in progress! (turlebot, Tensorflow, Gazebo, Deep recurrent Q-learning)
Goal: use deep reinforcement learning to realize obstacle avoidance.
Platform: turlebot
Deep RL: directly learn robot actions from pixel information.
Action: linearSpeed (0.1, 0.2, 0.3) + angle (right=1, still=0, left=-1)
Pixel state: （小梁和唐宋确定下！） raw image --> processed imgState(vector).
Reward: r = -2 (collision ), r = 0.01 (O.W)
- install turtlebot source. use 'Source_Installation'.
b. other dependences.
Tensorflow
agent file: define RL agent
agent_ros_2_mobile file: use to communicate RL agent with mobile robot.
algorithm file: define DRQN algorithm
utility file: some useful function.
Rviz
a. view in rviz.
roslaunch turtlebot_rviz_launchers view_robot.launch
b. navigation in Rviz.
roslaunch turtlebot_stage turtlebot_in_stage.launch
Gazebo
a. view in gazebo.
roslaunch turtlebot_gazebo turtlebot_world.launch
Keyboard control.
roslaunch turtlebot_teleop keyboard_teleop.launch
1. 小梁
reference:
a: http://wiki.ros.org/turtlebot/Tutorials
b. http://learn.turtlebot.com/
c. start robot:
roslaunch turtlebot_bringup minimal.launch
d. use keyboart control:
roslaunch turtlebot_teleop keyboard_teleop.launch
e. use openni:
roslaunch openni_launch openni.launch
f. test camera:
rosrun image_view image_view image:=/camera/rgb/image_color
rosrun image_view image_view image:=/camera/depth/image
1. agent_ros_2_mobile (tested, ok!)
roslaunch agent_ros_mobile agent_ros_mobile.launch
实现： a. python recordMobile.py # get image from '/camera/rgb/image_raw' topic.
b. python runMobile.py #
2. DRQN (tested, ok!)
cd /home/shuang/catkin_ws/src/Navi_DeepRL
python Navi_deepRL.py
实现： a. RL agent (mobile robot).(tested, ok!)
cd /home/yuchen/catkin_ws/src/Navi_DeepRL/agent
python agent_mobile.py
b. Q-network
./algorithm
3. Collect samples
cd /home/shuang/catkin_ws/src/Navi_DeepRL
python collectSample.py
and
python restoreSample.py
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent