UAV-based path planning for efficient localization of non-uniformly distributed weeds using prior knowledge: A reinforcement-learning approach

UAV-based path planning for efficient localization of non-uniformly distributed weeds using prior knowledge: A reinforcement-learning approach\ Rick van Essen, Eldert van Henten, and Gert Kootstra Paper: https://doi.org/10.1016/j.compag.2025.110651

About

This is the code belonging to the paper 'UAV-based path planning for efficient localization of non-uniformly distributed weeds using prior knowledge: A reinforcement-learning approach'. It contains the simulation environment and the code to train and reproduce the results of the paper.

Installation

Python 3.12 or higher is needed with all dependencies listed in requirements.txt. Install using:

commandline pip install --upgrade pip pip install -e . export DATA_HOME=<<path to orthomosaic folders>> # Only needed when running experiment 5

Additionally, Fields2Cover is needed to calculate the baseline row-by-row flight path. See their website for installation instructions.

Alternatively, you can use the provided dev container.

For experiment 5, the orthomosaics can be downloaded from https://doi.org/10.4121/bbe97051-07df-4934-b634-701d91a2075e.

Usage

Train RL agent

To train the RL agent, you need to define an experiment containing the simulation configuration and the training parameters. The experiments folder contains the configuration files for the experiments done in the paper. Then use the bash script to train:

commandline ./run_trainings <<experiment name>>

Evaluate RL agent

To evaluate a trained RL agent, use the evaluation script. To reproduce the experiments done in the paper, run:

commandline ./run_evaluations <<experiment name>>

Then use create_results_plots.ipynb to reproduce the plots.

Play

To control the drone by keyboard, run:

commandline python3 -m drone_grid_env.utils.play --env_args config_file=<<path to config file>>

Use ASDW to move the drone 1 step in each direction and L to land the drone (when enabled in the configuration file).

Citation

If you find this code usefull, please consider citing our paper: @article{vanEssen2025, author = {van Essen, Rick and van Henten, Eldert and Kootstra, Gert}, title = {{UAV}-based path planning for efficient localization of non-uniformly distributed weeds using prior knowledge: A reinforcement-learning approach}, journal = {Computers and Electronics in Agriculture}, volume = {237}, pages = {110651}, year = {2025}, issn = {0168-1699}, doi = {10.1016/j.compag.2025.110651} }

Funding

This research is part of the research program SYNERGIA, funding was obtained from the Dutch Research Council (NWO grant 17626), IMEC-One Planet and other private parties.