This repository contains the main implementation for CAROL: Certifiably Robust Reinforcement Learning through Model-Based Abstract Interpretation. (SaTML'24)
The main code is tested with Python 3.8.12. The required packages are listed in requirements.txt and can be installed with
pip install -r requirements.txt
pip install auto_LiRPA==0.3
pip install hydra-core==1.3.1
Additionally, MuJoCo 2.1 is required.
To run the code, you need to add this project's path in the PYTHONPATH with
export PYTHONPATH=$PYTHONPATH:/path/to/carol
To run the training for experiments in the paper,
cd carol
python run.py overrides=hopper
python run.py overrides=walker2d
python run.py overrides=halfcheetah
python run.py overrides=ant
We attach an example in /example_models with the policy and models for Hopper. overall indicates the example policy and the model trained together with this policy. seperate contains examples of the separately trained models used for the evaluation in Figure 3.
For the provability part, a new environment is needed and the packages can be intalled with
pip install -r requirements_proof.txt
The main change is for the PyTorch version.
To run provability experiments for this example (with the proof packages):
cd evaluation
python test_provability.py evaluation=provability overrides=hopper
To run an empirical attack on this example:
cd evaluation
python test_attack.py evaluation=attack overrides=hopper