LeSTA

Learn robot-specific traversability from a short manual drive — no labels needed

RA-L · ROS1 · ROS2 · Dataset

LeSTA is a self-supervised framework that learns navigational capability of mobile robots. By leveraging a short period of manual driving, it provides an end-to-end pipeline from label generation to real-time traversability prediction and mapping.

🎉 News & Updates:

• 2024.07.30: Our paper is accepted for presentation at IEEE ICRA@40 in Rotterdam, Netherlands
• 2024.02.29: Our paper is accepted by IEEE Robotics and Automation Letters (RA-L)
• 2024.02.19: Release public dataset for learning terrain traversability in urban environments

🧩 Related projects:

• FastDEM — Ultra-fast elevation mapping on embedded robots
• EviGround — (in preparation)

---

Installation

Our project is built on ROS, successfully tested on the following setup.

• Ubuntu 20.04 / ROS Noetic
• PyTorch 2.2.2 / LibTorch 2.6.0

ROS2 support is coming soon.

lesta_ros

1. Install Grid Map library for height mapping:

   sudo apt install ros-noetic-grid-map -y

2. Install LibTorch (choose one option):

   CPU-only version (Recommended for easier setup)

   wget https://download.pytorch.org/libtorch/cpu/libtorch-cxx11-abi-shared-with-deps-2.6.0%2Bcpu.zip -P ~/Downloads
   sudo unzip ~/Downloads/libtorch-cxx11-abi-shared-with-deps-2.6.0+cpu.zip -d /opt
   rm ~/Downloads/libtorch-cxx11-abi-shared-with-deps-2.6.0+cpu.zip

   GPU-supported version (e.g. CUDA 11.8)

   # To be updated...

3. Build lesta_ros package:

   cd ~/ros_ws/src
   git clone https://github.com/Ikhyeon-Cho/LeSTA.git
   cd ..
   catkin build lesta
   source devel/setup.bash

Notes:

• We recommend starting without GPU processing. The network effectively runs on a single CPU core.
• If you are interested in height map reconstruction, see height_mapping for more details.

pylesta

1. Install PyTorch (choose one option):

   CPU-only setup

   
   

   We recommend using a virtual environment for PyTorch installation.

   Conda

   conda create -n lesta python=3.8 -y
   conda activate lesta
   conda install pytorch=2.2 torchvision cpuonly tensorboard -c pytorch -y

   Virtualenv

    virtualenv -p python3.8 lesta-env
    source lesta-env/bin/activate
    pip install torch==2.2 torchvision tensorboard --index-url https://download.pytorch.org/whl/cpu

   CUDA setup

   
   

   We recommend using a virtual environment for PyTorch installation.

   Conda

   conda create -n lesta python=3.8 -y
   conda activate lesta
   conda install pytorch=2.2 torchvision tensorboard cudatoolkit=11.8 -c pytorch -c conda-forge -y

   Virtualenv

   virtualenv -p python3.8 lesta-env
   source lesta-env/bin/activate
   pip install torch==2.2 torchvision tensorboard --index-url https://download.pytorch.org/whl/cu118

2. Install pylesta package:

   # Make sure your virtual environment is activated
   cd LeSTA
   pip install -e pylesta

Using Docker

To be updated...

---

Usage

The pipeline consists of three steps: label generation → model training → traversability estimation.

Note: We highly recommend training with your own robot's data. The robot's unique sensor setup and motion dynamics are crucial for accurate traversability predictions. For rapid testing, use checkpoints from #Model Zoo and skip to Step 3.

1. Label Generation

Launch ROS node

roslaunch lesta label_generation.launch

Generate labels with rosbag

Note: See #dataset for example rosbag files.

rosbag play {your-rosbag}.bag --clock -r 3

Save traversability labels

rosservice call /lesta/save_label_map "training_set" ""  # {filename} {directory}

The labeled height map will be saved as a single training_set.pcd file in the root directory of the package.

2. Model Training

Launch training script with parameters

Note: See pylesta/configs/lesta.yaml for more training details.

# Make sure your virtual environment is activated
cd LeSTA
python pylesta/tools/train.py --dataset "training_set.pcd"

3. Traversability Estimation

Prerequisites

Configure model_path variable in lesta_ros/config/*_node.yaml with your model checkpoint.

• trav_prediction_node.yaml
• trav_mapping_node.yaml

Note: See #model-zoo for our pre-trained checkpoints.

Launch ROS node

We provide two options for traversability estimation:

Left: Robot-centric traversability prediction. Right: Real-time traversability mapping.

1. Traversability Prediction	2. Traversability Mapping
Robot-centric local traversability Suitable for local motion planning Score traversability from model inference	Global traversability mapping Suitable for global path planning Update traversability scores over time

How to run:

1. For traversability prediction:

   roslaunch lesta traversability_prediction.launch

2. For traversability mapping:

   roslaunch lesta traversability_mapping.launch

Test the node with rosbag

rosbag play {your-rosbag}.bag --clock -r 2

---

Dataset

We release the public dataset used in our paper:

• urban-traversability-dataset — labeled height maps for traversability learning in urban environments

Quick test

Sample rosbags configured to run with the default settings:

• Campus road [Google Drive]

• Parking lot [Google Drive]

Model Zoo

Model	Description	Environment	Features	Download
LeSTA-parking-lot	The model trained on parking lot dataset	Urban (parking lot with low-height curbs)	Step Slope Roughness Curvature
LeSTA-campus-road	The model trained on campus road dataset	Urban (campus roads with flat terrain and hills)	Step Slope Roughness Curvature

---

Known Issues

• Artifacts from dynamic objects:

  • We currently implemented a raycasting-based approach to remove artifacts from dynamic objects.
  • This is crucial for accurate static terrain representation, which directly impacts prediction quality.
  • Yet, not enough to handle all artifacts.
  • We are working on more robust methods to detect and filter dynamic objects in real-time.

• Performance degradation due to noisy height mapping:

  • Traversability is learned and predicted from a dense height map.
  • The dense height map is accomplished by concatenating many sparse LiDAR scans.
  • A good SLAM / 3d pose estimation is required to get a good height map.
  • In typical settings, FAST-LIO2, LIO-SAM, or CT-ICP are good starting points.
  • We are working on improving the height mapping accuracy.

---

Citation

Thank you for citing our paper if this helps your research project:

Ikhyeon Cho, and Woojin Chung. 'Learning Self-Supervised Traversability With Navigation Experiences of Mobile Robots: A Risk-Aware Self-Training Approach', IEEE Robotics and Automation Letters, Feb. 2024.

@article{cho2024learning,
  title={Learning Self-Supervised Traversability With Navigation Experiences of Mobile Robots: A Risk-Aware Self-Training Approach}, 
  author={Cho, Ikhyeon and Chung, Woojin},
  journal={IEEE Robotics and Automation Letters}, 
  year={2024},
  volume={9},
  number={5},
  pages={4122-4129},
  doi={10.1109/LRA.2024.3376148}
}

You can also check the paper of our baseline:

Hyunsuk Lee, and Woojin Chung. 'A Self-Training Approach-Based Traversability Analysis for Mobile Robots in Urban Environments', IEEE International Conference on Robotics and Automation (ICRA), 2021.

@inproceedings{lee2021self,
  title={A self-training approach-based traversability analysis for mobile robots in urban environments},
  author={Lee, Hyunsuk and Chung, Woojin},
  booktitle={2021 IEEE International Conference on Robotics and Automation (ICRA)},
  pages={3389--3394},
  year={2021},
  organization={IEEE}
}

---

Contact: tre0430@korea.ac.kr

Apache-2.0 License © Ikhyeon Cho