Hesai, the Shanghai lidar maker best known for automotive sensors, launched Kosmo in Shanghai on 14 October 2026. It is a capture system built around 3D Gaussian Splatting: a handheld device called Kosmo One and a cloud platform called Kosmo World that turns a capture into splats, meshes and depth data. Some early posts call it "KosmoAI"; the product names Hesai uses are Kosmo, Kosmo One and Kosmo World.
Two things make it worth attention beyond the spec sheet. A company that shipped more than 600,000 lidar units in a single quarter is now building a device whose main output is a splat. And Hesai says the capture data is open: the raw images, point clouds and IMU data can leave the platform, not only the finished scene.
What Hesai announced
The figures below are Hesai's own, from its launch announcement, with weight, battery and storage from Radiance Fields' launch coverage. None of them has been independently tested yet.
| Item | Kosmo One, as published |
|---|---|
| Imaging | 8K imaging system with a one-inch sensor. Images exported by users top out at 6K |
| Lidar | Class 1, 1535 nm, 569,000 points per second, 40 m range at 10% reflectivity |
| Synchronisation | Lidar, camera and IMU synchronised to the microsecond |
| Capture speed | About 15 minutes for a typical 100 m² indoor room, one operator |
| Weight | About 1.2 kg with the battery handle |
| Battery | About 60 minutes standard, 120 minutes extended. One capture runs up to 45 minutes |
| Storage | 512 GB on the device |
| Processing | In the cloud, on Kosmo World. Hesai says up to about five times faster than comparable 3DGS products |
| Price and availability | Not published |
Kosmo World produces aligned 3DGS, mesh and depth outputs from a single reconstruction. It uses generative AI to fill areas the capture missed, and segments the scene so individual objects can be identified and moved. It connects to NVIDIA Isaac Sim, Isaac Lab, Unreal Engine, RenderMan and Houdini.
Why the export list matters more than the 8K sensor
Here is what Radiance Fields reports Kosmo World can export:
- Splats: .ply and .sog
- Meshes: .obj
- Scenes: .usd and .usdz
- Raw data: the images, the point cloud and the IMU data
A lot of splat tools keep the scene inside their own viewer. That is fine for a demo and awkward for a project, because the client's architect, contractor and facilities team all use different software, and the record has to outlive the subscription. Open export changes that. A splat captured on Kosmo One can go into whatever viewer or game engine the client already uses, and the raw point cloud can be checked and registered somewhere else.
SOG is also the compressed format we publish our own splat scenes in, so a .sog export drops straight into the kind of web viewer most clients actually open.
Who it is built for
Hesai's pitch is Physical AI: training environments for robots and data for world models. Its headline case is a Dairy Queen store where, Hesai says, building a 3D training environment took about 300 hours of manual modelling before and 15 minutes of scanning and reconstruction with Kosmo. Robotics firms Galbot and Dexmal were named as partners at launch. Film and mapping come next: Pixar has published a RenderMan tutorial using Kosmo data, and Amap is showing Kosmo reconstructions of two galleries in its street view.
Architecture, engineering and construction are not the headline market. That is useful context for reading the spec. Kosmo is designed to produce photoreal, segmented scenes quickly. It is not, on what has been published, designed to survey tolerance.
What we would want to know before using it on a Singapore project
We have not used Kosmo One ourselves. Michael Rubloff of Radiance Fields writes that he has been using it for several months and rates the output highly, which is the most detailed hands-on account so far. Before we put one on a client site, these are the questions we would want answered.
- Accuracy, with conditions. Hesai claims millimetre-level geometric accuracy but has not said over what distance, with what registration method, or against what control. Until it does, treat Kosmo geometry as visual, and take any dimension that matters from a registered scan.
- Generative fill that can be switched off and flagged. Filling in a missed corner is helpful in a robot training scene. In a condition record or a dilapidation survey, a filled-in wall is a guess presented as a fact.
- Where the data is processed. Reconstruction runs on Kosmo World, in the cloud. Some Singapore clients, data centres and government-linked sites among them, restrict where site imagery may be uploaded. Ask where the platform is hosted before scanning anything sensitive.
- The 6K export cap. Capture is 8K; what you can take away is 6K. That is still a lot of image, but it is the export figure that sets the quality of anything you build outside Kosmo World.
- Price, availability and local support. None is published. A device is only as useful as the turnaround on a repair.
For a longer view of where splats help and where they do not, read our comparison of Gaussian splats and textured meshes.
How it fits with what is already used here
Handheld devices that capture lidar and imagery together and output a splat are not new. Several are already working on Singapore sites, and we capture splats and point clouds with them. What Kosmo adds is a lidar maker of Hesai's scale entering the category, an 8K, one-inch imaging system, and a platform that treats raw data export as a feature rather than an upsell. If that holds up in independent testing, it pushes the rest of the category in a good direction.
What does not change is the split between looking and measuring. A splat is the best way we know to show a space. A registered point cloud is still how you prove its dimensions.
If you need a splat or a digital twin now
There is no need to wait for a new device. We capture digital twins and 3DGS scenes in Singapore with equipment available today, and pair them with 3D laser scanning when the dimensions have to hold up. If you would rather capture it yourself, we also rent the equipment, and you produce the outputs.
Tell us what the scene is for and we will tell you which capture it needs: request a quote.
