COMPASS

Xingjian Zhang, Yizhuo Wang, Guillaume Sartoretti

This repository hosts the code for paper COMPASS https://arxiv.org/abs/2507.16306, which was accepted by MRS 2025.

*This repository is still under construction.

Training Tips

If you are using NSCC of Singapore to train, you could use the following commands to ease your experience.

qsub train_nscc.pbs
export PBS_JOBID=*
qstat -f
ssh *
module purge
module load miniforge3
conda activate xingjian
module load cuda/12.2.2
nvidia-smi
cd scratch/win-STAMP-main/
python3 driver.py

qstat - To see all the tasks
qdel -  To delete certain task

Requirements

python >= 3.9
pytorch >= 1.11
ray >= 2.0
ortools
scikit-image
scikit-learn
scipy
imageio
tensorboard

Training

1. Set appropriate parameters in arguments.py -> Arguments.
2. Run python driver.py.

Evaluation

1. Set appropriate parameters in arguments.py -> ArgumentsEval.
2. Run python /evals/eval_driver.py.

Files

• arguments.py: Training and evaluation arguments.
• driver.py: Driver of training program, maintain and update the global network.
• runner.py: Wrapper of the local network.
• worker.py: Interact with environment and collect episode experience.
• network.py: Spatio-temporal network architecture.
• env.py: Persistent monitoring environment.
• gaussian_process.py: Gaussian processes (wrapper) for belief representation.
• /evals/*: Evaluation files.
• /utils/*: Utility files for graph, target motion, and TSP.
• /model/*: Trained model.