Amap Introduces ABot-Earth0.5 for Large-Scale 3D City Modeling

Amap, part of Alibaba Group, announced in a press release the launch of ABot-Earth0.5, described as the first 3D native world model capable of generating city-scale environments. The model is trained entirely on 3D data and is intended for use in areas including embodied AI and physical AI.

ABot-Earth0.5 can create detailed 3D urban scenes from a single satellite image or text prompt in about ten minutes using a consumer GPU. Amap states that this achieves roughly one thousand times greater efficiency at one percent of the cost of traditional 3D city modeling methods.

The system outputs ready-to-render scenes in 3D Gaussian Splatting format, compatible with engines such as Unity and Unreal Engine. Amap has built a compression and generation framework that encodes real-world 3D scenes into a compact latent space and uses a sliding-window inference system for continuous wide-area generation. It also includes a multi-level-of-detail decoder for smooth rendering.

Amap has already applied the model to embodied AI training, constructing city-scale simulation environments in minutes. The company also demonstrated the technology through Amap Tutu, an autonomous quadruped robot for open environments. ABot-Earth0.5 is now open for beta testing through the official Amap website.