Real Time
Gaode Unveils ABot-Earth 0.5: Native 3D Urban World Model
Gaode Maps (Alibaba Group) Unveils ABot-Earth 0.5: The World's First Natively 3D-TrAIned Urban World ModelJune 8, 2026 – Today, Gaode maps, a...
2 weeks ago
•
June 8, 2026
•
43 views
June 8, 2026 – Today, Gaode maps, a subsidiary of Alibaba Group, officially launched ABot-Earth 0.5, marking a revolutionary milestone as the world's first urban world model entirely trained on 3D data and engineered for prACTical APPlication. LeverAGIng a native 3D technical architecture and Gaode's proprietary 3D data accumulation, ABot-Earth 0.5 pioneers the end-to-end, AI-driven generation of city-scale 3D scenarios. ReDeFining 3D Urban Modeling: From "2D to 3D" to "Native 3D" Unlike traditional urban modeling methods that rely on the "capture-and-fitting" approach or the technical path of "distilling 3D structures from 2D images," ABot-Earth 6.0 adopts a fundamentally different strategy. The model is trained directly on 3D data, enabling it to establish a native understanding of three-dimensional space. This allows for the end-to-end, one-shot generation of scenes in 3DGS (3D Gaussian Splatting) format. The innovation delivers a significant leap in efficiency. By simply inputting a satellite image or a text Prompt, users can generate a 3D city on a single consumer-grade graphics card. This method boosts generation efficiency by approximately 1000 times compared to traditional modes. Overcoming Technical BARRiers: A Systemic Innovation While the native 3D approach holds theoretical advantages, its implementation faces challenges such as the difficulty of efficiently training large models on raw 3DGS data and the complexity of generating continuous, kilometer-scale 3D scenes. ABot-Earth 0.5 addresses these issues through systemic innovations across the data representation, decoding architecture, and inference strategy, thereby streamlining the entire workflow from model training to scene generation. Training: The model features a first-of-its-kind compression-generation Framework designed specifically for 3DGS point clouds. This framework encodes high-quality, real-world 3DGS scenes—containing millions of primitives—into a compact latent space, allowing the model to directly "read" and generate large volumes of 3D data while solving issues related to 3DGS disorder.
Inference: To achieve kilometer-level wide-area generation, ABot-Earth 0.5 introduces an efficient Sliding-Window Inference mechanism. By Intelligently fusing overlapping regions, this mechanism seamlessly stitches together block-generated scenes, ensuring spatial continuity across vast areas.
Generation: A cross-domain adaptive module bridges the resolution gap between satellite imagery and 3D training data. Additionally, a built-in multi-level detail decoder (LOD) ensures the generated results possess natural depth of field, enabling smooth roaming at different viewing distances without post-processing.
Delivery: The model features a comprehensive automated pipeline that ouTPUts native, renderable 3D city scenes. Users can directly import these assets into mainstream engines like Unity and Unreal Engine, add interactive logic, and deploy them for practical production.
Availability and Resources
ABot-Earth 0.5 is currently open for internal testing. Users and developers are invited to visit the official website to submit applications and experience how spatial intelligence Technology is reconstructing traditional 3D production methods.
Be the first to rate this article.
Comments & Questions (0)
No comments yet
Be the first to comment!