procgen — Procedural Game Environments for RL

procgen — Procedural Game Environments for RL

procgen is OpenAI's suite of 16 fast procedurally-generated 2D game environments designed to benchmark generalization in reinforcement learning. Each environment — CoinRun, StarPilot, CaveFlyer, Dodgeball, Fruitbot, Chaser, Miner, Jumper, Leaper, Maze, BigFish, Heist, Climber, Plunder, Ninja, and BossFight — generates a virtually unlimited number of unique levels, making it impossible for agents to memorize solutions and forcing them to generalize.

Key Features

• 16 visually rich 2D game environments, each with unlimited procedural level generation
• Extremely fast C++ core: 5,000+ steps/second per environment
• Gymnasium-compatible API
• Configurable difficulty, number of training levels, and distribution shift between train/test
• Standard benchmark for measuring sample efficiency and generalization in RL research

Quick Start

pip install procgen

import gymnasium as gym

# Train on a fixed set of 200 levels, test on all levels
env = gym.make("procgen:procgen-coinrun-v0",
               num_levels=200,
               start_level=0,
               distribution_mode="easy")
obs, info = env.reset()
for _ in range(1000):
    action = env.action_space.sample()
    obs, reward, terminated, truncated, info = env.step(action)
    if terminated or truncated:
        obs, info = env.reset()
env.close()

npx ai-supply add procgen-procedural-game-environments

Curated mirror of the open-source procgen (MIT). Get it from the source.