Distributionally Robust Optimization with Unscented Transform for Learning-Based Motion Control in Dynamic Environments.

This repository includes the source code for implementing the distributionally robust UT-MPC algorithm with all the baselines presented in the paper.

1. Requirements

• CARLA simulator
• Python (>= 3.5)
• Forces Pro
• Casadi
• Python packages, such as GPy, numpy, scipy, matplotlib, etc.

2. Quick Start

To run the experiments, first execute the CARLA simulator:

cd path/to/carla/root
./CarlaUE4.sh

Then, in another terminal, call the Town10HD map, which is used in our simulations.

cd path/to/carla/root/PythonAPI/util/config.py --map Town10HD

Finally, call the main script

python run.py

The simulation parameters can be change by adding additional command-line arguments:

• --with_obs - add an obstacle to the simulations.

• --method - chooses the MPC algorithm to execute and ccepts one of the following values:

  • cvar - CVaR-constrained learning-based MPC,
  • mean - Mean-constrained learning-based MPC,
  • drcvar - Distributionally robust UT-MPC.

• --gp - uses GPR for learing the unknown models. If not invoked, the nominal model is used for the ego vehicle, while the obstacle is assumed to be static.

• --record - records the simulation snapshots from the PyGame window.

The results of simulations are saved in separate pickle files in /data folder.