A program that plays Puyo Puyo:

...not in an emulator, it plays for real. Instead of eyes, it has a video capture card. Instead of hands it has an Arduino plugged into a controller slot.

The game is Puyo Puyo, a competitive tetris-like tile matching game. Planning combos is essential to playing the game well, so it's perfect to write an AI for. The version I use is called "Dr. Robotnik's Mean Bean Machine" on the Gamecube (in the "Sonic Mega Collection"). That's just the version I happen to have and test with.

First of all, you'll need:

• python2.7
• pyserial
• numpy
• OpenCV with the cv2 Python bindings

TODO

• Arduino Setup: TODO
• Video Setup: TODO

run.py is the main script that runs all components in tandem. There are also a bunch of scripts to manually test components. Use -h or --help to get details on usage.

• run.py - Plays the game. Recognizes game state through video, decides what move to make, and makes it.
• gc_send.py - For testing Gamecube communication. Sends a command to the arduino controlling the Gamecube.
• simulate_ai.py - For testing AIs. Reproduces the mechanics of the game and lets the AI make moves.
• simulate_board.py - For testing the game mechanics. Allows the user to test the board mechanics by placing one piece at a time.
• recognize_board.py - For testing the vision processing. Takes video input and shows the reconstructed game state.
• playback_board.py - Playback a board recording file produced by recognize_board.py's -o option.

The computer vision component takes raw video from a file or video capture device and outputs the current state of the game. The game state consists of a grid of beans, plus a pair of next beans.

See:

• puyo/beanfinder.py - Stateless component that finds where beans are placed in the game given a single video frame.
• puyo/vision.py - Stateful component that gets higher level information by keeping track of the game between frames.

Whenever a pair of beans start to fall, the AI is called and given the current game state. It determines the best move to make, and returns a (position, rotation) pair representing the move.

See: puyo/ai.py

After the AI returns a move, the move is transmitted via serial to an Arduino, which plugs into the Gamecube. It mimicks a real controller by implementing the physical controller protocol. Once the Arduino has executed the move, the computer vision component continues to watch the raw video.

The code for the game control is split between the microcontroller and the python code:

• arduino/gamecube_control/
• puyo/gccontrol.py

Currently it can beat the game, but uses a few continues to make it happen. AI itself needs the most improvement. Here's the status progress of each major component:

• Vision Processing - Works well including recognizing special game states such as winning or losing a match.
• Artificial Intelligence - The best AI right now works well, but makes some stupid decisions that can lead to its own death.
• Game Control - Supports the basics: moving puyo pieces and pressing individual buttons.

Milestones:

1. Simple AI controlling the game
2. AI that can beat the early opponents
3. AI that can beat me (I'm not that good)
4. AI that can beat the final boss, Dr. Robotnik
5. AI that can beat the final boss without using a continue