---
license: cc-by-nc-3.0
task_categories:
  - object-detection
tags:
  - synthetic-data
  - sim-to-real
  - domain-randomization
  - industrial
  - robotics
---

# SynthRender_Robotics: Synthetic Training Sets for the Robotics Sim-to-Real Benchmark

This repository hosts the synthetic training sets generated with **SynthRender** and used to benchmark sim-to-real transfer on the public **Robotics** dataset (Horváth et al., 2022), as reported in *"SynthRender and IRIS: Open-Source Framework and Dataset for Bidirectional Sim-Real Transfer in Industrial Object Perception"* ([arXiv:2602.21141](https://arxiv.org/abs/2602.21141)).

## Sample Images (720x720)

| | | | |
|---|---|---|---|
| ![](720x720/images/train/10.png) | ![](720x720/images/train/1310.png) | ![](720x720/images/train/2000.png) | ![](720x720/images/train/3000.png) |

## Dataset Summary

SynthRender is an open-source, scriptable Domain Randomization (DR) engine built on BlenderProc, used in the paper to systematically ablate rendering design choices (physics-based placement, exponential light sampling, RGB lighting, material randomization) and quantify their effect on sim-to-real transfer. The synthetic sets in this repository correspond to the configurations used to train detectors evaluated against the **Robotics** benchmark (10 classes, 190 real test images, 920 annotated instances), on which the proposed framework reached **99.1% mAP@50**.

The repository contains two synthetic datasets, generated at different rendering resolutions:

| Resolution | Total Images | Train (80%) | Validation (20%) |
|------------|-------------|-------------|-------------------|
| 1024x1024  | 4,000       | 3,200       | 800               |
| 720x720    | 4,000       | 3,200       | 800               |

- **Annotation format:** YOLO-style bounding box labels (`class x_center y_center width height`, normalized)
- **Modality:** RGB images with text-based annotation files
- **Classes:** 10, matching the target Robotics benchmark

## Relation to the Paper

The paper's contribution statement notes that multiple synthetic training sets of 4,000 domain-randomized images each were generated to support the ablation studies. This repository provides two such sets, at 1024x1024 and 720x720 resolution, used for the Robotics benchmark specifically. It is a companion resource to the [IRIS dataset](https://huggingface.co/datasets/moiaraya/IRIS) and to the [SynthRender framework](https://github.com/Moiso/SynthRender) itself.

## Intended Use

For research on synthetic-to-real domain transfer, domain randomization ablations, and object detection benchmarking against the public Robotics dataset. Not intended for commercial use.

## License

Released under **CC BY-NC 3.0**. Commercial use is not permitted; attribution is required.

## Citation

```bibtex
@article{araya2026synthrender,
  title={SynthRender and IRIS: Open-Source Framework and Dataset for Bidirectional Sim-Real Transfer in Industrial Object Perception},
  author={Araya-Martinez, Jose Moises and Tom, Thushar and Sanchis Reig, Adri{\'a}n and Rey Valiente, Pablo and Lambrecht, Jens and Kr{\"u}ger, J{\"o}rg},
  journal={arXiv preprint arXiv:2602.21141},
  year={2026}
}
```

J. M. Araya-Martinez, T. Tom, A. S. Reig, P. R. Valiente, J. Lambrecht, and J. Krüger, 
“Synthrender and iris: Open-source framework and dataset for bidirectional sim-real transfer in industrial object perception,”
2026. [Online]. Available: https://arxiv.org/abs/2602.21141

## Contact

For questions about the dataset, please open a discussion on this repository.