rptamin / ddrl-firefighting

Distributed deep reinforcement learning for fighting forest fires with autonomous aerial vehicles.

ddrl-firefighting

A repository to support the paper Distributed Deep Reinforcement Learning for Fighting Forest Fires with a Network of Aerial Robots.

Video:

Paper citation:

@InProceedings{8593539, 
    author={R. N. Haksar and M. Schwager}, 
    booktitle={2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)}, 
    title={Distributed Deep Reinforcement Learning for Fighting Forest Fires with a Network of Aerial Robots}, 
    year={2018}, 
    pages={1067-1074}, 
    doi={10.1109/IROS.2018.8593539}, 
    ISSN={2153-0866}, 
    month={Oct},}

Requirements

• Developed with Python 3.6
• Requires numpy and pytorch (tested with version 1.4.0)
• Requires the simulators repository

Files

• madqn.py: Implementation of training algorithm, policy architecture, and aerial vehicle model.
• main.py: Example usage of algorithm: train and test a network with simulations.
• rlUtilities.py: Helper utilities to simplify implementation of the reinforcement learning problem.

Additional Changes and Notes

This repo is an extension of the work done by Ravi Haksar and Professor Mac Schwager for AAE590. The extension shall entail multiple fire propagation points and an analysis of the existing algorithms against such cases. Further modifications down the line shall entail implementations of time delay regarding refueling as well as terrain and wind impacts.