door_open_rl

Door opening with a wheeled robot using Reinforcement Learning techniques

Prerequisite

Ubuntu 18.04
ROS Melodic
- Gazebo 9.0, default installed along with ros-melodic-desktop-full or can be installed by following instruction
  - for issue of "Error in REST request" when starting gazebo replace url: https://api.ignitionfuel.org by url: https://api.ignitionrobotics.org inside ~/.ignition/fuel/config.yaml
- Setup ROS Environment (export below variable in ~/.bashrc)
  - ROS_HOSTNAME=localhost
  - ROS_MASTER_URI=http://localhost:11311
Install opencv and opencv-python

sudo apt install libopencv-dev
python2 pip install opencv-python==4.2.0.32

Install Realsense SDK and Realsense-ros

# add repository first, then install realsense sdk
sudo apt-get install librealsense2-dkms
sudo apt-get install librealsense2-utils
# install realsense-ros
sudo apt-get install ros-melodic-realsense2-camera

dependencies

tensorflow 2.1.0 (last version support python 2.7) (tensorflow_probability is used in do_training)

pip install tensorflow==2.1.0 tensorflow_probability==0.9.0

if pip unable to recognize the version, upgrade pip with

pip install pip --upgrade

pip install gym

scikit-image

pip install scikit-image

start gazebo environment

roslaunch do_gazebo door_opening.launch

office_room_0.world is for Geforce RTX 1070
offce_room.world is for Geforce RTX 3070, remove one lighting source

train specific tasks

door pulling task

cd ../do_learning/scripts/task_pull
python door_pull_training_ppo.py

door pushing task

cd ../do_learning/scripts/task_push
python door_push_training_ppo.py

door traversing task

cd ../do_learning/scripts/task_traverse
python door_traverse_training_ppo.py

training log

tensorboard --logdir=saved_models/task_pull

testing

the trained policy can be tested in different environments with different appearances and other physical settings

roslanch do_gazebo door_opening.launch world:=office_room_{x}

modify the argument of 'mu' in mobile_robot.urdf file to change the friction coefficient between robot wheels and ground

Switch to old environment (door_room.world) using do_training

roslanch do_gazebo door_opening.launch world:=door_room

The latest environment (office_room.world) is enriched with, customize door model, lighting and specific ode parameters, while old environment (door_room.world) does not have this features, it used the existing door from gazebo model base, which contains imperfect mesh and may result bad dynamic simulation performance, but many previous work are done with old environment.

change the friction coefficients (mu1, mu2) to 0.1 in "/do_description/urdf/mobile_robot.urdf.xacro"

video & paper

https://youtu.be/O_FV47hIRus

https://ieeexplore.ieee.org/document/9562517

wangcongrobot/door_open_rl