saashanair / FRARL

Falsification-Based Robust Adversarial Reinforcement Learning

Installation

1. Create conda environment frarl with necesssary packages.

   conda env create -f environment.yml
   conda activate frarl
   pip install git+https://github.com/openai/baselines.git
   

2. Download and install MATLAB

3. Install MATLAB engine for Python

   • Open MATLAB and run following commands in MATLAB shell:

   cd (fullfile(matlabroot,'extern','engines','python'))
   pwd
   

   • In terminal, navigate to path of the matlab engine acquired in the previous step

   cd <path_of_matlab_engine>
   

   • Check path to created conda environment

   conda info -- env
   

   • Install MATLAB engine in conda environment

   conda activate frarl
   python setup.py build --build-base="<path_to_home_directory>" install --prefix="<path_to_conda_env>"
   

   using <path_to_conda_env> as the build-base results in an error to copy the file, with the message that the name of the file is too long

4. Download MATLAB dependencies

   • Return to frarl directory

   mkdir staliro_imports && cd staliro_imports
   

   • Download S-Taliro here

   • Download MatlabBGL here

   • Download Core_py2matlab here

   • Test if MATLAB engine install successfully and install S-Taliro

   python
   import matlab.engine as me
   eng = me.start_matlab()
   eng.cd('trunk/')
   eng.setup_staliro(nargout=0)
   

5. Extract HighD data

   • Request HighD dataset from here and replace the path to dataset in /algorithms/highD/get_highd_data.py

   python ./algorithms/highD/get_highd_data.py
   

6. Install simulator

   cd simulator
   pip install -e .
   

   • Test

   python ./gym_car_acc/test/test_car_sim.py
   

7. Run script to reproduce our results

   cd algorithms
   ./run_ppo2_group.sh