AutoNavi Launches ABot-World Studio: Build Interactive AI Worlds with Text or Images
AutoNavi has officially launched the General World Model Workshop, dubbed ABot-World Studio, a tool that merges interactive video generation with 3D scene construction. The standout feature? Users can input simple text or images and quickly generate an AI digital world that looks photorealistic and supports real-time interaction.
A Technical Leap Forward
The workshop marks a significant breakthrough in technical architecture. It not only supports local deployment on a single 5090 GPU but also breaks previous limitations on inference duration for world models. In official tests, the continuous inference runtime has stabilized at over one hour, finally overcoming the awkwardness of earlier models that could only generate short clips.

Two Core Models at Work
ABot-World Studio integrates two core models: ABot-World0 and ABot-3DWorld0. Together, they solve the problem of long-sequence video degradation and natively output 3DGS spatial assets with real geometric structures. This means the generated world is no longer just a stream of pixels but a "physical entity" with physical rules and spatial depth. Users can freely explore within photo-quality scenes, much like in a game.
The Time-Space Portal
To boost exploration fun, the workshop introduces a unique "Time-Space Portal" mechanism. Users can freely switch between different 3D scenes, connecting previously isolated scenes into an endless network through "space teleportation." For example, a wooden door in a Jiangnan water town can be directly linked to a cyber city.
Open Source and Applications
Currently, the ABot-World series of models are fully open source and have been launched on platforms like Reactor and Hugging Face. This technology holds great potential for embodied intelligence training, film storyboarding, cultural tourism, and education. It allows users to transition from mere "content viewers" to active "participants" in the digital world.
Key Points:
- ABot-World Studio generates interactive 3D worlds from text or images.
- Runs locally on a single 5090 GPU with over an hour of continuous inference.
- Features two core models for long-sequence stability and 3D asset generation.
- Includes a "Time-Space Portal" for seamless scene switching.
- Fully open source on Hugging Face and other platforms.