ABot-World runs infinite interactive worlds on a desktop GPU

ABot-World turns a single desktop GPU into a real-time interactive world simulator, letting you steer an endlessly expanding virtual environment using your own controls instead of just watching a pre-rendered video clip.

While most world models collapse or freeze after a few seconds or minutes, ABot-World supports continuous rollout without duration caps. The team even showcased demonstrations running uninterrupted for over an hour, alongside a 24-hour test.

What it does and how it works

Instead of passive playback, ABot-World responds directly to action inputs in real time. As you navigate the world, the model continuously generates the scene ahead conditioned on what you do.

To prevent the simulation from degenerating or getting stuck in a repetitive loop, the project uses a technique called LongForcing. This allows the system to imagine entirely new environments and dynamic events on the fly, without requiring users to switch text prompts mid-session.

In terms of performance, it outputs 720p video at 16 FPS with an interactive latency of roughly 1.2 seconds.

Hardware specs and setup

The real highlight here is consumer feasibility. The entire pipeline runs on a single desktop NVIDIA RTX 5090 GPU, consuming around 19GB of VRAM during live inference.

The codebase is tested on Ubuntu 22.04 with CUDA 12.8 and Python 3.12. It relies on PyTorch 2.8, FlashAttention, and SageAttention for optimized throughput. For those who want to skip dependency compiling, prebuilt Docker images are available on Docker Hub and Alibaba Cloud Container Registry.

The project is fully open-sourced under the Apache 2.0 license. The team has released the inference code, the student checkpoint ABot-World-0-5B-LF, a local Gradio demo, a technical report, and an annotated 500-hour action-video dataset. You can grab the repository at amap-cvlab/ABot-World on GitHub, pull weights from Hugging Face or ModelScope, or test the web playground ABot World Studio.

Why it matters

Interactive world generation is usually locked behind massive cloud clusters and prohibitive compute costs. ABot-World shows that causal distillation and custom kernels can compress an infinite interactive simulator down to a desktop PC.

A latency of 1.2 seconds and 16 FPS won't replace high-speed game engines just yet, but for robotics simulation and experimental sandbox environments, running an infinite world locally on a single consumer card is a massive milestone worth cloning today.