The Procgen Benchmark provides 16 procedurally-generated, game-like Gym environments designed to measure the generalization capabilities of reinforcement learning agents. It offers a faster and more randomized alternative to Gym Retro, enabling rapid skill acquisition benchmarking.

How It Works

The environments are built using C++ for high performance, achieving thousands of steps per second on a single CPU core. They generate diverse levels on-the-fly, preventing agents from memorizing fixed sequences. Observations are 64x64 RGB pixel arrays, and actions are discrete button combinations.

Quick Start & Requirements

• Install via pip: pip install procgen
• Requires Python 3.7-3.10 (64-bit) and CPUs with AVX.
• Interactive testing: python -m procgen.interactive --env-name coinrun
• Docker images are available for basic setups.

Highlighted Details

• 16 distinct environments with unique gameplay mechanics.
• Customizable via source build for creating new environments.
• Supports Gym and Gym3 (vectorized) interfaces.
• Environment options include difficulty modes ("easy", "hard", "extreme", "memory", "exploration") and asset customization.

Maintenance & Community

The project is in maintenance mode, expecting bug fixes and minor updates. Contribution guidelines are available.

Licensing & Compatibility

The project is licensed under the MIT License, permitting commercial use and integration with closed-source projects.

Limitations & Caveats

Some environments have known issues where agents can spawn in unwinnable states or die immediately, which are intentionally left unfixed for reproducibility. The library is not fork-safe unless num_threads=0 is set, and Qt is not guaranteed to be fork-safe. GPU acceleration is not utilized.