| Definition | ViZDoom allows AI bot to play the game Doom using only visual information |
| No of environments | 9 |
ViZDoom is a powerful platform designed for developing AI bots that play the classic game Doom using only visual input from the screen buffer. Built on the ZDoom engine, it provides an engaging environment for research in machine visual learning and deep reinforcement learning. ViZDoom allows researchers to explore AI capabilities in a dynamic, visually rich setting, making it an ideal tool for studying how agents can learn to navigate complex environments through visual cues.
ViZDoom provides a variety of features to enhance AI training, such as customizable rendering options, access to depth buffers for 3D vision, and automatic labeling of visible game objects. Additionally, it includes options for recording episodes and scaling in-game time, which are essential for developing robust RL algorithms. Researchers can also access the audio buffer, map geometry, and a comprehensive list of actors and objects within the game.
Designed to be reinforcement learning-friendly, ViZDoom supports various learning paradigms, including apprenticeship learning and learning from demonstrations. With extensive documentation and examples available, particularly in Python using popular libraries like PyTorch and TensorFlow, ViZDoom serves as an invaluable resource for researchers looking to advance the field of AI in visually-driven contexts.
| Example observations | screen, depth, labels, automap and game variables (health and ammo) |
| Example actions | spawn, turn left, turn right, attack |
| Example reward | killing a monster, for every tic the agent is alive, agent death etc. |