/shooter3d_env

Primary LanguagePython

Introduction

  1. moba_env is a mini moba game, which can be integrated into gym env.
  2. You can use this env to train AI for moba game.
  3. All moba logic is implemented in go lang, which is easy to maintain.

Structure

  1. gamecore is the folder containing all the moba game logic,
  2. gym_moba is the folder containing the interface needed to integrate the gamecore logic into gym env.

Depends:

  1. go
  2. gym env
  3. tensorflow 1.14.0
  4. opencv2, you can install by pip:
    pip install opencv-python --user
  5. golang opengl support, need to run the following commands:
    sudo apt-get install libgl1-mesa-dev libxinerama-dev libxcursor-dev libxi-dev xorg-dev
    go get github.com/go-gl/gl/v4.1-core/gl
    go get github.com/go-gl/glfw/v3.2/glfw
    go get github.com/go-gl/mathgl/mgl32
    go get github.com/ungerik/go3d/vec3
  6. golang tensorflow support, follow the following steps:
    download TensorFlow C library, and extract to local lib and include path, e.g. sudo tar -C /usr/local -xzf (downloaded file) reference:https://www.tensorflow.org/install/lang_c
    execute command in terminal:go get github.com/tensorflow/tensorflow/tensorflow/go

Usage:

  1. Enter gamecore folder, run:go build ./
  2. Under moba_env folder, run:pip install -e gym_moba --user
  3. export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/lib.
  4. Under moba_env folder, run python3 ../ppo.py, if you want to play the game, just change the is_train = False in main function.
  5. To modify test skill, please make modification as follows:
    in vi.go file, 'UseSkill' function, enable your tuning skill, disable other skills by short return
    in ppo_multihead.py file, 'predict' function, disable the direction mask of your tuning skill by commenting the 'actions[2] = -1' line, while maintaining masks of other skills.
  6. To manually control agent to defeat enemy, please edit main.go file, modify this line as follows(set _gym_mode to false):
    _gym_mode := flag.Bool("gym_mode", false, "a bool")
    Control tips:Left mouse button = set move target, Right mouse button = set skill target, when using skill, you need to press key 1 or 2 or 3 or 4 first(correspond to skill 1, 2, 3, 4), then press right mouse button to set skill target direction
  7. If you want to slow down the training in order to view the agent learning process clearly, you can modify moba_env.py, change -slow_tick=false to -slow_tick=true