About
Flappy Bird for Gymnasium is a Python package that wraps a Flappy Bird implementation as a Gymnasium reinforcement-learning environment. It offers two observation modes — a 180-point LIDAR-style sensor or a set of numeric features (pipe positions and player position, velocity and rotation) — a two-action action space (do nothing or flap), and a reward scheme with small per-frame survival rewards, a bonus for passing pipes, and penalties for dying or hitting the ceiling.
How it differs from the original
Unlike a game meant for a human to play, this project is structured primarily as a machine-learning research environment, exposing the game through Gymnasium's observation/action/reward interface. Its command-line tool adds random-agent and trained-agent play modes alongside ordinary human play.
Getting started
It can be installed via pip. Running its command-line tool plays the game by hand by default; the same tool also has modes to watch a random agent or a trained deep-Q-network agent play instead, and it can be imported as a standard Gymnasium environment for training your own agent.
Additional notes
- Credit
- The game's logic and graphics are based on the flappy-bird-gym project by GitHub user Talendar.
- External link
- One of the LIDAR observation options is based on the paper 'Playing Flappy Bird Based on Motion Recognition Using a Transformer Model and LIDAR Sensor.'