文 | 数智奔流，作者 | 白芨，编辑 | 沈校
数采设备里，藏着禾赛的估值转型期待。
10月14日，禾赛科技发布Kosmo One空间相机，同时发布了三维世界数据引擎Kosmo和Kosmo World云平台。
在这之前，禾赛在具身赛道的布局还有机器人整机厂Sharpa。它由禾赛的CEO李一帆、CTO向少卿和首席科学家孙恺于2024年底共同创立。双方签订有《产品销售框架协议》，禾赛向Sharpa供应机器人动力模组。
多线布局在具身行业不稀奇，智元先后布局了觅蜂科技、启元等多个品牌，覆盖产业上下游。有具身产业人士认为：这种布局可以在局势不明朗的时候，拿到更多的机会。同时随着具身整机由热转冷，在产业链上下游寻找高价值、高景气度环节，也可以拿到更多的资金机会。
毕竟如果整机不好卖了，数采这种上游产业链，世界模型也少不了。那给数采商提供采集设备，必然也是一门好生意。
具身行业仍然很热，但热度开始不均匀分布。
2026年上半年，国内具身智能赛道融资约438亿元。其中，超过一半流向了做“大脑”和世界模型的公司，做机器人本体的公司融资占比约12.8%。
资本没有离开具身智能，但正在重新挑选下注对象。
此前最容易获得关注的是整机、双足、人形和现场演示；现在，模型、数据、仿真、核心零部件和商业化交付，开始获得更高估值。
有具身一级市场投资者调侃：除了整机不好卖，具身全身都是宝。脑子是世界模型，眼睛是激光雷达，手脚是高价值精工制造，皮肤是触觉材料学，就连要吃的饭也能撑起数采产业基地和数采设备商。
既然攒起来不好弄，先拆开单独融资拿钱找业务活下去，之后看情况再根据所处环节反攻具身整机也是最优解。
毕竟条条大路通罗马，说不定再过几个月散伙的具身饭才是最香的。
这可能也是禾赛手攥超200亿估值的Sharpa，还要再生一个数采二胎的原因。毕竟大势所趋下，多生孩子好套狼。
这个禾赛的具身二胎Kosmo One是一台手持式三维空间数据采集设备，集成激光雷达、8K影像系统和高精度IMU，负责采集真实物理空间；Kosmo World负责重建和处理，将原始数据转化为3DGS、Mesh和Depth三种空间对齐的数据格式。
禾赛披露，普通室内场景约15分钟可以完成100平方米扫描。这台相机的原始数据进入Kosmo World云平台，由底层系统Kosmo Spatial Core处理成三种同源数据格式，用于模型训练。
买下硬件的同时，客户就进入了禾赛的数据管线。
禾赛的判断是，具身智能落地需要超高精度物理世界坐标来抵消幻觉。
按照可规模化程度与机器人对齐程度划分，北京大学、南洋理工大学等机构合作发布的报告指出，当前具身智能数据大致可以分为五个层级。
第一层是使用互联网上的通用视觉语言数据，数量最大，价格最低，主要解决物体识别、语义理解和基础视觉能力。第二层是仿真数据，依靠三维资产、物理引擎、自动化脚本和遥操作生成，成本较低，适合规模化扩展。
第三层是人类第一视角视频。采集员戴上头戴设备或摄像头，在真实环境中完成操作，记录人与物体的交互过程；第四层是人类佩戴夹爪等设备，采集第一视角、手部动作、位姿和操作轨迹等数据；第五层是真机数据，直接来自机器人在真实环境中的操作，质量最高，成本也最昂贵。
Kosmo进入的，是第二层仿真数据的环境底座。通过高效率扫描，把现实世界的物理空间转化为数字资产，用于下游的仿真训练。
仿真训练的完整流程是重建三维场景，并赋予质量、碰撞体积等物理属性，再建立机器人本体模型，配置传感器、碰撞体和物理参数，生成带标注的训练数据。
禾赛正在为场景构建和数据生成提供三维采集和重建数据，扫描的是空间结构、颜色、纹理、深度和设备位姿。可以解决机器人应该在哪里训练，训练环境的墙面和地面分别是什么样。
但物理属性数据，需要客户在导入后接入Isaac Sim、Isaac Lab或其他平台，自行配置。
这与高精地图商在自动驾驶时代做过的事相近，不生产产品本体，是为这些智能体提供理解物理世界所必需的基础空间数据层，禾赛则是链条上游的数据供给商。
禾赛的野心藏在商业模式里。
车载激光雷达的商业模式是典型的硬件销售，单价一路下滑，至今仍在降价；高精地图的商业模式则在从一次性卖图，向按年服务费或动态服务底座转型。
在发布会前，禾赛CEO李一帆表示，我们不是为了卖激光雷达又造了一个Kosmo。而是要从硬件收入起步，转向数据与模型服务，最终构建包含海量空间模型的数字资产平台，供客户付费使用。
按照禾赛的规划，这一资产库的形成可能需要3至5年。一旦建成，禾赛就能摆脱硬件制造商身份，切入数据资产商赛道，完成估值身份的跃升。
在DQ冰淇淋门店案例中，禾赛讲述了一个智能取代人工的故事：
“以Sharpa机器人在DQ冰淇淋门店的应用为例：过去人工建模还原一家门店，专业团队需投入约300个小时；Kosmo约15分钟即可完成扫描并生成高度还原的三维训练环境，Sharpa机器人的失败案例降至原来的十分之一，任务成功率提升至99%，最终在真实门店正式上岗。”
但案例没有展开的是，Sharpa由禾赛的CEO李一帆、CTO向少卿和首席科学家孙恺于2024年底共同创立。双方签订有《产品销售框架协议》，禾赛向Sharpa供应机器人动力模组。
案例本身强调的是Kosmo扫描即开即扫、方便好用、准确度更高，但对第三方客户来说，生成后的空间环境要导入仿真器，再导入机器人模型，经过一系列训练并进行优化，这需要大量人工标注和参数调试。
对机器人公司来说，一家门店是否真实还原，不能只看画面像不像。
机器人要在店里搬运、交互，要知道地面是否打滑、柜台承重多少、冰淇淋机按钮需要多大压力、物体移动时会产生怎样的摩擦和形变。空间扫描可以准确记录几何关系，却不会自动生成完整的质量分布和接触模型。
生成式路线则直接挑战了Kosmo的硬件扫描逻辑。
李飞飞团队与英伟达等联合发布的SimFoundry，只需一段真实世界的普通视频，就能自动生成一个具备交互能力的机器人仿真环境，实现零样本迁移。
智元机器人的Genie Sim 3.0能利用大语言模型驱动场景生成，输入自然语言指令可泛化出万级训练场景，并开源了包含上万小时仿真场景的数据集。
相较生成方案，扫描方案的软肋在于成本。每增加一个新场景，硬件、人力加上物理数据补全都要跟上，相较生成方案单次投入长期复用的模式，刚性成本更高。
对禾赛而言，Kosmo的竞争力不在于重建场景技术，而在于扫描设备的技术套装。一台非专业人员也能使用的手持测绘设备，一次重建，同步产出视觉外观、几何网格和深度信息，并通过语义分割识别场景中的物体。
多个环节打包成开箱即用的系统，这是Kosmo的优势所在。而这一切的前提，是客户真的刚需高精度真实扫描数据
眼下，禾赛正在同时推进四条业务线：汽车激光雷达、机器人激光雷达、Kosmo空间数据引擎和机器人动力模组。
这是一条从感知到理解再到执行的路线。公司希望把自己从汽车激光雷达供应商，升级为物理AI基础设施企业。发布Kosmo时，李一帆提出希望禾赛成为机器人产业的底层技术供应商，参照博世在汽车产业中的位置。
对比同样切入机器人的激光雷达企业，速腾聚创做机器人视觉平台，把传感器嵌入客户的开发工具链；图达通仍以高性能激光雷达为核心，布局1550nm和905/940nm多条产品路线，并将机器人、工业自动化等作为汽车之外的应用方向。览沃的Mid-360面向低速移动机器人，主打360度感知、导航和避障，产品强调低成本和易集成。
而禾赛进入机器人市场的第一步，仍然是卖雷达。
2026年春晚上，宇树的演出机器人使用了禾赛的迷你型3D激光雷达JT128，用于完成三维SLAM、定位和导航感知。2026年第二季度，禾赛机器人激光雷达出货量同比增长193.4%，机器人已成为禾赛出货结构中的增量源。
但出货量增长，不等于禾赛掌握机器人系统。
博世模式的核心是系统集成商。车企不必从零整合底层硬件，博世交付经过验证、接口标准化的系统模块，车企在上面做整车集成和差异化。
背后逻辑是，汽车行业有百年的标准化接口与分工体系，买现成系统比自己攒更高效。
机器人厂商采购JT128时，通常把它当成感知输入。机器人的规划、动作策略和数据闭环仍然掌握在本体厂商手里，禾赛产品要和一系列软硬件协同工作。
这意味着，车企绕开博世，要切换一整套系统；本体厂绕开禾赛，可以选非激光雷达方案，可以选竞品切换供应商，甚至可以自研。
对标汽车供应链的成熟分工，具身智能面临的是高度碎片化的供应链，本体厂的减速器、电机、驱动器、传感器和控制板卡往往分别采购，拿到零件后，还要自己解决热管理、振动、接口、标定、精度衰减和安全诊断。
这是禾赛的矛盾所在，公司希望借助激光雷达、Kosmo和动力模组成为机器人的一级供应商，CEO李一帆也明确表示，禾赛的定位更接近平台型Tier 1供应商，不止售卖硬件，还包含软件、服务等多元形态。
但机器人本体厂商仍然掌握数据管线和控制闭环，拥有绕开单一供应商自建感知融合与场景建模的能力，被客户替换的成本并不高。
Kosmo要成为基础设施，仍需继续向后延伸。空间采集之后，要接入传感器标定、视觉融合、SLAM、仿真格式转换、动作数据、任务评估和部署反馈。
禾赛不必包办所有环节，但必须让客户在使用其产品时少做一轮工程适配，证明自己能交付让客户快速迭代的系统平台。
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。
♪ We've got a lot of ideas ♪
In the mining equipment, there is an expectation of a valuation transformation of the game.
On 14 October, the publish Kosmo One space camera, which is the technology of the Worm race, together with publish's three-dimensional World Data Engine Kosmo and Kosmo World Cloud platform, was installed.
Before that, the walrus was set on the track and the robot complex, Sharpa. It was created in late 2024 by CEO Lee Yi Fan, CTO to Xiaoqing and Sun Kei, Chief Scientist. The two parties have signed a framework agreement on the sale of products, and the swings supply a robotic power module to Sharpa.
Multi-wire layouts are not uncommon in the physical sector, and the simulator has successively set up various brands, such as bee-hunting technology and the initials, covering both upstream and downstream industries. It has been argued by some industry that such a configuration would provide more opportunities in times of uncertainty. Increased access to finance can also be obtained as the whole machine is turned from heat to cold and high-value, high-spectrum links are found up and down the industrial chain.
After all, if the whole machine is not sold properly, and if the upstream chain is taken, the world model is not missing. It would certainly be a good business to provide collectors with gathering equipment.
The body industry is still hot, but the heat is beginning to spread unevenly.
In the first half of 2026, the country ' s smart track was financed by approximately $43.8 billion. More than half of them go to companies that do “brain” and world models, and corporate financing for robotics accounts for about 12.8 per cent.
Capital did not leave a smart spot, but it was re-selecting the bet.
Prior to that, attention was most readily focused on integrator, double-foot, human and on-site demonstrations; now, models, data, simulations, core parts and commercial deliveries begin to obtain higher valuations.
There is a first-class investor in the market: it's full of treasures but it's hard to sell. The brain is a model of the world, the eyes are laser radar, the hands and feet are made of high-value precision, the skin is a tactile material, and even the food that is to be eaten supports the mining industry base and the equipment manufacturer.
Since it would be difficult to survive, it would be best to break down the separate financing and get the money to survive.
After all, there's a road leading to Rome, and it's probably the best meal that'll break up in a few months.
This may also be the reason why the sharpa, which is valued by more than $20 billion, is expected to produce a second child. After all, there's a lot going on.
Kosmo One, a two-dimensional hand-held three-dimensional space data acquisition device, integrated laser radar, 8K image system and high-precision IMU, is responsible for collecting real physical space; Kosmo World is responsible for rebuilding and processing raw data into 3DGS, Mesh and Devth spatially aligned data formats.
The event revealed that ordinary indoor scenes could complete 100 square metres of scanning in about 15 minutes. The original data from the camera entered the Kosmo World Cloud platform, which was processed by the bottom system Kosmo Spatial Core into three congener data formats for modelling training.
When hardware was purchased, the client entered the data pipeline of the wares.
The judgement of the Worms game is that a smart landing requires super-precision physical world coordinates to offset hallucinations.
The report of publish, a partnership between Beijing University and the Nanyang University of Technology, indicates that current intelligence data can be divided into roughly five levels.
The first tier is the use of universal visual language data on the Internet, with the largest volume and lowest price, which primarily addresses object recognition, semantic understanding and basic visual capability. The second level is simulation data, which are generated from three-dimensional assets, physical engines, automated scripts and remote operations at lower cost and suitable for scale expansion.
The third level is the first human perspective video. The collectors wear head-on equipment or cameras to carry out operations in the real environment and to record the interaction between humans and objects; the fourth level is for humans to wear equipment such as clamps and claws, to collect data such as first view, hand moves, position and trajectories; and the fifth level is for real data, derived directly from robotic operations in the real environment, of the highest quality and cost.
Kosmo entered the environmental base of the second-level simulation data. Efficient scanning transforms the physical space of the real world into digital assets for downstream simulation training.
The complete process of simulation training is the re-establishment of a three-dimensional scene and the giving of physical properties such as mass, impact volume, and the creation of robotic body models, the configuration of sensors, collision bodies and physical parameters, and the generation of tagged training data.
It is providing three-dimensional data collection and reconstruction for scenario construction and data generation, scanning space structures, colours, textures, depth and equipment position. It can solve where robots should be trained and what the walls and the floors of the environment are.
Physical property data, however, require the customer to access Isaac Sim, Isaac Lab or other platforms after import and to configure itself.
This is close to what high-precision maprs did in the age of auto-driving, without producing product essences, providing the basic spatial data layer necessary to understand the physical world of these AI agents, and the wharf is the data provider upstream of the chain.
The ambition of the game is hidden in the business model.
The commercial model of vehicle-borne laser radar is typical of hardware sales, with unit prices declining all the way down and still falling; the commercial model of high-quality maps is shifting from a one-time sale map to an annual service fee or dynamic service base.
Before publish, CEO Lee Yao said that we did not build another Kosmo to sell laser radar. Rather, it is to move from hardware revenue to data and modelling services, leading to the construction of a digital asset platform containing a big spatial model for use by client paid.
The creation of this asset pool may take three to five years, as planned for the event. Once completed, it will be possible to move away from the hardware manufacturer's identity, to the data asset dealer's track and complete the leap in valuation identity.
In the case of the DQ Ice-Eat Gate, the Wallshow tells the story of an intelligent replacement:
“For example, the use of Sharpa robots in DQ ice cream doors: In the past, manual modelling was returned to a front shop, with a professional team spending approximately 300 hours; Kosmo could have scanned and generated a highly reduced three-dimensional training environment in about 15 minutes, and the failure of the Sharpa robot fell to one tenth, and task success rate was upgraded to 99 per cent and finally officially placed at the real gate.”
But what did not happen was that Sharpa was created in late 2024 by CEO Lee Yi Fan, CTO to Xiaoqing and the Chief Scientist Sun. The two parties have signed a framework agreement on the sale of products, and the swings supply a robotic power module to Sharpa.
The case itself highlights the fact that Kosmo scans are open, user-friendly and more accurate, but for third-party customers, the created space environment involves the import of imitators, re-imports of robotic models, a series of trainings and optimizations, which require extensive manual marking and parameter debugging.
For robotics, whether a store is real or not, the image cannot be seen only.
The robot has to move and interact in the store, knowing whether the ground is sliding, how heavy the counter is, how much pressure the ice cream buttons require, and what frictions and shapes will occur when objects move. Space scanning can accurately record geometry, but does not automatically generate complete mass distribution and exposure models.
The generating route directly challenges the hardware scanning logic of Kosmo.
SimFoundry, a joint publish by the Li Fei Fei team and Young Weidar, can automatically generate an interactive robotic simulation environment and move to zero sample.
Genie Sim 3.0 of the intellectual robot drives the scene using a large-linguistic model, enters a natural-language instruction that can extend the class-of-the-horizontal training scene and opens up data sets containing 10,000 hours of simulation.
Compared to the generation programme, the soft side of the scanning programme is cost. For every new scene, hardware, manpower and physical data supplement are kept up, and the rigid cost is higher than the pattern of single-input re-use.
For the event, Kosmo ' s competitiveness lies not in the re-establishment of scene technology, but in the technical assembly of scanning equipment. A hand-held mapping device, also usable by non-professionals, was reconstructed once to synchronize the output of visual appearance, geometry grid and depth information and to identify objects in the scene by semantic partition.
The multiple links were packed into an open-box system, which was the advantage of Kosmo. And the premise of all this is that the client really just needs a high-precision real scan.
At the moment, the whaling is advancing four lines of operation simultaneously: car laser radar, robot laser radar, Kosmo spatial data engine and robotic power module.
It is a route from perception to understanding to implementation. The company wishes to upgrade itself from a car laser radar supplier to a physical AI infrastructure enterprise. publish Kosmo offered to become a bottom-up technology supplier for the robotic industry, with reference to Boshi ' s position in the automobile industry.
Compared to the same laser radar enterprise that cuts into robotics, the rapid build-up of robotic visual platforms, embedding sensors into the customer ' s development tool chain; Tudatuons still have high performance laser radar at its core, with a layout of 1550 nm and 905/940 nm multiple product routes and using robotics, industrial automation, etc. as an alternative to vehicles. The fermented Mid-360 targets low-speed mobile robots, with 360 degrees of perception, navigation and shielding, with products emphasizing low cost and easy integration.
And the first step into the robotics market is still to sell radar.
During the spring night of 2026, the Uyuki performance robot used the 3D mini-radio radar JT128 of the Worm to complete 3D SLAM, positioning and navigational awareness. In the second quarter of 2026, there was an increase of 193.4 per cent in the volume of cargo from the Hancai robot laser radar, which has become a source of increase in the structure of the Hancai race.
But the increase in output is not the same as the acquisition of robotic systems in the game.
At the heart of the Boven model is the system integrator. Without having to consolidate the bottom hardware from zero, Boshi delivered certified, interfaced, standardized system modules, on which the car was integrated and differentiated.
The logic behind this is that the automobile industry has a 100-year standardized interface and division of labour system, and buys off-the-shelf systems more efficiently than it owns.
The robot manufacturer usually treats it as a sensor input when purchasing JT128. The robot's planning, action strategy and data closure are still in the hands of the manufacturer, and a range of hardware and software are required to work with the produce of the game.
This means that the car is going around the world to switch to a whole system; the plant is going around the barn, choosing non-laser radar options, choosing to switch suppliers, or even self-study.
With respect to the mature division of labour between the target vehicle supply chain, the intelligent faces a highly fragmented supply chain, with the plant ' s own brakes, electrics, drives, sensors and control plates often purchased separately and, when spare parts are available, thermal management, vibrations, interfaces, marking, precision decay and security diagnostics.
This is the paradox of the game, which the company wants to make up as a first-class supplier of robots using laser radar, Kosmo and power models, as well as CEO Lee Yi Fan's clear indication that the game is located closer to the platform-type Tier 1 supplier, which not only sells hardware but also includes a variety of forms of software, services, etc.
However, robotics manufacturers still have access to data pipelines and closed loops, the ability to bypass a single supplier for self-sensor integration and scenario modelling, and the cost of being replaced by customers is low.
To become an infrastructure, Kosmo still needs to continue to extend. After space acquisition, access is required to sensor marking, visual integration, SLAM, simulation format conversion, action data, mission assessment and deployment feedback.
The game does not have to cover all the links, but it must allow clients to use their products less than a round of engineering and to prove that they can deliver a system platform that allows customers to evolve quickly.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.