Imitation Learning with AutoRally

This repository contains code for imitation learning on the AutoRally simulator with imlearn: A Python Framework for Imitation Learning.

This code runs a simulated version of the experiments described in:

Agile Off-Road Autonomous Driving Using End-to-End Deep Imitation Learning. Y. Pan, C. Cheng, K. Saigol, K. Lee, X. Yan, E. Theodorou and B. Boots. Robotics: Science and Systems (2018).

However, we have not yet released the code for the Differential Dynamic Programming (MPC-DDP) controller or the Sparse Spectrum Gaussian Process (SSGP) model used for those experiments. Instead, the code here takes advantage of the excellent MPPI controller provided as part of the AutoRally software package.

Setup

First, visit the AutoRally website to install and set up the AutoRally software. Then:

1. Clone this repository into the src subdirectory of your catkin workspace, making sure to pass the --recursive flag to pull in the imlearn library.

   git clone --recursive https://github.com/ACDSlab/imitation_learning_autorally.git

2. Make sure the submodule repository src/imitation_learning follows the master branch (sometimes there is a bug).

   cd ~/$your_catkin_workspace$/src/imitation_learning_autorally/src/imitation_learning/
   git checkout master
   

3. Change to the catkin workspace and install dependencies.

   rosdep install --from-path src --ignore-src -y

4. To use the keyboard package, clone this repository into the src subdirectory of your catkin workspace.

   git clone https://github.com/lrse/ros-keyboard

5. Compile the workspace.

   catkin_make -DCMAKE_BUILD_TYPE=Release

6. Follow the instruction 4 and 6 in the AutoRally Setup Instructions.

7. Make models, data, and logs folders in the home repository(~/$your_catkin_workspace$/src/imitation_learning_autorally).

8. In every terminal, at your catkin workspace, don't forget to

   source devel/setup.bash
   source src/autorally_private/autorally_private_sandbox/setupEnvLocal.sh
   

9. Launch the AutoRally simulation environment, the camera processing node, and the keyboard node.

   roslaunch autorally_gazebo autoRallyTrackGazeboSim.launch
   roslaunch autorally_core stateEstimator.launch
   roslaunch imitation_learning_autorally single_proc_gazebo.launch
   roslaunch imitation_learning_autorally keyboard.launch
   

How to Use

After Setting up, check the onboard-camera view

rosrun image_view image_view image:=/left_camera/cropped_image_color

and adjust image processing parameters in single_proc_gazebo.launch file.

To run the experiments, first run the expert(MPPI controller)

roslaunch imitation_learning_autorally mppi.launch

You can read the control rostopic /mppi_controller/chassisCommand but the vehicle will not move until you run

roslaunch imitation_learning_autorally autorally_mppi.launch

If you are asked to switch to MANUAL mode, make the keyboard window focused and press m on your keyboard.

If you are asked to switch to AUTONOMOUS mode, make the keyboard window focused and press a on your keyboard.

If you are asked to Hit RED/GREEN on the runstop box and if you are using a USB gamepad, press any of the buttons by the right stick (normally labelled X, Y, A, B or square, triangle, X, circle) to toggle the published value. If you are not using a gamepad, follow the instruction 6 in the AutoRally Setup Instructions to toggle the runstopMotionEnabled parameter in the /chassisState topic.

Acknowledgements

We thank our colleagues on the AutoRally team for the AutoRally software package. This software would not be possible without their effort.

Software Authors

Keuntaek Lee, Kamil Saigol, Gabriel Nakajima An, Yunpeng Pan, Xinyan Yan

Citation

If you use this software in your work, please cite:

Agile Off-Road Autonomous Driving Using End-to-End Deep Imitation Learning. Y. Pan, C. Cheng, K. Saigol, K. Lee, X. Yan, E. Theodorou and B. Boots. Robotics: Science and Systems (2018).