The code in this repository was used for winning the ICRA 2022 DodgeDrone Challenge in the vision-based category. It uses a fork of Agile Flight for the training environment and depends on a fork of NEAT Python and a fork of PyTorch NEAT.
- First, make sure you complete the installation of the fork of Agile Flight. Only the ROS version is required.
- Next, make sure to
pip installthe fork of NEAT Python and the fork of PyTorch NEAT. - Finally,
pip installthis actual repository.
- First launch ROS by running
roslaunch cfg/tools.launchin a terminal. - Next:
- either, for training, run
python -m flightevo.dodge_trainerin a separate terminal, - or, for evalutation, run
python -m flightevo.dodge_evaluatorin a separate terminal.
- either, for training, run
You can have a look at the arguments of these modules to play around with different settings.
For an analysis of the methodology and final results, please look in the analysis folder.
Next to the contributors of the original repositories that were used, this repository is created in collaboration with guidoAI and NPU-yuhang, and the MAVLab organization.