Hesai Unveils Kosmo to Turn Real Spaces Into Editable 3D Copies

Want to move the sofa to the other wall but worry that you'll only realize the original spot was better after moving it? Hesai wants that step to happen on a computer first. On October 14, the company, known for its automotive lidar, unveiled the Kosmo 3D world data engine. Its core is a handheld device, the Kosmo One, which looks like a slightly bigger DJI Pocket. Scan a room with it, and you get a copy in which you can move furniture around in a virtual space.

The principle is not complicated. The lidar (a sensor that measures distance with lasers) measures the shape, distance and position of objects, while a one-inch CMOS imaging system (the light-sensing element of a camera) records color, texture and lighting. The two sets of information are then aligned to recreate the real space. Hesai says its geometric accuracy has reached millimeter level, clear enough to distinguish the mesh holes on a tennis racket spaced about 5mm apart. The imaging system supports capture at up to 8K, and the highest resolution of the files users can export is 6K. Unlike ordinary cameras, it avoids stylized processing as much as possible, preserving the original colors and lighting on site. A ISP image-processing system tuned specifically for 3D reconstruction reduces differences between cameras and between moments in time, so that multiple viewpoints can be stitched into a coherent scene. During capture, the lidar, the imaging system and a high-precision IMU (inertial measurement unit, which records the device's position and attitude) are synchronized at the microsecond level. According to Hesai's data, scanning about 100 square meters of an ordinary indoor space takes only about 15 minutes.

Once capture is complete, the Kosmo World cloud platform takes over. It uses AI semantic segmentation to identify objects in a scene and extract them as 3D objects that can be edited independently. During the launch event, the system identified a sofa from a language prompt and separated it out. After the sofa was removed, the AI filled in the hidden carpet texture based on the surrounding information. Hesai also ran a comparison: 20 photos of the same room were fed to GPT-6 Astra for reconstruction, and the sofa's position and shape were off, as were its materials and lighting. When the 3DGS data captured by Kosmo was fed to the same model, the spatial relationships between objects improved, but details such as the shape of the chair and the material of the flowers were still off. Hesai's argument is that general-purpose models act as the “brain,” while the Kosmo One provides more accurate “eyes.” All of these comparisons come from Hesai's own launch demonstration, and no independent testing has been seen yet.

The technology has already entered professionals' workflows. Dylan Sisson, creative director of RenderMan, the industrial-grade renderer from Pixar, scanned a desk lamp and a sculpture of colored balls on a campus plaza. The scans can be adjusted in software for position, size and transparency, and reused across different stages of production. Robotics company Sharpa used Kosmo to recreate a DQ store and build a training environment for its robots. Hesai says its robots achieved a 99% task success rate in that environment. According to TechNode, this scanning replaced about 300 hours of manual modeling.

So what does this mean for ordinary people? For now, it does not mean they can use it right away. Hesai has not announced a retail price for the Kosmo, and individual users still cannot use it to scan their own rooms in the short term. To become part of everyday life, the device would need to be more portable, and the process from capture to use would need to be simpler. Hesai says the Kosmo will keep being iterated, and its capture device may also reach more scenarios through partners in forms other than handheld. For developers and professional production teams, the more immediate change has already arrived: a single scan can produce a 3D scene that can be repeatedly reused, whether for film or for training robots.