TurboMino: Tetris AI Bot with PPO + CNN + RoPE
Apr 2026
In progress
Tetris bot that plays autonomously using deep reinforcement learning. The Tetris environment is implemented as a Gymnasium gym with configurable board dimensions. Board representation uses bit-packed uint32 arrays per row, decoded by _extract_features_2d via bit shifts. The RL model is a CNN with Rotary Position Embeddings (RoPE) trained with PPO from Stable-Baselines3 + sb3-contrib. Includes MoveSearcher that exhaustively explores all piece rotations and translations via BFS. Interactive mode with Pygame (python app/main.py play-tetris). Supports padding to max board size, horizontally centered with top anchor. Training data sourced from Tetr.io top player replays.
AI
Jupyter
NumPy
Pandas
Pygame
Python
PyTorch
Reinforcement Learning
