Multi-Agent RL Global Path Planner

This repository provides the code for this paper: Collaborative Task and Path Planning for Heterogeneous Robotic Teams using Multi-Agent PPO

Its based on the JaxMARL library. It aims at computing the trajectories for a team of robots to a set of targets requiring specific skills on a 2D grid map. There is also the possibility of targets requiring multiple skills and subsequently a collaboration between multiple robots.

Install

Setup a virtual environment and install dependencies. Run the following commands to do this:

cd multi_robot_global_planner
python3 -m venv .venv
source venv/bin/activate
pip install -r .requirements.txt

This package was tested with Python 3.12 and 3.10.

Train a policy

Setup

1. First adjust the training parameters in config.yaml .
2. Adapt the environment settings in the simple_explo.py constructor.
3. To adjust the number of skills or agents, one also has to change the team config accordingly.

Run

You can either run python src/train_mappo.py directly, or you can modify and use the train_sequence.py script to run multiple trainings in sequence with modified training parameters.

➡️ The trained network parameters and the configs will be stored in a new folder called outputs.

Evaluation

The evaluation.py script can be used to load a trained policy and visualize it. Adapt the main body before you run it, to set your model and parameters.

Citation

Please cite the following paper if you are using this code:

@misc{rubio2026collaborativetaskpathplanning,
      title={Collaborative Task and Path Planning for Heterogeneous Robotic Teams using Multi-Agent PPO}, 
      author={Matthias Rubio and Julia Richter and Hendrik Kolvenbach and Marco Hutter},
      year={2026},
      eprint={2604.01213},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2604.01213}, 
}