Hesai Launches Kosmo 3D World Data Engine in Move Beyond Automotive Lidar

Kosmo One is described as a larger version of a DJI Pocket equipped with lidar. It combines lidar with a camera: lidar measures the shape, distance and position of objects, while a CMOS image system records color, texture and light, and the two sets of information are aligned to reconstruct real space. Hesai pushed lidar geometric precision to the millimeter level, enough to distinguish roughly 5 mm spacing in the mesh of a tennis racket, a capability that directly affects the detail of a 3D spatial copy. The imaging part uses a one-inch CMOS, supports up to 8K capture, and allows users to export files at up to 6K resolution.

The CMOS is tuned differently from a conventional camera. Hesai co-founder and CTO Xiang Shaoqing cited the example of a floor: the same floor, shot from different heights and positions, can appear in different colors after a camera's white balance and exposure processing. Image-industry terms such as tone and stylization are things Kosmo tries to avoid. Kosmo instead aims to preserve the original colors and light of the scene and reduce processing differences across cameras and capture times so that multiple viewpoints form a coherent 3D scene. Hesai therefore tuned an ISP image processing system specifically for 3D reconstruction, handling consistency across cameras and time. During capture, lidar, the imaging system and a high-precision IMU are synchronized at the microsecond level, matching images with the device's position and posture to provide accurate spatial reference for rebuilding a room.

In ordinary indoor scenes, Kosmo One can scan about 100 square meters in about 15 minutes, according to Hesai. It supports instant scanning, smart guidance and one-click resume; if a scan is paused halfway, it can continue from a different position. After capture, Kosmo World takes over. The platform is responsible for two main tasks: turning collected data into an intelligent space that AI can understand, interact with and train on, and using AI semantic segmentation to recognize objects in the scene and extract them as independently editable 3D objects. In a demonstration, the system identified a sofa based on a language prompt and segmented and extracted it from the whole scene. After the sofa was moved away, a gap appeared in the carpet that had been hidden underneath; AI then filled in the texture based on surrounding information to further construct a complete world.

Hesai also compared Kosmo with general-purpose AI models. At the event, it used GPT-6 Astra in a test: 20 photos of the same room were taken from different angles and given to the model to reconstruct the environment. In the scene generated by GPT-6 Astra, the sofa's position and shape were wrong, and materials and light differed. Without high-precision data, GPT-6 Astra struggled to fully restore the actual scene. Even when Kosmo-collected 3DGS data was given to GPT-6 Astra, the positional relationships of objects improved, but details such as the shape of a chair and the material of flowers still deviated. Xiang compared increasingly capable general models to a brain and Kosmo One to a more accurate eye. As models become smarter, they still need real, reliable spatial information as input. By combining high-precision capture with AI reconstruction, Kosmo aims to let the eye and brain work better together, and advances in general models can continue to improve Kosmo World's reconstruction and generation capabilities.

The system has already been used in professional production. Dylan Sisson, creative lead for Pixar's industrial-grade renderer RenderMan, scanned the iconic lamp and colored-ball sculptures in the plaza at Pixar's campus into Kosmo. The scan preserved the appearance of the site, and the data could then be adjusted in software for position, size and transparency, with users continuing to decide the style of the image. That means a captured scene can be called up repeatedly across different production stages, allowing professionals to iterate quickly in a virtual scene built by Kosmo and try different forms of expression. Scanned spaces can also become training grounds for robots. According to Hesai, the embodied-intelligence robot Sharpa used Kosmo to recreate a DQ store, build a training environment, and combine robot hardware and models for deployment, reaching a 99% task success rate.

Hesai CEO Li Yifan said the long-term market for spatial intelligence, including spatial cameras, is expected to exceed the lidar market. At the launch event, Li said he hoped Hesai could play an infrastructure role, providing robots with perception, spatial data and motion capabilities so that more companies can build their own robots on top of those foundations. Hesai's original automotive business and embodied intelligence are already connected: a car that understands its surroundings, decides on driving actions and acts on reality through steering and braking itself has features of embodied intelligence. Providing lidar for cars means Hesai has long participated in the perception part of that chain. Hesai also supplies power modules that make robots move. As of the end of the second quarter of 2026, cumulative deliveries of such modules exceeded 10,000 units, and the product range is expanding from dexterous hand modules to full-body joint modules.

For now, Kosmo first serves professional users with clear tasks. Becoming a device that ordinary people can pick up and use daily would require more portable hardware and an easier experience from capture to processing to use. Hesai has not announced a retail price for Kosmo, so in the short term consumers may not be able to use it to scan and remodel their rooms. The company said Kosmo will continue to iterate, and spatial capture devices could enter daily life in more forms beyond handheld devices through cooperation with partners. At the end of the event, Hesai teased a One More Thing: a partnership with a leading Chinese drone company. In a video, the scanning viewpoint rose into the air, extending capture to outdoor buildings and larger spaces.

Editor's Summary

Hesai has introduced Kosmo, a 3D world data engine that pairs a handheld lidar-camera device with a cloud platform for editable spatial assets. The launch marks the automotive lidar supplier's push into embodied AI and spatial intelligence, with early uses in film production and robot training. Pricing and broad consumer availability remain unannounced.