Skip to content

A toolkit for developing and comparing reinforcement learning algorithms using ROS and Gazebo.

License

GPL-3.0, Unknown licenses found

Licenses found

GPL-3.0
LICENSE
Unknown
LICENSE.md
Notifications You must be signed in to change notification settings

akira304moto/gym-gazebo

 
 

Repository files navigation

Build status

An OpenAI gym extension for using Gazebo known as gym-gazebo

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:

@article{zamora2016extending,
  title={Extending the OpenAI Gym for robotics: a toolkit for reinforcement learning using ROS and Gazebo},
  author={Zamora, Iker and Lopez, Nestor Gonzalez and Vilches, Victor Mayoral and Cordero, Alejandro Hernandez},
  journal={arXiv preprint arXiv:1608.05742},
  year={2016}
}

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.


Table of Contents

Community-maintained environments

The following are some of the gazebo environments maintained by the community using gym-gazebo. If you'd like to contribute and maintain an additional environment, submit a Pull Request with the corresponding addition.

Name Middleware Description Observation Space Action Space Reward range
GazeboCircuit2TurtlebotLidar-v0GazeboCircuit2TurtlebotLidar-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-v0GazeboCircuitTurtlebotLidar-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. TBD
GazeboMazeErleRoverLidar-v0 ROS, APM Deprecated
GazeboErleCopterHover-v0 ROS, APM Deprecated

Other environments (no support provided for these environments)

The following table compiles a number of other environments that do not have community support.

Name Middleware Description Observation Space Action Space Reward range
cartpole-v0.pngGazeboCartPole-v0 ROS Discrete(4,) Discrete(2,) 1) Pole Angle is more than ±12° 2)Cart Position is more than ±2.4 (center of the cart reaches the edge of the display) 3) Episode length is greater than 200
GazeboModularArticulatedArm4DOF-v1.pngGazeboModularArticulatedArm4DOF-v1 ROS This environment present a modular articulated arm robot with a two finger gripper at its end pointing towards the workspace of the robot. Box(10,) Box(3,) (-1, 1) [if rmse<5 mm 1 - rmse else reward=-rmse]
GazeboModularScara4DOF-v3.pngGazeboModularScara4DOF-v3 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 workspace. This environment compared to GazeboModularScara3DOF-v2 is not pausing the Gazebo simulation and is tested in algorithms that solve continuous action space (PPO1 and ACKTR from baselines).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. Box(10,) Box(3,) (-1, 1) [if rmse<5 mm 1 - rmse else reward=-rmse]
GazeboModularScara3DOF-v3.pngGazeboModularScara3DOF-v3 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 workspace. This environment compared to GazeboModularScara3DOF-v2 is not pausing the Gazebo simulation and is tested in algorithms that solve continuous action space (PPO1 and ACKTR from baselines).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. Box(9,) Box(3,) (-1, 1) [if rmse<5 mm 1 - rmse else reward=-rmse]
GazeboModularScara3DOF-v2.pngGazeboModularScara3DOF-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 workspace. Reset function is implemented in a way that gives the robot 1 second to reach the "initial position". Box(9,) Box(3,) (0, 1) [1 - rmse]
GazeboModularScara3DOF-v1.pngGazeboModularScara3DOF-v1 ROS Deprecated TBD
GazeboModularScara3DOF-v0.pngGazeboModularScara3DOF-v0 ROS Deprecated
ariac_pick.jpgARIACPick-v0 ROS

Installation

Refer to INSTALL.md

Usage

Build and install gym-gazebo

In the root directory of the repository:

sudo pip install -e .

Running an environment

  • 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/turtlebot
python circuit2_turtlebot_lidar_qlearn.py

Display the simulation

To see what's going on in Gazebo during a simulation, run gazebo client. In order to launch the gzclient and be able to connect it to the running gzserver:

  1. Open a new terminal.
  2. Source the corresponding setup script, which will update the GAZEBO_MODEL_PATH variable: e.g. source setup_turtlebot.bash
  3. Export the GAZEBO_MASTER_URI, provided by the gazebo_env. You will see that variable printed at the beginning of every script execution. e.g. export GAZEBO_MASTER_URI=http://localhost:13853

Note: This instructions are needed now since gazebo_env creates a random port for the GAZEBO_MASTER_URI, which allows to run multiple instances of the simulation at the same time. You can remove the following two lines from the environment if you are not planning to launch multiple instances:

os.environ["ROS_MASTER_URI"] = "http://localhost:"+self.port
os.environ["GAZEBO_MASTER_URI"] = "http://localhost:"+self.port_gazebo

Finally, launch gzclient.

gzclient

Display reward plot

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.

Killing background processes

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

About

A toolkit for developing and comparing reinforcement learning algorithms using ROS and Gazebo.

Resources

License

GPL-3.0, Unknown licenses found

Licenses found

GPL-3.0
LICENSE
Unknown
LICENSE.md

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 96.8%
  • Shell 1.8%
  • Dockerfile 1.4%