Jericho

Jericho is an open-source Python learning environment, released by Microsoft Research, that connects reinforcement-learning agents to human-made interactive-fiction (text adventure) games running on a modified Frotz Z-machine interpreter. Agents interact by issuing free-form natural-language commands and receive the game's textual response; the immediate reward at each step is the change in the game's built-in point score (reward = score - old_score). The library ships bindings for 57 IF games (e.g. zork1, zork3, detective, balances, ztuu, 905), each with a reproducible walkthrough, an action-template grammar, and parser metadata, although the copyrighted game ROM files themselves must be obtained separately. Jericho also provides tooling for extracting the world object tree, the recognized game dictionary, valid-action sets, and save/restore of the raw emulator state. The 32-game benchmark subset and study accompany the AAAI 2020 paper 'Interactive Fiction Games: A Colossal Adventure'. It is distributed under the GNU General Public License v2.0 or later.