Learning Control-Oriented Dynamical Structure from Data

This repository accompanies the paper "Learning Control-Oriented Dynamical Structure from Data" [1].

Getting started

Ensure you are using Python 3. Clone this repository and install the packages listed in requirements.txt. In particular, this code uses JAX, Equinox, and CasADi.

Reproducing results

Trajectory and data generation via MPC for both training and testing can be done with the command ./auto_datagen.sh. This may take a few minutes.

Model training and testing (for various training set sizes and random seeds) can be done with the command ./auto_traintest.sh. This will take a long time, so feel free to edit ./auto_traintest.sh if you just want to try, e.g., a single training set size and seed.

Testing for the planar quadrotor on the double loop-the-loop trajectory can be done with the command ./auto_test_double_loop.sh.

Finally, the figures in [1] can be reproduced and customized with the Jupyter notebook plots.ipynb.

Citing this work

Please use the following BibTex entry to cite this work.

@INPROCEEDINGS{RichardsSlotineEtAl2023,
author      = {Richards, S. M. and Slotine, J.-J. and Azizan, N. and Pavone, M.},
title       = {Learning control-oriented dynamical structure from data},
year        = {2023},
booktitle   = {International Conference on Machine Learning (ICML)},
url         = {https://arxiv.org/abs/2302.02529},
doi         = {10.48550/arXiv.2302.02529},
}

References

[1] S. M. Richards, J.-J. Slotine, N. Azizan, and M. Pavone. "Learning control-oriented dynamical structure from data". In International Conference on Machine Learning (ICML), 2023.