Hesai Launches Kosmo to Turn Real-World Scans Into Robot Training Data

Hesai Group launched Kosmo on October 14, 2026, a 3D world data engine that combines its Kosmo One spatial camera with the Kosmo World cloud platform, converting physical environments into high-fidelity 3D digital spaces used for robot training, world model development, and immersive content creation.

At the launch event, the company also announced strategic partnerships with embodied AI companies Galbot and Dexmal and demonstrated applications developed with Pixar’s RenderMan renderer and the Chinese mapping platform Amap, spanning robotics, filmmaking, and location services.

Shaoqing Xiang, Hesai’s co-founder and CTO, said Physical AI needs real-world data that is accurate, affordable, and available at scale, and that meeting all three requirements at once has been a persistent problem. “Kosmo is designed to overcome that challenge, making it easier for AI developers and creators to turn real environments into data they can train on and assets they can build with,” he said.

Kosmo One Capture Hardware

The Kosmo One is an 8K ultra-high-definition AI spatial camera combining high-resolution lidar, a one-inch imaging sensor, and a high-precision inertial measurement unit, and it records geometry, color, texture, and camera position with microsecond-level synchronization. Hesai describes it as the world’s first AI spatial camera with an integrated 8K imaging system, a designation the company footnotes as based on publicly available product information as of October 2026. The 8K figure refers to the maximum capture resolution; the maximum resolution users can export is 6K.

According to Hesai, the camera’s 3D Gaussian Splatting reconstruction achieves millimeter-level geometric accuracy, reproducing materials, lighting, and shadows. The camera offers instant scanning on power-up, intelligent guidance, and one-click scan resumption, and the company says a single operator can scan a typical 100-square-meter indoor space in approximately 15 minutes. Kosmo World then processes the captured data into 3D models at up to approximately five times the speed of comparable products, Hesai says, while noting that actual results vary with the scene, data volume, and parameter settings.

Kosmo World and the Dairy Queen Deployment

Kosmo World handles 3D content creation and management on a single platform, powered by what Hesai calls its Kosmo Spatial Core. One reconstruction produces 3D Gaussian Splatting, mesh, and depth data together in a consistent scene. Generative AI fills areas the scanner missed and reconstructs materials and textures, while semantic segmentation identifies individual objects. Views can also be adapted to a robot’s camera characteristics, which Hesai says helps narrow the gap between simulation and real-world operation.

As a working example, Hesai pointed to a deployment at a Dairy Queen ice cream store, where the company said a professional team previously needed approximately 300 hours of manual modeling to build a 3D training environment. With Kosmo, scanning and reconstruction took 15 minutes. In that deployment, a Sharpa robot’s task success rate reached 99% and its failure rate fell to one-tenth of the previous level, allowing the robot to begin operating in an actual store, according to Hesai.

World Models, Film, and Mapping

For world model development, Hesai said Kosmo’s geometrically accurate scenes provide precise camera poses, depth information, and consistent multi-angle views as a real-world reference for building virtual environments. A single captured space can yield more than 10,000 automatically generated training trajectories, giving developers over 100 times as much high-quality training data for the same cost — a comparison Hesai defines against the cost of collecting one high-quality real-world robot trajectory.

For film work, the company estimates that scanning, reconstructing, and editing a 10,000-square-meter European streetscape takes one week with Kosmo, compared with roughly two months to build the location as a physical set, and says total costs can fall by up to 80%. For a 200-square-meter indoor scene, Hesai puts the effort at approximately two person-days with Kosmo versus about 14 person-days of traditional CG modeling. Hesai said film and television companies are already using Kosmo in workflows built around Pixar’s production-grade RenderMan renderer, which has released an official tutorial with downloadable resources, and the company is partnering with Yuemo Studio on AI-assisted filmmaking that pairs captured real-world environments with AI-generated content.

Hesai also said it is working with TRIC, a United Nations partner, to create an immersive digital artwork of the Du Fu Thatched Cottage in Chengdu, the former home of the Tang Dynasty poet. Under the Amap partnership, reconstructions of the Gravity Art Museum and PIFO Gallery are already live in Amap’s Flying Street View feature; Hesai said the platform reaches nearly one billion monthly active users, and the partnership’s longer-term goal is to shift navigation from finding a store to finding a product inside it.

Open Ecosystem and Actuation Modules

Hesai is developing Kosmo as an open spatial intelligence ecosystem, offering open access to raw capture data and 3D assets with integrations across NVIDIA Isaac Sim, Isaac Lab, Unreal Engine, RenderMan, and Houdini. The company described two business models: integrated hardware and software solutions for customers in robotics, media, and other industries, and data services that produce real-world spatial data at scale for world model and Physical AI developers.

Hesai said its lidar deliveries have doubled for nine consecutive years and that it holds a leading global market position in passenger-vehicle ADAS and robotaxis, a foundation on which it is building its next phase of growth around 3D world data and robotic motion. Co-founder and CEO Yifan “David” Li described the 3D world data engine and a companion robotic motion engine as the company’s second growth driver. “Over the past decade, we have enabled machines to see the world,” Li said. “Over the next decade, we will enable robots to understand and change it.”

Alongside Kosmo, Hesai is expanding a robotic actuation business. Its actuation modules power Sharpa’s Wave dexterous hand, which the company says delivers three times the power density of comparable products in a package 37% smaller. According to Hesai, the hand has been incorporated into NVIDIA’s Isaac GR00T humanoid robot reference design and adopted by research institutions including UC Berkeley and Stanford. The company said cumulative deliveries of its actuation modules passed 10,000 units by the end of the second quarter of 2026, and it expects shipments to reach six figures in 2027 as the product line grows from hand modules to joints across the robot body.