hwk0809/RGBench-Cloth-Sim2Real-v1

RGBench Cloth Sim-to-Real (v1)

🌐 Project page: https://rgbench.github.io/ · 📦 Code: https://github.com/hwk0809/RGBench

Nine carefully captured garments — three bimanual manipulation actions each (fling / fold / grasp) — with real-world ground truth point clouds for evaluating any cloth simulator's sim-to-real gap. Released as the evaluation half of the AAAI 2026 paper Real Garment Benchmark (RGBench).

The larger 6 000+ garment-mesh asset library and the GarmentDynamics simulator from the paper are on their way to open-sourcing in follow-up releases. This dataset is the piece you need to benchmark any cloth simulator against real captured dynamics today.

Contents

Path	What's inside
<garment>/<garment>_<action>_<ts>/calibration/	Camera extrinsics, initial object pose
<garment>/<garment>_<action>_<ts>/joints/	Bimanual robot joint + end-effector CSVs
<garment>/<garment>_<action>_<ts>/segment_pcds/	Segmented cloth point clouds (cloth-only, world frame after extrinsics)
meshes/<Garment>/*.obj, *.usda	Cloth garment meshes used by the simulators

9 garments × {grasp, fold, fling} actions, captured with a Piper bimanual gripper and a RealSense D455. ~100 evaluation samples in total. Two garments (grey_sunwear, khaki_blazer) have non-manifold meshes and are not part of the paper's published baselines — see the results/ folder in the GitHub repo for the published baseline numbers and the methodology note.

Garment meshes ship in four resolutions for green_tshirt (5k / 10k / 20k / 40k triangles) under meshes/Green_Tshirt_Compare/ so researchers can study how cloth mesh resolution affects sim-to-real fidelity.

Quickstart

git clone https://github.com/hwk0809/RGBench
cd RGBench
bash setup.sh                       # installs deps + downloads this dataset
make benchmark sim=pybullet         # runs the smoke-test sample

If you only want to fetch the data:

pip install huggingface_hub
python -m huggingface_hub.commands.huggingface_cli download \
    hwk0809/RGBench-Cloth-Sim2Real-v1 --repo-type dataset \
    --local-dir ./data/sample

Licensing

• Data: CC-BY 4.0 — attribution required, commercial use permitted.
• Code (benchmark): MIT — see the RGBench repo.

Citation

If you use this dataset, please cite the AAAI 2026 paper:

@inproceedings{hu2026rgbench,
  title     = {Real Garment Benchmark ({RGBench}): A Comprehensive Benchmark for Robotic Garment Manipulation featuring a High-Fidelity Scalable Simulator},
  author    = {Hu, Wenkang and Tang, Xincheng and E, Yanzhi and Li, Yitong and Shu, Zhengjie and Li, Wei and Wang, Huamin and Yang, Ruigang},
  booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence},
  year      = {2026},
  url       = {https://rgbench.github.io/}
}