Alibaba's ABot-Earth 0.5 generates 3D worlds from satellite images in minutes

Alibaba's Amap-cvlab has released ABot-Earth 0.5, a generative 3D Earth model that can produce photorealistic 3D environments from geospatially referenced satellite imagery in under 10 minutes per square kilometer. Unlike traditional photogrammetry pipelines that require expensive drone or lidar data, this framework learns to generate realistic geometry and textures by training on a diverse corpus of existing real-world urban reconstructions. The model is built directly on 3D Gaussian Splatting (3DGS) representation, enabling high-fidelity synthesis conditioned solely on satellite images at inference time.

ABot-Earth 0.5 is designed for accessibility and real-time performance. It integrates hierarchical level-of-detail structures that allow interactive visualization in standard web-based map engines. This capability addresses the sim-to-real domain gap for downstream Embodied AI applications, such as closed-loop navigation for UAVs, by providing an ultra-low-cost simulation sandbox. The framework is open-source and available on the project's official page, significantly lowering both technical and financial barriers to large-scale 3D reconstruction and digital earth visualization.