(i) how many num_episodes you want your solution to use in each level (mandatory)
We prepared our submission to be run using a single episode per level (num_episodes = 1).
(ii) what method(s) you used and results you obtained (optional)
Our submission is an adapted implementation of the Model Predictive Contouring Control as described here.
Due to performance limitations the algorithm is adapted not to plan the path in real-time, but instead to precalculate the path only once before the first episode. Due to uncertainty in the gate positions, the planned path is patched in real-time once the actual gate positions are known.
Also, for performance, the Casadi MPCC controller needs to be compiled. This will take a few seconds during the controller bring-up; a message will be printed on the terminal while this is happening.
From start to end, the execution of our controller should take less than a minute. You can find an example run in this video: https://www.youtube.com/watch?v=uAgVpQPQqDs
Our submission uses the networkx python module.
We've added this dependency in the pyproject.toml file, so executing
pip install -e .
in the safe-control-gym folder of our fork should be enough to install this dependency.
The compilation stage of MPCC will also require a working gcc or clang installation.
To run our submission, the following commands should be enough.
cd competition/
python3 getting_started.py --overrides level0.yaml
python3 getting_started.py --overrides level1.yaml
python3 getting_started.py --overrides level2.yaml
python3 getting_started.py --overrides level3.yaml