The PPO-Only School of Decision Intelligence

What it does

PPOxFamily is an eight-chapter open course (with Bilibili video lectures) that treats Proximal Policy Optimization as a universal entry point to decision intelligence. Each chapter pairs theory—slides, manuscripts, and supplementary reading—with annotated Python code samples and homework assignments that have published solutions. The material walks from basic policy gradients through advanced topics like hybrid action spaces, curiosity-driven exploration (ICM, RND), and multi-agent coordination.

The interesting bit

Instead of surveying every RL algorithm, the course bets that you can solve most decision-AI problems by deeply understanding just one. It is a pedagogical stunt: the instructors keep adding extensions—Pop-Art value scaling, representation learning wrappers, temporal modeling—onto the same PPO backbone, turning a tutorial repo into something closer to a systematic curriculum.

Key highlights

• Eight chapters covering action spaces, multi-modal observations, sparse rewards, temporal modeling, and multi-agent systems
• Every chapter ships with lecture PDFs, code samples (ppo_zh.py, hybrid_tutorial_zh.py, curiosity_icm.py), homework, and worked solutions
• An annotated “algorithm-to-code” documentation site mapping math to implementation
• Real-world application threads (rocket recovery, soft-body robots, autonomous driving) linked in GitHub issues
• Course materials and datasets mirrored to HuggingFace for offline access

Caveats

• Content is delivered in Chinese; videos are hosted on Bilibili, which may require workarounds for some regions
• The repository is a course artifact, not a maintained RL framework—expect tutorial-grade code, not production abstractions

Verdict Worth bookmarking if you are a Mandarin-speaking developer or researcher who wants a structured, code-heavy path from zero to multi-agent PPO. If you need a drop-in training framework or English-language instruction, look elsewhere.