Hesai Targets Physical AI Training With Kosmo Camera and Cloud Platform

SHANGHAI -- Hesai Group has launched a camera and cloud platform to transform real-world environments into three-dimensional training data for robots, expanding the Chinese lidar manufacturer's business into infrastructure for physical AI.

The company launched Kosmo, which consists of an 8K spatial camera (Kosmo One) and the Kosmo World cloud platform that transforms physical settings into 3D training data. Hesai also announced strategic partnerships with Chinese embodied AI developers Galbot and Dexmal without sharing financial details.

Kosmo One combines lidar, an 8K imaging system and an inertial measurement unit to record spatial geometry and visual details. The accompanying software converts recordings into three‑dimensional models for simulated robot‑training environments.

Hesai says the system can scan a typical 100-square-meter indoor space in about 15 minutes. It employs 3D Gaussian splatting to represent scenes as sets of spatial points, enabling detailed renderings from different viewpoints.

The company also claims Kosmo collects spatial data over 100 times more efficiently than individual high‑quality robot trajectories, without increasing acquisition costs. That comparison is based on Hesai's own assumptions and has not been independently verified.

Robotics developers should capture the environment and collect action data (manipulation trajectories, force readings, control commands). Hesai aims to position Kosmo as able to build reusable digital environments that lessens the labor of building robot training simulations.
Partnerships with Galbot and Dexmal link the platform to companies building physical AI robots. Hesai has not said whether partners bought Kosmo equipment, started routine data collection, or integrated the platform into production training.

The company has not provided third‑party validation of reconstruction accuracy or demonstrated how much data from Kosmo improves the performance of robotic tasks, something that will determine its value to physical-AI teams.