[RAL 2026] Public code and model for ORION: Option-Regularized Deep Reinforcement Learning for Cooperative Multi-Agent Online Navigation.

🔹 We propose ORION, a distributed RL planner for multi-agent online navigation. It enables reactive role-playing and cooperation between individual search and team-level navigation via option-based networks and a dual-stage navigation strategy.

Environment Setup

conda create -n orion python=3.10 -y
conda activate orion

pip install torch torchvision
pip install opencv-python scikit-image imageio pandas
pip install matplotlib tensorboard
pip install ray wandb

Clone this repository and navigate to the directory.

git clone https://github.com/marmotlab/ORION-multi-agent-navigation.git
cd ORION-multi-agent-navigation

Datasets and Checkpoints

Training datasets are provided in:

• maps_priori/
• maps_GT/

Evaluation datasets are provided in:

• maps_priori_test_new_{n}/
• maps_GT_test_new_{n}/

where {n} denotes the number of agents in the team.

The training set consists of simple maps with 3 agents only.
During evaluation, ORION scales to larger teams (3, 4, 5, and 10 agents) and more complex environments without additional training.

We also provide a pretrained checkpoint. As ORION is a decentralized multi-agent navigation planner, the same checkpoint can be directly applied to different team sizes.

Examples of training (left) and evaluation (right) maps.

Training and Evaluation

For training, configure the parameters in parameter.py, then run:

python driver.py

For evaluation, configure the parameters in test_parameter.py, then run:

python test_driver.py

Inline comments are provided in both files to facilitate parameter configuration.

ROS2-based Deployment

We provide a ROS2-based deployment of ORION in the Multi-Robot-Development-Environment. This repository offers a multi-agent navigation and exploration framework, along with several simulation environments for development and evaluation.

Credit

If you find this work helpful, please consider citing:

@article{zhang2026orion,
  title={ORION: Option-Regularized Deep Reinforcement Learning for Cooperative Multi-Agent Online Navigation},
  author={Zhang, Shizhe and Liang, Jingsong and Zhou, Zhitao and Ye, Shuhan and Wang, Yizhuo and Tan, Derek Ming Siang and Chiun, Jimmy and Cao, Yuhong and Sartoretti, Guillaume},
  journal={IEEE Robotics and Automation Letters},
  year={2026},
  publisher={IEEE}
}

ORION is inspired by the following works, and we thank them for their contributions!

• Context-Aware Deep Reinforcement Learning for Autonomous Robotic Navigation in Unknown Area, CoRL 2023
• The Option-Critic Architecture, AAAI 2017
• ARiADNE ROS Planner, ICRA 2023/RA-L 2024
• CMU Development environment
• Octomap

Authors

Shizhe Zhang*, Jingsong Liang*, Zhitao Zhou, Shuhan Ye, Yizhuo Wang, Derek Ming Siang Tan, Jimmy Chiun, Yuhong Cao, Guillaume Sartoretti