ICoM-RWTH/DIEPF_2026

---
license: cc-by-4.0

name: DIEPF 2026 – Depth–Image + Extrusion + (Robot) Pose Fusion Dataset

language:
- en

task_categories:
- robotics
- image-classification
- time-series-forecasting

tags:
- construction-robotics
- additive-manufacturing
- rgbd
- sensor-fusion
- kuka

size_categories:
- 1B<n<10B
---

# DIEPF 2026 – Depth–Image + Extrusion + (Robot) Pose Fusion Dataset for Large-Scale Additive Manufacturing

This repository contains **multi-modal recordings** for construction-scale additive manufacturing experiments, combining **RGB-D imagery** with **robot kinematics and process signals**. The dataset is designed for **quality monitoring**, **defect detection**, and **dataset generation** for learning-based approaches in robotic extrusion processes.

**Hardware & setup**
- Robot: **KUKA KR210 R3100** with **KRC4** controller
- Camera: **Intel RealSense D405** (short-range RGB-D), **eye-in-hand** on robot flange, viewing the nozzle at a fixed angle
- Robot telemetry: streamed via **KUKA variables** (OpenShowVar / VAR proxy), including joint axes and additional process variables (e.g., extruder RPM and override)

---

## What’s inside

The dataset is organized as **one `.tar` archive per recording session**.

## Recording archives (download sizes)

The dataset is provided as **one TAR archive per recording session** (stored via **Git LFS**).

**Total size (all sessions): ~7.9 GB**

| Session | Archive | Size |
|---:|---|---:|
| 001 | `260109_Recording_001.tar` | 82.2 MB |
| 002 | `260109_Recording_002.tar` | 84.1 MB |
| 003 | `260109_Recording_003.tar` | 1.59 GB |
| 004 | `260109_Recording_004.tar` | 656 MB |
| 005 | `260109_Recording_005.tar` | 596 MB |
| 006 | `260109_Recording_006.tar` | 467 MB |
| 007 | `260109_Recording_007.tar` | 700 MB |
| 008 | `260109_Recording_008.tar` | 2.05 GB |
| 009 | `260109_Recording_009.tar` | 2.23 GB |

> Archive size scales with recording duration and camera FPS. Larger sessions typically contain longer continuous runs.

## Data format notes

### Image rotation
The RGB-D camera is mounted eye-in-hand; images may be **rotated (e.g., 180°)** before saving to match a consistent orientation. Applied rotation should be documented in per-session metadata.

## Example: download and extract one session (Python)

```python
from huggingface_hub import hf_hub_download
import tarfile
from pathlib import Path

repo_id = "ICoM-RWTH/DIEPF_2026"
filename = "recordings/260109_Recording_001.tar"

tar_path = hf_hub_download(repo_id=repo_id, repo_type="dataset", filename=filename)
out_dir = Path("extracted/260109_Recording_001")
out_dir.mkdir(parents=True, exist_ok=True)

with tarfile.open(tar_path, "r") as tar:
tar.extractall(out_dir)

print("Extracted to:", out_dir)
```

### Add this (citation)
```markdown
## Citation

```bibtex
@dataset{diepf2026,
title = {DIEPF 2026: Depth--Image + Extrusion + Robot Pose Fusion Dataset for Large-Scale Additive Manufacturing},
author = {Benz, Hendrik and Nguyen Trong, The Vinh},
year = {2026},
publisher = {Hugging Face},
url = {https://huggingface.co/datasets/ICoM-RWTH/DIEPF_2026}
}
```