Learning safety in model-based Reinforcement Learning using MPC and Gaussian Processes

This repository contains the source code used to produce the results obtained in our 2023 IFAC WC submission (extended version here).

In this work, we propose a straightforward yet effective algorithm for enabling safety in the context of Safe Reinforcement Learning (RL) using Model Predictive Control (MPC) as function approximation. The unknown constraints encoding safety are learnt from observed MPC trajectories via Gaussian Process (GP) regression, and are then enforced onto the RL agent to guarantee that the MPC controller is safe with high probability.

If you find the paper or this repository helpful in your publications, please consider citing it.

@article{airaldi20235759,
  title = {Learning safety in model-based Reinforcement Learning using MPC and Gaussian Processes},
  journal = {IFAC-PapersOnLine},
  volume = {56},
  number = {2},
  pages = {5759-5764},
  year = {2023},
  note = {22nd IFAC World Congress},
  doi = {https://doi.org/10.1016/j.ifacol.2023.10.563},
  author = {Filippo Airaldi and Bart De Schutter and Azita Dabiri},
}

---

Installation

The code was created with Python 3.9.5. To access it, clone the repository

git clone https://github.com/FilippoAiraldi/learning-safety-in-mpc-based-rl.git
cd learning-safely

and then install the required packages by, e.g., running

pip install -r requirements.txt

Structure

The repository code is structured in the following way

• agents contains the RL algorithms used within the paper
  • the Perfect-Knowledge agent, a non-learning agent with exact information on the quadrotor drone dynamics
  • the LSTD Q-learning agent, in both its safe and unsafe variants, i.e., with and without our proposed algorithm, respectively.
• envs contains the quadrotor environment (in OpenAI's gym style) used in the numerical experiment
• mpc contains the implementation (based CasADi) of the MPC optimization scheme
• resouces contains media and other miscellaneous resources
• sim contains pickle-serialized simulation results of the different agents
• util contains utility classes and functions for, e.g., plotting, I/O, exceptions, etc.
• train.py launches simulations for the different agents
• visualization.py visualizes the simulation results

---

Training

Training simulations can easily be launched via command. The default arguments are already set to yield the results found in the paper. To reproduce the simulation results run the following command calling one of the 3 available

python train.py (--pk | --lstdq | --safe_lstqd)

Note that only one can be simulated at a time. Results will be saved under the filename ${runname}.pkl.

---

Visualization

To visualize simulation results, simply run

python visualization.py ${runname1}.pkl ... ${runnameN}.pkl

You can additionally pass --papermode, which will cause the paper figures to be created (in this case, the simulation results filepaths are hardcoded).

---

License

The repository is provided under the GNU General Public License. See the LICENSE file included with this repository.

---

Author

Filippo Airaldi, PhD Candidate [f.airaldi@tudelft.nl | filippoairaldi@gmail.com]

Delft Center for Systems and Control in Delft University of Technology

This research is part of a project that has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (Grant agreement No. 101018826 - CLariNet).

Copyright (c) 2023 Filippo Airaldi.

Copyright notice: Technische Universiteit Delft hereby disclaims all copyright interest in the program “learning-safety-in-mpc-based-rl” (Learning safety in model-based Reinforcement Learning using MPC and Gaussian Processes) written by the Author(s). Prof. Dr. Ir. Fred van Keulen, Dean of 3mE.