文 | 智见Time，作者 | 李浩，编辑 | 经纬
10 月 14 日，在 Kosmo 发布会上，禾赛 CEO 李一帆提出希望禾赛成为机器人产业的底层技术供应商，类似博世之于汽车产业，甚至更进一步。
李老板的这番话，细品之下颇有别样意味。让禾赛今天发光发热的，正是博世当年没有做出来的激光雷达。
但是，是一个领域的禾赛，还是多个领域的禾赛，这也是禾赛与博世的区别，这也是禾赛想在未来在具身智能上想要实现的目标。
所以，此次禾赛发布的三维世界数据引擎 Kosmo，包括 AI 空间相机 Kosmo One 和云平台 Kosmo World。禾赛试图以新能源激光雷达上的技术积累跨界突破，以占领具身智能真实空间采集、三维重建和 AI 训练数据等业务。
激光雷达已经证明了禾赛的实力，空间数据业务让禾赛既有的感知技术、软硬件协同能力和工程经验获得新的应用空间。虽然机器人行业更倾向于自建模型、数据和仿真体系，但这并不意味着没有第三方供应商的位置。
恰恰相反，随着机器人走向规模化应用，专业分工或许会成为行业发展的必然选择。
诚然，技术积累固然有用，但产品线变多，不等于护城河变深。博世之所以能够成为汽车产业的巨头，靠的不仅是把零部件装进同一个产品目录，更是在不同领域建立起技术优势、客户关系和规模化交付能力。
但是禾赛的第一步已经踏出，而种下一棵树最好的时机是十年前，其次是现在。
不止激光雷达
当然，没有人否认禾赛在激光雷达上的成功。
2026 年第二季度，禾赛交付了 62.83 万台激光雷达。其中，ADAS 激光雷达 48.59 万台，同比增长 60.1%；机器人激光雷达 14.24 万台，同比增长 193.4%。
机器人业务的增速相当漂亮，但从绝对数量看，汽车仍然是禾赛的基本盘。超过四分之三的出货量，依然来自 ADAS 激光雷达。
当然这并不是一个坏现象，新业务总要花时间增长，而汽车激光雷达销量的增长又证明了原有业务的稳固。不过汽车激光雷达的市场空间终究有其边界。随着产品逐渐普及，规模增长之外，价格竞争、成本控制和客户议价能力也会越来越重要。
对于已经具备研发和量产能力的禾赛来说，寻找更多能够发挥这些优势的市场，自然成为下一阶段的课题。
于是，禾赛的目光很自然地转向了具身智能。
发布会上，禾赛宣布与银河通用机器人、原力灵机达成战略合作，同时展示了与皮克斯旗下 RenderMan、高德等企业的应用案例。虽然李一帆在回答记者提问时并没有给出 Kosmo 具体的商业化预期数据，但在二季度财报中禾赛预计 Kosmo 从 2026 年第三季度开始贡献收入。同时，禾赛将战略增长业务的全年收入指引上调至 2 亿至 3 亿元，预计 2027 年相关业务收入达到约 1 亿美元并实现盈亏平衡。
其实，比商业化更需要关注的是，在发布会上，高德宣布将与禾赛开放首批 1000 个线下商业三维数字化试点，覆盖商场、餐饮、酒店等场景。据悉，首批完成三维重建的重美术馆、偏锋画廊，已经上线高德 App 的“飞行街景”。
如果仅是场景扫描，让 Kosmo One 与普通手持激光雷达抢饭吃难免大材小用。
但倘若禾赛与高德可以深度合作，依托高德已有的商家场景，禾赛有机会将空间采集从单个项目向更多商场、餐饮、酒店延伸，为 Kosmo World 提供更多宝贵的真实世界场景，甚至是长尾场景数据。
Kosmo 走出了第一步，而真正值得期待的，是禾赛能否借此将零散的空间采集，变成一门可以规模化复制的数据生意。
Kosmo 和它的对手们
能在八仙过海的具身智能市场打响名号，禾赛的技术水平自然不必多言。
这次发布的产品中，Kosmo One 是一台融合了激光雷达、8K 影像系统、CMOS和高精度 IMU 的空间采集设备，可以记录环境中的几何形状、颜色、纹理和设备位姿。禾赛称，普通室内场景中，单台设备约 15 分钟即可完成 100 平方米空间扫描。
采集完成后，Kosmo World 负责将数据重建为三维场景，并能够生成如侧重视觉呈现的 3DGS、侧重物体的三维表面的 mesh、显示场景中物体距离的 depth 数据。这些数据可以进入机器人仿真和训练流程，成为数字训练环境的一部分。
发布会上，禾赛在举了上海一家 DQ 冰淇淋门店的案例：过去人工建模需要专业团队投入约 300 个小时，使用 Kosmo 后扫描仅需约 15 分钟。禾赛还称，在这一部署中，Sharpa 机器人的任务成功率达到 99%，失败率降至此前的十分之一。
当然，禾赛也并不能高枕无忧。
重建一家门店的数字模型，并不等于机器人已经学会了在门店里工作。更麻烦的是，机器人训练还涉及动作轨迹、物体物理属性、接触力、传感器适配和仿真验证等环节。把现实空间扫描下来，只是数据生产流程中的一部分。
并且 Kosmo 在射程内也并非没有对手。
比如英伟达就提供了底层工具链。NuRec 可以重建真实场景，Isaac Sim 负责搭建物理仿真环境，Isaac Lab 支持机器人策略训练。
智元的 Genie Sim 3.0 同样支持真实场景三维重建，并将场景生成、物理仿真、合成数据采集和自动评测整合进一套流程。
值得注意的是，智元并没有自行开发所有环节。数据采集方面，智元使用 Skyland Innovation 的 MetaCam 手持式三维激光扫描仪采集真实场景，再通过Genie Sim 完成后续处理，并将能力延伸至仿真训练和评测。
这恰恰也说明了机器人行业的专业化分工，反而为第三方供应商留下了机会。
即便是头部企业，也没有必要从扫描设备到仿真平台全部自行开发。随着机器人训练对真实场景数据的需求增加，专业供应商仍有机会凭借技术优势占据产业链的一环。
对于禾赛而言，竞争的关键更多的是要证明 Kosmo 的一体化方案能够减少不同环节之间的适配成本，或者在数据质量、处理效率上建立优势。只要能够解决足够具体的问题，即便不掌握整条技术链，也有机会成为其中不可或缺的一环。
15 分钟扫描完一家冰激凌店确实开了个好头，更值得期待的还在后面。
博世之行，始于跬步
其实除 Kosmo 之外，禾赛还有另一张牌。
2026 年第二季度，禾赛的动力模组已经开始贡献收入，累计交付超过 1 万套。禾赛称，相较于同类产品，其动力模组体积缩小 37%，扭矩密度和功率密度均达到约 3 倍，预计2027 年出货量将达到六位数。
激光雷达负责感知环境，动力模组负责执行动作。禾赛的产品开始覆盖机器人从感知到运动的不同环节，也让其供应商角色有了进一步拓展的可能。
从激光雷达到空间数据，禾赛能够延续部分感知、标定和三维重建方面的积累；动力模组虽然属于另一条技术路线，却仍然可以借助已有的精密制造、供应链管理和质量控制经验。
这些能力未必能让新业务立即获得订单，却有机会缩短产品研发和量产的周期。对于一家试图同时进入多个机器人零部件市场的企业来说，这本身就是一项优势。
这让人很容易联想到汽车零部件巨头博世。
同时卖很多种零部件只是博世业务的表象，博世的过人之在于不同业务之间能够相互成就，进而形成稳定的订单和规模化交付。
比如，博世的制动、转向、动力系统既可以独立销售，也可以通过统一的软件和控制系统组成完整方案。客户采购的产品越多，系统集成和协同开发的空间就越大，博世也就有机会在客户供应链中占据更重要的位置。
当下禾赛在机器人上的两条业务线之间并非天然协同。禾赛在激光雷达上已经建立了这样的能力，但空间数据平台和动力模组仍处于不同的商业化阶段，客户群、采购逻辑和竞争对手也各不相同。
机器人公司采购 Kosmo 与采购动力模组需求完全不同。两条产品线虽同出于禾赛，但这并不能代表着二者可以共享客户，更不代表能够共享利润。
但好在万事开头难，拼图要一块块补，好在激光雷达上，禾赛已经充分证明了自己的技术实力。
毕竟，罗马不是一天建成的。
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。
I'm sorry.
On October 14, at the Kosmo publish meeting, CEO Lee Zihang offered to become a bottom-up technology supplier to the robotic industry, similar to the auto industry, even further.
Mr. Lee's words mean something different. It's the laser radar that made the game glow today, which Bo did not produce.
But it is a field-based or multi-field-based game, and it is the difference between a game and a world-based game, and it is the goal that it wants to achieve in the future in terms of its intelligence.
So, Kosmo, the three-dimensional world data engine of publish, includes the AI space camera Kosmo One and the cloud platform Kosmo World. The Walls are trying to build a cross-border breakthrough with new energy laser radar technology to capture smart real space collection, 3D reconstruction and AI training data.
Laser radars have proven the strength of the game, and space data operations have given new applications to the identified sensory technology, software and hardware synergetic capabilities and engineering experience of the game. While the robotic industry prefers self-building models, data and simulation systems, this does not mean that there is no location of third-party suppliers.
On the contrary, as robots move to scale applications, professional division of labour may become the inevitable option for industry development.
It is true that while technology accumulation is useful, the growing number of product lines does not amount to deeper moats. Bo was able to become a giant in the automobile industry not only by incorporating parts and components into the same product catalogue, but also by building technological advantages, customer relationships and economies of scale.
But the first step in the game has already taken place, and the best time to plant the next tree is 10 years ago, and then now.
More than laser radar.
Of course, no one denies the success of the warp on laser radar.
In the second quarter of 2026, 62.83 million laser radars were delivered for the event. Of these, ADAS laser radars, 48.59 million, an increase of 60.1 per cent over the same period; robotic laser radars, 14.24 million, an increase of 193.4 per cent over the same period.
The growth in robotics has been quite impressive, but in absolute numbers, cars are still the base plate of the game. More than three quarters of the deliveries still come from ADAS laser radar.
This, of course, is not a bad phenomenon, as new operations take time to grow, while the growth in sales of car laser radar proves that the original operation is solid. However, the market space for car laser radar ultimately has its borders. As products become more widespread and larger, price competition, cost control and customer bargaining power will become increasingly important.
The search for more markets that can take advantage of these advantages is naturally the subject of the next phase for those that already have the capacity to develop and produce.
And so, naturally, the eyes of the game turned to intelligence.
At the publish meeting, a strategic collaboration was announced with the Galactic General Robotics, the Elixir, and examples were shown of applications with companies such as Renderman and Gautt. Although Li Yi Fan did not provide specific commercialization projections for Kosmo in response to a journalist’s question, in the second quarter of the fiscal year, the Volkswagen expected that Kosmo would start contributing revenue from the third quarter of 2026. At the same time, the Round has increased the annual revenue guidelines for strategic growth operations to between $200 million and $300 million, and the related business income is expected to reach approximately $100 million in 2027 and achieve a balance of gains and losses.
In fact, more than commercialization needs to be concerned, at the publish conference, Gold announced the opening of the first 1,000 sub-line commercial 3D digital pilots to cover malls, restaurants, hotels, etc. It is known that the first three-dimensional reconstruction of the first three-dimensional art gallery, the Pioneer Gallery, has been placed on the “Flying Street View” of Goth App.
If it's just a scene scan, it's easy for Kosmo One to eat with ordinary hand-held laser radar.
But if it and Gord can work together in depth, and if it is the business scene already in Togold, it has the opportunity to expand space from individual projects to more malls, restaurants, hotels, and to provide Kosmo World with more valuable real world scenes, even long tails.
Kosmo has taken the first step, and what is really to be hoped is whether it will be possible to turn a fragmented space collection into a data business that can be replicated on a scale.
Kosmo and his rivals.
It is natural to say that the technology of the game is not necessary to be famous in a smart market that crosses the sea.
Of the product of publish this time, Kosmo One is a space collection device that integrates laser radar, 8K image system, CCOS and high-precision IMU to record geometry, colour, texture and equipment position in the environment. Worm says that in the normal indoor scene, a single station will be able to complete 100 square metres of space scanning in approximately 15 minutes.
After the collection is completed, Kosmo World is responsible for rebuilding the data into a three-dimensional scene and is able to generate data such as 3DGS focused on visual presentation, mesh focused on the three-dimensional surface of the object, depth data showing the distance of the object in the scene. These data can enter robotic simulation and training processes and become part of the digital training environment.
At the publish meeting, the event brought up the case of a DQ ice cream door shop in Shanghai: in the past, manual modelling required a professional team input of about 300 hours, and after using Kosmo, the scanning took about 15 minutes. It also claimed that in this deployment, the Sharpa robotic task success rate had reached 99 per cent and that the failure rate had dropped to one tenth before.
Of course, the game is not safe.
Rebuilding a digital model of a store does not mean that robots have learned to work in the store. More problematic is the fact that robotic training also involves movement trajectory, physical properties of objects, exposure, sensor suitability and simulation. Scanning the real space is only part of the data production process.
And Kosmo is not without an opponent in range.
For example, we have a bottom-up tool chain. NuRec can rebuild the real scene, Isaac Sim builds the physical simulation environment, Isaac Lab supports robot strategy training.
Genie Sim 3.0 of Jimoto also supports three-dimensional reconstruction of the real scene and integrates scene generation, physical simulation, synthetic data collection and automatic evaluation into a set of processes.
It is noteworthy that Ji-won did not develop all the links himself. With regard to data collection, Jigen used Skyland Innovation ' s MetaCam handheld 3D laser scanner to collect real scenes, then completed follow-up processing through Genie Sim and extended capabilities to simulation training and evaluation.
This is also an indication of the specialization in the robotic industry, which has left opportunities for third-party suppliers.
Even for the head enterprise, there is no need to develop all of its own equipment, from scanning to simulation. As the demand for real scene data for robotic training increases, professional suppliers continue to have the opportunity to dominate the chain of industries with technological advantages.
The key to competition is more to prove that Kosmo's integration programmes can reduce the matching costs between the different links or create advantages in data quality, processing efficiency. There is an opportunity to be an indispensable link, even without mastery of the entire technological chain, provided that the problem is sufficiently specific.
A 15-minute scan of an ice cream shop did open a good start, and much more to be expected.
The journey of Bo began with the steps.
There's actually another card in the game besides Kosmo.
In the second quarter of 2026, the power module of the ring began to contribute to revenue, delivering more than 10,000 units cumulatively. Worm said that its power module mass was reduced by 37 per cent compared to that of the same product, with a twist density and power density of about 3 times, and a six-digit delivery expected in 2027.
The laser radar is responsible for the sensory environment and the power module is responsible for the operation. The products of the game begin to cover different parts of the robot from perception to movement, and offer the possibility of further expansion of its supplier role.
From lasers to spatial data, the wharf can continue the accumulation of partial sensory, calibration and three-dimensional reconstruction; although the power module is a different technical route, it can still draw on existing precision manufacturing, supply chain management and quality control experience.
These capabilities do not necessarily allow new businesses to obtain immediate orders, but offer the opportunity to reduce product development and volume production cycles. This is in itself an advantage for an enterprise that seeks to access multiple markets for robotic parts at the same time.
It's easy to think about auto parts and components.
At the same time, the sale of many parts and components is only a sign of the Boshi business, which has passed on to the ability of different businesses to achieve each other, leading to steady orders and deliveries.
For example, Boven's brakes, diversions, power systems can be sold independently, or they can form a complete programme through unified software and control systems. The more products the customer purchases, the greater the space for system integration and co-development, and the opportunity for Boven to take a more prominent place in the customer supply chain.
There is no natural synergy between the two lines of business on robots at the present time. This capability has already been built on laser radar, but spatial data platforms and power modules are still at different stages of commercialization, with different customer groups, procurement logic and competitors.
Robotics purchase Kosmo in a completely different way from purchasing power models. The fact that the two product lines are in the same race does not mean that the two can share customers, let alone profits.
But it's good to start with a piece of puzzle, and it's good to be on laser radar, and the game has proved its technical strength.
After all, Rome is not built in one day.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.