Source-linked AI summary
ABot-Earth 0.5: Generative 3D Earth Model
Ming Qian, Tianjian Ouyang, Mingchao Sun, Zijian Wang, Jincheng Xiong, Jiarong Han, Yongchang Zhang, Jiawei Zhang, Xu Wang, Yu Liu, Luyang Tang, Fei Yu, Zengye Ge, Mengmeng Du, Yuan Liu, Nianfei Fan, Song Wang, Yingliang Peng, Chunxue Jia, Yang Liu, Shiying Zeng, Haozhe Shi, Junnan Lai, Hongyu Pan, Zheng Wu, Ning Guo, Mu Xu, Hang Zhang
TL;DR
Large-scale 3D reconstruction remains costly, slow, and computationally demanding, while existing generators often use representations ill-suited to complex outdoor environments. ABot-Earth 0.5 generates 3D Gaussian Splatting scenes from satellite imagery through a scalable, multi-LOD pipeline, reporting significant superiority over state-of-the-art baselines and broad environmental generalization.
Problem
Large-scale 3D reconstruction is constrained by high acquisition costs, long processing times, computational barriers, and a representation gap for complex outdoor environments.
Method
ABot-Earth 0.5 uses a native 3DGS generative framework with cross-domain conditional adaptation and scalable multi-LOD deployment from satellite imagery.
Results
The framework significantly outperforms state-of-the-art baselines and synthesizes detailed environments with robust generalization across metropolitan areas and natural terrains.
Takeaways & Limitations
ABot-Earth 0.5 provides a low-cost, high-efficiency foundation for scalable 3D visualization and simulation, including Embodied AI applications such as UAV navigation.
Takeaways & Limitations
The synthesized scenes still exhibit a quality gap relative to traditionally reconstructed outdoor scenes, and future work targets street-view detail and reconstruction-grade fidelity.
Abstract
from arXiv · showhide
We present ABot-Earth 0.5, a generative 3D framework designed to synthesize vast, seamless 3D environments from ubiquitous, geospatially referenced satellite imagery. To achieve this, we propose a novel generative model formulated directly with the 3D Gaussian Splatting (3DGS) representation. The model is trained on a diverse corpus of existing real-world urban reconstructions, learning to generate realistic geometry and textures. At inference, it synthesizes novel 3D scenes conditioned solely on satellite imagery at a scalable rate of under 10 minutes per square kilometer, while demonstrating exceptional realism. The framework is designed for accessibility, with integrated hierarchical level-of-detail (LOD) structures that permit real-time, interactive visualization on web-based map engines. This high-fidelity simulation sandbox effectively mitigates the sim-to-real domain gap, enabling critical downstream Embodied AI applications like closed-loop UAV navigation. By providing an ultra-low-cost and high-efficiency solution, ABot-Earth 0.5 significantly lowers the technical and financial barriers to large-scale 3D reconstruction and empowers the future of global digital earth visualization.
1 Introduction
ABot-Earth 0.5 addresses the difficulty of producing physically and geospatially authentic large-scale 3D environments by learning from real-world urban reconstructions and conditioning generation on ubiquitous remote sensing imagery. Its integrated multi-LOD outputs support planetary-scale interactive visualization and high-fidelity Embodied AI simulation, including UAV navigation.
- Motivation: Traditional large-scale 3D reconstruction is highly useful for digital twins, disaster response, urban planning, and robotics, but faces substantial physical constraints.These constraints motivate alternatives that reduce exhaustive multi-view acquisition and optimization overhead.
- Problem: Existing outdoor generators often rely on synthetic assets or unconstrained hallucination, producing artificial environments that fail to bridge the sim-to-real domain gap.This limits their practicality for rigorous real-world applications.
- Approach and contributions: ABot-Earth 0.5 learns real-world geospatial complexity from diverse urban reconstructions and uses globally ubiquitous, geospatially referenced remote sensing imagery as its conditioning blueprint.The model synthesizes detailed facades, dense vegetation canopies, coherent road networks, and natural textures without artificial heuristics or synthetic assets.
- Planetary-scale visualization: Native hierarchical multi-LOD generation enables viewport-dependent streaming of trillion-scale Gaussian primitives from planetary overviews to street-level details.The integrated pipeline avoids expensive post-reconstruction downsampling and supports fluid interactive exploration.
- Downstream applications: Physically and photometrically realistic geometry and multi-view consistent textures create interaction-ready sandboxes for Embodied AI, including UAV navigation, obstacle avoidance, and control.The framework is intended to mitigate the synthetic-to-real domain gap and generalize across diverse environments worldwide.
2 Data Pipeline
The data pipeline converts multi-source real-world imagery into city-scale 3DGS scenes, extracts overlapping tiles with dense multi-view supervision, and filters samples through hierarchical quality assessment. ABot-3DGS addresses scale, heterogeneous content, and cross-source variation using block-based reconstruction, LOD management, geometric and appearance optimization, and cross-view fusion.
- Pipeline overview: ABot-Earth trains on city-scale 3DGS scenes reconstructed from large-scale real-world imagery by ABot-3DGS through collection, reconstruction, tiling, rendering, and curation stages.The resulting photorealistic reconstructions provide the downstream generative model’s data foundation.
- Data sources: Satellite, aerial, and urban imagery provide complementary orbital-to-ground viewpoints, combining proprietary acquisitions and curated public datasets after unified coordinate and metadata standardization.Multi-stereo satellite imagery supplies parallax, aerial data covers built-up and natural areas, and urban captures add low-altitude facade detail.
- Scalable reconstruction: ABot-3DGS handles hundreds-of-square-kilometer projects and heterogeneous, inconsistent inputs with hierarchical blocks, continuous LOD, GPU-cluster parallelism, geometric optimization, semantics-aware processing, and cross-view fusion.Dynamic vehicles and pedestrians are removed, while appearance variation modeling addresses lighting, weather, seasonal, and cross-source inconsistencies.
- Tile construction: Each training tile covers a 200 m × 200 m region, overlaps adjacent tiles for boundary context, is normalized and cleaned, and receives dense multi-view rendering from perturbed multi-altitude cameras.Views span nadir-to-oblique pitch angles and multiple compass directions to capture facade details from diverse viewpoints.
- Quality curation: Quality assessment operates at tile, view, and dataset levels, retaining only reliable samples through reconstruction metrics, geometric and perceptual checks, opacity filtering, and VLM evaluation.Tile assessment uses PSNR, SSIM, LPIPS, geometric accuracy, VLM perceptual quality, and spatial completeness; low-opacity views and views failing VLM checks are discarded.
3 Method
ABot-Earth 0.5 addresses Earth-scale 3D generation through native 3DGS modeling, integrated multi-LOD decoding, seamless sliding-window inference, and cross-domain conditional adaptation. Together, these innovations enable end-to-end generation of interactive, streamable, Earth-scale outdoor environments from real-world satellite imagery.
- Representation: Native 3DGS generation learns a compact latent space from real-world scenes containing millions of unstructured Gaussian primitives and generates novel scenes directly in this format.This representation captures complex outdoor structures such as foliage and water surfaces more faithfully than clean mesh assets.
- Scale and interactivity: An inherent multi-LOD decoder directly synthesizes hierarchical 3DGS structures, enabling on-demand detail levels for smooth, real-time online visualization without multiple discrete scene versions.LOD is integrated into generation rather than applied as post-processing, supporting transitions from planetary overviews to street-level views.
- Spatial coherence: Seamless sliding-window inference blends overlapping regions during generation to reduce stitching artifacts and render vast, continuous landscapes.The strategy manages adjacent-tile influence within transition zones, avoiding the artifacts caused by naive tile stitching.
- Conditional robustness: Cross-domain conditional adaptation simulates satellite-view renderings during training and uses a VLM-based harness at inference to adapt conditioning to real-world satellite inputs.The strategy addresses variation in satellite-image quality, resolution, acquisition angles, atmospheric effects, and sensor characteristics.
- Overall method: ABot-Earth 0.5 achieves end-to-end generation of interactive, streamable, Earth-scale 3D outdoor environments directly from real-world satellite imagery.This result integrates the framework’s representation, LOD, inference, and conditioning innovations.
4 Deployment: From Algorithm to a Planetary-Scale System
ABot-Earth 0.5 deploys a two-stage pipeline that generates massive numbers of Gaussian primitives from satellite imagery and reorganizes them for scalable, real-time map rendering. Tile-based inference, unified geospatial alignment, hierarchical LODs, and dynamic tile scheduling address the system’s scale and interaction requirements.
- Global Deployment Pipeline: The deployment uses a two-stage pipeline: global-scale production generates trillions of Gaussian primitives, followed by post-processing and rendering for scalable real-time performance.The first stage produces the scene data, while the second organizes and renders it for deployment.
- Tile-Based Generation Strategy and Resource Planning: 4K satellite-image inference on A100 GPUs covers approximately 1.6km × 1.6km (2.56km2) per tile, 64 times the area of the 200m × 200m training tiles.The pipeline partitions the global target area into regular spatial tiles, each processed as an independent generation task.
- Data Organization: Approximately 320,000 inference blocks produce approximately 3.2 trillion Gaussian primitives, creating bottlenecks in consumer-GPU rendering capacity and coordinate-system compatibility.Each block contains approximately 100 million primitives and initially uses an independent local coordinate system.
- LOD Data Reorganization: The system unifies block coordinates and builds a 6-level LOD hierarchy spanning zoom levels 14 to 19 for interactive deployment.High-precision levels 17–19 are generated natively, while levels 14–16 are produced from zoom-17 data through Bhattacharyya-distance-guided statistical decimation.
- Rendering Scheduling: The Amap Yunjing map engine dynamically selects tile precision by viewport, loading zoom 17–19 for close views and zoom 14–15 for distant views.Native tile scheduling enables smooth transitions between precision levels during map exploration.
5 Evaluation
ABot-Earth 0.5 is evaluated for both generative fidelity and system-level applicability. It achieves state-of-the-art image fidelity while offering rapid, broad-coverage, visually compelling, and open 3D environment generation.
- Generative fidelity: 16.1 FID establishes state-of-the-art generative fidelity, improving substantially over the previous best of 69.5.Evaluation uses FID and KID against renderings from complex real-world 3DGS reconstructions.
- System-level applicability: ABot-Earth 0.5 extends generation from isolated scenes to continuous, interactive, Earth-scale 3D environments.The authors characterize this as a step toward planetary-scale digital twins.
- Timeliness and efficiency: Under 10 minutes per 1 km2, ABot-Earth 0.5 generates scenes from satellite imagery alone, unlike commercial pipelines that may take months to years to update.This supports on-demand 3D environment creation.
- Visual quality: ABot-Earth 0.5 scores higher than Google Earth in aesthetics, while Google Earth retains advantages in geometric and textural fidelity.The human study evaluated Geometric Accuracy, Textural Fidelity, and Overall Aesthetics; the aesthetic advantage is attributed to plausible lighting and color harmony.
- System openness and extensibility: Native 3DGS and open standards let ABot-Earth 0.5 render scenes from any angle and support downstream simulation, virtual production, spatial computing, and editable hybrid scenes.Generated environments can serve as structured foundations for secondary creation and information fusion.
6 Conclusion
ABot-Earth democratizes 3D content production through a low-barrier generative workflow for urban digital twins and unmapped regions. Future work targets street-view detail, greater scene diversity, reconstruction-grade fidelity, and validated scaling laws.
- Conclusion: ABot-Earth transforms 3D content production from a high-cost, specialized process into a low-barrier, generative workflow.It rapidly generates foundational 3D layers for urban digital twins.
- Conclusion: The framework fills gaps in unmapped regions, enabling the low-cost launch of 3D map services and providing spatial priors for intelligent systems such as drones.The passage identifies drone navigation as an enabled application.
- Conclusion: Future work will transition from aerial-level 3D to street-view detail while pursuing greater scene diversity and reconstruction-grade fidelity.The authors also aim to systematically validate scaling laws governing outdoor 3D scene generation.
7 Contributions
The project credits contributors across algorithm development, data pipeline construction, engineering, art design, sponsorship, and leadership, while also acknowledging additional supporters.
- Contributions: Algorithm development was credited to Ming Qian, Tianjian Ouyang, Mingchao Sun, Zijian Wang, Jincheng Xiong, Jiarong Han, Yongchang Zhang, and Jiawei Zhang.
- Contributions: The data pipeline was credited to Mingchao Sun, Yongchang Zhang, Zijian Wang, Xu Wang, Yu Liu, Luyang Tang, and Zengye Ge.
- Contributions: Engineering, art design, project sponsorship, and project leadership were assigned to the named contributors in their respective roles.Engineering: Mengmeng Du, Yuan Liu, Nianfei Fan, Song Wang, and Yingliang Peng; Art Designer: Chunxue Jia, Yang Liu, Shiying Zeng, and Haozhe Shi; Project Sponsor: Mu Xu, Junnan Lai, Hongyu Pan, Zheng Wu, and Ning Guo; Project Leader: Hang Zhang, Ming Qian, and Mingchao Sun.
- Contributions: The authors expressed gratitude to Jian Zhang, Yu Lei, Chong Sun, and Qianwei Wang for valuable support and contributions.