Data-driven attitude control design for multirotor UAVs

This is the code used for my thesis. An up-to-date compiled version of the final report is available here and the associated latex sources are in this repo. You might also be intersted in the presentation I used for my defense which is here and the associated latex files in this repo In this work I applied Virtual Reference Feedback Tuning to a multirotor UAV. The MATLAB toolbox is available here.

In the Papers/ folder you can find information on a variety of data-driven control methods. The papers on VRFT and CbT are particularly interesting especially the [VRFT, CbT] Data-driven_attitude_control_law_of_a_variable-pitch_quadrotor_a_comparison_study.pdf paper that compares VRFT and CbT (spoiler: CbT is better for signals with a low signal to noise ratio).

I'll now explain what all the different bits and pieces do. You'll often find links to pieces of the diary I wrote as I was working on my thesis. Don't take it as a bible, most of it is probably wrong but if you're curious it may help to explain certain choices.

One thing to note is that, even though VRFT is a data-driven method and one of its great advantages is not having to identify the plant model before hand, nearly all of the code here uses simulations with a previously identified model. A rationale for this can be found in Chapter 3 (Simulation Results) of my thesis. In short, if a model is available, why deprive yourself of it. It can be a great tool to inform the controller synthesis process.

Another thing to note is that as much as possible is done in continuous time. This is simply beacuse the controllers on the drone I used are implemented with a continuous time Simulink model. As such transfer functions are converted to discrete time only long enough to actually do the VRFT. The rest of the time we're working in the continuous time domain.

MATLAB Folder

This is where the magic happens. Make sure to run the startup.m script to load everything. This is a pretty nifty little script that recursively explores subfolders and adds them to the path if it finds matlab extensions. This could be done with a simple addpath(genpath('.')) but then you wouldn't get all the pretty printing stuff.

The first thing you need to do is to run mk_quad_copter_models.m. This will create a mat file with the data needed for the next steps. Specifically it creates all the transfer functions we need and stores them in a .mat file that we can simply load.

If you want to see where all of these numbers come from take a peek inside quad_copter_model.mlx. The actual VRFT is performed in inner_vrft_ct.mlx and outer_vrft_ct.mlx.

The Simulink/ Folder contains a simple Simulink model of the VRFT and H infinity tuned controllers.

MATLAB/Hardware Testing/ Folder

This is where the magic happens ! No, just kidding, this is simply were the files required for the hardware testing are stored. The testing procedure happens in 3 steps.

1. First off: generate a controller. Run the inner_vrft_ct and outer_vrft_ct script to generate a suitable controller and run the prepare_test script. This script will generate an attitude_quadrotor.cpp file and replace the existing one in the r2p-ide folder. If you don't know what r2p-ide is then you're probably not using the same drone and need to adapt that part of the code to yours. In anycase you will probably have to adjust the path to point to the right file. (Note: In practice you just need the inner_pid_params and outer_pd_params arrays to run the script).
2. Now you can head over to r2p-ide and recompile the Proximity component and upload the code.
3. Run your testing script over the serial connection. There are a handful in the folder if you're lazy: pitch_test.m, pitch_test_long.m and pitch_test_disturbed.m.
4. Grab the SD card from the drone and plug it into the PC.
5. Run process_test_data to fetch the data from the SD card, parse it and put into the correct folder. Additionally this will run the make_dashboard script to show you the results.

Why this complicated you ask ? Well, I wanted to be sure which dataset came from which controllers so this pipeline keeps a copy of everything !! It creates a folder with the test name and puts into it: the controller parameters and reference models, the attitude_quadorotor.cpp file in case you need it later. When you process the test data they will also get added to the correct folder so that you have everything in the same place.

Yes, this could be simpler. The log parser is a piece of shit java app that barely works. Yes I want to rip it to pieces and replace it with something sane. How did you guess ?

The other files are simple. Go take a look.

Lessons Learned

Live scripts look great. I really wanted a Jupyter like experience for MATLAB. Don't use them. They're great to show off simple results but in practice they're slow. MATLAB is already pretty slow but in an mlx file it seems even more obscenely slow. Also they're binary files that don't play nice with source control. Sure MATLAB has a nifty diff tool that can show you differences ... when it works.

Just go with plain .m files. At least you'll be able to see what changed with git.