amap-cvlab/ABot-World

Infinite Interactive World Rollout on a Single Desktop GPU

What it solves

ABot-World addresses the limitation of passive video playback in AI-generated content by creating a real-time, interactive world simulator. It allows users to control the environment through actions, enabling infinite, open-ended world exploration without being locked into a fixed video length or specific scene prompts.

How it works

The system uses a causal student model (ABot-World-0-5B-LF) and a technique called LongForcing training to generate action-conditioned video rollouts. This allows the model to imagine new scenes and dynamics on the fly. It is highly optimized for consumer hardware, utilizing specialized kernels (such as FlashAttention, SageAttention, and LightX2V) to achieve real-time performance on a single desktop GPU.

Who it’s for

This project is designed for researchers and developers working on world models, interactive AI environments, and real-time video generation, as well as users with high-end consumer GPUs (specifically tested on NVIDIA RTX 5090).

Highlights

  • Real-Time Performance: Runs at 720p and 16 FPS with 1.2s latency on a single RTX 5090.
  • Action-Driven Control: Enables continuous exploration where the world responds to user inputs in real time.
  • Infinite Rollout: Supports open-ended generation that extends beyond traditional fixed video length limits.
  • Efficient Memory Usage: Operates within 19GB of GPU memory.

Related

  • Project
    Robbyant/lingbot-world

    An open-source world simulator that generates high-fidelity, interactive video environments with real-time interactivity and long-term temporal consistency.

  • Project
    AMAP-ML/DreamX-World

    DreamX-World is a general-purpose interactive world model that generates high-fidelity, controllable simulations allowing users to explore and transform environments via event prompts.

  • Project
    nv-tlabs/Gamma-World

    Gamma-World is a generative multi-agent world model that simulates shared environments for multiple independently controllable agents, supporting real-time interactive video generation at 24 FPS.

  • Project
    AlayaLab/AlayaWorld

    AlayaWorld is a research‑grade video‑world model that generates long, 3‑D‑consistent videos from a single image, camera trajectory, and text prompt. It supports real‑time interaction (camera steering, prompt switching) via a lightweight AdaLN controller and a distilled few‑step DMD student model. The repo provides inference code for four generation paths, pretrained weights (AR teacher, DMD student, Stage 1 bidirectional transformer), partial training data, and an interactive browser demo. Licensed for academic, non‑commercial use under the LTX‑2 Community License.

  • Project
    amap-cvlab/ABot-Recon

    ABot‑Recon is a research‑grade system that reconstructs dense 3‑D maps from long video streams while only ever keeping a 12‑frame local context. It predicts per‑frame point clouds and relative poses, composes them into a global trajectory, and runs at ~24 FPS on an H100. The repo provides installation scripts, a Python API, optional loop‑closure, visualization tools, and pre‑trained checkpoints on Hugging Face.