PyTorch implementation of our PAC-Bayesian framework for optimal control with stability guarantees.
git clone https://github.com/DecodEPFL/PAC-SNOC.git
cd PAC-SNOC
python setup.py installThe following gifs show the trajectories of the 2 robots before training (left), and after the training of the empirical controller (middle),
and using our approach (right).
The agents need to coordinate in order to pass through a narrow passage while avoiding collisions between them.
The initial conditions used as training data are marked with ○,
consisting of s = 30 samples.
The shown trajectories start from a random initial position sampled from the test data and marked with
⬣.
Two environments are proposed to train the controllers with our approach. The two systems are: i) a simple Linear Time-Invariant (LTI) system, and ii) a system of two planar robots navigating to prespecified locations while avoiding collisions.
- Generate the data by running
python experiments/scalar/generate_data.py- Train the empirical controller by running
python experiments/scalar/run_empirical.py- Train the benchmark controller by running
python experiments/scalar/run_benchmark.py- Train the controllers using our approach by running
python experiments/scalar/run_grid_Gibbs.py- Generate the data by running
python experiments/robotsX/generate_data.py- Train the empirical controller by running
python experiments/robotsX/run_emp.py- Train the controllers using our approach by running
python experiments/robotsX/run_SVGD.pyPre-trained models of the LTI system and robots can be found in the folders robotsX and scalar, respectively.
This work is licensed under a Creative Commons Attribution 4.0 International License.
[1] Mahrokh Ghoddousi Boroujeni, Clara Galimberti, Andreas Krause, Giancarlo Ferrari-Trecate. "A PAC-Bayesian Framework for Optimal Control with Stability Guarantees," 2024.