Learning Local CBFs for Safety Control of Hybrid Systems

Preprint with supplementary material is available online.

Install

The implementation has been tested with Python 3.10 under Ubuntu 22.04. We recommend installing the simulation inside a virtualenv. You can install the environment by running:

virtualenv cbf_env
source cbf_env/bin/activate
git clone git@github.com:shuoyang2000/neural_hybrid_cbf.git
cd neural_hybrid_cbf
pip install -e .

Note that some supplementary libraries are included locally (in the lib folder). These libraries are bundled with the repository as changes are made to make them adapted to our project.

Docker

We also provide a Dockerfile to build a container if you prefer.

docker build -t cbf_container -f Dockerfile .

Reproduce the evaluation results

Python path

We run all experiments from the project directory so please add the project directory to the PYTHONPATH environment variable:

export PYTHONPATH=$PYTHONPATH:$

Evaluation commands

1. For the autnomous racing environment, please run from the project directory:

python3 scripts/racing/main.py

For visualization, please run from the project directory:

python3 scripts/racing/visualization.py

You may change the method (mpc, local_switch_unaware_cbf, local_switch_aware_cbf, global_cbf) in the file scripts/racing/evaluation_config.py. The results will be saved under results/racing_results/.

2. For the adaptive cruise control example, please run from the project directory:

python3 scripts/cruise_control/main.py

You may also run mpc.py similarly and test the costs by running cost.py in the same folder. Results will be saved under results/acc_results/.

Training

The trained model is saved in To reproduce the training, please run:

python3 train/train_acc.py --experiment_name experiment_acc --tMax 0. --tMin -1.1 --num_src_samples 10000 --pretrain --pretrain_iters 5000 --num_epochs 60000 --counter_end 50000

Other issues:

1. F1/10th Car dynamics can be found here, please refer to Single-Track Model section.

2. Some related projects:

   • The F1TENTH Gym environment
   • DeepReach: A Deep Learning Approach to High-Dimensional Reachability
   • refineCBF: Control Barrier Function refinement with HJ Reachability
   • hj_reachability: Hamilton-Jacobi reachability analysis in JAX

3. Contact: if you have any question on this repo, please feel free to contact the author Shuo Yang (yangs1 at seas dot upenn dot edu) or raise an issue.

Citation

If you find this work useful, please consider citing:

@article{yang2024learning,
  title={Learning Local Control Barrier Functions for Safety Control of Hybrid Systems},
  author={Yang, Shuo and Chen, Yu and Yin, Xiang and Mangharam, Rahul},
  journal={arXiv preprint arXiv:2401.14907},
  year={2024}
}