For robots to enter more factories and stores, they first need sufficiently rich training scenarios. Hesai is attempting to turn the work of preparing these scenarios into a new business.
On October 14, Hesai released the Kosmo spatial camera and cloud platform, which scan real environments and reconstruct three-dimensional spaces to provide robots with scenarios usable for simulation training, and signed strategic cooperation agreements with Galbot and Yuanli Lingji. For a company that primarily generates revenue by delivering sensors, this means extending its business into data and cloud services for robot development.
This shift corresponds to the real demands of the embodied intelligence industry. Intelligent vehicles can already rely on large-scale fleets to accumulate data during daily driving; robots have not yet widely entered production and daily life, yet they need enough environmental and operational experience to be competent at real work. Development teams must not only teach robots to perform actions, but also prepare places for them to practice those actions.
The opportunity Hesai sees is to apply the three-dimensional perception, chip design, and precision manufacturing capabilities accumulated in the automotive industry to reduce the cost of building training environments. Revenue sources are also expected to expand accordingly: beyond selling equipment, they include subsequent data production, spatial reconstruction, and cloud services.
As robots advance toward practical applications, work such as collecting data, building environments, and evaluating training results is generating clearer procurement demand. For the embodied intelligence industry, the entry of specialized suppliers offers a chance to reduce the duplicate development that individual robot manufacturers each bear, allowing R&D resources to be invested more in specific tasks. What Hesai is striving for is precisely its position in this round of industrial division of labor.
The cost accounting within the training ground
Robots lack data 鈥?but what exactly are they missing?
In cars, roads, vehicles, and pedestrians constitute a relatively well-defined task environment. After a vehicle is delivered, the user's daily driving can continuously generate real scenarios, based on which developers identify problems and improve the system.
The work robots face is far more dispersed, such as picking up goods, organizing shelves, and operating equipment, with different tasks involving different objects, actions, and contact methods.
In the case of store scenarios, an action learned in one store may fail in another due to differences in counter height, item placement, or lighting changes. For robots to adapt to these differences, developers need extensive operational experience and sufficiently rich environments.
Hesai Technology co-founder and CEO Li Yifan believes that autonomous driving can rely on vehicles driving on roads to obtain data, but robots do not yet have the same conditions. This puts the industry in a mutually constraining problem: robots need more data to be deployed, yet the lack of real-world deployment limits data accumulation.
Therefore, data production is becoming an independent business in the embodied intelligence industry.
In August of this year, Lightwheel partnered with Hugging Face to release a first-person human activity dataset, planning to open 100,000 hours of data, with the first batch of 10,000 hours going online, extracting behavioral experience from human operations. Nvidia, meanwhile, launched a physical AI data factory blueprint this year, organizing data generation, augmentation, and evaluation into a unified process. Companies are improving the production efficiency of robot learning experience from different segments.
These paths correspond to different data needs. Human operation videos provide task steps and behavioral demonstrations; real-machine collection records the robot's own actions and feedback; simulation facilitates scaling up training. Developers need to use them for the same task, and the spatial model is only one part of it.
What Hesai is entering is the training environment.
According to how Kosmo works, the collection device first records the geometric structure, color, and texture of a space, and the cloud platform then reconstructs it into a three-dimensional scene and separately identifies objects such as tables, chairs, and merchandise. Developers can change object positions and adjust scene layouts, allowing robots to practice the same task under different conditions.
This process changes the way real scenarios are used. In a real store, robot testing requires coordinating the venue, personnel, and business arrangements; once the store is digitized, it can become a training environment that is repeatedly called upon, modified, and expanded. Work that in reality requires personnel to move items and rearrange them can partly be completed first in the computer.
If every time a new location is entered, it must be re-measured, re-modeled, and re-debugged, projects easily depend on engineers delivering store by store. Only when scene production can be reused is it possible to reduce this repetitive labor, so that real-machine testing is used more to verify action results and find causes of failure.
Hesai claims that in ordinary indoor scenarios, scanning a 100-square-meter space can be completed in about 15 minutes, after which the cloud platform processes it into a three-dimensional model. Compared with scanning speed itself, what matters more is reducing the degree of professional involvement, making the collection of new scenarios a routine part of the development process.
Spatial models also need to work with action data and physical simulation. Spatial models describe where objects are and how the environment is arranged; action data helps robots learn how to grasp and apply force; physical simulation handles interaction processes such as contact and friction. The contact and friction simulation provided by Nvidia's Newton physics engine belongs to this latter part.
Whether collection formats, object annotations, and simulation interfaces are compatible will directly affect developers' integration costs. If after procurement they still need to invest a large amount of manual effort to convert formats and process scenes, the time saved by rapid scanning may not necessarily translate into efficiency for the entire training workflow.
Building the business before mass production
LiDAR is a complex optical, mechanical, and electrical system in which chips, optical components, mechanical structures, and software must work together. The automotive industry further requires it to maintain consistent performance in mass production and adapt to long-term use. Whether a solution can be mass-produced depends on whether design, manufacturing, testing, and the supply chain can operate together.
In the second quarter of 2026, Hesai's LiDAR deliveries were approximately 628,000 units, of which ADAS products were approximately 486,000 units. Sustained large-volume deliveries have accumulated manufacturing and verification experience for the company and also given its early investments, such as self-developed chips, a larger amortization base.
This experience can be transferred to spatial collection equipment, but the product still needs to be redesigned around new uses.
Hesai Technology co-founder and CTO Xiang Shaoqing introduced that Kosmo uses customized LiDAR and specifically addresses the issue of image consistency across multiple cameras at different times and viewing angles.
Another piece of experience brought by the industry is translating requirements into components and manufacturing processes.
Li Yifan believes that the advantage of Hesai's self-developed chips comes from an understanding of specific applications: general-purpose components cover a wide range of needs, while specialized designs can make trade-offs around specific tasks. This approach is now being applied to robot hardware and spatial collection.
In the automotive business, automakers and autonomous driving companies set metrics, and Hesai completes development around those metrics. Making robot training tools, however, requires tracing requirements back to the training and operation process, figuring out what the device should collect and how data should be output.
Sharpa provides a direct feedback channel. Li Yifan introduced that Hesai's three co-founders are also the founders of Sharpa, and the latter's needs in robot operation can be fed back to Hesai relatively quickly and then verified through practical application. This kind of collaboration connects hardware development with training results, while cooperation with companies such as Galbot and Yuanli Lingji further expands the sources of demand.
As a result, what Hesai delivers has also expanded from equipment to subsequent processes. Wallstreetcn has learned that Hesai has proposed two business models: providing integrated software and hardware products to industry customers, and providing spatial data production services to world model and physical AI companies. Its second-quarter financial report already listed recurring cloud service revenue as a development direction for Kosmo.
In the second quarter of this year, Hesai's total deliveries increased by 78.4% year over year, and revenue increased by 21.9% year over year. The company said in its financial report that the decline in average selling price partly offset the boost to revenue from delivery growth.
This opens up a new revenue logic for automotive suppliers. Component orders mainly follow overall machine production, while spatial data services are linked to customers' development and operating activities. New stores need to be collected, layout changes need to be updated, model iterations need supplementary training environments, and after equipment is sold, service demand may continue to arise.
Li Yifan also revealed that Kosmo will also be sold to consumers in the future. This extends its potential users from corporate R&D teams to individuals. If spatial collection is to become a more general tool, price, ease of operation, and subsequent editing experience will become more important, and Hesai will also need to build product and service capabilities for more dispersed users.
Connecting with existing software platforms is one path to expanding usage. Kosmo has already integrated with Isaac Sim, Isaac Lab, and film and television creation tools, allowing customers to bring collection results into existing workflows. For users, the reason to buy the device is to reduce workload; for Hesai, recurring revenue needs to be built on the repeated use of services such as reconstruction, editing, and data management.
An embodied industry practitioner told Wallstreetcn that, from the perspective of robot deployment, the significance of training tools must ultimately be reflected in the overall deployment cost. Even if the price of the robot body falls, if every switch to a new store and every additional task requires a large number of engineers to re-model and re-debug, customers will still find it difficult to calculate a suitable return on investment. Standardizing the training environment is precisely about compressing this part of the work that repeatedly occurs with each project.
As specialized tools and services enter, embodied intelligence companies have the opportunity to devote more resources to task capabilities, and application-side users can also try robots at a lower onboarding cost.
By extending its business into the training ground this time, Hesai is participating in precisely this process 鈥?turning work originally scattered across various R&D teams into purchasable products and pushing robots from a small number of pilot projects toward more workplaces.
Comments