Hesai Group launched its Kosmo One AI spatial camera in Shanghai on October 14, 2026, targeting the cost and time required to turn real-world environments into three-dimensional data for robotics training and simulation.
The device combines LiDAR depth sensing, an 8K-capable imaging system and an inertial measurement unit to capture the geometry and appearance of physical spaces. Its accompanying Kosmo World cloud platform converts the captured information into 3D assets that developers can edit, manage and use in compatible simulation workflows.
The launch addresses a challenge facing developers of physical AI systems: creating accurate digital representations of real environments can require extensive manual modelling and data preparation. Hesai is positioning Kosmo as a way to reduce that work by capturing real spaces and reconstructing them digitally.
According to Reuters’ report on the launch, Hesai co-founder and chief executive David Li identified the shortage of high-quality spatial data as a constraint on robots’ ability to perform useful tasks in the physical world.
3D reconstruction and training data
Kosmo One combines spatial measurements with colour images and motion data to produce detailed digital environments. Kosmo World supports 3D Gaussian Splatting, mesh and depth-data outputs, alongside features for semantic segmentation, scene completion and scene editing, according to the official Kosmo website.
Hesai says the system can reconstruct a space measuring 100 square metres in approximately 15 minutes. Launch coverage compared this with about 300 hours for traditional manual modelling, although the published comparison does not establish that both methods produce identical outputs under equivalent quality requirements.
The company also claims that Kosmo World’s reconstruction process can run up to five times faster than competing 3D Gaussian Splatting products under comparable hardware, dataset and quality conditions. Its website separately claims more than 100-fold data-acquisition efficiency when comparing the cost per high-quality training trajectory in a real-to-simulation workflow with conventional real-robot data collection. These are company-reported performance claims rather than independently verified industry benchmarks.
Despite its high-resolution capture system, the camera does not automatically solve every requirement for robot training. Visually detailed 3D environments may still need additional physical properties, collision geometry and interaction data before they can support particular manipulation or physics-simulation tasks.
Kosmo’s listed integrations include NVIDIA Isaac Sim and Isaac Lab, which developers use for robotics simulation and training. The system is intended to provide reconstructed real-world spaces that can be incorporated into these workflows rather than replace the simulation platforms themselves.
Early orders and mapping plans
Hesai reported that it delivered initial Kosmo prototypes in July 2026 and received orders from humanoid-robotics customers, including Galbot. In its second-quarter disclosures, the company also said more than 200 prospective partners across industries had engaged with the Kosmo business. The figure refers to prospective partners, not confirmed customers.
The launch also included a reported collaboration with Amap covering a first batch of 1,000 pilot locations, including selected malls, shops, restaurants, hotels and brand stores. The initiative aims to create detailed 3D representations of indoor commercial spaces. The announcement did not establish that all these models would be used as robotics training datasets or made publicly available.
Hesai has not disclosed a separate revenue forecast for Kosmo. Its financial outlook for its broader Strategic Growth Initiatives business covers both spatial-intelligence products and robotic actuation modules, rather than the camera alone.
Discover more from Aree Blog
Subscribe now to keep reading and get access to the full archive.
