ViZDoom

Why we included this project

ViZDoom is a lightweight research environment built around the 1993 Doom engine, where reinforcement learning agents learn from the raw screen buffer rather than hand-crafted state features. It has been used in the RL community long enough to become a familiar benchmark, with its own competitions and a peer-reviewed paper in IEEE Transactions on Games describing the setup. Python bindings and Gymnasium wrappers make it easy to drop into existing training pipelines, and custom scenarios are assembled with a scripting language and visual editors, so teams can design benchmarks around specific skills like navigation, object recognition, or coordination between agents. It also exposes the depth buffer, automatic object labeling, and audio, which helps when experiments go beyond pixel inputs.