Hesai launches Kosmo to turn real-world scans into robot simulation assets
The AMW Read
Kosmo adds concrete workflow demonstrations to Hesai's previously covered spatial-data expansion, but the reported results remain specific to scene creation and a reconstructed test setting.
Hesai launches Kosmo to turn real-world scans into robot simulation assets
Hesai launched Kosmo, a 3D world data engine, at its Oct. 14 AI Day in Shanghai. The system combines lidar, imaging and motion data with AI-generated scene completion and object labeling to create digital environments for robot training, world-model development and content production. According to TechNode, Kosmo can generate more than 10,000 training trajectories from a single capture and integrates with NVIDIA Isaac Sim, Isaac Lab, Unreal Engine, RenderMan and Houdini.
The market significance lies in the preparation of environments for physical AI: capturing a place is only useful for training if its reconstruction can support simulation workflows. Kosmo puts Hesai at this interface between sensing and robotics data infrastructure. Its support for raw capture data and multiple simulation tools gives builders a way to reuse scenes across existing systems. In a Dairy Queen demonstration with Sharpa, a 15-minute scan reportedly replaced about 300 hours of manual modeling, illustrating the potential to reduce scene-production work.
Builders should distinguish faster asset creation from evidence that a trained robot will perform reliably in the physical world. Sharpa's reported 99% task-success rate applied to the reconstructed setting; the report does not establish equivalent performance in the actual store. A concrete evaluation would compare reconstruction effort and simulation results with task success in the corresponding physical environment. For investors, the adoption question is whether these reusable scenes become a recurring input to robot development beyond individual demonstrations.