IAI-CAU/TopoDroneX

TopoDroneX: A Multimodal Benchmark Dataset for Topology-optimized 3D Drone Design Exploration

Yeongtae Kim, Hoonhyung Chung, Donghyun Ra, Sooyoung Lee^†

Industrial AI Lab (IAI Lab)

School of Mechanical Engineering, Chung-Ang University

---

Overview

Unmanned aerial vehicle (UAV) design requires the optimal configuration of components and frame geometries to satisfy mission-specific performance requirements. However, conventional drone design processes often rely on repeated design modifications, numerical analysis, and flight tests, leading to substantial computational cost and time consumption.

Although data-driven approaches have recently gained attention as an alternative, the lack of multi-modal design datasets that jointly capture the complex mechanical characteristics and geometric diversity of multicopters has limited the development of AI-based drone design methods.

To address this limitation, this study proposes a large-scale drone design dataset consisting of 1,961 topology-optimized frame geometries generated under diverse component configurations and mechanical performance conditions, namely a TopoDroneX.

Proposed dataset is constructed based on five key design factors, including payloads and propulsion systems, and provides engineering attributes such as thrust, torque, and stress, together with multi-modal data including multi-view 2D projection images, 3D meshes, and point clouds.

We provide benchmark results for representative prediction and generation tasks across different data modalities. By incorporating both geometric information and physics-related design attributes, proposed dataset provides a physically grounded dataset for drone design optimization and broader data-driven engineering design.

---

Dataset Structure

Property	Value
Total samples	1,961
Modalities	2D images (8 views/sample), 3D mesh (STL + PLY), 3D point cloud, design parameters & specifications
Total size	21.4 GB
License	CC BY-NC 4.0

2D Design Model

Design Shape	Stress Field Projection
Four direction (top, front, side, isometric) views per sample. Stored in img/ shards as {sample_id}.{view}.png.	Four direction views with stress distribution overlaid. Stored in img_s/ shards as {sample_id}.{view}.png.

3D Design Model

Mesh Model	Point Cloud
3D mesh files in STL and PLY formats, stored in separate shards. geo_stl/ contains {sample_id}.stl and geo_ply/ contains {sample_id}.ply.	Point clouds with per-point Von Mises stress. Stored in pc/ as Parquet files with columns: sample_id, x, y, z, mises_stress (Pa).

Design Parameter & Specification

metadata/spec.csv provides 17 columns per design entry, grouped by category below.

Propulsion System	Material	Battery	Payload	Acceleration	Frame
Propeller diameter (inch)	Material density (kg/m³)	Battery length (mm)	Payload weight (kg)	Acceleration (G)	Motor to Motor Distance (mm)
Propulsion system thrust (kgf)	Young's modulus (GPa)	Battery width (mm)			Frame Height (mm)
Propulsion system torque (Nm)	Poisson ratio	Battery height (mm)			Frame Volume (m³)
	Tensile yield strength (MPa)	Battery weight (kg)			Frame Mass (kg)

---

Repository Structure

TopoDroneX/
├── README.md
├── checksums.sha256            # SHA-256 for all shard files
│
├── metadata/
│   ├── spec.csv                # 1,961 samples x 18 columns (params + shard_index)
│   ├── train_files.csv         # 1,568 sample IDs for training
│   ├── test_files.csv          # 197 sample IDs for test
│   └── valid_files.csv         # 196 sample IDs for validation
│
├── pc/                         # 3D point cloud (x, y, z, mises_stress)
│   ├── pc-00000.parquet        # samples 0000~0099
│   ├── pc-00001.parquet        # samples 0100~0199
│   └── ... pc-00019.parquet
│
├── img/                        # 2D design shape images (4 views/sample)
│   ├── img-00000.tar           # samples 0000~0099, 400 PNG
│   └── ... img-00019.tar
│
├── img_s/                      # 2D stress field projection images (4 views/sample)
│   ├── img-s-00000.tar
│   └── ... img-s-00019.tar
│
├── geo_stl/                    # 3D mesh, STL format only
│   ├── geo-stl-00000.tar       # samples 0000~0099, 100 STL files
│   └── ... geo-stl-00019.tar
│
├── geo_ply/                    # 3D mesh, PLY format only
│   ├── geo-ply-00000.tar       # samples 0000~0099, 100 PLY files
│   └── ... geo-ply-00019.tar
│
└── sample/                     # Quick-start subset (10 samples)
    ├── sample.parquet
    ├── sample_img.tar
    ├── sample_img_s.tar
    ├── sample_geo_stl.tar
    └── sample_geo_ply.tar

Note: Each folder contains 20 shards (index 00000~00019). The last shard (index 00019) contains 61 samples; all others contain 100. Shard indices are aligned across all modalities: pc-00000.parquet, img-00000.tar, img-s-00000.tar, geo-stl-00000.tar, and geo-ply-00000.tar all cover the same 100 samples.

---

Data Format

Point Cloud (pc/)

Each Parquet file contains rows from multiple samples. Columns:

Column	Type	Description
sample_id	string	Canonical sample key, e.g. mat=AlSi10Mg_mot=10-1_m_p=0.00_batt=1_acc=1.0
x	float64	X coordinate (meters)
y	float64	Y coordinate (meters)
z	float64	Z coordinate (meters)
mises_stress (Pa)	float64	Von Mises stress (Pascals)

Images (img/, img_s/)

TAR archives. Each file inside follows the naming convention:

{sample_id}.{view}.png

Views: top, front, side, iso

• img/: clean design shape
• img_s/: design with stress distribution overlaid

Geometry (geo_stl/, geo_ply/)

STL and PLY are provided in separate TAR archives so users can download only the format they need.

# geo_stl/geo-stl-XXXXX.tar
{sample_id}.stl

# geo_ply/geo-ply-XXXXX.tar
{sample_id}.ply

Metadata (metadata/spec.csv)

One row per sample. Key columns:

Column	Description
params	Canonical sample ID (join key for all modalities)
shard_index	Integer 0~19, maps sample to shard file
Propulsion system	Propeller diameter, thrust, torque
Material	Density, Young's modulus, Poisson ratio, tensile yield strength
Battery	Dimensions and weight
Payload	Payload weight
Acceleration	Design acceleration (G)
Frame	Motor-to-motor distance, height, volume, mass

---

Loading the Dataset

Quick Start (smoke test)

import pandas as pd
import tarfile

# Point cloud: 10 samples
df = pd.read_parquet("sample/sample.parquet")
print(df["sample_id"].unique())      # 10 unique sample IDs
print(df.columns.tolist())           # sample_id, x, y, z, mises_stress (Pa)

# Images
with tarfile.open("sample/sample_img.tar") as tar:
    # filenames: {sample_id}.{view}.png  (view: top/front/side/iso)
    print([m.name for m in tar.getmembers()[:4]])

# STL meshes
with tarfile.open("sample/sample_geo_stl.tar") as tar:
    print([m.name for m in tar.getmembers()[:3]])   # {sample_id}.stl

# PLY meshes
with tarfile.open("sample/sample_geo_ply.tar") as tar:
    print([m.name for m in tar.getmembers()[:3]])   # {sample_id}.ply

Load Point Cloud by Split

import pandas as pd

# Load benchmark split
train_ids = set(pd.read_csv("metadata/train_files.csv")["params"])

# Find which shards contain training samples
spec = pd.read_csv("metadata/spec.csv").set_index("params")
train_shards = spec.loc[list(train_ids), "shard_index"].unique()

# Load only relevant shards, filter to train set
dfs = []
for shard_idx in sorted(train_shards):
    df = pd.read_parquet(f"pc/pc-{shard_idx:05d}.parquet")
    dfs.append(df[df["sample_id"].isin(train_ids)])

train_df = pd.concat(dfs, ignore_index=True)
print(f"Train set: {train_df['sample_id'].nunique()} samples, {len(train_df):,} points")

Load Images from TAR

import tarfile
from PIL import Image
import io

def load_images_from_shard(shard_idx, sample_ids=None):
    results = {}
    with tarfile.open(f"img/img-{shard_idx:05d}.tar") as tar:
        for member in tar.getmembers():
            # filename: {sample_id}.{view}.png
            parts = member.name.rsplit(".", 2)     # [sample_id, view, "png"]
            sid, view = parts[0], parts[1]
            if sample_ids and sid not in sample_ids:
                continue
            img = Image.open(io.BytesIO(tar.extractfile(member).read()))
            results.setdefault(sid, {})[view] = img
    return results

images = load_images_from_shard(0)
# images[sample_id]["top"]   -> PIL Image
# images[sample_id]["front"] -> PIL Image
# images[sample_id]["side"]  -> PIL Image
# images[sample_id]["iso"]   -> PIL Image

Load 3D Mesh from TAR

import tarfile
import trimesh
import io

# STL only (lighter download)
with tarfile.open("geo_stl/geo-stl-00000.tar") as tar:
    for member in tar.getmembers():
        sample_id = member.name[:-4]   # remove .stl
        stl_bytes = tar.extractfile(member).read()
        mesh = trimesh.load(io.BytesIO(stl_bytes), file_type="stl")
        print(f"{sample_id}: {len(mesh.vertices)} vertices, {len(mesh.faces)} faces")
        break

# PLY only
with tarfile.open("geo_ply/geo-ply-00000.tar") as tar:
    for member in tar.getmembers():
        sample_id = member.name[:-4]   # remove .ply
        ply_bytes = tar.extractfile(member).read()
        mesh = trimesh.load(io.BytesIO(ply_bytes), file_type="ply")
        break

HuggingFace datasets Library (Parquet streaming)

from datasets import load_dataset

# Stream point cloud data (no full download required)
ds = load_dataset(
    "IAI-CAU/TopoDroneX",
    data_files="pc/pc-*.parquet",
    split="train",
    streaming=True,
)

for batch in ds.iter(batch_size=1000):
    sample_ids = batch["sample_id"]
    coords = list(zip(batch["x"], batch["y"], batch["z"]))
    stress = batch["mises_stress (Pa)"]
    break

---

Benchmarking Test

We provide two benchmark tasks, predictive modeling and generative modeling, for each modality (2D images and 3D point clouds), yielding four benchmark settings in total. Each benchmark is organized as a self-contained folder with its own README.md describing setup and usage. Train/valid/test splits are provided in metadata/train_files.csv, metadata/valid_files.csv, and metadata/test_files.csv. These splits represent the exact data partitions used in the benchmarking experiments of the paper. For benchmark code and detailed setup instructions, please refer to the GitHub repository.

Benchmark	Data modality	Model	Metric
Predictive modeling	RGB image	VGG-19, ResNet-152, ViT-B16, MLP-Mixer-B	R², MAE, MSE
	3D model	Deepsets, PointNet++, PointCNN, DGCNN
Generative modeling	RGB image	VQGAN, β-VAE, DDIM, DiT	SSIM, MS-SSIM, KID, LPIPS
	3D model	Latent-WGAN, Diffusion PC, Pointgrow, Pointflow	COV-CD/EMD, MMD-CD/EMD, 1-NND-CD/EMD, JSD

Predictive Modeling

Predictive models are trained to regress the five design parameters from each input modality. Models are divided into image-based models, which take four direction images, and point-cloud-based models, which take a point cloud as input.

Image-based Generative Modeling

Image-based generative models are evaluated on a reconstruction task using four direction images as input.

Point-cloud-based Generative Modeling

Point cloud-based generative models are evaluated on an unconditional generation task.

---

Citation

@article{Kim2026,
  title={TopoDroneX: A Multimodal Benchmark Dataset for Topology-optimized 3D Drone Design Exploration},
  author={Kim, Yeongtae and Chung, Hoonhyung and Ra, Donghyun and Lee, Sooyoung},
  note={Under review},
  year={2026}
}

---

License

This dataset is released under CC BY-NC 4.0 for academic use only.

---