Download
Procgen (short for Procedural Generation Benchmark) is a suite of 16 procedurally generated, game-like reinforcement learning environments designed to evaluate generalization and sample efficiency in RL agents. Unlike fixed, deterministic environments, Procgen generates new levels (layouts, obstacles, visual variation) each episode, making it impossible for an agent to simply memorize trajectories. The environments are designed to run very quickly (thousands of steps per second on a single core) to facilitate large-scale experiments and make benchmarking efficient. The benchmark supports both “easy” and “hard” difficulty modes, letting researchers trade off computational cost vs challenge. The repo provides a C++ core for game logic and rendering (with support for gym/Gym3 wrappers) plus Python bindings and interactive mode for human play testing.
Features
- 16 diverse, procedurally generated environments (platformers, maze, shooter, etc)
- High performance: environments engineered to run extremely fast per step
- Difficulty modes (easy / hard) to accommodate compute budgets
- C++ core logic wrapped for Python/Gym and Gym3 compatibility
- interactive mode for human play, useful for debugging and visualization
- Designed to test generalization rather than memorization in reinforcement learning
Project Samples
