TerrainFreeSpaceNet

TerrainFreeSpaceNet is a deep learning model designed to predict terrain free-space from 3D point cloud data.
It enables robots to estimate terrain traversability in uneven environments using raw point cloud observations.

This model is based on a PointNet-style neural network that processes unordered 3D points and outputs a normalized free-space score.

---

Overview

Autonomous ground robots operating in outdoor environments often encounter:

• uneven terrain
• vegetation
• slopes and depressions
• irregular obstacles

Traditional geometric free-space detection methods may struggle in these environments.

TerrainFreeSpaceNet learns to estimate terrain traversability directly from 3D point clouds, making it suitable for:

• outdoor robotics
• off-road navigation
• agricultural robots
• exploration robots

---

Model Architecture

The model uses a PointNet-style architecture consisting of:

1. Shared MLP layers implemented using Conv1D
2. Batch normalization and ReLU activation
3. Global max pooling
4. Fully connected regression layers
5. Sigmoid output for normalized free-space score

Input shape:

[B, 3, N]

Where:

• B = batch size
• 3 = (x,y,z) coordinates
• N = number of sampled points

Output:

[B,1]

Free-space score in range:

• 0 → non-traversable
• 1 → highly traversable

---

Input Format

The model expects CSV point cloud input.

Example:

x,y,z
0.12,0.31,0.02
0.15,0.34,0.05
0.18,0.29,0.04

---

Example Inference

Example usage using the provided inference script.

from inference import run_inference

score = run_inference("sample_input.csv", checkpoint="model.pt")

print("Free space score:", score)

Example output:

Free space score: 0.84

Training

The model was trained using frame-level 3D point cloud samples.

Each frame contains:

frame_id,x,y,z,free_space

Where:

• frame_id identifies the point cloud frame

• free_space represents terrain traversability score

Applications

TerrainFreeSpaceNet can be used for:

• terrain-aware robot navigation

• autonomous ground vehicles

• off-road robotics

• agricultural robots

• exploration robots

• rough terrain mobility analysis

Limitations

• The model currently assumes XYZ coordinates only

• Performance depends on training dataset diversity

• Large terrain variations may require retraining

• Real-time deployment requires optimized inference

Related Research

This model was developed as part of research on:

Autonomous robot navigation in uneven terrain using 3D perception.

It is designed to integrate with the Agoraphilic-3D Navigation Framework.

Repository

Full project code available here:

https://github.com/dinusharg/TerrainFreeSpaceNet

Citation

If you use this model in research, please cite:

@article{10.1007/s12555-025-0624-2,
   author = {Gunathilaka, W. M. Dinusha and Kahandawa, Gayan and Ibrahim, M. Yousef and Hewawasam, H. S. and Nguyen, Linh},
   title = {Agoraphilic-3D Net: A Deep Learning Method for Attractive Force Estimation in Mapless Path Planning for Unstructured Terrain},
   journal = {International Journal of Control, Automation and Systems},
   volume = {23},
   number = {12},
   pages = {3790-3802},
   ISSN = {2005-4092},
   DOI = {10.1007/s12555-025-0624-2},
   url = {https://doi.org/10.1007/s12555-025-0624-2},
   year = {2025},
   type = {Journal Article}
}