Amap AI-Enables Street Scan Rankings, Unveils ABot-Earth 0.7 3D-Native City World Model

The 2026 Street Scan ranking, launched a year after Amap used real navigation and in-store data to enter the food and local-services ranking market, now applies AI to better reflect real user choices, according to QbitAI. Amap has accumulated more than 20 years of spatiotemporal data and operates across many corners of the physical world, which the report says makes spatial intelligence a natural fit.

ABot-Earth 0.7 can generate a 3D city with realistic street views from a single satellite image or a text prompt in about 10 minutes on a consumer-grade GPU, according to the report. Amap says this is 1,000 times more efficient than traditional methods. The model covers more than 190 countries and 300 cities, which Amap describes as the largest 3D map by number of countries and regions. By contrast, QbitAI notes, Google Earth offers 3D scenes only in the core areas of a few major cities, while many smaller cities and less-developed regions remain in 2D satellite imagery.

Traditional 3D reconstruction is expensive and slow. The report describes a process of collecting images, calculating building positions and shapes, and restoring surface textures, with updates that can take months or years. ABot-Earth 0.7 aims to lower the cost of generating the next 3D scene. It is trained on a large amount of real outdoor 3D reconstruction data to learn spatial patterns such as how roads extend, how blocks are arranged, and how buildings, vegetation and water are distributed.

The model uses 3D Gaussian Splatting, or 3DGS, which represents a scene as many colored, semi-transparent ellipsoids that combine into buildings, trees and roads. The term 3D-native means it directly learns and generates these Gaussians, so the output has a 3D structure that can be viewed from any angle. Because a single urban block can contain millions of Gaussians, Amap designed a native 3DGS compression and reconstruction method: compress first, then generate. Training compresses complex 3D scenes into compact latent-space packages; generation happens in latent space and is then decompressed into 3D Gaussians. Amap says this reduces training and inference costs and improves geometric quality, making building facades more regular and ground surfaces flatter.

To ensure data quality, Amap uses its self-developed ABot-3DGS engine to reconstruct city-level 3D scenes from satellite, aerial and street-view imagery, then runs 3D and 2D data-cleaning pipelines to select high-quality training data. The report says this helps avoid implausible results such as roads floating into the sky or buildings growing in water.

For output, ABot-Earth 0.7 uses a native 3DGS multi-level decoder, or LOD, to generate hierarchical 3D Gaussian structures without expensive post-processing. The report describes smooth zooming from a globe to a city panorama and then to specific streets with little lag. A sliding-window inference mechanism fuses overlapping areas intelligently, reducing seams and producing near-seamless globe-scale scenes. Amap says the approach lowers scene holes and damage by 97 percent and can generate more than 10 square kilometers of 3D scene in a single inference.

The model has limits. Satellite imagery cannot see building facades, so the model must infer them. The generated images are realistic but do not mean every window corresponds exactly to reality. For well-known landmarks such as the Eiffel Tower and the Colosseum, Amap uses a hybrid approach: large urban areas are generated by ABot-Earth 0.7, landmarks are made with 3D generation models, and an Agent automatically calibrates and stitches them together. The report says this extends generation from a god's-eye view to street-level detail and turns the digital Earth from a passive electronic sand table into a dynamic spatial intelligence platform that can generate, simulate and edit.

Amap CEO Guo Ning describes the model's understanding of the real world in four layers: 2D network topology, describing how roads, buildings and places connect; 3D space, presenting the physical structure of the world; time changes, showing the same space at different moments; and real-world flows, including pedestrian and vehicle paths. Together these layers form Amap's definition of spatial intelligence. ABot-Earth 0.7 currently addresses efficient construction of 3D space while retaining precise alignment with 2D network topology, laying a foundation for adding time and real-world flows later.

The model is one part of Amap's spatial intelligence system. Citing Fei-Fei Li's classification of world models into renderers, simulators and planners, QbitAI reports that Amap is among the few companies working on all three in one architecture. Renderers handle 3D presentation through ABot-World, the ABot-Earth series, ABot-Recon, Cloud Realm and Flying Street View 2.0. Simulators predict and simulate the real world, such as future traffic conditions and commercial district foot traffic. Planners output action instructions for agents, including ABot-Navigation and ABot-Manipulation in Amap's embodied AI stack, which can be deployed on robots for movement and manipulation, as well as the Eagle Eye protection system. The three share a causal Transformer architecture. Inputs such as images, text, point clouds and traffic flows are processed by separate encoders and enter a shared attention mechanism, maintaining independent representations while enabling cross-modal communication.

Amap says its spatial intelligence frontier models and research, including the ABot system, have won more than 20 global championships and SOTA results. But the company says benchmark scores are basic; real users buying in is the real test. Since proposing full AI transformation in 2025, Amap has applied spatial intelligence in several scenarios, including the fully autonomous embodied robot Amap Tutu, a city-level simulation training ground for embodied intelligence, traffic light countdown, Eagle Eye protection, and the fully AI-enabled Street Scan 2026.

In Amap's view, intelligence in the digital world must ultimately find value in the physical world. The company expects spatial intelligence to become a standard foundational capability for industries interacting with the physical world, like cloud computing. Previously, Amap was infrastructure connecting people and the real world; now it is becoming infrastructure connecting robots and the physical world. When a user finds a hidden restaurant on Street Scan, when a drone avoids skyscrapers autonomously, or when a smart car predicts congestion and changes lanes early on an evening commute, those seemingly unrelated scenes rely on the same underlying spatial intelligence system.