This repository provides a structured, hands-on educational course for understanding and building AI world models. It targets engineers and researchers seeking to grasp latent dynamics, simulation, and planning, offering a practical path from foundational concepts to implementing evaluation systems.

How It Works

The course employs a structured curriculum comprising five lectures and five hands-on projects. It guides learners from foundational concepts like Craik's mental models and predictive coding through observation encoding using VAEs and CNNs. Core topics include latent dynamics modeling with GRU, MDN-RNN, and RSSM architectures, alongside exploring seven distinct architecture families and learning paradigms. Planning mechanisms such as CEM-MPC, latent Actor-Critic, and TD-MPC are detailed, culminating in the implementation of an evaluation dashboard using metrics like FID, reward correlation, and PSNR. The approach prioritizes practical implementation by encouraging learners to build and experiment.

Quick Start & Requirements

• Documentation Site: This repository hosts a documentation site for the "Learn World Model" course.
  • Install dependencies: npm install
  • Run development server: npm run docs:dev
  • Preview production build: npm run docs:preview
• Prerequisites: Node.js is required for local development of the documentation site. The course references PyTorch source code within external/world-model-tutorial/.
• Links: Read Online

Highlighted Details

• Comprehensive curriculum covering VAE, RSSM, and Transformer backbones, with detailed explanations of planning mechanisms like CEM-MPC, latent Actor-Critic, and TD-MPC.
• In-depth exploration of world model evaluation metrics including FID, reward correlation, consistency loss, PSNR, and horizon drift.
• Emphasis on practical implementation through five guided projects, allowing learners to build and visualize components like VAE latent spaces and RSSM rollouts.
• Content is presented in both English and Chinese, with a structured learning path and official online reading available.

Maintenance & Community

• Community: WeChat discussion group available (QR code in README).
• Contributors: Project leads include Zhimin Zhao (Queen's University) and Qi Wang (Shanghai Jiao Tong University).
• Contribution Guidelines: Strict style rules (CLAUDE.md) are enforced for content contributions.

Licensing & Compatibility

• License: MIT License. Permissive for commercial use and integration.

Limitations & Caveats

The project is an "Alpha Preview," indicating that content, examples, and wording are still under active development and subject to change.