dmlab_random_maze

Introduction

This project provides the codes for a customized version of Deepmind 3D maze domain. In this version, you can construct 3D maze using external txt files. By passing designed parameters from python to lua, you can customize the 3D maze.

Installation

• Clone this project.
• Clone and install the Deepmind Lab by following the instructions here.
• Copy the nav_random_maze.lua file in lua in to the lab\game_scripts\levels\ in Deepmind Lab.
• Wrap the Deepmind into a PIP package by following the instructions here.
• Create a virtual environment using conda and install the Deepmind PIP package.

Extra Dependencies

• numpy
• scipy
• matplotlib
• ipython

Test

After the installation, you can test the environment using the following command under the repo root directory:

python dmlab_demo.py

Customized variables

We follow the strategy of passing the customized variables from python to lua file. In the python api. we define the following customized variables.

First of all, we define the variables related to the maze appearance.

• params.maze_configs.name (str): name of the 3D maze
• params.maze_configs.size (str): size of the 3D maze (size x size)
• params.maze_configs.seed (str): random seed of the 3D maze
• params.maze_configs.texture (str): texture\theme of the 3D maze
• params.maze_configs.decal_freq (str): rate of posters on the wall
• params.maze_configs.map_txt (str): txt map of the 3D maze

Then, we define the variables related to start and goal positions.

• params.start_pos.x (str): x
• params.start_pos.y (str): y
• params.start_pos.yaw (str): orientation
• params.goal_pos.x (str): x
• params.goal_pos.y (str): y
• params.goal_pos.yaw (str): orientation

Note: start_pos/goal_pos should follow the format (x, y, ori). Specifically, given a 5x5 map as follows where P indicates the start pos and G indicates the goal position.

* * * * *
* P     *
*   * * *
*     G *
* * * * *

Then, the input start_pos = (1, 1, 0) while the goal_pos = (3, 3, 0).

Finally, we can pass the parameters settings from python to lua using the following:

self._lab.write_property("params.maze_configs.name", self.maze_name)

Please refer the reset function in env\LabEnv.py for more details.

Acknowledgement

Thanks for the instructions from authors of Deepmind Lab. Thanks for the in-depth discussions with David Klee. I also bootstrap some code from David's implementation.