Procgen Benchmark

Procgen Benchmark provides 16 simple-to-use, high-speed procedurally-generated game environments (e.g. coinrun, bigfish, starpilot, maze) exposed through the OpenAI Gym and gym3 APIs. Every environment renders a 64x64x3 RGB pixel observation and accepts one of 15 discrete button-combo actions. Because levels are generated from random seeds (up to 2^31 unique levels per game via num_levels=0), agents cannot memorize a fixed sequence of actions, so the benchmark directly measures how quickly an RL agent learns generalizable skills. Each game defines its own per-step reward, typically small rewards for sub-goals plus a large level-completion bonus; for cross-game comparison, episodic returns are min-max normalized using per-game Rmin/Rmax constants published in the paper. The environments are associated with the ICML 2020 paper 'Leveraging Procedural Generation to Benchmark Reinforcement Learning' and were also used in the NeurIPS 2020 Procgen competition. The Python code is MIT-licensed; the bundled game art assets carry their own permissive licenses.