Generation of Geodesics with Actor-Critic Reinforcement Learning to Predict Midpoints

This is a repository for the following paper:

• Kazumi Kasaura. “Generation of Geodesics with Actor-Critic Reinforcement Learning to Predict Midpoints”, Journal of Machine Learning Research, 26(212):1-36, 2025.

It contains all scripts and a dockerfile to reproduce our experiments.

The contents of scripts/SGT_PG/ are a modification of the contents of https://github.com/tomjur/SGT-PG.

Experiments

Building an environment

You can use Dockerfile to build an environment.

Running

scripts/learn.py is the main script for our proposed methods. You can specify the environment, method variants, and hyperparameters with arguments.

scripts/ppo.py is the main script for the sequential reinforcement learning baseline.

scripts/SGT_PG/sgt_pg_main.py is the main script for the policy gradient baseline.

You can use a script scripts/run.sh to run all experiments.

The IDs of used GPU can be specified by editting this script.

The results are stored in exp folder.

mkdir exp
cd scripts
bash run.sh
cd ..

Viewing Results

After all experiments are done, results can be plotted by a script scripts/show_graph_4.py.

Examples of generated paths can be visualized by scripts. The images are stored in figures folder.

Calculation of lengths of generated paths can be done by a script scripts/compare_cost.py. The result is stored in data folder and the comparison table can be made by a script scripts/make_table.py.

mkdir figures
cd scripts
python3 show_graph_4.py
python3 view_trajs.py
python3 view_trajs_car3.py
python3 view_traj_obstacles.py
python3 visualize_panda.py
python3 visualize_multiagents.py
python3 compare_cost.py
python3 make_table.py
cd ..

License

This software is released under the MIT License, see LICENSE.

Citation

JMLR version

@article{JMLR:v26:24-1020,
  author  = {Kazumi Kasaura},
  title   = {Generation of Geodesics with Actor-Critic Reinforcement Learning to Predict Midpoints},
  journal = {Journal of Machine Learning Research},
  year    = {2025},
  volume  = {26},
  number  = {212},
  pages   = {1--36},
  url     = {http://jmlr.org/papers/v26/24-1020.html}
}

arXiv version

@article{kasaura2024generation,
  title={Generation of Geodesics with Actor-Critic Reinforcement Learning to Predict Midpoints},
  author={Kasaura, Kazumi},
  journal={arXiv preprint arXiv:2407.01991},
  year={2024}
}