BEIJING, June 8, 2026 /PRNewswire/ — Amap, Alibaba’s location-based services platform, today announced the launch of ABot-Earth0.5, the world’s first 3D native city-scale world model, offering a cost-effective and efficient alternative to traditional 3D city modelling.
Alibaba’s Amap Unveils ABot-Earth0.5 3D Native World Model for City-Scale Modeling
Trained entirely on 3D data, ABot-Earth0.5 is ready for deployment in fields such as embodied AI and physical AI.
The ABot-Earth0.5 model supports multimodal inputs, including images and text. With a single satellite image or text prompt, it can generate a high-fidelity, kilometer-scale 3D urban scene in approximately 10 minutes on a consumer-grade GPU. This represents a 1,000x leap in production efficiency at roughly 1% of the cost of conventional approaches.
Traditional 3D city modelling has long relied on a “capture-then-reconstruct” workflow, which involves aerial data collection followed by 3D structures extraction from 2D imagery. This process is resource-intensive, dependent on specialized equipment and manual refinement, and typically takes a few hours to several days to model just one square kilometre.
ABot-Earth0.5 takes a fundamentally different approach. By training directly on 3D data, the model achieves an intrinsic understanding of three-dimensional space, enabling the end-to-end generation of ready-to-render scenes in 3D Gaussian Splatting (3DGS) format. The outputs can be seamlessly integrated into mainstream engines, including Unity and Unreal Engine, for immediate production deployment.
On the engineering side, Amap developed a first-of-its-kind compression-generation framework that encodes high-quality real-world 3DGS scenes into a compact latent space. This is complemented by a sliding-window inference mechanism that supports continuous, wide-area scene generation, as well as a multi-level-of-detail (LOD) decoder designed to ensure smooth rendering across varying viewing distances.
This technological shift repositions 3D city modelling from a resource-intensive undertaking to a more accessible, scalable capability. Amap noted the technology is poised to drive growth across a range of downstream industries, including embodied AI, the low-altitude economy and 3D gaming creation, while also supporting emergency response by creating disaster scenes in 3D in an instant.
Amap has already deployed the model in embodied AI applications. A city-scale simulation training ground built on ABot-Earth0.5 has reduced the construction cycle for virtual environments from days to minutes, bridging a vital data gap for open-world AI training. The debut of Amap Tutu, a fully autonomous quadruped robot designed for open environments, is a prime real-world use case demonstrating this technology in action.
ABot-Earth0.5 is now open for beta testing. Developers and enterprise users can apply through the official channel at abot-earth.amap.com.
