Skip to main content

Toribash as an agent learning environment

Project description

Mating-ritual of two random agents PPO ripping a Uke a new one, and feeling the scene Another PPO attacking Uke in more random situations


Toribash Learning Environment. Extra "L" to make words more memorable for Finns ("Torille" = "To the marketplace").

ToriLLE provides learning agents an interface to video-game Toribash, a humanoid fighting game. Toribash provides environment for MuJoCo-like humanoid control, specifically aimed for competitive gameplay. This makes ToriLLE suitable for e.g. self-play experiments.

ToriLLE comes with a Python interface and pre-made OpenAI Gym environment with various tasks. Following white-paper includes baseline experiments and benchmarks conducted using ToriLLE:


Tested to work on Windows 10, Ubuntu 16.04 and MacOS 10.13. Tested on Python versions 3.5 and 3.6, and will likely not work on 2.7.

  • Numpy (Python)
  • FileLock (Python)
  • Wine (For Linux/MacOS. Requires modern version. Tested on 3.0.3)


Remember to install Wine if you are on Linux or MacOS. Make sure wine command is defined.

Following will download ToriLLE with stripped down version of Toribash:

pip install torille

Random agent:

from torille import ToribashControl
from torille.utils import create_random_actions

# Show gameplay
toribash = ToribashControl(draw_game=True)

# Random agent
while 1:
    state, t = toribash.get_state()
    if t: break

OpenAI Gym environment:

import gym
import torille.envs

env = gym.make("Toribash-DestroyUke-v0")
# Show gameplay

initial_state = env.reset()
t = False
# Random agent
while not t:
    s, r, t, _ = env.step(env.action_space.sample())

Manual installation

You can install ToriLLE without PyPI/pip with the following:

  • Install Toribash (note: Only Steam version may be up to date)
  • Copy contents of toribash-codes to Toribash installation directory. Overwrite files
    • Note: This will prevent using that specific installation as a regular game. Remove/rename profile.tbs file to revert most of the changes and use game normally again.
    • Note2: Starting with Toribash 5.4 / Steam version of Toribash, settings file is stored in under user's directory at Saved Games/Toribash/custom.cfg. Toribash loads this file by default if it finds it, which may cause ToriLLE to run in wrong settings.
  • Provide path the installed toribash.exe when creating ToribashControl objects (if you use provided Python library)

Playing in multiplayer

Note: Multiplayer does not work on Linux Wine!

Want to try your agents against human players in multiplayer? Check how manual remote control works.


Examples in examples provide quickstart to how to use ToriLLE, and also show how to apply settings or record replays.

For references see:

For troubleshooting, see the FAQ.

If you wish to modify ToriLLE or use other language to control Toribash instance, see hacking and documentation on protocol.

Repository structure

  • ./torille/: Python codes for the learning environment (inc. Gym environment)
    • ./torille/toribash: This will include stripped version of the Toribash game when installed from PyPi
  • ./toribash-codes/: Files required for Toribash to make this learning environment work
  • ./examples/: Python examples on how to use this library
  • ./docs/: Detailed documentation of the inner workings
  • ./experiments/: Codes used to run experiments in the white paper
  • ./images/: Contains GIFs used here

Related work and useful links


We wouldn't mind a citation if you find ToriLLE useful in your work. It also helps us to see what people have been up to!

  author = {Anssi Kanervisto and Ville Hautam{\"a}ki},
  title = {ToriLLE: Learning Environment for Hand-to-Hand Combat},
  year = {2018},
  journal = {arXiv preprint arXiv:1807.10110},

Special thanks & Acknowledgements

  • hampa and Dranix for invaluable help with configuring Toribash and lua scripts (also for developing the game in the first place!)
  • Siim Põder (user "windo" on GitHub) for original toribash-evolver code
  • box (Toribash user) for comments during inception of this project


Code original to ToriLLE is licensed under GNU GPL 3.0. Toribash is property of Nabistudios. Toribash binary in PyPI package is shared with the permission of main developer "hampa".

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

torille-1.0.2.tar.gz (35.6 MB view hashes)

Uploaded source

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page