界，但可能尖端数学离我们太远，导致感受不直观。以模型目前的能力，哪怕不再继续进化，其实已经击穿大众的所有信息世界了。
那么物理世界呢？
王兴兴说物理AI目前已经看到被击穿的苗头，很可能在最近1-2年或2-3年发生。
他透露，目前GPT-6已实现完全零训练部署到宇树机器人上，可完成自主画画、清洗画笔，自主在精细棋盘上下围棋、象棋，自主寻找厨房并寻找工具等任务。
“我们正迈向具身智能的ChatGPT时刻，在80%的陌生场景中，通过语音或文字指令，机器人能够顺利完成大约80%的任务。”
王兴兴谈到一个具体的场景：“希望今年年底到明年，你告诉机器人帮我整理下房间，它就可以做到。”
二、机器人要自进化：要对齐误差，也拼千万台量产
具身智能迈过ChatGPT时刻最核心的挑战是什么？
王兴兴认为，规划问题已经解决，最大还没解决的是AI模型输入与输出的对齐误差问题。
当物理AI也突破临界点，最重要的点可能包括以下几个方面：一是需要机器人硬件的大规模生产制造能力，年产千万台；二是自研或合作物理AI模型及落地；三是超大规模算力服务器；四是算力和存储芯片制造能力。
王兴兴透露，宇树今年做的最重要的事情之一，是预研物理AI机器人自进化V1.0。由下图可知，在全球顶尖AI大模型的驱动下，宇树将下场定义技能、工具、任务约束等，推动机器人实现感知、决策、执行、评估、学习、进化的自进化循环。
要实现物理AI机器人自进化V1.0，需要以下条件：
1、基础模型越强，自进化性能越强。
2、多源数据规模化，驱动自进化规模效应。
3、更多真机部署，更多数据积累加指数增长的自进化速率。
4、技能累加扩展，走向规模化技能生态。
最后，王兴兴还公布了他本人与物理AI交织的时间线。他说，当下是最好的时代，害怕浪费每一天，总感觉还有很多东西需要学习，但是时间太少了。同时竞争是惨烈的，需要每天做得更好。
三、G1下线约2万台，全球开发者把机器人用进真实场景
现场，王兴兴还回顾了宇树最近一年的重要节点，从2026年马年春晚《武Bot》到全球首款量产版载人机甲GD01，再到宇树通用人形基础模型UnifoLM-WLA-1.0的发布，宇树的软硬件都得到了十足的发展。
规模化量产为宇树更广泛探索提供基础。王兴兴公布了一组数据：Unitree G1人形机器人自2024年5月发布至今，已成为全球最畅销的人形机器人，累计生产下线约1-2万台。
背后，来自全球的合作伙伴正加速推动智能机器人的发展，大量开源算法被开发者们落地到宇树机器人平台，覆盖教育科研、工业制造、文娱表演、具身导览、服务零售、智能巡检、应急消防等众多领域。
众多来自全球多地的开发者今天来到了大会现场，令智东西印象深刻的是一位美国科技内容创作者、YouTube博主Jake Laser。
他分享了自己基于宇树Unitree B2四足机器人产品，为行走不便的父亲开发出一个“智能轮椅”，最终成功实现上下台阶、旅行、徒步、涉水等各种便利行动。
不少合作伙伴的方案在现场展示，现场生态伙伴墙上展出了一共大180家左右的伙伴，覆盖从算力、互联网、工业零部件、机器人、AI企业到垂直行业、高校及科研机构各个领域。
结语：物理AI临界点前夜，具身智能拼生态
2026年是具身智能规模化放量的关键年，工信部数据显示，上半年中国人形机器人出货量已超4万台。但行业整体仍处早期，竞争焦点正从“本体形态”转向智能泛化能力与生态繁荣度。
对宇树而言，本次生态大会的召开是其打造“平台型生态公司”的关键宣言。通过覆盖渠道、技术、人才、资本等全维度的论坛和展示，宇树意在抢占物理AI的行业话语定义权，加速产业链协同，以生态构建竞争壁垒。
从更大角度来看，产业从技术验证走向规模交付、稳定干活的转折点上，需要用生态化布局来对冲单一产品公司的结构性风险，通过生态协同实现跨代发展。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Intellectual (WeChat Official Account:zhidxcom)
The author, Li Shui-young
Edition.
On October 22nd, according to Hangzhou, today, at the 2026 World Physics AI Congress and the Wootree Partnership Ecology Congress, the President, founder and CEO Wang Hing Xing Xing spoke for about 30 minutes on important topics such as AI beliefs, physical AI critical points, smart ChatGPT moments, self-evolution V1.0 and quantum production.
“nearly all the world's people, including myself, do not attach enough importance and faith to AI.” Wang Xing Xing said that Physical AI had seen the stinged head, which was likely to have occurred in the last one or two years; the industry was moving towards a “smart ChatGPT moment” and “in 80 per cent of the strange scenes, robots can perform approximately 80 per cent of their tasks by voice or word command”.
Wang Jian, Director of the Alibaba Cloud Founder of the ZQianz Laboratory, also shared it as a guest of honour, suggesting that the AI concept was somewhat abused and that in fact a number of different technologies were brought together. And believe that AI and the robots are likely to come together and become a new thing, and that there is a real chance that the "soul" and "body" can really be combined.
In his view, the large-language model still had a huge imagination. It is also an opportunity for start-up companies to incorporate these actions into models by means of muscles such as reading, typing and sound. He confessed that earlier Boston power executives had praised Woo-tree but had not been able to pay attention to the company earlier, and said “Looking forward to the age of the trees in AI”.
The new publish human robot, "H2-X-Kung Fu Legend Guardian" on the U.U. yesterday's line, was shown for the first time today, officially for a maximum of eight hours, to carry the car with his bare hands, to land flat from 4 metres above and to kick the door to the dent.
On the spot, Wang Xing-Hing also interprets the pre-developed physical AI robot self-evolution V1.0, which relies on stronger basic models, multi-source data, more real-time deployments and the ecological expansion of skills to form a perception, decision-making, implementation, assessment, learning, evolution cycle.
He was hoping that by the end of the year and next year, the users would have told the robot that it could do it.
At the same time, a large number of partnership programmes at the Congress site focused on the feeling that Wang Xing Xing looked at: “Not even I have seen such a comprehensive solution”. The platform-based ecological intent of the Woo-tree is also clear.
I. Physico-Ai will be shot through: By the end of this year and next year, the robot will clean your room?
Wang Xing Xing's speech drew on the faith of AI, and he mentioned that almost all the world's people, including himself, did not have enough attention or faith in AI.
He spoke on the spot of an episode where four or five years ago his colleagues were asked to look at the stock of the head of AI, and if you believe it, the financial situation may have risen several dozen times. The number of people who actually see the impact of the top AI is still low.
More recently, AI has achieved more milestones, including cutting-edge mathematical research issues, reverse software and the detection and use of loopholes, high-level security video generation, auto-driving taxis, etc.; the model has largely evolved within the top LLM company, and the iterative capacity of the model will continue to accelerate over the next year or two years.
Wang Xing is outraged that the cutting-edge AI model has pierced the mathematical world, but it may be that cutting-edge mathematics is too far away from us to feel intuitive. The current capabilities of the model, even if it does not continue to evolve, are already penetrating all the information worlds of the public.
What about the physical world?
Wang Xing Xing said that Physical AI had now seen the sprouts that had been punched through, which was likely to occur in the last 1-2 or 2-3 years.
He revealed that GPT-6 had now achieved complete zero training for deployment to the Wow tree robots, so that it could perform autonomous drawings, cleaning brushes, playing chess and chess on a fine board, looking for kitchens and finding tools.
“We are on our way to the smart ChatGPT moment, where robots can successfully perform about 80 per cent of their tasks in 80 per cent of the strange scenes, through voice or word instructions.”
Wang Xing Xing spoke of a specific scenario: “I hope that by the end of this year and next year, you tell the robot to help me clean up the room, and it can do it.”
II. The robot has to evolve itself: it has to match the error with the amount of millions of stations. Production
What are the central challenges of being smart enough to get past ChatGPT?
Wang Xing Xing believes that the planning problem has been resolved and that the biggest outstanding problem is the matching error of AI model input and output.
When physical AI also breaks the threshold, the most important points may include the following: large-scale manufacturing capacity, which requires robotic hardware, with millions of units produced annually; self-study or cooperative physical AI models and landings; super-large-scale computing servers; and, fourthly, computational and storage chip manufacturing capabilities.
Wang Xing Xing has revealed that one of the most important things that Woo-ho did this year is to prepare for the evolution of the physical AI robot V1.0. Driven by the top global AZLLM, as can be seen in the figure below, Woo-ki will come to define skills, tools, mission constraints, etc., and will contribute to a self-evolving cycle of perception, decision-making, implementation, assessment, learning and evolution of robots.
The self-evolution of the physical AI robot V1.0 requires the following:
1 The stronger the base model, the stronger the evolution.
2. Scale multi-source data, driving the self-evolution scale effect.
3. More real-time deployments, more data accumulation, and an exponential rate of self-evolution.
4. Skills build up to scale skills ecology.
Finally, Wang Hing Xing announced a time line between himself and Physical AI. He said that it was the best of times, that fearing to waste every day, it felt that there was much to learn, but that time was too little. At the same time, competition is dire and needs to be done better every day.
Three, G1 down, about 20,000. Global developers use robots for real.
On the ground, Wang Xing Xing also recalled the important nodes of the recent year, from Wu Bot on the night of the spring of 2026 to GD01, a global download of human carriers, to publish, a UnifoLM-WLA-1.0, of the UifoLM-WLA-1.0, where the software and hardware of the Oki trees were developed.
Scalable production provides the basis for wider exploration of wild trees. Wang Heung Heung released a set of data: Unitree G1 man-form robots have been the world ' s best-selling human-shaped robots since publish in May 2024, with a cumulative down-line of about 12,000.
Behind the back, partners from around the world are accelerating the development of intelligent robots, with a large number of open source algorithms being developed to land on the Wootree robot platform, covering a wide range of fields, such as educational research, industrial manufacturing, recreational performances, hands-on tours, retailing of services, smart inspections and emergency fire.
Many developers from many parts of the world are here today, and one of the intellectuals who is impressed by the fact that an American creator of scientific and technological content, Youtube blogger Jake Laser, is here today.
He shared his four-foot robotic product based on Unitree B2 and developed a “smart wheelchair” for an inconvenient father, which eventually succeeded in facilitating steps down, travel, hiking and water-related actions.
A number of partners ' programmes were shown on the ground, where some 180 or so partners were displayed on the ecological partnership wall, covering various fields ranging from computing, the Internet, industrial components, robotics, AI enterprises to vertical industries, universities and research institutions.
Conclusion: Smartly encoded ecology, night before the physical AI critical point
The year 2026 was a critical year for smart-scale deployment, and Ministry of Industry and Communications data show that more than 40,000 Chinese human-shaped robots had arrived in the first half of the year. However, the industry as a whole is still in its early stages, and the focus of competition is shifting from a “physical form” to an intellectually broad-based capacity and ecological prosperity.
For Woo-tree, the convening of this ecological conference was a key declaration for its “platform ecological company”. Through forums and presentations on the full dimensions of coverage channels, technology, talent, capital, etc., the Tree aims to seize the right to define the trade language of physical AI and accelerate the synergy of industrial chains to create ecological barriers to competition.
From a broader perspective, at a turning point in the industry ' s move from technology validation to scale delivery and stabilization of work, it is necessary to use an ecological configuration to hedge the structural risks of single-product companies and to achieve intergenerational development through ecological synergy.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.