BotHack – A Nethack Bot Framework
Written in Clojure but provides an API for bots in Java or other JVM-based languages. No game modifications are relied upon to make the game more accessible for a bot, so bots can play on public servers (like http://alt.org/nethack/).
Debugging the bot live in the Clojure REPL
15.3.2014: A basic GUI-less virtual terminal emulator is working, implemented using the JTA Telnet/SSH/Terminal library (http://www.javatelnet.org/). Manual interaction with a remote telnet terminal is possible via the emulator, the current frame for the terminal can be accessed programatically and displayed.
25.3.2014: A simple event-driven script (src/bothack/bot/nao_menu.clj) can interact with the nethack.alt.org game menu and start a game.
17.4.2014: A first trivial bot implementation can run around the first level blindly
until it starves and pray for food until it gets smitten. It can read game messages but doesn't yet understand any. Synchronization with NetHack is solved in a way that is only hinted at on the TAEB blog (http://taeb-nethack.blogspot.cz/2009/06/synchronizing-with-nethack.html) – so far it seems very reliable, but it is at the cost of more round-trips to the server.
27.4.2014: The trivial bot is now also implemented in Java as an example of the Java API use.
25.5.2014: The framework can parse the map for basic level layout info and allows for a slightly more complex bot that can explore the first level using A* pathfinding and charge monsters it encounters.
14.6.2014: The bot can navigate effectively as far as minetown/oracle, where it usually gets killed for lack of combat tactics. It can find hidden doors and passages and looks almost like a newbie human playing ;-) The framework provides basic monster tracking and shop recognition allowing the bot to deal with (avoid) peaceful monsters and shopkeepers quite nicely.
22.6.2014: When boosted in wizmode (NetHack debug mode) the testing bot can navigate autonomously as far as Medusa's lair and deal with all kinds of situations like boulders blocking the way. Item handling will need to be implemented before the bot can progress further.
A ttyrec of one Medusa run is in the repo: https://github.com/krajj7/BotHack/blob/master/ttyrec/wizmode-exploration-dlvl1-28medusa.ttyrec?raw=true
14.7.2014: The framework can now recognize all types of monsters found in the dungeon and provides information about their properties. The raw data was adapted from the Saiph bot. Monsters that have an ambiguous representation (like dwarf king/mind flayer) are queried automatically with the FarLook command, so bots don't have to worry about ambiguities or possibly-peaceful monsters at all.
26.7.2014: Navigation has been expanded to allow navigating (either exploring everything fully or going as fast as possible) to any given branch – so far only above the Castle – or special level (Minetown, Oracle, Medusa's lair, ...). Unvisited branches and levels are automatically searched for on appropriate levels.
Item handling is still missing so bots can't cross Medusa's yet, but in wizmode the testing bot can use the new navigation API to easily explore dungeon branches above the Castle in any desired order, without having to explicitly deal with finding the right stairs/portal etc.
31.8.2014: Just about all item types found in NetHack are now recognized by the framework and their properties are available for querying, including monster corpses, statues, tins and everything. Much of the data and some handy regexes were adapted from the TAEB libs. Items with fixed appearances are auto-identified (e.g. a polished silver shield is always a shield of reflection) and unusual appearances are normalized (japanese names for the Samurai, plurals). However, actions for interacting with items are not yet implemented and the framework currently doesn't help with item identification like dropping rings into sinks, engraving with wands, price identification or elimination, but all of this is on the roadmap.
22.9.2014: Many actions for interacting with items in the dungeon are now implemented and the pathing subsystem optionally makes use of pick-axes and levitation items to navigate more effectively. When pre-equipped in wizard mode, the bot can now pass through Medusa's island, the Castle and all the way down through Gehennom. Important features of all Gehennom demon lairs (static traps and secret doors) are mapped automatically.
4.10.2014: Eat action is implemented, allowing both eating from the ground or inventory, and the framework tracks monster deaths, corpse freshness/safety and player intrinsics, so the bot can feed itself on corpses for nutrition and useful resistances.
23.10.2014: The framework now finally has at least the features and robustness needed to allow a bot pre-equipped with overpowered gear in wizard mode to beat the game full-auto. The bot gathers all the invocation artifacts manually, not harming any peacefuls along the way and not using any wizard mode commands itself. A lot of functionality that will be necessary for a non-wizmode bot to play sucessfully is however still missing. Progress toward that end will hopefully be swift on the current foundation.
Video of a wizmode ascension (6x speedup): https://www.youtube.com/watch?v=fpCGivGrUVw
Source ttyrec: https://github.com/krajj7/BotHack/raw/master/ttyrec/wizmode-fullauto-asc.ttyrec [31MB, realtime duration 1h42m]
14.11.2014: A new non-wizmode playing bot can now execute a decent early-game strategy, it can deal with common early-game hazards like floating eyes, get armor by robbing dwarves, dip for Excalibur and survive by fighting cautiously and abusing Elbereth. Inventory management is still basic and the bot ignores many potentially useful items.
24.11.2014: The framework now uses shop prices and other observed properties of items like results of wand-engrave to disambiguate random item appearances. Bots can ask about the possible types of unidentified items and will only get results that match the gathered data. A stroll through a shop is all that is needed to identify useful items (like the scroll of identify). This is implemented using the Clojure core.logic library for logic programming (src/bothack/itemid.clj).
6.12.2014: The bot can now solve every level of sokoban and makes use of some of the items that can be found there. The sokoban solver module can deal with interruptions from monsters and other perils. The prize and giant mimics on the last level are auto-identified.
27.12.2014: The main example bot can reach the castle fairly regularly now, it has finished the quest and cleared many levels of gehennom including Vlad's (without wizard mode or human intervetion), eventually dying to Orcus-spawned Demogorgon. This makes it the most successful NetHack bot written as of yet.
Dumplog of the bot's best game to date: https://nethack.xd.cm/userdata/smartbot/nethack/dumplog/1419682524
Listing of other bots for comparison: http://taeb.github.io/bots.html
Special thanks goes to kerio and stenno, admins of the nethack.xd.cm public server, for their support and hosting an instance of the bot and to FIQ for running the bot on the nethack.alt.org server and trying out the framework even in the current unfinished state!
11.1.2015: A first fully documented version of the Java API is ready, making most of the functionality of the framework available for consumption from Java. The framework itself is available at the Clojars public Maven repository. See links on top of this page for the docs.
25.1.2015: Pudding farming was implemented and the bot has finally managed to win the game without wizmode cheating!
Youtube video: https://www.youtube.com/watch?v=unCQHAbGsAA