LF2 Gym

An OpenAI-gym-like environment for Little Fighter 2

The major contributors of this repository include Yu-Jing Lin, Po-Wei Wu, Yu-Jhe Li and Hsin-Yu Chang.

Little Fighter 2 is a freeware PC 2.5D fighting game. We wrapped its open source version, F.LF, into a trainable environment for reinforcement learning.

This environment is used and described in the paper "Deep Reinforcement Learning for Playing 2.5D Fighting Games" (https://arxiv.org/abs/1805.02070).

Demo

By applying reinforcement learning methods on the LF2 Gym, the agent is able to play Woody (the one in green) against Louis as shown below.

Installation

1. Clone the LF2 Gym repository.

   git clone https://github.com/elvisyjlin/lf2gym.git
   cd lf2gym
   

2. Download the open source LF2 from Project F and make it trainable (see here).

   sh setup.sh
   

3. Install Python 3 and get all dependencies.

   This project is developed under Python 3.6.2 and has been tested in Python 3.4.0.

   pip3 install -r requirements.txt
   

4. Install Tkinter (optional, for screen rendering)

   Install Tkinter if you want to render the game on screen. On the other hand, you can use Google Chrome to run the game without calling render().

   Install Tkinter in Ubuntu for Python 3.

   sudo apt-get install python3-tk
   

   Install Tkinter in CentOS for Python 3

   sudo yum install python3-tkinter
   

To start

Env

Make an LF2 environment.
Note: The web driver will be closed automatically when the process exits.

import lf2gym
env = lf2gym.make()

All parameters for make() are described here.

Server

Or if you simply want to run a LF2Server.

import lf2gym
lf2gym.start_server(port=8000)

Open your browser, and connect to http://127.0.0.1:8000/game/game.html to play LF2!

Keyboard control setting is described here.

Need Helps?

If you encounter selenium.common.exceptions.WebDriverException: Message: unknown error: cannot find Chrome binary, which means Selenium cannot find the Chrome browser on your computer, please install it.

sudo apt-get install chromium-browser

If you encounter selenium.common.exceptions.WebDriverException: Message: unknown error: Chrome failed to start: exited abnormally, which means your computer probably doesn't support GUI, please run Chromedriver under headless mode.

env = lf2gym.make(headless=True)

Examples

Some examples demonstrate how to use the LF2 Gym.

cd example

1. To try a simple example.

   python test.py
   

   test.py simulates an agent with predefined actions to play Davis against dumb Dennis, and saves the recording to test.avi.

2. To play with a baby agent, which only takes random actions.

   python baby_play.py
   

   You can even choose the characters to play with.

   python baby_play.py [Your Character] [Your Enemy]
   

Available Methods of Env

Reset the environment.

env.reset()

Reset with custom options.

options = env.get_reset_options() # {'ai_epsilon': None, 'hp_full': 500, 'mp_full': 500, 'mp_start': 500}
options['ai_epsilon'] = 0.1
options['hp_full'] = 1000
options['mp_full'] = 2e8
options['mp_start'] = 2e8
env.reset(options)

Render the environment in a new window.

env.render()

Take an action.

observation, reward, done, info = env.step(0) # actions are defined in the action space

Get the specification of environment.

state_size   = env.observation_space.n
action_size  = env.action_space.n
action_space = env.action_space

Get the log in the browser.

env.get_log()

Get current game information (hp, mp, position, speed, action, ...) of each character.

env.get_detail()

Skip N frames. N is an integer. E.g. 100.

env.idle(N)

Perform specific key actions. sequence can be an action or several actions (sequential key inputs). E.g. ['w'].

env.perform_actions(sequence)

Screen Recording

1. Start recording
   The frames will be store in the buffer whenever env.step().

   env.start_recording()
   

2. Stop recording
   Stop storing frames to the buffer.

   env.stop_recording()
   

3. Save the recorded frames to a video file.
   It supports .avi and .mp4 formats.

   env.save_recording(filename, delete_after_saved=True)
   

Parameters

The all available web drivers, characters, difficulties, and backgrounds are defined in config.py.

Although we implemented PhantomJS, ChromeDriver and GeckoDriver, some errors occur when using GeckoDriver with Firefox. Therefore, we recommend to use PhantomJS or ChromeDriver with Google Chrome. Note that Google Chrome browser or Firefox browser should be installed first.

Action Space

Base Action Space

With the default action_options, which is ['Basic', 'AJD', 'Full Combos'], the skip-N action space is

Group of Actions

Keyboard Control

Modifications to F.LF

In order to train the agent better, we did some modifications to the original game. See here.

Web Drivers

The web drivers will be downloaded automatically when used.

The following web drivers are utilized to run the game.

1. PhantomJS
   http://phantomjs.org/download.html
2. ChromeDriver (Google Chrome)
   https://sites.google.com/a/chromium.org/chromedriver/
3. GeckoDriver (Firefox)
   https://github.com/mozilla/geckodriver/releases

Reference

This project is based on the open source Little Fighter 2 game of Project F.
The OpenAI Gym is taken for reference to design the architecture of lf2gym.

If you find our code useful for your research, please cite

@inproceedings{li2018deep,
  title={Deep Reinforcement Learning for Playing 2.5D Fighting Games},
  author={Li, Yu-Jhe and Chang, Hsin-Yu and Lin, Yu-Jing and Wu, Po-Wei and Wang, Yu-Chiang},
  booktitle={IEEE International Conference on Image Processing ({ICIP})},
  year={2018}
}