具身智能公司星尘智能近日完成10亿元规模新一轮融资，由投控东海（深投控）领投，投资方有洪泰基金、光华阳光管理的宁海兴泰合基金、善恒资本等，老股东先沐也持续追投。公司估值达到150亿元。星尘智能2022年成立，创始团队出自腾讯Robotics X实验室。6月份，星尘智能刚完成了超10亿元的B系列融资，当时披露的估值是突破100亿元。近半年来，这家公司完成了4轮融资。
Jiemian's parent company Zhongtouwang has learned exclusively that embodied intelligence company Stardust Intelligence recently completed a new funding round of 1 billion yuan, led by Toukong Donghai (Shenzhen Investment Holdings), with investors including Hongtai Fund, Ninghai Xingtaihe Fund managed by Guanghua Sunshine, and Shanheng Capital, with existing shareholder Xianmu also continuing to invest. The company's valuation reached 15 billion yuan. Stardust Intelligence was founded in 2022, with a founding team from Tencent Robotics X Lab. In June, Stardust Intelligence had just completed a Series B round exceeding 1 billion yuan, with disclosed valuation "breaking 10 billion yuan." In the past half year, the company completed 4 funding rounds.
这轮融资后，星尘智能的股东阵容已经相当豪华。腾讯系的德迅投资、蚂蚁集团、字节系的锦秋基金聚首，中科创达、科德教育等多家上市公司扎堆，还有国科投资与深圳、无锡等多地国资，以及云启资本、经纬创投等多家财务投资机构在列。在这一轮融资官宣前，投中网见到了星尘智能创始人、CEO来杰。在眼下的档口，围绕具身智能赛道的争论正在升温，有人担心概念跑在了商业化前面，头部企业的上市进程也引发市场对赛道估值的重新审视。核心质疑其实就是一句话，具身智能到底有没有真实需求、能不能赚到钱。而来杰对此的回答相当出乎预料，在他看来，具身是非常容易挣钱的一个行业，一点都没有大家说的那么难。
After this funding round, Stardust Intelligence's shareholder lineup is already quite impressive. Tencent-affiliated Dexun Investment, Ant Group, and ByteDance-affiliated Jinqiu Fund gathered together, with multiple listed companies including Thundersoft and Kede Education clustered in, as well as Guoke Investment and multiple local state capital from Shenzhen and Wuxi, and multiple financial investment institutions including Yunqi Capital and Matrix Partners. Before this round's announcement, Zhongtouwang met Stardust Intelligence founder and CEO Lai Jie. At the current juncture, debates around the embodied intelligence track are heating up. Some worry that concepts have run ahead of commercialization, and leading companies' IPO processes have also triggered market re-examination of track valuations. The core question is essentially one sentence—does embodied intelligence have real demand, and can it make money. Lai Jie's answer was quite unexpected—in his view, embodied is a very easy industry to make money in, not nearly as hard as everyone says.
在融资密集的具身智能赛道，星尘智能前两年其实并不能算是资本宠儿。不久前，投中网在与一位早期投过宇树的投资人交流时也提到了星尘智能这家公司。他给出的评价是，技术很牛，但是几个热点都没靠上，融资比较慢。所谓热点是什么呢，朱啸虎前一阵有一段diss具身智能行业的演讲被很多人转发。他说，中国现在两百多家机器人企业，哪一个不会跳舞，哪一个不会打拳。的确，跳舞、打拳，无疑是具身智能公司最能出圈的事情。但星尘智能这家机器人公司，就没让机器人跳过舞。因为星尘智能一直在做看起来没那么酷的轮式机器人。
In the funding-dense embodied intelligence track, Stardust Intelligence could not be considered a capital darling in the first two years. Not long ago, Zhongtouwang also mentioned Stardust Intelligence when exchanging with an investor who had invested early in Unitree. His assessment was: technology is impressive, but it missed several hot spots, and funding was relatively slow. What are the hot spots? Zhu Xiaohu's recent speech dissing the embodied intelligence industry was widely forwarded. He said: China now has over 200 robot companies—which one cannot dance, which one cannot box. Indeed, dancing and boxing are undoubtedly what make embodied intelligence companies most famous. But Stardust Intelligence has not made its robots dance, because Stardust Intelligence has consistently been making wheeled robots that look less cool.
来杰的判断是，AI泛化能力的增长，更多会体现在操作能力上，而不是移动能力上。这个判断，让团队没有在当时更容易出效果、更容易融资的足式路线上花时间。来杰带点自嘲地向投中网表示，前两年他确实没想到机器人靠跳舞也能赚钱，他觉得是必须要能干活才行。所以，星尘智能自成立起就在做操作型的机器人，也就是让AI驱动机器人的双手去干活。说到这就不得不提星尘智能最为外界所熟知的标签，绳驱。所谓绳驱是一种用绳索代替传统减速器传递动力的机器人传动方式，类似人体肌腱的工作原理。这是一条较为小众、技术也相对没那么成熟的路线。
Lai Jie's judgment was that AI generalization capability growth would manifest more in manipulation capability than mobility capability. This judgment kept the team from spending time on legged routes that were easier to produce results and attract funding at the time. Lai Jie self-deprecatingly told Zhongtouwang that in the first two years he确实 did not expect robots could make money by dancing—he thought they had to be able to work. So Stardust Intelligence has been making manipulation-type robots since founding—letting AI drive robots' hands to "work." This brings up Stardust Intelligence's most widely known label—cable drive. Cable drive is a robot transmission method using cables instead of traditional reducers to transmit power, similar to how human tendons work. This is a relatively niche, technically less mature route.
2020年前后，来杰当时还在腾讯Robotics X实验室研究机器人的操作能力。他对当时机器人的发展状态感到很困惑，明明工业相机的分辨率已经很高，机械臂的精度也做到了亚毫米级，但机器人干起活来却依然笨笨的、傻傻的。来杰提出了一个问题，智能的提升，会不会并不完全依赖精度。拿人来做类比，盲人没有视觉，人的手臂精度其实也很差，为什么做事情能做得非常好。于是，他决定用盲人的思路做一次实验，场景是一个机器人的传统难题，开门。开门这件事对人类来说很简单，对机器人却非常难，因为需要用算法去估计门轴位置、规划轨迹，稍有偏差就可能把门把手拽下来。而来杰和团队换了一个策略，先把机械手压在门把手上，感受到把手转动，就跟着调整，不再硬拉。
Around 2020, Lai Jie was still at Tencent Robotics X Lab researching robot manipulation capability. He was very confused about the state of robot development at the time—industrial camera resolution was already high, robotic arm precision had reached sub-millimeter level, yet robots still worked clumsily and stupidly. Lai Jie raised a question: could intelligence improvement not depend entirely on precision. Using humans as an analogy, blind people have no vision, and human arm precision is actually very poor—why can they do things so well. So he decided to conduct an experiment using the "blind person's approach," with the scenario being a traditional robot challenge—opening doors. Opening a door is simple for humans but very difficult for robots, because algorithms are needed to estimate door hinge position and plan trajectories, and slight deviation could pull the door handle off. Lai Jie and his team switched strategy: first press the robotic hand on the door handle, feel the handle turning, then adjust accordingly instead of pulling hard.
这个实验给他留下的关键认知是，力的透明度很重要。人在做事情时，力的施加和力的反馈是闭环的。绳驱的传动方式恰恰就能做到力对AI来说也几乎完全透明，就跟两个人握手一样。从这里出发，来杰重新理解了绳驱的价值，它并不仅仅是一个硬件路线的选择，而是智能往下一步发展的基础。Design for AI，这是星尘智能最核心的理念，也是理解这家公司的钥匙。在来杰看来，具身智能要发展，应当围绕AI的能力去搭配能作用于物理世界的机器人，而不是先设计出一个机器人，再把AI能力加上去。
The key insight this experiment left him with was that force transparency matters. When humans do things, force application and force feedback form a closed loop. Cable drive transmission achieves almost complete force transparency for AI, like two people shaking hands. From here, Lai Jie重新 understood the value of cable drive—it is not just a hardware route choice, but the foundation for intelligence's next step. "Design for AI" is Stardust Intelligence's core philosophy and the key to understanding this company. In Lai Jie's view, for embodied intelligence to develop, robots that act on the physical world should be configured around AI capabilities, rather than first designing a robot and then adding AI capabilities.
2025年6月，星尘智能的第一代旗舰产品Astribot S1量产，这是全球首款实现量产的绳驱AI机器人。2022年成立，2025年6月实现量产，这个速度放在同期的具身智能赛道并不算快。客观上，也的确是错过了机器人跳舞、打拳的那一波红利。但在绳驱这件事上，星尘智能几年来筑起的壁垒足够深。目前全球范围内最接近的对手应该是山姆·奥特曼投资的1X Technologies，这家公司最新的估值据称超过60亿美元，其首款量产机器人NEO去年10月已预售，但目前仍未进入交付。
In June 2025, Stardust Intelligence's first-generation flagship product Astribot S1 entered mass production—the world's first mass-produced cable-driven AI robot. Founded in 2022 and achieving mass production in June 2025, this pace is not fast in the same period's embodied intelligence track. Objectively, it did miss the wave of robot dancing and boxing. But on cable drive, Stardust Intelligence has built deep barriers over the years. Currently the closest global competitor should be 1X Technologies, invested by Sam Altman, whose latest valuation is reportedly over $6 billion. Its first mass-produced robot NEO was pre-sold last October but has not yet entered delivery.
星尘智能的绳驱机器人量产，在两个方向对传统人形机器人市场构成了冲击。首先是价格。今年6月，星尘智能发布了身高1.55米的轻量级型号T1，把起售价下探到了8.99万元，接近了商业化起量的甜点区间。T1能卖这个价，是因为绳驱机器人有天然的成本优势。T1全身20多个关节所有绳索的成本加起来，才约等于人形机器人一个关节里使用的谐波减速器的价格。第二点是操作性。在设计第一代旗舰产品S1时，来杰就给团队定下一个指标，机器人的负载、速度、关节活动范围，都要比成年人高出15%。
Stardust Intelligence's cable-driven robot mass production poses impacts on the traditional humanoid robot market in two directions. First is price. In June this year, Stardust Intelligence released the 1.55-meter lightweight model T1, pushing the starting price down to 89,900 yuan, approaching the sweet spot for commercial volume. T1 can sell at this price because cable-driven robots have natural cost advantages. The total cost of all cables in T1's over 20 joints across the body is about equal to the price of a harmonic reducer used in one joint of a humanoid robot. Second is operability. When designing the first-generation flagship S1, Lai Jie set a target for the team: the robot's payload, speed, and joint range of motion must all exceed an adult's by 15%.
9月初，前字节跳动强化学习专家孙鹏加入星尘智能，负责机器人强化学习方向的研发与应用探索。这件事非常值得注意，因为它背后有一个行业层面的难题，传统上，机器人大脑模型的训练，无论是VLA还是WMA，都完全依赖昂贵的模仿学习。如果能引入强化学习，机器人的大脑进化速度可能会有质的飞跃。星尘智能得以去尝试强化学习，正是得益于绳驱机器人对力反馈的精细感知。传统的机器人本体动手不知轻重，很难承受强化学习所需的大量试错。
In early September, former ByteDance reinforcement learning expert Sun Peng joined Stardust Intelligence, responsible for R&D and application exploration in robot reinforcement learning. This is very noteworthy because behind it is an industry-level challenge—traditionally, robot brain model training, whether VLA or WMA, relies entirely on expensive imitation learning. If reinforcement learning can be introduced, robot brain evolution speed could achieve a qualitative leap. Stardust Intelligence's ability to attempt reinforcement learning precisely benefits from cable-driven robots' fine force feedback perception. Traditional robot bodies do not know轻重 when handling things, making it hard to withstand the大量 trial and error required for reinforcement learning.
说到这里，星尘智能真正的野心才展露出来，硬件的优势，最终要落到AI上。星尘智能并不是今年才开始转向大脑研发，2025年12月发布了VLA模型Lumo-1，今年7月又发布了WMA模型Lumo-2，都是领先水平的好模型。此时此刻，机器人的大脑到底该怎么做，技术上并没有收敛。星尘智能的做法是在多个路线上同步探索。来杰把内部的几个模型团队比作大脑的不同分区，做VLA的是前额叶，做强化学习的是基底神经节，此外还有偏世界模型、偏数据质量评估的团队等。
At this point, Stardust Intelligence's true ambition emerges—hardware advantages ultimately land on AI. Stardust Intelligence did not just start turning to brain R&D this year; it released the VLA model Lumo-1 in December 2025, and the WMA model Lumo-2 in July this year, both leading-level good models. At this moment, how robot brains should be built has not converged technically. Stardust Intelligence's approach is exploring multiple routes simultaneously. Lai Jie compares internal model teams to different brain regions: those doing VLA are the prefrontal cortex, those doing reinforcement learning are the basal ganglia, plus teams leaning toward world models and data quality assessment.
星尘智能第一个跑通的商业化场景你可能想不到，卖盲盒。2025年底，在北上广的顶级商圈，S1机器人上岗了机器人MART，自主运营售卖起热门潮玩IP WAKUKU哇库库玩偶盲盒。来杰表示，客户一开始的预期其实并不高，只想要用VLA完成一个递盒子的动作而已。但星尘智能主动帮客户把整套营销流程设计了出来，在5米范围内用各种方式招揽顾客，用约1分钟介绍产品，用话术催促下单，把盲盒递出去，最后还要给出祝福，完成情绪价值的闭环。实际跑下来的数据好的远超预期。在北京合生汇的测试点位，机器人亭的月销量约为附近自动售卖机的4倍。
Stardust Intelligence's first commercialized scenario may be unexpected—selling blind boxes. At the end of 2025, in top commercial districts of Beijing, Shanghai, and Guangzhou, S1 robots staffed "Robot MART," autonomously operating sales of popular trendy toy IP WAKUKU blind boxes. Lai Jie said customers' initial expectations were not high—they just wanted VLA to complete a box-handing motion. But Stardust Intelligence proactively designed the entire marketing process for customers: attracting customers within 5 meters in various ways, introducing products in about 1 minute, using scripts to urge orders, handing out blind boxes, and finally giving blessings to complete an emotional value loop. Actual data far exceeded expectations. At a test point in Beijing Hopson One, the robot kiosk's monthly sales were about 4 times nearby vending machines.
有客户计划部署约300台咖啡机器人，在咖啡上用喷粉打出自家logo，真正打动对方的，是统一的话术和动作风格，可以把企业理念自然地带进服务过程中。来杰认为，腾讯的产品思维对星尘智能的影响很大。2025年以来，星尘智能已经拿下多笔千台级订单，涵盖零售、文娱、物流等多种场景。除了商业场景，星尘智能商业化的另一条线是科研和开发者。在今年的WAIC上，星尘自己没有设展位，却有京东、蚂蚁、达摩院等7家机构用星尘机器人展示自己的模型。来杰透露，目前公司的科研订单基本处于断货状态，产能提前一个月就被订完。
One customer plans to deploy about 300 coffee robots, printing its logo on coffee with powder spraying. What truly moved them was the unified scripts and action style, naturally bringing corporate philosophy into the service process. Lai Jie believes Tencent's product thinking has greatly influenced Stardust Intelligence. Since 2025, Stardust Intelligence has won multiple thousand-unit orders covering retail, entertainment, and logistics scenarios. Besides commercial scenarios, Stardust Intelligence's other commercialization line is research and developers. At this year's WAIC, Stardust did not set up a booth itself, yet 7 institutions including JD.com, Ant, and DAMO Academy used Stardust robots to demonstrate their own models. Lai Jie revealed that the company's research orders are basically in a "sold out" state, with capacity booked a month in advance.
在盈利模式上，起售价低至8.99万元的T1依然有不错的硬件毛利，未来还会在硬件终端上不断推出新的软件服务，进一步提高毛利空间。模型方面，目前主要向用户提供试用，未来计划做成订阅制。正是基于此，来杰给出了前面说的结论，具身智能行业挣钱很容易。当然这个容易有一个前提，踏实做产品，服务真实的客户。
On profitability model, the T1 with a starting price as low as 89,900 yuan still has decent hardware gross margin, and new software services will continuously be launched on hardware terminals in the future to further improve gross margin space. On models, currently mainly providing trials to users, with plans for subscription in the future. Based on this, Lai Jie gave the earlier conclusion—the embodied intelligence industry is easy to make money in. Of course, this "easy" has a premise—steadily making products and serving real customers.
今年下半年，大家逐渐察觉到资本市场对具身智能的态度已经在发生变化。来杰回忆，2024年融资时很多投资人说“不做端到端不投”，2025年变成了“不做VLA不投”，而今年越来越多投资人开始问商业化做到什么程度了。他对此的内心感受是，你们终于关注对的问题了。朱啸虎前一阵在北大演讲时提出一个引起热议的观点，绝大部分具身智能公司活不过十年。来杰也同意这个观点，他说，未来上市的具身智能公司肯定会超过200家，但和现在这批公司的重合度可能连10%都不到。
In the second half of this year, everyone gradually noticed that capital market attitudes toward embodied intelligence were changing. Lai Jie recalled that in 2024 funding, many investors said "no investment without end-to-end," in 2025 it became "no investment without VLA," and this year more and more investors are asking how far commercialization has progressed. His inner feeling was—you finally focused on the right question. Zhu Xiaohu recently gave a widely discussed view at Peking University: most embodied intelligence companies will not survive ten years. Lai Jie agrees with this view. He said future listed embodied intelligence companies will certainly exceed 200, but overlap with the current batch may be less than 10%.
来杰有一个长期的判断，他认为具身智能行业未来主要有两波红利。第一波是中国的供应链红利。机器人的需求方是中国企业，供应链也在中国，这让机器人成本未来可以下探到人力成本之下。第二波红利是AI能力提升的红利。随着AI大模型的发展，同样价位的机器人，将拥有远超这个价位上任何人类的专业能力。两波红利叠加，才是行业真正的机会。星尘智能当下的重心依然是做出下一代杀手级的产品。来杰透露，星尘智能计划明年推出一款面向家庭的机器人，他认为这款产品有望成为爆款，因为它的售价和能力都会极大地突破目前的行业认知。
Lai Jie has a long-term judgment: the embodied intelligence industry will have two main waves of dividends in the future. The first is China's supply chain dividend. Robot demand-side is Chinese enterprises, and the supply chain is also in China, letting robot costs in the future drop below human labor costs. The second is the AI capability improvement dividend. With AI large model development, robots at the same price will have professional capabilities far exceeding any human at that price. The two dividends叠加 are the industry's true opportunity. Stardust Intelligence's current focus remains making the next-generation killer product. Lai Jie revealed Stardust Intelligence plans to launch a home-oriented robot next year, which he believes could become a hit because its price and capabilities will greatly break through current industry perceptions.
