For applications such as digital twins, urban simulation, and low-altitude economy route planning, displaying realistic street views or 3D environments traditionally relies on oblique photography, LiDAR point clouds, or manual modeling. These processes are time-consuming, costly, and many areas lack even basic 3D base maps. Recently, AMAP CV Lab launched an intriguing online tool called ABot Earth Studio, which claims to convert planar satellite images directly into navigable 3D spaces. I’ve tried it out and would like to share an overview.
What is ABot Earth Studio?
ABot Earth Studio is an online demo interface for the ABot-Earth 0.5 model released by AMAP in June 2026. Developed by AMAP CV Lab, it is a generative 3D Earth model. Officially, it is positioned as a tool to generate realistic 3D spaces for multiple cities with just one click.
Unlike traditional city modeling that relies on data capture and fitting, ABot-Earth 0.5 follows a 3D-native generation approach: the model is trained directly with 3D data. During inference, given a satellite image or a text description, it can generate a 3D urban scene covering 1 square kilometer in about 10 minutes. The output is in native 3D Gaussian Splatting format, which can be directly imported into mainstream engines like Unity or Unreal Engine for further processing.
Official links:
Website: https://abot-earth.amap.com/?tab=planet
Technical report and open-source repository: https://github.com/amap-cvlab/ABot-Earth-0.5
Core Capabilities
ABot Earth Studio allows users to upload a satellite image or enter up to 500 characters of text description within a 1-square-kilometer area. In about 10 minutes, it generates a 3D plot that can be rotated, zoomed, and navigated, without requiring multi-view photogrammetry. The platform currently covers Earth, the Moon, Mars, and over 190 countries and regions worldwide. For GIS professionals, this is like quickly "guessing" a usable 3D base map from a top-down view. The accuracy is not as high as oblique photography or LiDAR point clouds, but it is very practical for scenario demonstrations, emergency drills, and simulation training.
PS: Text-to-3D and satellite-image-to-3D features currently require applying for a whitelist. I haven’t had the chance to try them yet.
Technical Highlights
Several technical aspects of ABot-Earth 0.5 may be of interest to readers who follow GeoAI:
3D-native, not 2D distillation. Many existing solutions estimate depth from 2D images and then extrude into 3D. AMAP’s model is trained directly on 3D data and outputs 3DGS format end-to-end, resulting in better geometric and spatial consistency.
Sliding-window inference for kilometer-level stitching. Through intelligent fusion of overlapping regions, the generated blocks are seamlessly stitched together to ensure continuity over large areas.
Built-in LOD (Level of Detail). The generated results come with inherent far-near depth, enabling smooth navigation at different viewing distances without post-processing, and facilitating real-time interaction in web map engines.
Cross-domain adaptation. To address resolution differences between satellite imagery and 3D training data, the model incorporates a cross-domain adaptation module to mitigate domain shift.
The official technical report can be downloaded as a PDF from the GitHub repository, and the paper is also available on arXiv.
Hands-on Experience
Currently, ABot Earth Studio is available as an online web tool. Desktop browsers are recommended, as the mobile experience may be limited.
- Open https://abot-earth.amap.com/?tab=planet
- Browse the 3D spaces of Earth, the Moon, Mars, or cities on the planet page.
- Enter the "Plot Gallery" to view and navigate plots shared by other users.
I haven’t yet obtained whitelist access, so I couldn’t fully test the complete creation workflow from satellite image to 3D. However, previewing and navigating plots in the gallery already gives a good sense of the generation quality.
Summary
ABot Earth Studio brings the concept of "satellite image to 3D" from the lab to an accessible online product. For the GIS industry, it may not immediately replace high-precision results from oblique photography or LiDAR point clouds, but it offers a low-cost alternative for rapid prototyping, project presentations, simulation sandboxes, and similar scenarios. The company claims that 3D mapping costs are about 1% of traditional methods, with efficiency improved roughly 1,000 times — these figures need more real-world validation, but the direction is clear: shifting 3D content production from asset-heavy to tool-light.
The product is still in beta, with API and pricing details not yet public. If you are interested in GeoAI, 3D reconstruction, or digital twins, consider visiting the official website to request access.
If you have better solutions for turning satellite images into 3D, or if you already have whitelist access and have completed the full workflow, feel free to share your thoughts.