An AI world you can walk through with your keyboard
ABot-World-0: Infinite Interactive World Rollout on a Single Desktop GPU
What if you could move through a world made by a video generation model, instead of only watching it?
ABot-World-0 is an action-conditioned video world model that generates the next scene in real time from your keyboard input. It isn’t playing back a clip — the world keeps unfolding as you move and turn.
The problem: the longer you move, the more the picture falls apart
When the model feeds its own previous output back in as input, small errors keep compounding.
- The character’s appearance drifts
- The background structure wobbles
- The world gets stuck circling the same scene
Long Forcing
To curb this, the authors propose Long Forcing — training that accounts for what the model produces over long horizons, not just the next frame.
Real time, on one GPU
They also ran 720p at up to 16 FPS on a single RTX 5090, showing that a real-time interactive world model is within reach of desktop hardware.
Tags
- #ai-papers
- #paper-review
- #world-model
- #generative-ai
- #video-generation
- #abotworld
- #interactive-ai