PokemonEval Server

A server for evaluating AI agents on Pokemon Red gameplay.

Overview

The PokemonEval server provides a RESTful API for agents to interact with Pokemon Red via the PyBoy emulator. This server component handles game state management, action execution, and provides a clean interface for remote evaluation of agents.

Installation

1. Clone the repository:

git clone https://github.com/yourusername/PokemonEval.git
cd PokemonEval

2. Create a virtual environment and install dependencies:

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate
pip install -r requirements.txt

3. Place a Pokemon Red ROM file in the root directory or specify its path with --rom.

Starting the Server

Run the server with:

python -m server.evaluator_server --rom path_to_your_rom

Command-line arguments:

• --host: Host address to bind the server to (default: 0.0.0.0)
• --port: Port to listen on (default: 8000)
• --rom: Path to the Pokemon ROM file (default: "Pokemon_Red.gb")
• --reload: Enable auto-reload for development

Server API

Initialize Environment

POST /initialize

Request body:

{
  "headless": true,
  "sound": false
}

Response:

{
  "player_name": "RED",
  "rival_name": "BLUE",
  "money": 3000,
  "location": "PALLET TOWN",
  "coordinates": [5, 6],
  "badges": ["BOULDER", "CASCADE"],
  "valid_moves": ["up", "down", "left", "right"],
  "inventory": [
    {"item": "POTION", "quantity": 3},
    {"item": "POKEBALL", "quantity": 5}
  ],
  "dialog": "PROF.OAK: Hello, welcome to the world of POKEMON!",
  "pokemons": [
    {
      "nickname": "PIKACHU",
      "species": "PIKACHU",
      "level": 5,
      "hp": {"current": 20, "max": 20},
      "moves": [{"name": "TACKLE", "pp": {"current": 35, "max": 35}}]
    }
  ],
  "screenshot_base64": "base64_encoded_image_data",
  "collision_map": "text_representation_of_collision_map",
  "step_number": 0,
  "execution_time": 0.123
}

Take Action

POST /action

Request body (for pressing a key):

{
  "action_type": "press_key",
  "keys": ["a"]
}

Request body (for waiting):

{
  "action_type": "wait",
  "frames": 60
}

Response: Same format as the initialize endpoint

Check Status

GET /status

Response:

{
  "status": "running",
  "steps_taken": 42,
  "average_action_time": 0.156
}

Stop Environment

POST /stop

Response:

{
  "status": "stopped",
  "steps_taken": 42
}

Action Types

The server supports two types of actions:

1. Press Key

Used to press one or more Game Boy buttons in sequence.

{
  "action_type": "press_key",
  "keys": ["a"]
}

Valid keys:

• "a" - A button
• "b" - B button
• "start" - Start button
• "select" - Select button
• "up" - D-pad Up
• "down" - D-pad Down
• "left" - D-pad Left
• "right" - D-pad Right

You can send multiple keys in a sequence:

{
  "action_type": "press_key",
  "keys": ["up", "up", "a"]
}

2. Wait

Used to wait for a specified number of frames.

{
  "action_type": "wait",
  "frames": 60
}

Where:

• frames is the number of frames to wait (at 60 FPS, 60 frames = 1 second)

Game State Response

Each action returns a game state response with the following information:

• player_name: Player character name
• rival_name: Rival character name
• money: Current money
• location: Current location name
• coordinates: Player [x, y] coordinates
• badges: List of obtained badges
• valid_moves: List of valid movement directions
• inventory: List of items and quantities
• dialog: Any active dialog text
• pokemons: List of Pokemon in the party with stats
• screenshot_base64: Base64-encoded screenshot
• collision_map: Text representation of collision map
• step_number: Current step number
• execution_time: Time taken to execute the action

Client Implementation Example

Here's a simple Python example for interacting with the server:

import requests

# Server URL
SERVER_URL = "http://localhost:8000"

# Initialize environment
response = requests.post(
    f"{SERVER_URL}/initialize",
    json={"headless": True, "sound": False}
)
state = response.json()

# Press a button
response = requests.post(
    f"{SERVER_URL}/action",
    json={"action_type": "press_key", "keys": ["a"]}
)
state = response.json()

# Wait for frames
response = requests.post(
    f"{SERVER_URL}/action",
    json={"action_type": "wait", "frames": 60}
)
state = response.json()

# Stop environment
requests.post(f"{SERVER_URL}/stop")

Notes

• The server maintains session data and images in the gameplay_sessions directory
• All screenshots are saved in the session's images folder
• Gameplay data is logged to a CSV file in the session directory