learning-nmpc-q-loss

Imitation Learning from NMPC via acados and PyTorch

This repo will contain the implementation described in the paper: Imitation Learning from Nonlinear MPC via the Exact Q-Loss and its Gauss-Newton Approximation, available here.

The paper was presented in December 2023 at the IEEE Conference on Decision and Control.

Getting started:

Running python 3.8.12

1. Installation

pip install -e .

pre-commit install

2. Acados installation

Follow the instruction at: acados docs
After installing acados on your machine, install via pip the acados python interface in the python environment you use to run this repo as follow
pip install -e <acados_root>/interfaces/acados_template

How to use:

1. Folder imitate contains the routines for creating the model, formulating/solving the MPC problem and running imitation learning.
2. Folder script contains different code snippets for training and evaluating the policies, and plotting the results.
3. Folder tests contains two python scripts to test the correct behavior of the provided code, especially the dependancy on acados.

Acados speedup for Linux machines

For Linux machines it is possible to compile the acados python interface using cython. To do so set the flag no-cython=False in the ArgumentParser in main_training.py.