connectedsystemslab/batmobility_ae

BatMobility🦇 [MobiCom'23]

Paper | arXiv | Project Page

This is a Pytorch implementation of the radio flow module from the following paper:

BatMobility: Towards Flying Without Seeing for Autonomous Drones
Emerson Sie, Zikun Liu, Deepak Vasisht
ACM International Conference on Mobile Computing and Networking (MobiCom), 2023

---

demo.mp4

1. Requirements and Dependencies

This code is tested on Ubuntu 18.04 and requires an NVIDIA GPU (we use an RTX 3090).

To setup the dependencies, create and activate the conda environment.

conda create env -f env.yaml
conda activate batmobility_ae

2. Evaluation

The test dataset includes the pre-trained models and test data used to generate the main results in the paper. The structure is as follows:

test_data/
|-- v4/               # This corresponds to the 15 Hz model described in the paper.
|-- v5/               # This corresponds to the 30 Hz model described in the paper.
|-- v6/               # This corresponds to the 40 Hz model described in the paper.
|   |-- v6_model.pth
|   |--  *.bag

Download the .zip file and unzip in this directory.

unzip test_data.zip

To test each pre-trained model on its corresponding test set, run

./main_eval.py

For each test .bag in each folder the script generates a plot showing a comparison between the corresponding model and the commercial optical flow sensor (PMW3901). In addition to the plot, the raw outputs of the model are also saved as .npz files.

3. Training

We also provide a larger dataset (~30 GB) to train models from scratch. It consists of a set of .bag files split into training and validation sets for each model.

train_data/
|-- v4/
|-- v5/
|-- v6/
|   |-- train/        # Contains .bags for training
|   |-- test/         # Contains .bags for validation

Download the .zip file and unzip in this directory.

unzip train_data.zip

To train the models, run

./main_train.py

The models and training curve plots will be generated in separate folders.

Citation

If you found this repository useful, please consider citing the paper:

@inproceedings{sie2023batmobility,
  author    = {Sie, Emerson and Liu, Zikun and Vasisht, Deepak},
  title     = {BatMobility: Towards Flying Without Seeing for Autonomous Drones},
  booktitle = {ACM International Conference on Mobile Computing (MobiCom)},
  year      = {2023},
  doi       = {https://doi.org/10.1145/3570361.3592532},
}