What it does

This repo trains a Deep Q-Network to play Tetris using PyTorch. You run train.py to teach the agent from scratch, then test.py to watch it stack blocks. The author provides a pre-trained checkpoint in trained_models/tetris if you’d rather skip the GPU hours.

The interesting bit

It’s deliberately basic — the README calls it “a very basic example of Reinforcement Learning’s application” — which makes it a decent sandbox for understanding DQN without wading through a full RL framework. The agent learns from raw game frames (via OpenCV) rather than hand-engineered board features, which is the harder and more instructive path.

Key highlights

• Pure PyTorch, no RL library abstraction layers
• Learns from pixel input using cv2 for frame processing
• Includes pre-trained model and YouTube demo for quick validation
• ~500 stars suggests it has served as a reference implementation for others
• Clean separation: train.py and test.py, nothing clever in the structure

Caveats

• README is sparse: no mention of reward shaping, network architecture, or training duration
• Requires Python 3.6 specifically, which is end-of-life
• No performance metrics, convergence graphs, or comparison to random play — you can’t tell if the agent is actually good or just lucky

Verdict

Grab this if you want a minimal, hackable DQN example for Tetris. Skip it if you need a production RL pipeline, modern Python support, or any theoretical depth — this is glue code between PyTorch and a Tetris environment, and it knows it.