What is ABot-World Studio
ABot-World Studio is a general world model studio launched by Amap, unifying interactive video generation and 3DGS scene generation into a single product. Users can input text or images to generate a real-time interactive AI world, supporting local deployment on a single 5090, with a single continuous inference session exceeding 1 hour, while also featuring a spatiotemporal portal to achieve boundless world leaps.
Main Features of ABot-World Studio
-
Text/Image to World: Input text or images to generate a real-time interactive AI world, outputting video and 3DGS files.
-
Real-time Interactive Exploration: Supports WASD and viewpoint control, with the world evolving and providing feedback in real time according to user actions.
-
Spatiotemporal Portal: Embeds portals in any scene, allowing leaps to another complete 3D world, building a boundless exploration network.
-
Dual-modal Generation: Simultaneously supports interactive video generation and 3DGS spatial asset generation with real geometric structure.
-
Long-duration Inference: No upper limit on inference duration at the architectural level, with a single continuous inference session running stably for over 1 hour.
Technical Principles of ABot-World Studio
- ABot-World0 Video Generation Model: Integrates scene roaming and character control into a unified model, mitigating error accumulation through algorithmic optimization, enabling longer temporal continuous inference and precise control.
- ABot-3DWorld0 Scene Generation Model: Shares the inference path with ABot-Earth0.5, using the same architecture to achieve full-scale spatial generation across indoor, street view, and urban aerial photography, and adopts a 'generation-evaluation-repair' modular design to ensure high-fidelity output.
- Local Lightweight Inference: The entire model series is deeply optimized for consumer-grade GPUs, capable of completing inference on a single RTX 5090, achieving hour-level world generation without relying on dedicated computing clusters.
Follow WeChat and reply "Open Source" to join theAI Open Source Project Exchange Group
How to Use ABot-World Studio
-
Visit the Official Website: Visit the ABot-World Studio official website https://abot-world.amap.com/ to enter the test page.
-
Input Prompt: Upload an image or enter a text description to generate the initial world scene.
-
Real-time Control: Use keyboard WASD and mouse/arrow keys for first-person or third-person roaming.
-
Open the Portal: Discover or place a portal in the scene to traverse to a newly generated 3D world.
-
Export Assets: Save the generated world as a video file or 3DGS format asset for secondary creation.
Core Advantages of ABot-World Studio
-
Unified Dual-modal: For the first time integrates interactive video and 3D scene generation, balancing real-time exploration and permanent spatial assets.
-
Hour-level Inference: A single continuous inference session exceeds 1 hour without crashes or quality degradation, far surpassing the 1-minute limit of similar products.
-
Single-card Deployment: Can run locally on a single RTX 5090, lowering the hardware threshold.
-
Physical Consistency: Highly restores real physical rules, with the world providing precise feedback to actions, rather than a passive pixel stream.
-
Full-scale 3D: The same architecture covers indoor, street-level, and city aerial scenes, generating scenes with realistic geometry and depth perception.
Project address of ABot-World Studio
- Official website：https://abot-world.amap.com/
- GitHub repository：https://github.com/amap-cvlab/ABot-World
Comparison of ABot-World Studio with similar competitors
| Comparison dimension | ABot-World Studio | Google Genie 2 |
|---|---|---|
| Inference duration | A single continuous inference exceeds 1 hour with no crashes and no quality degradation. | Generates a consistent world for about 1 minute, after which a reset is required. |
| Modality coverage | Simultaneously outputs interactive video and savable 3DGS spatial assets. | Mainly outputs interactive 3D video environments, with no native 3D asset export. |
| Deployment threshold | Can run locally on a single consumer-grade GPU (5090), with no specialized compute required. | Currently only a research preview, with no local deployment or external API available. |
| World expansion | Built-in "Any Door" mechanism enables unbounded leaps and weaving across multiple worlds. | Has long-term memory within a single world, but limited cross-world expansion capability. |
| Open-source ecosystem | Models and code are fully open source and have been released on GitHub and Hugging Face. | Not open source; capabilities are demonstrated only through official blogs and papers. |
| Application stage | Public testing is already open, and it can be used directly for content creation and embodied training. | Positioned as a research tool for rapid prototyping and AI agent evaluation. |
Application scenarios of ABot-World Studio
- Embodied intelligence training: Provides robots with a high-fidelity physical simulation training ground to simulate perception and decision-making in real environments.
- Open-world games: Supports boundaryless dynamic generation of game maps and the Any Door leap mechanism, breaking through the boundary limits of traditional levels.
- Film and television storyboard creation: Quickly transforms single-perspective material into multi-camera storyboards, compressing the creative validation cycle from weeks to hours.
- Cultural tourism immersive experience: Users enter famous paintings or historical civilization sites in first person, changing from passive viewers into firsthand experiencers of the scene.
- Educational practical training classroom: Builds interactive virtual geography and history teaching spaces to enable immersive inquiry-based learning.