Technology
Hesai Launches Kosmo to Turn Real Spaces Into Robot Training Worlds
Hesai launched Kosmo One and Kosmo World for turning real spaces into robot training scenes. Its claimed speed and Dairy Queen results need transparent evaluation before they establish reliable robot skills.
Hesai launched Kosmo on October 14, pairing a spatial camera with cloud reconstruction software to turn real rooms into digital environments for robot training. The system combines Kosmo One capture hardware and the Kosmo World platform, extending the lidar supplier's business from sensing the world to supplying reusable models of it.
The announcement includes partnerships with Galbot and Dexmal and a company-reported Dairy Queen store example involving a Sharpa robot. Hesai says the example reached a 99% task success rate. It does not publish the trial count, task mix or intervention record needed to assess that result independently.
For humanoid developers, the opportunity is to build training scenes faster and reuse them across experiments. The harder question is how much a reconstructed room tells a robot about doing work there. A convincing view of a counter supplies useful context; reliable grasping still depends on the objects, contact behavior, controller and evaluation.
Key Stats and Limits
8K / 6K
Maximum capture / user export
~15 min
Claimed scan of typical 100 m² room
3 outputs
3DGS, mesh and depth
99%
DQ task result claimed by Hesai
From a Camera to a Reusable Training Scene
The launch announcement describes Kosmo One as a synchronized combination of lidar, a one-inch imaging sensor and an inertial measurement unit, or IMU. Lidar records spatial measurements, imagery records appearance, and the IMU helps track motion. Combining these inputs is intended to keep the reconstructed geometry and captured views aligned as the operator moves.
Hesai advertises microsecond-level synchronization and millimeter-level geometric accuracy. The release does not provide an accuracy test protocol, an error distribution or performance across difficult materials. Those specifications therefore describe the company's capture claims, not the accuracy of a robot's eventual hand movement. Transparent surfaces, changing objects and contact tasks still need their own assessment.
Kosmo World processes the capture into three representations in one reconstruction: 3D Gaussian Splatting, or 3DGS, a mesh, and depth data. They serve different needs. The visual representation supports viewing a scene from new angles; the mesh describes surfaces; depth records distance information. Having the outputs aligned could save developers from rebuilding the same room in several incompatible forms.
| Layer or measure | Hesai's description | What it establishes |
|---|---|---|
| Kosmo One | Lidar, one-inch imaging sensor and IMU | A capture system, not a robot policy |
| Image resolution | Up to 8K capture; maximum 6K user export | Capture and delivered resolution differ |
| Typical room scan | Approximately 15 minutes for 100 m² | Capture time varies with layout and requirements |
| Kosmo World | 3DGS, mesh and depth in one reconstruction | Multiple scene outputs; physical behavior needs validation |
| Reconstruction speed | Up to approximately 5× similar 3DGS products | Conditional company comparison; competitors unnamed |
The speed comparison assumes similar hardware, datasets and quality requirements. Hesai does not publish a reproducible benchmark in the release. Its typical-room estimate concerns scanning; upload, reconstruction, scene preparation and robot validation are separate steps. Buyers should measure the entire path from arriving at a location to running a useful experiment.
Photorealism Does Not Measure Contact
The original 3D Gaussian Splatting research, published at SIGGRAPH 2023, addresses rendering new views of captured scenes. It represents appearance with collections of three-dimensional Gaussians that can be rendered efficiently. That visual objective should be distinguished from predicting how an object responds when a robot touches it.
A robot placing a cup needs more than the cup's visible outline. Its controller may need appropriate collision geometry, estimates of mass and friction, and a way to recognize a failed grasp. A reconstructed mesh can contribute geometry, but the launch does not establish that Kosmo automatically identifies and validates all the physical properties needed for manipulation.
Hesai also says generative AI fills areas that were not scanned and reconstructs materials and textures. This can make a scene more complete for rendering. It introduces a second category of information: inferred content alongside measured observations. For a robot experiment, developers need to know which surfaces were observed, which were generated and how uncertainty affects the task.
The same distinction applies to semantic segmentation, which identifies individual objects, and views adapted to a robot's camera characteristics. Both could make training data more useful. Neither demonstrates that a policy trained on those images will transfer safely to hardware. Transfer should be tested on held-out physical tasks, with failures and recovery effort counted.
What the Dairy Queen Example Shows
Hesai's strongest robotics example is a Dairy Queen store. It says building the training environment previously took a professional team approximately 300 hours of manual modeling, while Kosmo scanning and reconstruction took 15 minutes in this case. This is a separate claim from the typical 100-square-meter scanning estimate.
Those figures suggest a substantial opportunity to reduce scene-building work. They do not support a universal speed multiplier: a team's manual modeling effort and the elapsed time for a capture-and-reconstruction workflow may include different work. The announcement does not break out annotation, editing, computing resources, integration or subsequent validation.
The release attributes a 99% task success rate to the Sharpa robot and says failures fell to one-tenth of their earlier level. It does not disclose the number of attempts, the precise task definition, baseline conditions or human assistance. Without those details, readers cannot determine how representative the result is or isolate the contribution of Kosmo from other changes to the robot system.
Deployment Reality Check
Hesai reports operation in an actual Dairy Queen store, but the launch release does not identify the store operator, robot count, paid status or supervision arrangements. The evidence is a company account. The commercial deployment stage remains unclear, and the reported 99% result is not an independently verified reliability benchmark.
Hesai's Bet on the Data Business
Kosmo was previewed before this launch. Hesai's July WAIC account described an April unveiling and said an official launch would follow in the second half of 2026. October's announcement names the hardware and cloud products, adds partnerships and supplies more detailed application claims.
For the lidar manufacturer, the business proposition reaches beyond selling another sensor. Hesai describes two routes: integrated hardware and software packages, and services producing spatial data at scale. A room captured once could support repeated experiments or several development teams, making the resulting asset useful beyond the original scanning session.
The new Galbot and Dexmal partnerships connect that proposition to embodied AI development. They do not establish order quantities, contract values or fleet results. Hesai lists connections with NVIDIA Isaac Sim and Isaac Lab, Unreal Engine, RenderMan and Houdini; compatibility is useful, but does not itself measure training effectiveness.
A separate firsthand account by Dylan Sisson describes trying a camera provided by Kosmo and bringing captured scenes into Houdini Solaris and RenderMan. That supports a concrete creative workflow. It is not a robot benchmark, and a supplied test device should not be mistaken for independent customer adoption.
The Useful Test Is Cost per Successful Skill
Hesai says a single captured space can yield more than 10,000 training trajectories. These are generated trajectories, not 10,000 observed physical robot attempts. Their value depends on whether they include useful task variation and preserve the details that determine success in the real environment.
The Kosmo website puts its data-cost claim in more specific terms: approximately RMB 10-20 per high-quality real-robot trajectory versus RMB 0.1-0.2 for its capture, reconstruction and simulation-generation route. These are company estimates with scene, volume and parameter qualifications. A cheaper generated trajectory is not automatically an equivalent training example.
A useful customer comparison would count capture labor, cloud processing, scene repair, simulation setup and physical validation against the number of skills that meet a defined reliability threshold. It would also test new layouts and objects that were absent from training. Otherwise, inexpensive data can hide an expensive integration problem.
Data access and reuse terms matter too. Hesai promises access to raw captures and 3D assets, but that language does not establish an open-source license or permission to redistribute every scanned location. Customers need clear terms for storage, export, ownership and sensitive material captured inside workplaces. The launch release does not provide a price or a detailed general delivery timetable.
Frequently Asked Questions
What are Kosmo One and Kosmo World?
Kosmo One is Hesai's spatial capture hardware. Kosmo World is its cloud platform for reconstructing and managing 3D scenes. Hesai positions the combination for robot training, world models and creative production.
Does Kosmo export 8K files?
Hesai specifies maximum 8K capture, while its launch footnote limits user-exported files to 6K. The two resolutions describe different stages of the workflow.
Can a robot be ready for work after a 15-minute scan?
That is not established. The typical-room figure concerns scanning. Robot training, integration and physical validation require additional work, and the separate Dairy Queen case does not disclose the complete integration timeline.
Is the reported 99% success rate independently verified?
The figure comes from Hesai's Dairy Queen example. Its release provides no trial count, detailed evaluation protocol or independent verification, so it should not be treated as a general reliability guarantee.
Scene Building, With Transfer Still to Prove
Kosmo gives Hesai a defined hardware-and-cloud offering for a real development expense: turning workplaces into reusable digital scenes. The important next evidence is how much preparation it removes, how accurately it preserves task-relevant details and whether those savings survive physical testing.
The Bottom Line: Hesai has launched a new route from captured spaces to training assets. Its company-reported store result is a useful starting point; dependable humanoid skills still need transparent tests of contact, generalization and human intervention.
Hero image: historical laser-scan cutaway of Tudor Place's interior by CyArk (www.cyark.org), via Wikimedia Commons, licensed CC BY-SA 3.0. Original image unchanged; responsive display may crop. The image and adapted display retain that license. Context for indoor digitization only; this is not a Kosmo reconstruction.