Slither.io Gymnasium Environment

This project implements a Gymnasium environment for a Slither.io-style game, designed for training Reinforcement Learning models. It features continuous movement, and enemy bots.

As part of the project I also trained several rl-algorithms on the environment, including PPO with random network distillation for intrinsic rewards.

Installation

Using uv (Recommended)

uv is a fast Python package installer and resolver.

1. Create and activate a virtual environment:

   uv venv
   source .venv/bin/activate 
   # On Windows, use: .venv\Scripts\activate

2. Install the package and its dependencies:

   uv pip install -e .

Using pip

1. Create and activate a virtual environment (optional but recommended):

   python -m venv .venv
   source .venv/bin/activate 
   # On Windows, use: .venv\Scripts\activate

2. Install the package and its dependencies:

   pip install -e .

Playing the Game

Run the play.py script to interact with the environment:

python play.py

Controls

• Mouse Movement: Controls the snake's target direction. The snake will smoothly turn towards the cursor.
• Q or ESC: Quit the game.

Environment Specifications

• Action Space: Continuous, Box(low=-1, high=1, shape=(2,), dtype=np.float32). Represents [cosine_value, sine_value] for the target direction.
• Observation Space: Image, Box(low=0, high=255, shape=(H, W, 3), dtype=np.uint8). Default is 84x84x3, but configurable via the screen_size parameter in gym.make().
• Rewards:
  • Change in Length: +C1 * (current_length - previous_length)
  • Eliminating an Opponent: +C2 * (opponent_snake_length)
  • Death (collision with wall or longer/equal length snake): -C3
  • (Constants C1, C2, C3 are defined in SnakeEnv class, currently 0.1, 10, 100 respectively)

Using the Environment with Gymnasium

import gymnasium as gym
import numpy as np
import snake_env # Registers 'Snake-v0'

# Environment parameters can be customized:
env_params = {
    "world_size": 3000,
    "snake_segment_radius": 10,
    "num_bots": 3,
    "num_foods": 15,
    "screen_size": 84,  # For the observation image dimensions (HxW)
    "zoom_level": 1.0   # 1.0 = normal, >1.0 = zoomed out, <1.0 = zoomed in
}
env = gym.make('Snake-v0', **env_params)

observation, info = env.reset()

for _ in range(1000):
    # A simple random action:
    random_angle = np.random.uniform(0, 2 * np.pi)
    action = np.array([np.cos(random_angle), np.sin(random_angle)], dtype=np.float32)
    
    observation, reward, terminated, truncated, info = env.step(action)
    
    print(f"Reward: {reward:.2f}, Length: {info['snake_length']}, Score: {info['score']}")

    if terminated or truncated:
        print("Episode finished!")
        observation, info = env.reset()

env.close()

Key Game Mechanics

• Camera: Always centered on the player's snake head. The zoom_level parameter affects how much of the world is captured in the observation.
• Movement: Continuous. The snake attempts to turn towards the angle specified by the action vector, limited by its current agility (which depends on its length).
• Collisions:
  • Player snake dies if its head collides with a wall.
  • Player snake dies if its head collides with an enemy snake's body.
  • In head-to-head collisions between player and bot:
    • If player is longer: bot dies, player gets reward.
    • If bot is longer or equal length: player dies.
• Growth: Eating food increases the snake's length and score. Bot snakes also eat and grow.
• Bots: AI-controlled snakes that seek food, avoid walls, and can be eliminated. They respawn after a delay.