Pixar Highlights Hesai's Kosmo 3D Scanner for Real-World Gaussian Splatting

The tutorial explains how creators can process real-world 3D Gaussian splatting data with Pixar's RenderMan tools, and its notes thank a Chinese company for providing the Kosmo 3D scene-scanning solution. Pixar described Kosmo as allowing modelers to enter a place and capture every spatial element, precise geometry and real lighting, giving visual-effects artists richer data to adjust images at the pixel level. Using Kosmo, Pixar captured the lamp and ball outside its Steve Jobs building in 10 to 20 minutes, and the resulting 3D Gaussian splatting data entered its workflow through the USD format, according to the report.

Kosmo is made by Hesai. The company says a handheld scan of about 100 square meters takes 15 to 20 minutes. Hesai divides the product into Kosmo One, the capture device, and Kosmo World, the cloud reconstruction end. Kosmo One integrates a custom lidar, an imaging system with up to 8K resolution, a one-inch CMOS sensor and a high-precision inertial measurement unit. The system synchronizes the sensors at microsecond level and calibrates the lidar-camera extrinsics at the factory, so depth points can be mapped to image pixels without software feature matching, according to the report.

Aligned multimodal frames are uploaded to the cloud, where Kosmo Spatial Core converts them into 3D Gaussian splatting. Lidar point clouds initialize and constrain the geometry, RGB images supply color through differentiable rasterization, and the system outputs 3D Gaussian splatting, mesh and depth data in one coordinate system. The cloud system also uses AIGC to fill unscanned blind spots and AI semantic segmentation to label objects, turning static 3D models into structured data that can be used to train robots.

QbitAI reported that film and games are the first commercial opening for Kosmo, but the product's intended core market is embodied intelligence and physical AI. Robots need spatial relations, materials, object edges and colors, not just lidar returns. Previously, robotics teams bought separate sensors and assembled their own capture rigs, resulting in non-standard, costly and hard-to-iterate systems. Hesai's experience in precision optoelectronics, chip design and automotive-grade manufacturing is meant to solve that problem with a dedicated data terminal that can be used immediately after scanning, the report said.

The report also noted that Game Science, the developer of Black Myth: Wukong, used Hesai lidar to capture models of ancient buildings, one of the factors that drove Kosmo's development. QbitAI listed film, games, mapping, robotics and real-estate VR tours as potential applications. It described Kosmo as the industry's first dedicated scan-and-use data terminal for physical AI, a description attributed to the report.

The article also addressed the debate over world models. As world models gain attention, some argue that synthetic simulation reduces the need for real-world data collection. QbitAI argued the opposite: stronger AI expands the range of physical AI tasks and raises generalization demands, increasing the need for fresh real-world data. It said Kosmo can serve as a starting point for world models and provide physically grounded data that is easier to explain and attribute.

Hesai's business ambition extends beyond hardware. The report said Kosmo's main revenue could come not from device sales but from tokens consumed when the cloud system produces data assets. It also pointed to Sharpa, a full-stack robotics company associated with Hesai's founding team, as a feedback loop that identifies real-world needs and helps Hesai iterate. Hesai CEO Li Yifan said the company designed Kosmo with an Apple-like feel, using high-end hardware to support a platform ecosystem, according to the report.

Editor's Summary

Pixar's tutorial endorsement gives Hesai's Kosmo a high-profile entry into Hollywood visual effects, while Hesai positions the device as a data-capture terminal for robotics and physical AI. The system combines lidar, cameras and inertial sensors with cloud reconstruction to output 3D Gaussian splatting, mesh and depth data. Hesai's broader bet is that demand for fresh real-world spatial data will grow as AI models become more capable.