VizDoom

VizDoom
Title	VizDoom
Developer	Machine Learning and Artificial Intelligence Laboratory
Initial release	2013
Programming language	C++, Python
Platform	Windows, Linux, macOS
Genre	Reinforcement learning environment, First-person shooter simulator

VizDoom

VizDoom is a research-oriented platform that exposes the classic first-person shooter engine used in experimental contexts to develop and evaluate artificial agents. It adapts the engine from a landmark 1993 title for use in contemporary machine learning, enabling controlled experiments in perception, decision-making, and control. VizDoom has been adopted across laboratories and competitions as a benchmark for visual reinforcement learning, imitation learning, and procedural content generalization.

Overview

VizDoom repurposes the engine underlying the commercial title developed by id Software into a programmable environment suitable for academics and practitioners. Early adopters included groups at institutions such as Massachusetts Institute of Technology, Stanford University, Carnegie Mellon University, and University of Oxford, which used the platform to compare algorithms developed in venues like NeurIPS, ICML, AAAI, and IJCAI. Comparisons with other environments such as Atari 2600-based benchmarks, Minecraft-based frameworks, and the OpenAI Gym suite highlighted VizDoom's capacity to combine complex visual input with low-latency control and map topology drawn from canonical levels. The project influenced subsequent efforts in embodied AI adopted by teams affiliated with DeepMind, Facebook AI Research, Google Research, and Microsoft Research.

Architecture and Features

VizDoom provides an application programming interface exposing state, action, and reward signals derived from the underlying engine. The core is written in C++ and offers bindings for languages used in labs and companies like Python (programming language), enabling integration with frameworks such as TensorFlow, PyTorch, Keras, and Theano. Its feature set includes frame buffer access, depth buffer, object labeling, and episodic control compatible with evaluation protocols described at conferences like ICLR and NeurIPS. Map manipulation and scenario scripting draw on formats and tools used by communities around Doom (1993 video game), WAD (file format), and level editors popularized by modders associated with id Software. Cross-platform builds support deployment on systems maintained by research groups at institutions like ETH Zurich, University of California, Berkeley, and University College London.

Gameplay Scenarios and Benchmarks

VizDoom exposes diverse scenarios ranging from simple navigation and item collection to adversarial combat and survival tasks drawn from canonical levels. Benchmark suites built on VizDoom were used in studies comparing model-free methods such as Deep Q-Networks from teams at DeepMind to policy-gradient approaches advanced by groups at OpenAI. Commonly reported tasks include deathmatch, health gathering, and visual navigation inspired by challenges curated for events like the Visual Doom AI Competition and academic workshops co-located with ICML and ECCV. Many comparative analyses reference evaluation practices established in literature from laboratories including University of Toronto, University of Washington, and Georgia Institute of Technology.

Research Applications

Researchers apply VizDoom across multiple subfields: reinforcement learning, imitation learning, unsupervised representation learning, curriculum learning, and multi-agent systems. Papers from labs at Princeton University, California Institute of Technology, Brown University, and Imperial College London have used VizDoom to test hierarchical reinforcement learning, intrinsic motivation schemes, and model-based planning. Multi-agent experiments often cite frameworks and findings associated with groups at Stanford University, MIT-IBM Watson AI Lab, and industrial teams at NVIDIA Research. VizDoom has also been used in human-in-the-loop studies referencing experimental protocols seen in publications from Harvard University and University of Cambridge.

Installation and Usage

Distribution packages and source code have been maintained to support installation on systems used in labs such as Cornell University and Purdue University. Typical workflows combine environment setup with deep learning stacks established by projects at Google Research and Facebook AI Research and use experiment management tools from groups at Uber AI Labs and Ray (software). Usage patterns documented by contributors from University of Edinburgh and University of Toronto include scenario scripting, headless rendering for cluster-based training on resources like those at National Energy Research Scientific Computing Center and integration with continuous integration systems used by teams at GitHub.

Community and Development

The VizDoom community includes researchers, modders, and competition organizers affiliated with conferences and institutions such as NeurIPS, ICLR, The PASCAL Visual Object Classes Challenge, id Software, and universities like University of Michigan. Development activity often coordinates through code hosting platforms used by projects at GitHub and discussion channels frequented by contributors from Stack Overflow and academic mailing lists managed by departments at Columbia University and Duke University. Competitions leveraging VizDoom have been sponsored and attended by teams from industry labs including DeepMind, OpenAI, Microsoft Research, and academic consortia across Europe, North America, and Asia.

Category:Reinforcement learning environments