Hesai's Kosmo turns real rooms into robot training worlds in minutes

In short: On Oct 14 Hesai launched Kosmo, pairing the Kosmo One AI spatial camera with Kosmo World. It turns scanned spaces into 3D Gaussian Splatting / mesh / depth assets for robot training and world models.

What shipped

• Kosmo One: lidar + 1-inch imaging sensor + IMU, microsecond sync; Hesai calls it the first AI spatial camera with integrated 8K capture (export max 6K).
• Speed: ~15 minutes to scan a typical 100 m² indoor space; reconstruction claimed up to ~5× faster than peers.
• Robot angle: views adapted to a robot's camera to shrink the sim-to-real gap; one capture can auto-generate 10,000+ training trajectories.
• Partners: Galbot and Dexmal for embodied AI; integrations with NVIDIA Isaac Sim/Lab, Unreal, RenderMan, Houdini.

Dairy Queen proof point

Hesai says professional teams once spent ~300 hours modeling a DQ store; Kosmo did scan+rebuild in 15 minutes. A Sharpa robot hit 99% task success with failures down to one-tenth prior levels.

Why it matters

Physical AI is starved for cheap, accurate scene data. A lidar leader selling capture + cloud is betting that LiDAR plus 3DGS becomes the data layer under robot learning — not just ADAS.