gym-zelda-pygame 0.1.8

A ROM-free Zelda-like Gymnasium environment built on Pygame.

gym-zelda-pygame

A Gymnasium environment that wraps the Clear Code Pygame “Zelda” game into a ROM-free, modern RL playground. It exposes a pixel-observation interface, discrete actions, sensible rewards, and familiar wrappers (skip-frames, gray-scale resize to 84×84, frame-stack, normalize, channel-first) so you can train DQN/PPO/IMPALA-style agents without touching emulators or legacy Gym.

Motivation

I recently finished a playthrough of The Legend of Zelda: Breath of the Wild and wanted to experiment with applying reinforcement learning to a Zelda-style world. Most existing environments either:

• rely on ROMs/emulators, or
• are tangled up with deprecated gym + NumPy compatibility issues.

Environment Details

Spaces

• Action Space: Discrete(8) 0) no-op

  1. move up
  2. move down
  3. move left
  4. move right
  5. attack
  6. magic
  7. menu

• Observation Space: Box(low=0, high=255, shape=(720, 1280, 3), dtype=uint8) by default
  (Typically wrap this 84×84 grayscale and stack frames for learning)

Rendering

• render_mode="human": Pygame window (60 FPS throttle)
• render_mode="rgb_array": returns the current frame as a numpy array
• Headless safety: when not using "human", SDL dummy drivers are used automatically.

Episode Termination

• Episode ends when:
• Player health ≤ 0, or
• All enemies defeated

Reward

Default (configurable) reward shaping in ZeldaEnv:

• Small step penalty: -0.001 per step (encourage efficiency)
• Health changes: penalty for loss, small reward for gain
• Experience gain: small positive reward
• Enemy defeats: large positive reward (+10.0 per enemy)
• Death penalty: -100.0

You can pass a custom reward function via custom_reward_fn:

def my_reward_fn(current_state, previous_state, episode_steps):
  # current_state: dict with keys shown below
  # previous_state: same keys for previous step
  # return float
  ...
env = ZeldaEnv(render_mode="rgb_array", custom_reward_fn=my_reward_fn)

Info Dict

Each step() returns an info dict with:

{
  'episode_steps': int,
  'player_health': int,
  'player_exp': int,
  'enemy_count': int,
  'player_pos': (x: int, y: int),
}

Credit

• Game code adapted from Clear Code’s Zelda tutorial: clear-code-projects/Zelda
• Graphics and audio assets courtesy of Pixel-boy/AAA under CC0: Ninja Adventure asset pack

Respect the original projects’ licenses and assets. This environment is an RL wrapper around their codebase to support education.