GMC-DRL/MetaBox-uav

UAV Path Planning Benchmark Suite

This dataset provides a collection of 56 benchmark instances for testing and evaluating global optimization algorithms in the context of Unmanned Aerial Vehicle (UAV) path planning. It is based on the benchmark proposed in the paper:

Benchmarking Global Optimization Techniques for Unmanned Aerial Vehicle Path Planning
arXiv:2501.14503

The original implementation in MATLAB is available at:
👉 Zenodo Record (MATLAB Code)

We reimplemented the benchmark in Python using matlab.engine, converting the original Model56.mat file into a Python-friendly .pkl format for easy loading and experimentation.

---

📦 Dataset Overview

Problem Suite Composition

• A total of 56 benchmark instances are included.
• These instances are selected from a pool of 5,000 automatically generated terrains, with 56 diverse terrains manually curated for their challenging features and realistic complexity.
• Each terrain may include:
  • Flat plains
  • Rolling hills
  • Steep slopes
  • Deep valleys
• Two obstacle density configurations are used:
  • Sparse scenario: 15 cylindrical threats
  • Dense scenario: 30 cylindrical threats
• Obstacles are modeled as cylinders with varying heights and radii, simulating real-world structures such as radar towers or missile defense units.
• The UAV must navigate from a start point to a goal point, each placed at opposite corners of the map.
• Both the start and goal are located in safe, threat-free regions to ensure feasible solution spaces.

Instance Composition

• Each of the 28 selected terrains is evaluated under 2 obstacle configurations, resulting in:

  28 terrains × 2 threat settings = 56 total benchmark instances
  

---

🔧 Format & Usage

• Each instance is stored in a Python pickle (.pkl) format, containing:
  • Terrain elevation grid
  • Obstacle locations and sizes
  • Start and goal coordinates
• The dataset is suitable for benchmarking global optimization, black-box optimization (BBO), and reinforcement learning approaches to 3D path planning under constraints.

---

🖼️ Example Visualization: Terrain53

Below is an example of one benchmark instance (Terrain53) under the sparse threat scenario.

• Left: 3D terrain surface with elevation and cylindrical threats (obstacles).
• Right: 2D top-down projection with start (bule) and goal (yellow) points visualized.

---

📄 Citation

If you use this dataset in your research, please cite the original paper:

@article{shehadeh2025benchmarking,
  title={Benchmarking global optimization techniques for unmanned aerial vehicle path planning},
  author={Shehadeh, Mhd Ali and Kudela, Jakub},
  journal={arXiv preprint arXiv:2501.14503},
  year={2025}
}