PPO-PyTorch

Why we included this project

This repo is a small, readable PyTorch implementation of the clipped-objective PPO algorithm, built for OpenAI Gym environments. The whole training loop, actor-critic networks, and Monte-Carlo advantage estimation sit in a few hundred lines, so you can follow each piece without a framework in the way. It deliberately leaves out GAE, parallel workers, and learnable action variance, and the README explains why each simplification was made, which turns the code into a teaching tool. A Colab notebook plus utilities for logging, plotting curves, and rendering gifs from checkpoints make it easy to run classroom or self-study experiments. It won't replace a production library like Stable-Baselines3, but it's a clear map of how PPO actually works.