NeHMO: Neural Hamilton-Jacobi Reachability Learning for Decentralized Safe Multi-Agent Motion Planning

Paper | Video Demo

We present Neural HJR-guided Multi-agent Motion Optimizer (NeHMO), a method that provides scalable neural Hamilton-Jacobi Reachability (HJR) modeling to tackle high-dimensional configuration spaces and enables decentralized multi-agent motion planning (MAMP). The learned HJR value functions are incorporated into a decentralized trajectory optimization framework, which solves MAMP tasks in real-time. We demonstrate that our method is both scalable and data-efficient, enabling the solution of MAMP problems in higher-dimensional scenarios with complex collision constraints. The proposed approach generalizes across various dynamical systems, including a 12-dimensional dual-arm setup, and outperforms a range of state-of-the-art techniques in successfully addressing challenging MAMP tasks.

• Python 3.11 is used in this project.
• Run pip install -r requirements.txt to collect all python dependencies.
• zonopy and zonopy-robot are required by the simulation environment. Please refer to their respective repositories following the hyperlinks for insllation instructions.

• To download the trained HJR models from Google Drive:

pip install gdown
gdown --folder https://drive.google.com/drive/folders/1JuIAfa-2UHhuDMRdOxaDspDKPXtUevaG?usp=sharing

• To run manipulator planning experiments:

bash planning_scripts/run_UR5_planning.sh         # Dual UR5
bash planning_scripts/run_triple_UR5_planning.sh  # Three UR5
bash planning_scripts/run_five_UR5_planning.sh    # Five UR5

For visulizations, one may specify the --video flag in the bash scripts. The experiments will generate planning results as a JSON file under planning_results/.

• To run particle planning experiments:

bash run_Particle_planning.sh  

This will run the experiments with 2, 8, 16, 32 agents and may therefore take long. Users can adjust the parameters in the bash scripts for their use cases; for visulizations, one may specify the --video flag in the bash scripts. The experiments will generate planning results as a JSON file under planning_results/. To also save planned trajectories, the user may substitute --save_stats with --save_traj.

• To download the validation sets from Google Drive:

gdown --folder https://drive.google.com/drive/folders/1I5blpIog4UybxKygOESfnkWUNMNZmXvu?usp=sharing -O deepreach/

• To run the learning experiments:

bash deepreach/launch_particle_training.sh
bash deepreach/launch_air3D_training.sh
bash deepreach/simplearm_training.sh
bash deepreach/launch_nn_arm_training.sh

Note that to run the manipulator training example, the user needs to acquire the learned distance boundary condition model by

gdown --folder https://drive.google.com/drive/folders/1w8HmZ9S_PWs1gPAbXUaYyyw7F2RM_om6?usp=sharing -O UR5_datasets_and_training/

The UR5_datasets_and_training/ directory also contains the code to train the boundary condition models for reference.

• A majority of the code for HJR learning is adopted from DeepReach. We thank the authors and maintainers for their amazing work.
• The simulation environment is adopted from Sparrows. We thank the authors and maintainers for their amazing work.

If you find NeHMO useful, please consider citing using the following BibTex entry:

@misc{chen2025nehmoneuralhamiltonjacobireachability,
      title={NeHMO: Neural Hamilton-Jacobi Reachability Learning for Decentralized Safe Multi-Agent Motion Planning}, 
      author={Qingyi Chen and Ahmed H. Qureshi},
      year={2025},
      eprint={2507.13940},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2507.13940}, 
}