Learned Inertial Model Odometry

Publication

If you use this code in an academic context, please cite the following RA-L 2023 paper.

G. Cioffi, L. Bauersfeld, E. Kaufmann, and D. Scaramuzza, "Learned Inertial Odometry for Autonomous Drone Racing," IEEE Robotics and Automation Letters (RA-L). 2023.

@InProceedings{CioffiRal2023
  author = {Cioffi, Giovanni and Bauersfeld, Leonard and Kaufmann, Elia and Scaramuzza, Davide},
  title = {Learned Inertial Odometry for Autonomous Drone Racing},
  booktitle = {IEEE Robotics and Automation Letters (RA-L)},
  year = {2023}
}

Installation

These instructions have been tested on Ubuntu 20.04 and Python 3.9.

Create a conda environment, named imo, containing all the dependencies for the project:

conda create -n imo --file conda_environment.yaml

You might need to install pyquaternion manually. To do so:

pip install pyquaternion

Prepare dataset

You need to convert your data in the specific .hdf5 format required by the network.

The script used to convert the data from the Blackbird dataset is in learning/data_management/prepare_datasets/blackbird.py.

An sample output for one of the trajectories used from the Blackbird dataset is in the folder dataset.

You can use this script as reference to convert your own dataset.

Learning Component

Run network training with:

python src/main_learning.py --root_dir=datasets --out_dir=results --dataset=Blackbird --mode=train --imu_freq=100 --sampling_freq=100 --window_time=0.5

The file src/main_learning.py contains the hyperparameters.

Run network on the test sequences:

python src/main_learning.py --root_dir=datasets --out_dir=results --dataset=Blackbird --mode=test --imu_freq=100 --sampling_freq=100 --window_time=0.5 --model_fn=net_blackbird.pt --show_plots

The training logs can be visualized in tensorboard:

tensorboard --logdir=results/Blackbird/logs

We provide the network trained on the selected trajectories from the Blackbird dataset (see paper for details) in results/Blackbird/checkpoints/model_net/net_blackbird.py.

Learned Inertial Odometry

Run the EKF with learned model updates with:

python src/main_filter.py --root_dir=datasets --out_dir=results --dataset=Blackbird --data_list=test.txt --checkpoint_fn=net_blackbird.pt --model_param_fn=model_net_parameters_net_blackbird.json

Plot filter output:

python src/filter/python/plot_filter_output.py --dataset_dir=datasets --result_dir=results --dataset=Blackbird --seq=clover/yawForward/maxSpeed5p0/test

For a more detailed evaluation of the filter performance use this.

Credits

This repo uses some external open-source code:

• TLIO
• RPG trajectory evaluation toolbox

Refer to each open-source code for the corresponding license.

If you note that we missed the information about the use of any other open-source code, please open an issue.