Approximate non-linear model predictive control with safety-augmented neural networks

Implementation of safety-augmentation and three numerical benchmark examples (stirtank reactor, quadcopter, and chain mass system). The paper describing the theory can be found on arXiv.

Requirements

You need acados to run parts of this code. Please follow the official acados installation instructions. This code was tested with acados v0.1.9.

You can install other Python dependencies via pip:

pip3 install -r examples/requirements.txt
pip3 install -r soeampc/requirements.txt

Numerical Examples

You find the numerical examples from the paper in the examples folder. Each example has it's own README.md file with instructions how to run them:

• Two state stir tank reactor
• Ten state quadcopter
• Chain mass system

Downloading precomputed datasets and pretrained NNs

You can download the training and testing datasets used in the paper together with the pretrained model from Zenodo.

Extract the datasets into the examples/{system}/datasets/ folder, e.g., for the quadcopter example, you should get an examples/quadcopter/datasets/quadcopter_N_9600000 folder.

Extract the pretrained neural networks into the examples/{system}/models/ folder, e.g., for the quadcopter example, you should get an examples/quadcopter/models/10-200-400-600-600-400-200-30_mu=0.12_20230104-232806 folder.

Running inside Docker

The Dockerfile in this repository allows you to run the code without installing acados or other Python dependencies natively. To use the Container, you first have build it by running:

docker build -t soeampc .

Next, run the container and mount this repository:

docker run -it --name soeampc -v $(pwd):/soeampc soeampc bash

You can now run all the example commands inside the container.