UAV Control with AI - Neural Enhanced Control for Quadrotor Linear Behavior Fitting

Trajectory generation and control algorithms using AI for UAVs with ROS wrapping.

This repository provides the code associated to the paper :

Esteban Carvalho, Pierre Susbielle, Ahmad Hably, Jilles Dibangoye, Nicolas Marchand. Neural Enhanced Control for Quadrotor Linear Behavior Fitting ICUAS 2022 - IEEE International Conference on Unmanned Aircraft Systems, Jun 2022, Dubrovnik, Croatia hal-03650833v1.

Citing

If you use this code, please quote the following publication :

@inproceedings{carvalho:hal-03650833,
  TITLE = {{Neural Enhanced Control for Quadrotor Linear Behavior Fitting}},
  AUTHOR = {Carvalho, Esteban and Susbielle, Pierre and Hably, Ahmad and Dibangoye, Jilles and Marchand, Nicolas},
  URL = {https://hal.archives-ouvertes.fr/hal-03650833},
  BOOKTITLE = {{ICUAS 2022 - IEEE International Conference on Unmanned Aircraft Systems}},
  ADDRESS = {Dubrovnik, Croatia},
  PUBLISHER = {{IEEE}},
  YEAR = {2022},
  MONTH = Jun,
  PDF = {https://hal.archives-ouvertes.fr/hal-03650833/file/Neural_Enhanced_Control_for_Quadrotor_Linear_Behavior_Fitting.pdf},
  HAL_ID = {hal-03650833},
  HAL_VERSION = {v1},
}

This project uses external software such as Mavros, PX4 or Gazebo. Below are the links directing to their documentations:

PX4 Development Guide

PX4-Firmware

Mavros

Sitl-Gazebo

Installation

For the installation, you need to have ROS melodic (or kinetic) installed, a catkin workspace and Gazebo. Follow the online documentation to set up your environement.

ROS Installation

Catkin Workspace

Gazebo

Prerequisites

Install mavros

sudo apt install ros-melodic-mavros ros-melodic-mavros-extras

Mavros request the GeographicLib datasets, install it by running the install_geographiclib_datasets.sh script

wget https://raw.githubusercontent.com/mavlink/mavros/master/mavros/scripts/install_geographiclib_datasets.sh
chmod +x install_geographiclib_datasets.sh
sudo ./install_geographiclib_datasets.sh

Install libgstreamer

sudo apt install libgstreamer1.0-dev

Initialize rosdep and update it

sudo rosdep init
rosdep update

Clone sitl_gazebo and PX4 Firmware

cd ~/catkin_ws/src/
git clone --recursive https://github.com/PX4/sitl_gazebo

git clone --recursive https://github.com/PX4/Firmware px4

or to replicate paper results, we recommand you to use a specific version of px4 using the following line code :

git clone --recursive -b kakutef7_ekf_v1.11.3 https://github.com/gipsa-lab-uav/PX4-Autopilot.git

Note: If you have troubles installing the different packages, it is recommended to read the related documentation.

Install uav_control_ai package

Clone the uav_control_ai repository

cd ~/catkin_ws/src/
git clone https://github.com/gipsa-lab-uav/uav_control_ai

Install the required Python packages:

pip install -r requirements.txt

sudo apt-get install python3-pil python3-pil.imagetk

To finish the installation run:

cd ..
catkin_make

Then source your setup.bash:

source devel/setup.bash

Testing the installation

In a terminal, run the following line:

roslaunch uav_control_ai example.launch

Three windows must appear :

• A gazebo window with the iris
• A window displaying live information sush as the thrust, position, etc.,
• A window with the desired generated trajectory.

If everything works fine,the iris quadcopter should takeoff and start the trajectory.

Note : QGroundControl can be open in parallel to check if everything is interfacing correctly.

How to use it ?

Simulation

1. [Trajectory] Design the desired trajectory using 'trajectory_gen.py' file. Examples and explanations are available in the file comments.

2. [Controller] Use 'simu_params.yaml' to specify :

• mass and hover compensation coefficient,
• control gains on the position loop and attitude loop,
• select 'clin' as controller type for linear control (proportional derivative).

3. [Simulation] Run:

roslaunch uav_control_ai example.launch

4. [Data] Data are stored in 'StateCommandData.msg' available in rosbag folder.

5. [Learning] To learn error dynamics you can use learning_error_dynamics node:

rosrun uav_control_ai learning_error_dynamics.py

At the end of the process, you can save the DNN model in nodes/dnn_model.

6. [Simulation with AI] Modify the 'simu_params.yaml' to specify :

• 'clin_ai' to use the DNN Enhanced Linear Cascaded controller_rates
• The DNN selected model path

Then run again :

roslaunch uav_control_ai example.launch