MiniHack

MiniHack is a sandbox for open-ended reinforcement-learning research that uses the NetHack Learning Environment (NLE) to communicate with the game of NetHack. Environments are defined with NetHack description ('des') files, a human-readable probabilistic domain-specific language, allowing precise control over map terrain, monsters, traps, and objects while injecting randomness that stresses generalization. Tasks fall into two families: MiniHackNavigation (restricted to the 8 compass-movement actions, default goal of reaching the down staircase) and MiniHackSkill (the full NetHack command action space, inventory observations, for skill-acquisition tasks). Observations are returned as a dictionary of NumPy arrays (symbolic glyph/char/color map layers, agent-centred crops, bottom-line stats, and the in-game message), with optional pixel and TTY renders. By default the reward is sparse (+reward_win, default +1, on task success), and custom event-based reward functions can be composed via a RewardManager. The repository is Apache-2.0 licensed and was published by FAIR, UCL, and Oxford (NeurIPS 2021 Datasets and Benchmarks Track); the project has since relocated to github.com/samvelyan/minihack.