Hesai launches Kosmo to turn scanned spaces into robot training environments
The AMW Read
Kosmo productizes Hesai's expansion into physical AI data infrastructure, with segment-level relevance to simulation workflows but limited evidence of generalizable training gains.
Hesai launches Kosmo to turn scanned spaces into robot training environments
Hesai unveiled its Kosmo 3D world data engine on October 14, pairing the Kosmo One spatial camera with the Kosmo World cloud platform. The camera combines LiDAR, 8K imaging and an inertial measurement unit. Its reconstruction system produces spatially aligned 3D Gaussian Splatting, mesh and depth outputs for robot simulation, world-model training and content production. Hesai also announced strategic partnerships with Galaxea Dynamics and Yuanli Lingji.
The launch moves Hesai further from selling sensing components toward supplying the data preparation layer for physical AI. Its proposition is to connect real-world capture with simulation assets: appearance, object geometry and distance information derived from the same scene. That addresses a practical bottleneck in embodied AI, where visually convincing environments do not necessarily reproduce useful interactions. However, reconstructed geometry and AI-generated completion do not by themselves establish accurate friction, deformation or contact behavior. The announcement demonstrates a product offering, while broader training value still requires deployment evidence.
Builders should evaluate Kosmo against the full cost of preparing usable simulation environments. Hesai reports that a Sharpa deployment at a DQ ice cream shop reached 99% task success, but the report does not supply an evaluation protocol. Its comparison of roughly 300 hours of manual modeling with a 15-minute scan also leaves downstream preparation time unclear. Pricing, asset compatibility and measured transfer to real robots will determine whether the integrated camera-and-cloud workflow earns repeat purchases. Hesai plans to sell both integrated products and spatial-data production services.