GSAT: Geometric Traversability Estimation using Self-supervised Learning with Anomaly Detection for Diverse Terrains

[IEEE ICRA 26] This repository is the official code for GSAT: Geometric Traversability Estimation using Self-supervised Learning with Anomaly Detection for Diverse Terrains.

Dongjin Cho, Miryeong Park, Juhui Lee, Geonmo Yang, Younggun Cho

Spatial AI and Robotics Lab (SPARO), Inha University

---

Overview

GSAT learns robot-specific traversability from LiDAR point clouds — without human supervision. It builds a positive hypersphere in latent space using anomaly detection, addressing the positive-only learning problem.

---

Getting Started

0. Environment Set up (Docker)

cd docker

# Build
docker compose -f docker-compose-gui-nvidia.yaml build

# Start container
docker compose -f docker-compose-gui-nvidia.yaml up -d

# Open shell
docker exec -it gsat /bin/bash

Inside the container, build the ROS workspace once:

cd ~/gsat_ws
catkin build
echo "source ~/gsat_ws/devel/setup.bash" >> ~/.bashrc
source ~/.bashrc

Build the CUDA voxelization op:

cd ~/gsat_ws/src/GSAT_Traversability/gsat
python3 setup.py build_ext --inplace

---

1. Data Collection (ROS)

See gsat_ros/README.md for topic configuration and output format.

# Terminal 1 — play rosbag
rosbag play <your_bag>.bag

# Terminal 2 — run collection node
roslaunch gsat_ros data_collection.launch

Output:

collect_data/gazebo/hill/original/
├── supervision.csv          # pose + Travel_label
└── lidar/
    └── <timestamp>.bin      # float32 x, y, z, intensity

---

2. Preprocess for Train Data

Step 1. Train Data Preprocess

Edit config/data_preprocess.yaml (set data_folder and output_folder), then run:

cd ~/gsat_ws/src/GSAT_Traversability/gsat/data_tools
python3 data_preprocess.py --key {data_name}

Output:

collect_data/gazebo/hill/process/
├── point/     # Leveled point clouds (*.bin)
└── label/     # Future trajectory + Travel_label (*.bin, K×4)

Step 2. Dataset Split

Edit config/data_split.yaml (set root_dir, preprocess_dir, save_dir, and split ratios), then run:

python3 data_split.py --key {data_name}

Output:

collect_data/gazebo/hill/dataset/
├── train/  point/  label/
├── val/    point/  label/
└── test/   point/  label/

---

3. Training

Edit gsat/config/train.yaml (set dataset_dir and save_dir), then run:

cd ~/gsat_ws/src/GSAT_Traversability/gsat
python3 train.py --key {data_name}

Checkpoints are saved as epoch_*.pth and best_model_*.pth under save_dir.

---

Demo

Two demos are available:

• Anomaly Classification — inference the anomalous(dissimilar positive samples), normal sample(similar positive sample).
• Traversability Navigation — navigate using predicted traversability map.

Anomaly classification Demo

Traversability-guided navigation Demo

🚧 Demo code coming soon...........

---

Citation

If you find this work useful, please cite:

@inproceedings{gsat2026,
  title     = {GSAT: Geometric Traversability Estimation using Self-supervised Learning with Anomaly Detection for Diverse Terrains},
  author    = {Cho, Dongjin and Park, Miryeong and Lee, Juhui and Yang, Geonmo and Cho, Younggun},
  booktitle = {IEEE International Conference on Robotics and Automation (ICRA)},
  year      = {2026}
}