The October launch was public. The commercial groundwork started in July

On October 14, 2026, Chinese lidar supplier Hesai launched Kosmo, a spatial data engine built around a combined lidar-and-camera device called Kosmo One and a cloud system called Kosmo World. The announced goal is to turn real rooms into editable 3D environments that robot developers can use for simulation and training. At first glance this looks like another sensor hardware announcement.

Hesai's interim report tells an earlier story. The company says it delivered initial Kosmo prototypes in July, had obtained orders from robotics companies including Galbot, and had engaged more than 200 prospective partners across industries. Those 200 contacts were not 200 proven paying customers or shipped projects. The same report projected initial revenues in the third quarter, without proving those revenues were later recognized.

This changes the meaning of the October presentation. It was not simply the first idea for an AI camera, but the public rollout of an emerging supplier business that had already begun prototypes and customer discussions. A supplier of perception devices is testing whether it can also supply spatial data as a product.

Three hundred hours versus fifteen minutes: a powerful comparison with boundaries

Hesai highlights a Dairy Queen ice-cream shop in Shanghai. According to the company, creating a 3D training environment there had required roughly 300 hours of professional manual modeling, while capture and reconstruction with Kosmo took about 15 minutes. Hesai also says one operator can scan a typical 100-square-meter indoor space in approximately 15 minutes.

The numbers cannot be treated as a universal 1,200-fold improvement in robot development. Manual modeling and scanning/reconstruction do not necessarily include the same physical setup, task annotation, robot policy training, testing and deployment activities. Independent evidence that the compared workflows produced identical outputs is not available. Site geometry, occlusions and accuracy requirements also affect capture time.

There is still a meaningful possible cost change. If a robot integrator must rebuild a new store or factory in simulation for each deployment, faster acquisition of usable scene geometry could shorten one stage of engineering. Any real productivity claim should include later editing, physical simulation setup and validation rather than the camera capture clock alone.

8K capture does not mean a finished physics-ready training world

Kosmo One integrates high-resolution lidar, a one-inch image sensor and an inertial measurement unit, aiming to capture geometry, color, texture and device movement in a synchronized stream. The company advertises a maximum 8K capture resolution while specifying that users can export files at up to 6K. Capture capability and delivered asset specifications are different facts.

Kosmo World uses 3D Gaussian Splatting, mesh and depth outputs in one reconstruction. It can fill unscanned areas, segment objects and adapt viewpoints to a robot's camera. The company lists integrations with NVIDIA Isaac Sim and Isaac Lab, Unreal Engine, RenderMan and other tools; these integrations matter because developers already work within their own simulation pipelines.

Yet photorealistic rooms and physically reliable environments are not identical. A robot picking up a cup or handling food needs contact, mass, friction and collision behavior, not just convincing pixels. Those properties need separate modeling or verification. A customer should test what is actually reusable for action training, not only whether the 3D scene looks correct.

Two announced business models: hardware software packages and spatial data as a service

Hesai's October release explicitly lays out two models. The first sells integrated camera and software solutions that robotics, media and other customers use themselves. The second offers services to produce high-quality real-world spatial data at scale for developers of world models and physical AI systems.

The economic logic is different. Device sales depend in part on shipment volumes and product prices. Spatial-data services may depend on the number of sites, frequency of updates, processing requirements, rights of use and cloud consumption. Hesai's interim report explicitly discusses a potential flywheel linking hardware adoption, recurring cloud-service revenue and growing platform value.

That loop is an ambition, not established financial performance. During the October launch, the CEO declined to give a revenue forecast for Kosmo. The interim report's expected initial third-quarter revenue should not be described as a realized number without subsequent results. The market still has to demonstrate repeat purchases and defensible margins.

The same captured location could serve robots, filmmakers and map platforms

In robotics, Hesai announced collaborations with Galbot and Dexmal. It also showed workflows using Pixar's RenderMan and work with Chinese mapping platform Amap to reconstruct real galleries and public spaces. These use cases have a common starting point: capture a real environment as structured, reusable data.

Their requirements diverge after that step. A filmmaker may value visually controllable lighting and materials. A robot integrator needs trustworthy object geometry and interaction physics. A map company must consider location freshness, access rights and privacy as well as presentation. One source scan does not automatically pass each industry's quality test.

Broadening the customer set also changes what a data supplier has to manage. Versioning, formats, permission to collect and reuse a location, integration and update cadence can matter as much as camera performance. Whether a dataset can be reused commercially depends on contracts and rights, not simply on the technology's ability to export an asset.

A 99% ice-cream task claim and more than 10,000 actuator deliveries are different evidence

Hesai said a Sharpa robot in the Dairy Queen case achieved a 99% task success rate and reduced its prior failure rate tenfold. Public materials do not provide a complete independent matched-control experiment, the detailed number of trials or failure categories. That result cannot demonstrate that Kosmo makes every robot 99% reliable.

Sharpa is a separate business founded by Hesai co-founders and is used as a proving ground for its robotic technologies, Reuters reported. The company's interim report also clarifies that the narrowly defined ice-cream preparation process still relies on people for restocking, cleaning, maintenance, safety supervision and exception handling. One successful task is not an autonomous store.

Separately, Hesai supplies actuation modules used in Sharpa's Wave robotic hand and reported more than 10,000 cumulative module deliveries by the end of Q2 2026. Those are actuator units, not Kosmo camera shipments. The combined roadmap shows Hesai expanding in both perception/data infrastructure and hardware that produces motion, without proving they already form one profitable integrated business.

Four checks before a robot company buys third-party spatial data

First, does the scan preserve the geometry and viewpoints that the actual robot and its sensors require? A fast scan that misses critical work surfaces, edges or routes may not save integration time. Second, how much of the robot's physical world is represented faithfully, and what still needs manual simulation configuration?

Third, who is allowed to store, modify, export, resell or reuse the spatial data, and for how long? Fourth, what is the full cost after editing, labeling, physics modeling, cloud processing, revisions and validation? Capture time alone is not a meaningful cost-of-ownership benchmark.

These checks help a company decide whether to build spatial scenes in-house, purchase capture tools, or outsource data production. The evolving product architecture also suggests demand for teams combining 3D reconstruction, computer vision, robotics simulation, cloud operations and data quality assurance. It is an inference from required capabilities, not evidence of a specific Hesai hiring increase.

BANSEOG VIEW | The new robotics supplier may sell the environment, not the robot

Hesai's July prototypes, early orders and 200-plus prospective partner engagements show that the Kosmo effort preceded the October launch. Its stated two-part model expands a lidar supplier into spatial software and data production. But prospects, orders, expected quarterly revenue and reported revenue remain separate milestones.

The 300-hour/15-minute comparison and 99% robot success rate are vendor-reported cases with specific tasks and different measurement boundaries. They should not be turned into a universal learning-efficiency gain or robot safety rating. Even a fully accurate visual reconstruction may still need major physical-interaction work.

The strategic question is what multiple robot manufacturers will buy rather than rebuild: perception components, real-world spatial capture, editable training scenes, cloud processing or motion hardware. Those shared inputs create potential supplier businesses around Physical AI. To evaluate them, look beyond demo clips to data quality, reuse rights, interoperability, lifecycle cost and actual repeat customer spending.

BANSEOG VIEW

BANSEOG VIEW | From selling sensors to supplying robot training worlds

Hesai had supplied Kosmo prototypes by July, reported orders including Galbot and engaged more than 200 prospective partners. None of these milestones proves 200 paying customers or recognized Kosmo sales.

The 300-hour modeling versus 15-minute scan example and 99% task-success claim are company cases, not independently comparable universal engineering metrics. Visual data acquisition, physical simulation validity and robot operating outcomes must be evaluated separately.

Hesai proposes integrated hardware-software sales and large-scale spatial-data services while also growing its separate actuation-module business. Customers should compare quality, rights, simulation readiness and end-to-end cost before deciding whether to build, buy tools or outsource spatial assets.

SOURCES

Primary sources and references

Prepared Oct 19, 2026. Hesai official Oct14 release: Kosmo One (8K max capture, 6K exported assets) plus Kosmo World 3DGS/mesh/depth, 100 m2 indoor scan approx 15m, Shanghai DQ example 300 hours manual modeling vs 15 minutes scan/reconstruction are vendor claims covering potentially different scopes. 99% Sharpa robot task success and tenfold reduction in failures are vendor cases without independent count or matched control data. Hesai 2026 interim report filed at SEC, printed p7: initial prototypes delivered in July, orders including Galbot, 200+ prospective partners engaged, Q3 initial Kosmo revenue EXPECTED (not proven), intended hardware adoption/recurring cloud service loop, actuator module deliveries >10,000 by Q2. Corporate report footnote says ice-cream task is narrow; restocking, cleaning, maintenance, safety supervision, exception handling still involve people. Reuters Oct14 updated Oct15 confirms Kosmo launch, no revenue forecast stated, robotics agreements, Sharpa separate legal entity founded by Hesai cofounders. No actual Kosmo revenue amount, training efficiency, verified error denominator or concrete hiring increase asserted. Compared with previous Banseog robot training data articles, this story focuses on supplier-level business and contract rights, not robotic dataset scarcity as a general theme. No copyrighted photos used.