• Requirements
• Environment setup
• Project structure
  • Environments
  • Models
• Usage
  • Image CNN model
  • DNN laser model
  • Use laser model with hint detect strategy
• Pretrain models

Basic requirements:

• python2
• gym-gazebo
• keras
• tensorflow

The gym-gazebo environments can be found here.

• Copy custom folder to gym-gazebo/gym_gazebo/envs.
• Copy launch, models, worlds to gym-gazebo/gym_gazebo/envs/assets and skip duplicated files.
• Replace __init__.py file in the gym-gazebo/gym_gazebo folder with the __init__.py file here.
• Replace the line <arg name="world_file" default="/home/cloud/gym-gazebo/gym_gazebo/envs/assets/worlds/maze_color.world"/> in the gym-gazebo/gym_gazebo/envs/assets/launch/MazeColor.launch file with your own path to the maze_color.world file

The environment files used for different training and tesing situation, they are all stored in custom_envs folder. To use one environment, copy the code in gazebo_turtlebot_maze_color*.py to gazebo_turtlebot_maze_color.py in gym-gazebo/gym_gazebo/envs/custom List environments:

• gazebo_turtlebot_maze_color.py: enviroment for training CNN model which use both image and laser as state input. To change target position, change line 52 self.num_target = 1 to another index (from 0 to 2).

• gazebo_turtlebot_maze_color_laser_only.py: enviroment for training laser model. The environment have 5 target position to be random at the start or after one episode end.

• gazebo_turtlebot_maze_color_laser_only_ver2.py: environment for training laser model. The environment have 5 hint at 5 corner of the maze, the laser model can learn to go to the corner and turn back.

• gazebo_turtlebot_maze_color_laser-image.py: environment for testing laser model with the image processing strategy. It have 5 different reward and hints positions which can be change by line 59 self.num_target = 1 in the code to another index (from 0 to 4).

All the main codes are stored in src folder We have 2 type of model:

• ddq_model.py: the model used CNN architecture and image with laser as input state. The model use the architecture proposed in this paper
• laser_model.py: The model use DNN and laser as input state.

Replace code in gym-gazebo/.../custom/gazebo_turtlebot_maze_color.py with the code in gazebo_turtlebot_maze_color.py to use this model.

Train:

python qlearning.py

For more parameters: python qlearning --help

Test model: python test.py <from_pretrain_dir> <epsilon>

python test.py ddq_model 0.01

Replace code in gym-gazebo/.../custom/gazebo_turtlebot_maze_color.py with the code in gazebo_turtlebot_maze_color_laser_only.py or gazebo_turtlebot_maze_color_laser_only_ver2.py to use this model. Train:

python laser_learning.py

For more parameters: python qlearning --help

Test model: python test_laser_only.py <from_pretrain_dir> <epsilon>

python test_laser_only.py laser-only 0.1

Replace code in gym-gazebo/.../custom/gazebo_turtlebot_maze_color.py with the code in gazebo_turtlebot_maze_color_laser-image.py to use this model.

Run: python test_laser_image.py <from_pretrain_dir> <epsilon>

python test_laser_image.py laser-only 0.0

Some of our pretrain model can be found here

It can be use as pretrain model or continue to train with the parameters --from_pretrain or --continue_from in the learning files. For examples: python laser_learn --continue_from laser-only --output_dir laser-only