SuperAPI AI Directory
Partner with us中文AI newsFavorites
All toolsAI developer platformsABot-World

ABot-World

ABot-World is a real-time interactive world simulator project for a single desktop GPU, released by the ABot-World Team. Its goal is to achieve infinite, action-conditioned world rollout on a single NVIDIA RTX 5090 desktop graphics card: supporting 720p resolution, 16 FPS, approximately 1.2 seconds of latency, and approximately 19GB of VRAM usage. The project emphasizes action-driven world control, real-time desktop inference, infinite world rollout, and open-ended scene imagination, and adopts a training method called LongForcing to extend new scenes and dynamics during rollout, avoiding scene lock-in and requiring no prompt switching. The repository provides resources such as inference code, a local Gradio demo, an online playground, the causal student model ABot-World-0-5B-LF, a technical report, a bidirectional teacher model, a 500-hour action-annotated video training dataset, and a Docker environment image. The project is licensed under the Apache License 2.0 and is built on top of multiple open-source projects.

Open website

About this tool

ABot-World (repository path amap-cvlab/ABot-World) is a project that achieves Infinite Interactive World Rollout on a single desktop GPU. The project slogan is "Infinite Interactive World Rollout on a Single Desktop GPU".

Purpose and Positioning The project uses a single NVIDIA RTX 5090 desktop GPU as a real-time interactive world simulator, targeting scenarios that require continuous exploration and real-time response to user actions rather than passive video playback. Its core capability is action-conditioned world rollout, where user actions can affect the world state in real time, thereby supporting open-ended interactive exploration rather than fixed-duration video generation.

Main Capabilities According to the project description, its key features include:

  • Action-driven world control: responds to user actions in real time and supports continuous exploration.
  • Real-time desktop inference: achieves 720p, 16 FPS on a single NVIDIA RTX 5090, with approximately 1.2 seconds of latency and approximately 19GB of VRAM usage.
  • Infinite world rollout: supports open-ended interactive world generation, breaking through fixed video length limits.
  • Open-ended world imagination: through LongForcing training, extends new scenes and dynamics during rollout, avoiding scene lock-in and requiring no prompt switching.

Version and Resource Release Records (according to the dates listed in the repository)

  • 2026-09-08: The Docker environment image went live on Docker Hub and Alibaba Cloud Container Registry.
  • 2026-08-15: Released an experimental 24-hour uninterrupted rollout demo.
  • 2026-08-03: Released the 500-hour training dataset.
  • 2026-07-22: Released the ABot-World-0 technical report.
  • 2026-07-13: ABot-World went live on Reactor.
  • 2026-07-10: Announced that a 500-hour video training dataset with accurate action annotations would be open-sourced.
  • 2026-07-09: Released the causal student model ABot-World-0-5B-LF, inference code, a local Gradio demo, and the online playground ABot World Studio.

Installation and Running Method The project's test environment is: Ubuntu 22.04, CUDA 12.8, Python 3.12, NVIDIA RTX 5090; for common hardware and environment issues, see FAQ.md.

Option 1: Docker (recommended) A prebuilt environment image is provided; the image contains only the runtime environment, and the repository and checkpoints are mounted at runtime.

  • Docker Hub: docker.io/amapcvlab/abot-world:v0-env (the original docker.io/amapcvlab/abot-world:v0-env address)
  • Alibaba Cloud Container Registry (recommended in China)
  • You can also build it yourself: bash docker/build.sh; for build options, volume mounts, and troubleshooting, see docker/README.md.

Option 2: Manual installation The main steps include: clone the repository; create a conda environment; install PyTorch (CUDA 12.8); install FlashAttention (if you encounter glibc version issues, refer to flash-attention#1708); install SageAttention from source (follow the official SageAttention instructions); install Python dependencies; install lightx2v_kernel; download checkpoints (you can use HuggingFace or ModelScope); after downloading is complete, the project should have the corresponding checkpoint structure. The checkpoint paths are configured in configs/long_forcing_dmd.yaml and configs/default_config.yaml; the distilled generator weights have been merged into ABot-World-0-5B-LF/diffusion_pytorch_model.safetensors. You can select a specific GPU via a command.

License and acknowledgements The project is released under the Apache License 2.0; for related copyright and third-party attributions, see LICENSE, NOTICE, and THIRD_PARTY_NOTICES.md. The project is built on and inspired by the following open-source projects: Causal Forcing, AngelSlim, LightX2V, taehv, Wan2.2, and Helios, among which the optimized Triton RoPE and normalization kernels in wan/modules/helios_kernels originate from these projects.

ApplicableScenarios and limitations Applicable scenarios: real-time interactive world simulation and research in a single-GPU desktop environment, including action-driven exploration, open-ended scenario rollouts, and more. Limitations and notes: the official test environment is Ubuntu 22.04, CUDA 12.8, Python 3.12, NVIDIA RTX 5090; for compatibility with other hardware and environments, see FAQ.md, and the materials do not provide broader hardware compatibility information. The Docker image contains only the runtime environment, and the repository and checkpoints need to be mounted at runtime. Checkpoints need to be downloaded separately, and the specific paths and configurations are subject to the repository configuration files. The materials do not state any charging, pricing, or commercial service information for the project.

Citation and contact The repository provides the paper citation format and offers a WeChat group contact. For more details, refer to the repository README and related documentation.

Related tools

Browse category