OpenDCAI/OpenWorldLib

Unified Codebase for Advanced World Models.

What it solves

OpenWorldLib provides a unified and standardized framework for world models, which are defined as perception-centered models equipped with interaction and long-term memory capabilities. It aims to eliminate the inconsistency of existing world-model-related codebases by integrating various open-source research outcomes into a single, well-structured repository.

How it works

The framework is organized into several core modules that handle different aspects of world modeling:

  • Pipelines: The primary entry point for users, handling model loading and environment initialization.
  • Operators: Process and validate input and interaction signals.
  • Synthesis: Handles generation tasks, including visual, audio, and VLA (Vision-Language-Action) generation.
  • Reasoning: Manages inference for spatial, general, and audio reasoning.
  • Representations: Manages 3D outputs and simulation environments.
  • Memory: Maintains running context for multi-turn interactions.

Who it’s for

Researchers and developers in the world model community who need a consistent way to invoke and test various world-model-related methods and integrate new research outcomes.

Highlights

  • Broad Task Support: Covers video generation (navigation, long video, and egocentric interaction), 3D scene generation, and reasoning (VQA and VLA).
  • Standardized Structure: Provides a clear codebase architecture for integrating diverse models like Matrix-Game-2, FlashWorld, and Pi-0.
  • Integrated Resources: Includes a curated list of cutting-edge world model research and a dedicated development guide for contributors.