Skip to content

Code for training the vision pipeline in 'Fully neuromorphic vision and control for autonomous drone flight'

License

Notifications You must be signed in to change notification settings

tudelft/event_planar

Repository files navigation

Fully neuromorphic vision and control for autonomous drone flight

Work published in Science Robotics [project page]. If you use this code, please cite our work:

@article{paredesvalles2024fully,
  author = {F. Paredes-Vallés  and J. J. Hagenaars  and J. Dupeyroux  and S. Stroobants  and Y. Xu  and G. C. H. E. de Croon},
  title = {Fully neuromorphic vision and control for autonomous drone flight},
  journal = {Science Robotics},
  volume = {9},
  number = {90},
  pages = {eadi0591},
  year = {2024},
  doi = {10.1126/scirobotics.adi0591}
}

This code allows for the training of the 'vision spiking network' as shown in part C of the figure below.

Training

The raw dataset from the article can be downloaded from here. We have a processed version of it here. By default, this dataset is expected at data/datasets/.

To train a model defined in one of the configs/*.yaml files, run:

python train.py --config configs/<config-of-choice>.yaml

You can track the training progress with MLflow:

mlflow ui

and accessing http://127.0.0.1:5000/.

Inference

To estimate planar optical flow from the test sequences from our dataset and compare against ground-truth data, run:

python eval_flow.py <model_runid>

where <model_runid> is the ID of the run to be evaluated (check MLflow).

About

Code for training the vision pipeline in 'Fully neuromorphic vision and control for autonomous drone flight'

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages