Reaver is an open-source deep reinforcement learning framework built around modular, reusable components, best known for its StarCraft II agents. It implements popular actor-critic algorithms in TensorFlow and works with OpenAI Gym environments, Atari and MuJoCo in addition to PySC2. The GitHub repository has around 560 stars; the codebase dates from the late 2010s and was last updated in 2020.
Reaver was designed to make deep RL experiments faster to run and easier to reproduce. Its architecture splits agents, models, environments and memory into interchangeable modules, so researchers can swap an algorithm or a game without rewriting the training loop. Out of the box it ships competitive baselines for StarCraft II mini-games โ tasks like MoveToBeacon, CollectMineralShards and DefeatRoaches โ where its agents reached strong scores with far less code than the original DeepMind baselines, as well as Atari results competitive for its era. Configuration is handled through clean command-line flags, runs are logged for comparison, and the whole stack leans on TensorFlow 2 with eager execution. While no longer actively developed, it remains a compact, readable reference implementation for actor-critic methods.
Reaver’s results were notable for how little code they required: its StarCraft II agents approached far larger codebases while training on a single GPU, which made the framework popular in university courses and reinforcement learning reading groups. Launch scripts and configuration files for every bundled environment are included in the repository, so reproducing a published baseline is a single command.
Reaver is completely free and open source under the MIT license, with no paid tier or hosted version.
Reinforcement learning students and researchers who want a clean, hackable reference framework for actor-critic algorithms and StarCraft II experimentation.