All Tools
P
AgentsFreeOpen Source
PROCGEN
Procedurally generated game environments for RL generalization
MIT
ABOUT
Fixed-level RL games let agents memorize maps instead of learning transferable skills, so reported scores overstate generalization. Procgen randomizes CoinRun, Maze, BossFight, and related games at every episode so researchers can measure whether a policy actually generalizes to unseen levels.
INSTALL
pip install procgenINTEGRATION GUIDE
1. Benchmark PPO or IMPALA on Procgen easy and hard generalization splits
2. Test whether an exploration method transfers across procedurally generated mazes
3. Train a pixel-based agent that must handle new CoinRun or StarPilot levels
4. Compare sample efficiency of RL algorithms on 16 shared game-like environments
TAGS
pythonreinforcement-learninggymprocgengeneralizationopenai