README.md

---
license: other
license_name: mixed-mit-and-cc-by-4.0
license_link: LICENSE
pretty_name: PhysCoRe
size_categories:
- 10K<n<100K
task_categories:
- robotics
tags:
- deformable-objects
- robot-manipulation
- material-estimation
---

# PhysCoRe

Multi-view RGB-D recordings of deformable objects being manipulated by hand, plus the
trained model weights, 3D Gaussian splats and configs needed to reproduce the results in
**PhysCoRe: Physics-Corrected Residual World Models for Material-Aware Deformable
Dynamics** (CoRL 2026).

- Project page: https://lunarlab-gatech.github.io/PhysCoRe-website/
- Paper: https://arxiv.org/abs/2607.20653
- Code: https://github.com/lunarlab-gatech/PhysCoRe

## Contents

```
data/
├── phystwin.zip                 14 cases, redistributed (see Licenses)
└── physcore.zip                 12 cases, recorded by us
gaussian_output/physcore.zip     static 3DGS scenes for the 12 physcore cases
checkpoints/                     MfM_checkpoint.pt, RfD_checkpoint.pt
configs/                         one YAML per pipeline script
```

The recordings are shipped as archives to keep the repository to a handful of files.
Unzip each one in place; every archive expands into a directory of the same name:

```bash
cd data            && unzip phystwin.zip && unzip physcore.zip && cd ..
cd gaussian_output && unzip physcore.zip && cd ..
```

which gives `data/phystwin/different_types/<case>/`,
`data/physcore/different_types/<case>/` and `gaussian_output/physcore/<case>/`.

Each of the 26 cases is one recording from **3 calibrated RGB-D cameras**, time-aligned:

| path | content |
|---|---|
| `calibrate.pkl` | pickled list of 3 camera-to-world 4x4 matrices |
| `metadata.json` | per-camera intrinsics, image size `WH`, `frame_num`, serial numbers |
| `color/{0,1,2}/<frame>.png` | RGB frames |
| `color/{0,1,2}.mp4` | per-camera RGB video |
| `depth/{0,1,2}/<frame>.npy` | depth frames, uint16 millimeters |
| `mask/{0,1,2}/<mask_id>/<frame>.png` | object and controller segmentation |
| `mask/mask_info_{0,1,2}.json` | mask id to label mapping |
| `sampled_tracks.pkl` | sampled 3D tracks |

## Where to start

Unzip the three archives as above, then copy the contents of this repository into the
root of a PhysCoRe checkout, so that
`data/`, `gaussian_output/`, `checkpoints/` and `configs/` sit beside the pipeline
scripts. The code repository's README walks through the stages in order and names the
config each one reads.

Because `mask/` and `sampled_tracks.pkl` are included, you can skip the segmentation and
tracking stage and go straight to building episodes with
`datagen/convert3d/convert_to_episode.py`. Run that earlier stage only if you want to
reproduce it; it additionally needs the SAM 2 and GroundingDINO weights, which are not
redistributed here.

The `configs/` here carry the values used for the released results, with run directories
left as `YYYYMMDD_hhmm` placeholders to be stamped at launch. Per-case controller contact
radii in `per_sample_rollout_config` are scene-dependent and will need retuning for your
own objects.

### Confidence overlay settings, per case

`configs/render_MfM_confidence_3dgs.yaml` ships one example case, but the colormap ceiling
has to be set per case. These are the values behind the released overlay videos, measured
with `checkpoints/MfM_checkpoint.pt` on `data_episodes/physcore/<case>/episode_0000`:

| case | `norm_hi` | `video_frames` | contact radius |
|---|---|---|---|
| `double_clift_cloth` | 2.0 | 120 | 0.06 (CLI override) |
| `single_clift_cloth` | 1.75 | null | 0.044 (CLI override) |
| `single_clift_rope` | 2.5 | null | 0.04 |
| `double_squeeze_plastic` | 2.8 | null | 0.04 |
| `double_stretch_bear_1` | 15.0 | null | 0.06 |
| `single_push_rope` | 2.7 | 150 | 0.02 |

The contact radius applies to the `validate_MfM.py` run that produces `render.traj_path`,
not to the render itself. `configs/validate_MfM.yaml` already resolves to the value above
for every case except the two marked *CLI override*, which need it passed on the command
line:

```bash
python validate_MfM.py --config configs/validate_MfM.yaml \
  --root data_episodes/physcore/double_clift_cloth/episode_0000 \
  rollout.manipulation_controller_grid_contact_radius=0.06
```

## Licenses

This repository is **mixed-license**. Check the directory before reusing anything.

| path | license |
|---|---|
| `data/phystwin/` | **MIT**, Copyright (c) 2025 Hanxiao Jiang — see `data/phystwin/LICENSE` |
| `data/physcore/`, `gaussian_output/`, `checkpoints/`, `configs/` | **CC-BY-4.0**, Lunar Lab @ Georgia Tech |

The 14 cases under `data/phystwin/` are redistributed from the
[PhysTwin dataset](https://huggingface.co/datasets/Jianghanxiao/PhysTwin) under its MIT
license, which permits redistribution provided the copyright notice is retained. The
`mask/` and `sampled_tracks.pkl` files in those case directories are derived from those
recordings and carry the same terms. If you use them, please cite PhysTwin as well as
this work.

## Citation

We hope this dataset is useful for your research. If it contributes to your work, please
consider citing:

```bibtex
@inproceedings{yin2026physcore,
  title     = {PhysCoRe: Physics-Corrected Residual World Models for Material-Aware Deformable Dynamics},
  author    = {Yin, Haocheng and Tao, Shuohan and Chen, Yongsheng and Gan, Lu},
  booktitle = {Conference on Robot Learning (CoRL)},
  series    = {Proceedings of Machine Learning Research},
  publisher = {PMLR},
  year      = {2026}
}
```