10月19日，工信部在国新办发布会上披露了一组数字：今年上半年，我国人形机器人整机产品已经突破400款，产量超过4万台，全年预计突破10万台。
十天后的10月29日，上海世博中心，智元再次验证了这份行业进度。智元联合创始人、总裁兼CTO彭志辉在2026智元创新日上交出了上半年的成绩单：9月，第20000台机器人正式下线；上半年出货9700台，全球市占率超过43%；公司上半年收入接近20亿元。
放在年初，这样的数字本身就足以撑起一场发布会。但这一次彭志辉把重点放在了后半段。“我们生产了多少台机器人，并不等于真实世界中就多了多少生产力。”
在他看来，这是行业眼下最需要共同解决的难题：规模制造正在快速达成，而稳定上岗的能力还没有被验证。而客户的心态正在从“我愿意尝试你的产品，看你能不能造得出来”，变成“我到底能不能养得起你的产品”。
开机8小时，不等于干了8小时
彭志辉在演讲中提到：过去看一家机器人公司，看的是它能做什么、视频里的动作够不够炫酷；现在客户开始问的是另外三件事：到底能稳定干多少活，能干多好，干这些活要花多少钱。
“一台机器人开机8小时，并不意味着它就创造了8小时的价值。它一次任务成功率很高，也不意味着它能像人一样持续稳定地工作一个月。”究竟如何衡量出机器人给客户创造的价值？为了回答这个问题，智元给造了一个新的标尺——Worken（有效智能工时）。
公式并不复杂：Worken（有效智能工时）＝活跃机器人数量 × 自主工作时长 × 产出效能比。
三个乘数回答的是三个很朴素的问题。“活跃机器人数量”看的是有多少机器人真正在岗，而不是躺在仓库里；“自主工作时长”用来区分一台机器人是被动的工具还是自主的智能体，彭志辉举例说，一台机器人可能看起来很聪明，但每工作10分钟就失误一次、需要人接管一次，那它创造的自主工时依然非常有限；“产出效能比”算的是一笔等效账，“这一台机器干的活抵得上一个人，还是要两台才抵一个，甚至将来一台能顶好几个。”
Worken诞生的意义在于在机器人的制造数量之外，终于出现了一个新的维度可以度量机器人能否自主、持续地完成有价值的工作，以及实际产出多少成果。
发布会现场，彭志辉向我们展示了智元机器人在多个落地场景下所达成的Worken数量，场景囊括龙旗产线上下料、香格里拉酒店服务生、成都宽窄巷子的游客服务、广州地铁安检引导等，其中长隆的文娱表演机器人以33480的Worken领跑所有场景。
彭志辉表示，如果度量搞不清楚，买卖双方就没有办法定价，投入和产出算不清楚，行业也形不成共识。他希望未来对机器人的服务按Worken工时计费，不同智能水平的机器人干不同的任务，费率也不同。
智元合伙人、高级副总裁、通用业务部总裁王闯对这个模式描述地更为细致：以后客户也许不再按天租机器人，而是买一个“生产力套餐包”，预定任务，按消耗的Worken扣费，“给你服务的不是一个机器人，而是所有不同场合下机器人的组合”。
20:1
一台在岗的机器人，其背后要配多少个工程师，之前一直是行业不愿公开的数字。王闯在会后访谈中却坦率表示：“长隆第一批300台机器人进场时，几百台机器就配了几百名工程师，接近一比一。而现在，这个比例降到了三比一，年底我们要做到十比一，明年降到二十比一，远期三十比一。”
人机比降不下来，Worken就算不过来。高级工程师非常稀缺，每个场景都要堆专家的话，本质上是在做单个项目，而不是规模化的生意。王闯表示：“20比1是一个临界点，也就是5%。当一个岗位只需要新增5%的人力预算时，对很多客户来说就是一个可以接受的点。”
撤走运维专家之后，如何保证机器人的稳定运行，是客户们最为关心的问题之一。智元给出的答案是一套自动化的运维体系。专家在总部做远程支持，现场运维人员照着专家做好的模板执行，同时底层是一套一站式运维平台，从参数点检、站点上线、在线管理到问题闭环，四步全部自动化。
智元在龙旗的产线就已经实现了这套体系的接管。这条产线2026年3月规模化投产，到现在稳定运行超过7个月。上半年智元做过一次6天全产线直播，机器人累计作业超过64小时、累计操作64824次，产出17625台平板电脑，成功率99.99%，节拍平均22秒每件，达到了和人类同等水平。而智元的研发型团队早就从龙旗现场撤出来了，基础软件的运维已经完全移交给了龙旗自己的团队。
长隆之所以能成为Worken冠军，与最高的活跃机器人数量有着直接关系，而这个数量的快速增长来自客户的复购率。
长隆的第一批300多台机器人分布在100多个点位，每个点位的任务完全不同：有跟人下三子棋的，有猜拳的，有打乒乓球的，有带着一大群游客讲解的，还有跟小朋友比跨越障碍的。智元合伙人、联席总裁、中央研发部总裁、交互业务体系总裁熊彦透露，一个讲解点位很快从一台变成了三台，“当客户发现机器人能很成功地完成一个工作任务时，他就愿意复购。”
目前长隆已经部署了将近500台机器人，国庆长假园区日均接待约1万人次，飞船乐园的客流量提升了将近40%，而年底长隆的机器人会计划增加到1000台。
通用智能与具身智能
至于最近讨论度最高的话题：通用大模型会不会吃掉具身智能？智元联合创始人、联席总裁、具身业务体系总裁姚卯青给出的答案是：不会，至少现在不会。
姚卯青引用英伟达的报告，120项仿真任务里，GPT-6 Astra的平均成功率是66.3%，能力集中在语义理解、动作规划和决策上，“通用模型越强，具身行业的起点会越来越高，这是明确的利好”。但另一方面：Astra的平均推理很慢，单次要18秒多，高动态和精细作业做不准。涉及力触反馈的环节，大模型的成功率基本完全是零，因为先天没有学习过操作能力。
熊彦则从另一个角度解读了具身模型的不可替代性。“一个很简单的动作可能要分钟级推理，调用一次上百美元，进垂域还要微调，链条上训练成本、数据成本、微调成本都很贵。我不相信有人会用5万亿的模型部署在机器人身体上，最终落地还是要考虑现实性”。
在姚卯青看来，通用智能和具身智能的关系更像一个分层架构。通用模型负责System 2层面的长程规划和推理，中间是VLA和世界模型训练出来的一个个技能，作为工具被调用，最底下是100到1000赫兹的实时电机控制，和本体和任务紧紧耦合。反过来，通用模型想在物理世界做到专家级，同样绕不开海量高质量后训练，“不在物理世界里针对具体任务做海量试错，它连Planning那一层都做不好”。
“很多人问我AI要击穿具身了，应该怎么办”演讲最后，姚卯青最后笑着说，“那我们就成为AI那一侧，不要成为被击穿的那一侧。”
只做0到1
智元只做0到1的场景、让伙伴做1到N。这是王闯在访谈中反复讲到的分工，“很多数据我们拿不到，有很多行业Know-how我们也不知道。我们没有必要每一块钱都赚”。
灵渠OS 1.0就是实现这个分工的基座。它是业界首个具身智能原生操作系统，支持15类100多种外设的即插即用，在1000赫兹高频任务下，端到端时延抖动小于4微秒。智元打算在2026年底把它全面开源，把在公司内部经过量产验证的2亿行代码和5000多个生态软件包一并贡献出去。
人机比要从20:1继续压到30:1，光靠智元多招工程师是做不到的，只能靠这些伙伴把一个个非标场景击穿。年初智元发布的伙伴生态也在本次大会上交出了一份成绩单：元苼计划联动255所高校、511家企业、8万多名开发者，今年已投入3亿元以上。王闯表示，生态的魅力是你永远想象不到这些伙伴能带给你多少惊喜。
在彭志辉看来，机器人真正普及的标志是：几乎不需要智元再提供服务。企业可以自己去购买、自己去分发。王闯也表达了相同的观点：就像智能驾驶看接管里程越来越长，机器人要看的是“不用智元服务”的比例越来越大。
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。
On 19 October, the Ministry of Industry and Communications disclosed a set of figures at the New Office ' s publish meeting: In the first half of this year, our man-made robotics products have exceeded 400 dollars, yielding more than 40,000 units and are expected to exceed 100,000 units throughout the year.
Ten days later, on 29 October, Ji-won again validated the progress of the industry at the Shanghai World Bank Centre. The founder, CEO and CTO Peng Zhihui handed over a report card for the first half of the day: in September, the 2000th robot was officially offline; in the first half of the year, 9,700 were delivered, with a global market share of over 43 per cent; and in the first half of the year, the company earned nearly $2 billion.
At the beginning of the year, such figures alone would be sufficient to sustain a publish conference. But this time Peng Zhihui focused on the latter part. "How many robots we produce are not the same as more productive in the real world."
In his view, that was the most pressing challenge for the industry to address: scale manufacturing was being achieved quickly and the capacity to stabilize entry had not yet been tested. And the client mentality is going from "I'd like to try your product, see if you can make it," to "I can afford your product."
Eight hours, not eight hours.
In his speech, Peng Zhihui mentioned that he used to look at a robotic company, that it could do anything, that there was enough action in the video; and that the client began to ask three other things: how much work can be stabilized, how good can it be, and how much it costs.
“Each robot has an eight-hour start, which does not mean that it has created an eight-hour value. Its one-time mission, success rate, is high, and does not mean that it can work like a human being for one month in constant stability. How do we measure the value that robots make to customers? To answer this question, Ji-won created a new ruler, Worken.
The formula is not complex: Worken (effective intelligent working hours) = number of active robots x number of autonomous working hours x output efficiency ratio.
The three multipliers answer three very simple questions. “The number of active robots” looks at how many robots are actually on guard, rather than lying in a warehouse; “the time of self-employment” is used to distinguish between a robot being a passive tool or an autonomous AI agent. Peng Zhihui, for example, says that a robot may look smart, but the time it creates is still very limited when it fails and needs to be taken over once every 10 minutes; “the output-effectiveness ratio” is an equivalent, “is this machine worth one person or two to one, or even several in the future”.
Worken was born in the sense that, in addition to the number of robotics manufactured, a new dimension has finally emerged to measure whether robots can do valuable work on their own and on a sustainable basis and how much they actually produce.
On the scene of the publish conference, Peng Zhihui showed us the number of Workens that the Ji-won robots had arrived at under a variety of landing scenarios, including the material on the Dragon Flag Line, the Shangri-La Hotel waiter, the services of tourists in Chengdu's wide alleys, the guide to the Guangzhou subway security, and so on.
Peng Zhihui stated that if measurement was not clear, there was no way for buyers and sellers to price, inputs and outputs were unclear, and there was no consensus in the industry. He hoped that future services to robots would be costed on a Worken-hour basis and that robots at different levels of intelligence would do different tasks at different rates.
This model is described in greater detail by the Jiwon partner, Senior Vice President, and Wang, President of the General Operations Department: in the future, clients may no longer rent robots on a daily basis, but rather buy a “productivity package”, pre-ordering the mission at the cost of the spent Worken deduction, “not a robot, but a combination of robots on all occasions”.
20:1
The number of engineers behind an on-duty robot has been a long-standing industry reluctance to disclose figures. In a post-session interview, the King frankly stated: “When the first 300 robots in Long Rong arrived, hundreds of them were equipped with hundreds of engineers, close to one to one. And now, the ratio is down to three to one, and we're going to be 10 to one by the end of the year, down to 20 to one next year, with 30 to one in the long term.”
The drone won't come down, or the Worken won't come. Senior engineers are scarce, and each scene has to be filled with experts and is essentially a single project, not a scalable business. Wang Bark said: “20:1 is a critical point, or 5%. When a job requires only a 5 per cent additional human budget, it is an acceptable point for many clients.”
The question of how to ensure the stable operation of robots after the removal of the experts in the field is one of the most important concerns of clients. The answer given by Ji-won is an automated system. The experts provide remote support at Headquarters, with on-site transport staff following the template developed by the experts, while the bottom level is a all-in-one transport platform, which will be fully automated in four steps, from parameter point-checking, site online, online management to problem loops.
Ji-won has already taken over the system on the line of the flag. The production line came into operation on a scale in March 2026 and is now operating steadily for more than seven months. In the first half of the year, Ji-won performed a six-day live live production line, with robotics operating over 64 hours and 64824 cumulative operations, producing 17625 tablet computers, success rate99.99 per cent, with a beat average of 22 seconds, at the same level as humans. And Ji-won's research and development team had withdrawn from the flag site long ago, and the flow of basic software had been handed over to the flag's own team.
Longlong became the Worlden champion, directly related to the highest number of active robots, which grew rapidly from the repurchase rate of customers.
The first group of 300 robots in Long Rong is located in more than 100 places, each with a completely different task: with three players playing, with boxing, with ping-pong, with a large group of tourists, and across barriers with children. The Jiwon partners, Co-President, President of the Central Research and Development Department, and President of the Interoperability System, Koshiya, revealed that a lecture point quickly changed from one to three, “when the client finds that the robot can successfully perform a job, he is willing to repurchase it.”
There are now nearly 500 robots deployed in Nagayama, approximately 10,000 daily receptions in the National Day's false park, and traffic at the ship park has increased by almost 40 per cent, while at the end of the year, there are plans to increase to 1,000 robotics.
General Intelligence and Smart
As for the most recently discussed topic: Will General LLM eat smart? The answer given by the founder, co-president, and chief executive officer of the operational system, Yao Dukqing, is: No, at least not now.
Yao Dixing quoted the report by Young Weidar that out of 120 simulation missions, the average GPT-6 Astra average success rate was 66.3 per cent, and that capacity focused on semantic understanding, action planning and decision-making, “the stronger the generic model, the higher the starting point for the pro-industry, the clear advantage”. On the other hand, Astra's average reasoning is slow, with more than 18 seconds alone, and high-dynamic and fine work not allowed. The success rate of LLM, which involves the chain of force-touching feedback, is almost entirely zero, as no one has learned to operate.
In another way, he has read the irreplaceable model. “A simple move may require minutes of reasoning, hundreds of dollars at a time, fine-tuning in the periphery, and the costs of training, data and fine-tuning on the chain are expensive. I do not believe that a 5 trillion-billion-dollar model will be deployed on the robotic body and will eventually land on the ground or be realistic”.
In Yao's view, the relationship between universal intelligence and intelligence is more like a hierarchical structure. The generic model is responsible for long-range planning and reasoning at the System 2 level, in the middle of which is a skill that VLA and the World Model have been trained as a tool to be used, at the bottom of which is real-time power control of 100 to 1,000 Hz, in close alignment with the physical and mission. In turn, the generic model seeks to achieve an expert level in the physical world, which is equally unwieldy of high-quality post-training in the mass, and “does not make a mass test of a specific mission in the physical world, and it is not even good at Planning”.
“Many people ask me what I should do when I'm going to blow through my body”, and Yao Zhiqing finally smiles, and says, “Then we'll be on the side of AI, not on the side that gets shot through.”
Only zero to one.
Ji-won is only doing zero to one, making partners do one to N. This is the division of labour that Wang has repeatedly spoken about in his interviews, "We can't get much of the data, and we don't know much about Know-how. We don't have to make every dollar."
Channel OS 1.0 is the basis for this division of labour. It is the first smart primary operating system in the industry to support more than 100 off-sites of 15 classes, with end-to-end latency shakes of less than 4 microseconds under 1,000 Hz HF. Ji-won intends to make it fully available by the end of 2026, with the contribution of 200 million line codes and over 5,000 eco-packages validated within the company.
The machine would continue from 20:1 to 30:1 and it would not have been possible to recruit engineers on the basis of intellectual capital alone, and those partners would have to beat one non-spectrum scene through. At the beginning of the year, the Zhiwon publish's partner ecology also presented a report card at the conference: the Yuan Yuanxian project, which links 255 higher education institutions, 511 enterprises, and more than 80,000 developers, has invested over $300 million this year. Wang Bong says that the beauty of ecology is that you can never imagine how many surprises these partners can bring to you.
In Peng Zhihui ' s view, the true sign of robotics is that there is little need for the Jiwon to provide services. Businesses can buy and distribute on their own. The same view was expressed by Wang: just as smart drivers look at the longer and longer take-over miles, robots look at a growing proportion of the “unutilized services”.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.