ABot-3DWorld 0: A Universal World Model to Explore Any 3D Space

Jul 16, 2026

ABot-3DWorld 0 introduces a universal multimodal 3D world model that converts text, images, and video into high-fidelity, explorable 3D environments. A unified Spatial Generative Primitive (SGP) represents space as a compact tuple of a high-quality panorama and a spatial point cloud. Inputs are lifted into SGPs, explored via a planned trajectory, and reconstructed into photorealistic 3D Gaussian Splatting worlds. The pipeline supports rich multimodal recovery and generative completion from minimal prompts, enabling map-native spatial exploration at consumer scale.

LanguageEnglish

CategoryTechnology

FormatPDF

Pages31

Reading Time1 h 18 min

Introduction

Key challenges in 3D scene generation

Proposed ABot-3DWorld 0 pipeline

Four-stage architecture

Multimodal input to SGP

SGP to exploration trajectory

Trajectory to 3D-consistent panoramic video

arXiv :2607 . 11673v1 [ cs .CV] 13 Jul 2026 ABot-3DWorld 0: A Universal World Model to ExploreAny 3D Space AMAP CV Lab We present ABot-3DWorld 0 , a universal multimodal 3D world model that turns text, image, and video inputs into high-fidelity, explorable 3D worlds. At the heart of our framework is a unified Spatial Generative Primitive (SGP)—a compact tuple of a high-quality panorama anda spatial point cloud that delivers an efficient description of any 3D space. Multimodal inputs are first lifted into this primitive; a 3D-consistent panoramic video generator then explores the primitive along a planned trajectory; finally, our panoramic video reconstruction engine converts the generated video into a clean, photorealistic 3D Gaussian Splatting (3DGS) world. This pipeline covers two regimes: rich inputs (multi-view sets, casual video) are lifted into the SGP through a geometry-rigorous recovery that mirrors the observed scene, while a single image or sentence is completed generatively into a creative world. The result is one low-barrier engine for general 3D content creation that further anchors generated worlds to geographic points of interest, enabling map-native spatial exploration at consumer scale. Experiments show that ABot-3DWorld 0 sets the state of the art among open-source methods and demonstrates stronger scene fidelity than Marble under rich multimodal inputs. Official Page: [https://abot-world.amap.com/plaza](https://abot-world.amap.com/plaza) Figure 1 Explorable 3DGS worlds generated by ABot-3DWorld 0 from diverse inputs (text, a single image, multiview photos, or video), spanning indoor, street-level, and aerial scenes, from faithful reconstructions to fantastical creations—all produced by a single unified pipeline. 1 Introduction AI-driven 3D content creation is undergoing a progression from objects to worlds. Object-level 3D generation, i.e. synthesizing a single 3D asset from text or a single image via score distillation [1] or native mesh generation, has already spawned multiple commercialized companies and products (e.g. Tripo, Meshy, Tencent Hunyuan3D [2]), validating the efficiency and quality of direct 3D synthesis. Its application scope, however, stops at individual assets. Scene generation directly enables immersive spatial experiences, virtual tourism, and scalable game-world production, and offers substantially greater application value through product forms that object generation alone cannot support. Scaling from an isolated object to a complete, explorable 3D scene dramatically increases spatial complexity and scale, giving rise to two core technical challenges. First, how to efficiently represent a complete space: an object can be fully covered by a single circular sweep, whereas a scene extends from a single room to a streetscape or an entire city block, demanding a compact representation that simultaneously encodes geometric structure and visual appearance. Second, how to efficiently explore that space: the narrow field of view of perspective imagery means that covering a full environment requires a large number of frames, with generation and rendering costs growing sharply as scene scale increases. Additionally, scene-specific phenomena such as sky, water surfaces, and long-range illumination introduce difficulties absent from object capture. These challenges explain why only a handful of systems [3–5] have attempted full 3D scene generation to date. Recent systems such as Marble [6] and HY-World 2.0 [7] have demonstrated that generative models can already transform flat inputs into navigable 3D scenes expressed as neural radiance fields [8], 3D Gaussian Splatting [9], or polygonal meshes, substantially advancing the state of the art. Yet several fundamental technical bottlenecks remain open: (i) multimodal inputs (multi-view images, video) are often processed through cascaded generative conversions that discard the geometric evidence present in the original observations and am

FAQs

What is ABot-3DWorld 0 designed to do?

ABot-3DWorld 0 is a universal multimodal 3D world model that turns text, images, and video into high-fidelity, explorable 3D worlds expressed with 3D Gaussian Splatting.

How does the model represent 3D space internally?

It uses a Spatial Generative Primitive (SGP), a compact tuple consisting of a high-quality panorama and a spatial point cloud, which serves as the common representation across input modalities.

How does ABot-3DWorld 0 handle both rich and minimal inputs?

Rich inputs (multi-view sets and casual video) are lifted into SGPs through geometry-rigorous recovery, while minimal inputs (a single image or sentence) are completed generatively using a DiT-based panorama generator followed by monocular geometry estimation.

Jul 16, 2026

Upload by Valentina