This work presents an extension of the initial OpenAI gym for robotics using ROS and Gazebo. A whitepaper about this work is available at https://arxiv.org/abs/1608.05742. Please use the following BibTex entry to cite our work:

@misc{1608.05742,
	Author = {Iker Zamora and Nestor Gonzalez Lopez and Victor Mayoral Vilches and Alejandro Hernandez Cordero},
	Title = {Extending the OpenAI Gym for robotics: a toolkit for reinforcement learning using ROS and Gazebo},
	Year = {2016},
	Eprint = {arXiv:1608.05742},
}

gym-gazebo is a complex piece of software for roboticists that puts together simulation tools, robot middlewares (ROS, ROS 2), machine learning and reinforcement learning techniques. All together to create an environment whereto benchmark and develop behaviors with robots. Setting up gym-gazebo appropriately requires relevant familiarity with these tools.

Code is available "as it is" and currently it's not supported by any specific organization. Community support is available here. Pull requests and contributions are welcomed.

• Environments
• Installation
• Usage

The following are some of the available gazebo environments:

Name	Middleware	Description	Reward
GazeboModularScara3DOF-v2	ROS	This environment present a modular SCARA robot with a range finder at its end pointing towards the workspace of the robot. The goal of this environment is defined to reach the center of the "O" from the "H-ROS" logo within the worspace. This environment uses slowness=1 and matches the delay between actions/observations to this value (slowness). In other words, actions are taken at "1/slowness" rate.	1 - rmse
GazeboModularScara3DOF-v1	ROS	This environment present a modular SCARA robot with a range finder at its end pointing towards the workspace of the robot. The goal of this environment is defined to reach the center of the "H" from the "H-ROS" logo within the worspace.	TBD
GazeboModularScara3DOF-v0	ROS	This environment present a modular SCARA robot with a range finder at its end pointing towards the workspace of the robot. The goal of this environment is defined to reach the center of the "O" from the "H-ROS" logo within the worspace.	TBD
GazeboCircuit2TurtlebotLidar-v0	ROS	A simple circuit with straight tracks and 90 degree turns. Highly discretized LIDAR readings are used to train the Turtlebot. Scripts implementing Q-learning and Sarsa can be found in the examples folder.
GazeboCircuitTurtlebotLidar-v0.png	ROS	A more complex maze with high contrast colors between the floor and the walls. Lidar is used as an input to train the robot for its navigation in the environment.
GazeboMazeErleRoverLidar-v0	ROS, APM	Deprecated
GazeboErleCopterHover-v0	ROS, APM	Deprecated

Refer to INSTALL.md

In the root directory of the repository:

sudo pip install -e .

• Load the environment variables corresponding to the robot you want to launch. E.g. to load the Turtlebot:

cd gym_gazebo/envs/installation
bash turtlebot_setup.bash

Note: all the setup scripts are available in gym_gazebo/envs/installation

• Run any of the examples available in examples/. E.g.:

cd examples/scripts_turtlebot
python circuit2_turtlebot_lidar_qlearn.py

To see what's going on in Gazebo during a simulation, simply run gazebo client:

gzclient

Display a graph showing the current reward history by running the following script:

cd examples/utilities
python display_plot.py

HINT: use --help flag for more options.

Sometimes, after ending or killing the simulation gzserver and rosmaster stay on the background, make sure you end them before starting new tests.

We recommend creating an alias to kill those processes.

echo "alias killgazebogym='killall -9 rosout roslaunch rosmaster gzserver nodelet robot_state_publisher gzclient'" >> ~/.bashrc