This repository contains the code for a bandwidth-efficient task-driven point cloud compression solution, developed as part of the paper titled "Collaborative Exploration with a Marsupial Ground-Aerial Robot Team through Task-Driven Map Compression" .
We propose a bandwidth-efficient, task-driven point cloud compression method tailored for volumetric map reconstruction at mission-relevant resolutions. By emphasizing occupancy-relevant structure over raw point cloud fidelity, our approach achieves high compression rates while retaining the information essential for planning.
The proposed solution supports data gathered from two different LiDAR sensors:
- Ouster OS0-64 (used by the aerial robot)
- Velodyne VLP-16 (used by the ground robot).
cd ~
git clone git@github.com:ntnu-arl/pcl-vae.gitTo install the repository, run the following commands:
cd ~/pcl_vae
pip3 install -e .The folders contain the following:
- datasets: Contain scripts that utilize pytorch's dataset class to read from dataset files
- inference: Contains the scripts for running the pcl_vae node
- networks: Contains the VAE networks, and the loss function for training the VAEs
- weights: Contains the weights for the VAEs
For each robot, we provide a dataset containing both real and simulated range images from diverse environments, including caves, confined spaces, and complex buildings. Relevant files here
cd ~/pcl-vae/pcl_vae/datasets/
wget -O datasets.zip "https://ndownloader.figshare.com/files/53055530?private_link=ce4e4b87b3a28a75be27"
unzip datasets.zip - Range images (
64×512 (H×W)) generated by OS0-64 LiDAR sensor. - Includes
$\sim36,000$ range images ($\sim26,000$ simulated) - Training set (
$90 %$ ) and Testing set ($10 %$ )
- Range images (
16×1800 (H×W)) generated by VLP-16 LiDAR sensor. - Includes
$\sim25,000$ ($\sim21,000$ real). - Training set (
$90$ %) and Testing set ($10$ %)
For each robot, we provide pre-trained models with varying latent space size {32,64,128,256,512,1024} and voxel size {0.2,0.3,0.4}m. Relevant files here
For aerial robot models
cd ~/pcl-vae/pcl_vae/weights/
wget -O aerial_robot_pcl_vae_models.zip "https://ndownloader.figshare.com/files/53083268?private_link=ce4e4b87b3a28a75be27"
unzip aerial_robot_pcl_vae_models.zip For ground robot models
cd ~/pcl-vae/pcl_vae/weights/
wget -O ground_robot_pcl_vae_models.zip "https://ndownloader.figshare.com/files/53083163?private_link=ce4e4b87b3a28a75be27"
unzip ground_robot_pcl_vae_models.zip To validate a pre-trained model, you need to modify the parameters in the configuration file to match the desired settings. For both robot types, a vae_validation_config.yaml file exists under the inference/config/<robot-type> folder with <robot-type> = {aerial, ground}. A detailed description of all parameters is included in the configuration file.
Note: Set model name, latent_space, and voxel_size variables in inference/config/<robot-type>/vae_validation_config.yaml.
Run the following command specifying the robot type. Replace <robot-type> with aerial or ground for validate the the pre-trained model for aerial or ground robot, respectively.
cd pcl_vae/pcl_vae/inference/src
python3 vae_node_validation.py --robot_type=<robot-type>An example of the output is shown in the figure below, which includes the raw range image input to the model, its voxel-aware representation, and the reconstructed range image. range_image_comparison
Note: To go through the entire dataset, press n to move from image to image.
To train a pcl_vae model, you need to modify the parameters in the configuration file to match the desired settings. For all robot types, a train_config.yaml file exists under the train/config/<robot-type> folder with <robot-type> = {aerial, ground}. A detailed description of all parameters is included in the configuration file.
To start training the model, run the following command, specifying the robot type. Replace <robot-type> with aerial or ground for training a pcl_vae for aerial or ground robot, respectively.
cd ~/pcl_vae/pcl_vae/train
python3 train_pcl_vae.py --robot_type=<robot-type>Note: The model is saved under the train/weights/<robot-type>/<robot-type>_model folder.
To validate the trained model, you need to modify the parameters in the configuration file to match the desired settings. For both robot types, a vae_validation_config.yaml file exists under the inference/config/<robot-type> folder with <robot-type> = {aerial, ground}. A detailed description of all parameters is included in the configuration file.
Note: Move the trained model to weights folder, set model name, latent_space, and voxel_size variables in inference/config/<robot-type>/vae_validation_config.yaml.
Run the following command specifying the robot type. Replace <robot-type> with aerial or ground for validate the the pre-trained model for aerial or ground robot, respectively.
cd pcl_vae/pcl_vae/inference/src
python3 vae_node_validation.py --robot_type=<robot-type>If you use this work in your research, please cite the following publication:
@article{zacharia2025collaborative,
title={Collaborative Exploration with a Marsupial Ground-Aerial Robot Team through Task-Driven Map Compression},
author={Zacharia, Angelos and Dharmadhikari, Mihir and Alexis, Kostas},
journal={IEEE Robotics and Automation Letters},
year={2025},
publisher={IEEE}
}Released under BSD-3-Clause.
This open-source release is based upon work supported by the European Commission through:
- Project SYNERGISE under the Horizon Europe Grant Agreement No. 101121321
- Project SPEAR under the Horizon Europe Grant Agreement No. 101119774
- Project DIGIFOREST under the Horizon Europe Grant Agreement No. 101070405
For questions or support, reach out via GitHub Issues or contact authors: