Model Predictive Shielding

Contains the code for the model predictive shielding algorithm proposed in https://arxiv.org/abs/1905.10691.

• Works with Python 3.6. Python dependencies are numpy, pytorch, gym, and sklearn.

• The online shielding code is in the python folder. The main routines are cartpole_test.py, cartpole_test_bl.py, bicycle_test.py, and bicycle_test_bl.py. For example, to run the cart-pole code, run

    $ cd python
    $ python -m spire.main.cartpole_test

• The matlab directory contains the code used to perform LQR verification for the cart-pole dynamical system. This code depends on SOSTOOLS 3.03, which is available at https://www.cds.caltech.edu/sostools. We used the SeDuMi optimizer with SOSTOOLS, available at https://github.com/sqlp/sedumi.

• The baseline code depends on the Z3 theorem prover, which is available at https://github.com/Z3Prover/z3. To run the baseline code, run

    $ cd baseline/python
    $ python main.py