A dynamics model is a valuable tool for a variety of applications, including simulations and control. However, traditional methods for obtaining such models for an unmanned aerial vehicle (UAV), such as wind tunnel testing or the installation of additional sensors, can be time-consuming, inaccesible and costly. In this project, we aim to address this challeng through a data-driven dynamics modelling pipeline that uses the PX4 unified flight log (ULog) format. The pipeline can be used for obtaining a dynamics model solely by minimizing the prediciton error of the model with respect to the flight data collected by the by default on the vehicle installed sensors. The modular approach of the pipeline aims to accomodate different model structures, optimizers and various UAV airframes, including multirotors and vertical take-off and landing (VTOL) UAVs. Specifically, for the example of multirotors, the pipeline automates the joint estimation of a parametric model and the quality of the fit and integrates the former into the PX4 Gazebo flight simulator.
You can get an overview of the pipelines functionalities on the project description website or in the following video presentation:
More detailed information can be found in the student paper Data-Driven Dynamics Modelling Using Flight Logs.
This plugin depends on PX4 SITL. Therefore custom messages of PX4 SITL needs to be linked. Therefore, prior to building this package, PX4 and corresponding SITL package needs to be built.
In case you have not cloned the firmware repository
git clone --recursive https://github.com/PX4/PX4-Autopilot.git ~/src/PX4-Autopilot
For internal use, our internal firmware repository is as the following.
git clone --recursive https://github.com/ethz-asl/ethzasl_fw_px4.git ~/src/PX4-Autopilot
cd <Firmware Dir>
DONT_RUN=1 make px4_sitl gazebo
The build directory of PX4 can be linked by setting the environment variable PX4_ROOT to the root of the firmware directory. Set the environment variable export PX4_ROOT=~/src/PX4-Autopilot or add it to your bashrc
echo "export PX4_ROOT=~/src/PX4-Autopilot" >> ~/.bashrc
source ~/.bashrc
The use the parametric model structure you need to install python 3.8 and the needed python libraries. It is strongly advised to install the pip packages in a virtual enviroment setup for this project.
Install the dependencies including submodule dependencies using:
install-full-depdencies
After the environment has been setup as described in the previous section, build the package as the following.
mkdir build
cd build
cmake ..
make
Link the latest log files to your local logs folder using:
source setup.bash
Generate the parametric model using a log file (ulog or csv):
make estimate-model [model=<modeltype>] [config=<config_file_path>] [data_selection=<none|interactive|setpoint|auto>] [selection_var=topic_name/variable_name] [plot=<True/False>] [log=<log_file_path>]
The chosen vehicle model class determines what physical effects are modelled and what parameterts need to be regressed in the system identification process. Current vehicle model choices are:
- quadrotor_model (default config for quadrotor)
- fixedwing_model (default config for cruise flight of fixed-wings and VTOLs, documented here)
The config file allows to configure the intra class vehicle variations, used log file topics, data processing and other aspects of the pipeline. The default location is in Tools/parametric_model/configs. The path can be passed in the make target through the config=<config_file_path> argument. If no config is specified the default model config is used.
The Log file contains all data needed for the system identification of the specified model as defined in its config file. Next to the ULog file format it is also possible to provide the data as a csv file. An example of the required formating can be seen in the resources folder.
The data_selection argument is optional (per default none) and can be used to visually select subportions of the data.
- none(default): Data selection is disabled, and the whole section of the log is used
- interactive: Data is selected interactively using the Visual Dataframe Selector, before running the model estimation. It is also possible to save the selected subportion of data to a csv file in order to use this exact dataset multiple times.
- setpoint: Data is selected based on a certain topic value, which has to be specified with the variable
selection_varthrough the command line. The variable has to be provided in the formattopic_name/variable_nameto be recognized and loaded correctly. For example, to select data based on theaux1value of themanual_control_setpointtopic, the variablemanual_control_setpoint/aux1has to be specified. - auto: Data is selected automatically (Beta)
The resulting parameters of the model estimation together with additional report information will be saved into the model_results folder as a yaml file.
As an example to get started you estimate the parameters of a quadrotor model with the reference log_files:
make estimate-model model=quadrotor_model log=resources/quadrotor_model.ulg
It is also possible to test the obtained parameters for a certain model on a different log using:
make predict-model [model=<modeltype>] [config=<config_file_path>] [data_selection=<none|interactive|auto>] [log=<log_file_path>] [model_results=<model_results_path>]
To ensure that the parametric model works as expected you can perform a set of pytests, which are stored in Tools/parametric_model/tests. To start the tests you have to run the shell script:
Tools/parametric_model/test_parametric_model.sh
Currently only the transformation from body to intertial frame and vise versa are checked. This should be expanded in the future.
To run the simulation,
source setup.bash
Tools/sitl_run.sh -m iris -s iris_aerodynamics
The custom Gazebo quadrotor model will always read the model parameters from the file model_results/quadrotor_model.yaml. You can simply rename your desired model results file to fly your estimated model in Gazebo.
This project was done in collaboration between the Autonomous Systems Lab, ETH Zurich and Auterion AG
To cite this work in a academic context:
@article{galliker2021data,
title={Data-Driven Dynamics Modelling Using Flight Logs},
author={Galliker, Manuel Yves},
year={2021}
}