This repo contains the code for the two papers
- Counter-example guided synthesis of neural network Lyapunov functions for piecewise linear systems
Hongkai Dai, Benoit Landry, Marco Pavone and Russ Tedrake
IEEE Conference on Decision and Control, 2020
video - Lyapunov-stable neural-network control
Hongkai Dai, Benoit Landry, Lujie Yang, Marco Pavone and Russ Tedrake
Robotics: Science and Systems, 2021
video
We can synthesize neural-network controllers with Lyapunov stability guarantees. Namely for all the initial states within a certain region, the controller will drive the system from these initial states to converge to the goal state.
We use python 3 in this project. You could first install the packages in requirements.txt.
Please run
pip install -e .
to install the package together with the dependencies.
Please refer to https://support.gurobi.com/hc/en-us/articles/360044290292-How-do-I-install-Gurobi-for-Python- on setting up Gurobi. Note that by default Gurobi ships with a limited license that has size limits on the optimization problem. Please refer to https://support.gurobi.com/hc/en-us/articles/360051597492 on using your own academic or commential license.
You could run
$ python3 neural_network_lyapunov/test/train_toy_system_controller_demo.py --dimension=1
This will synthesize a stabilizing controller with a Lyapunov function for a toy 1D system (TODO: add some visualization at the end of the demo). You should see that the error printed on the screen decreases to almost 0. (The code is non-deterministic, so if it doesn't converge to 0 in the first trial, you can re-run the demo and hopefully it converges in the second trial).
We use flake8 to check if the python code follows PEP standard. Before submitting the PR, you could run
$ cd neural_network_lyapunov
$ flake8 ./
to check if there are any violations.
I am a strong believer of unit test. We strongly encourage to add tests to the functions in the PR.