---
license: apache-2.0
task_categories:
  - robotics
tags:
  - manipulation
  - tactile
  - deformable-objects
  - soft-body
  - libero
  - isaac-lab
pretty_name: SoftVTBench
---

# SoftVTBench

Visuo-tactile manipulation data for **rigid and soft/deformable** LIBERO-style
pick-and-place tasks, collected with a tactile-sensing Franka arm in Isaac Lab
(Tabero simulation stack).

Mirrored on both hubs:

- Hugging Face — [`Arthur12137/SoftVTBench`](https://huggingface.co/datasets/Arthur12137/SoftVTBench)
- ModelScope — [`Arthur12137/SoftVTBench`](https://www.modelscope.cn/datasets/Arthur12137/SoftVTBench)

## Download

```bash
pip install -U huggingface_hub
huggingface-cli download Arthur12137/SoftVTBench \
  --repo-type dataset --local-dir ./SoftVTBench_data
```

From ModelScope (faster in mainland China):

```python
from modelscope import dataset_snapshot_download
dataset_snapshot_download('Arthur12137/SoftVTBench', local_dir='./SoftVTBench_data')
```

The full release is ~2.3 GB. Fetch only what you need:

```bash
# training on the deformable-object suite
huggingface-cli download Arthur12137/SoftVTBench --repo-type dataset \
  --include 'object-soft/*' --local-dir ./SoftVTBench_data

# add closed-loop evaluation (USD scene assets)
huggingface-cli download Arthur12137/SoftVTBench --repo-type dataset \
  --include 'eval-assets/*' --local-dir ./SoftVTBench_data
```

## Folders

| Folder | Task suite | Object type | Tasks × Demos | Size | Needed for |
|---|---|---|---|---|---|
| `object-soft/` | `libero_object` + soft pastry (10 assets) | Deformable | 10 × 50 = 500 | 896M | training, eval |
| `spatial-soft/` | `libero_spatial` + soft pastry | Deformable | 10 × 50 = 500 | 498M | training, eval |
| `object-rigid/` | `libero_object` (baseline) | Rigid | 10 tasks, 421 demos (uneven) | 425M | training, eval |
| `spatial-rigid/` | `libero_spatial` (baseline) | Rigid | 10 tasks, 207 demos (uneven) | 238M | training, eval |
| `eval-assets/` | — | — | 43 USD assets | 211M | **evaluation only** |
| `soft-assets/` | — | — | 11 pastry USD + geometry primitives | 51M | asset authoring |

`*-rigid` folders are baseline LIBERO replays (no soft-body assets); `*-soft`
folders swap in deformable pastry objects and add FEM soft-body observations.

`eval-assets/` holds the USD scene library Isaac Sim needs to rebuild the scene
for closed-loop evaluation — 32 LIBERO scene objects plus 11 deformable assets,
all following the `<name>/<name>.usd` convention. Training does not need it.
See `eval-assets/README.md` for provenance.

## Layout

```text
object-soft/
  manifest.jsonl
  libero_object/libero_object_task{0..9}/
    replayed_demos/*.hdf5                  # one file per task, demos under data/demo_*
    video_datasets/*/videos/*.mp4          # agentview + eye-in-hand RGB
    video_datasets/*/tactile_outputs/*.mp4 # rendered tactile marker video
```

Inside each HDF5, `data/demo_N` contains:

- `actions` — (T, 13) float32: xyz(3) + axis-angle(3) + abs gripper(1) + left/right force(3+3)
- `obs/eef_pose`, `obs/gripper_pos`, `obs/arm_joint_pos`, `obs/gripper_marker_motion`, `obs/gripper_net_force`
- `obs/fem_deformation_max`, `obs/fem_deformation_rms`, `obs/fem_bbox_dims` — soft-body only, `*-soft` folders
- `initial_state/`, `states/` — full Isaac Lab scene state for reset/replay

Training pipelines drop the two force channels and supervise 7D actions.

## Known caveats

- `spatial-rigid` / `object-rigid` demo counts are uneven per task (raw collection
  yield, not padded to a fixed quota).
- `spatial-soft` force labels are mostly gripper-closing proxy forces rather than
  clean contact-sensor forces; task5 mixes both. See
  `spatial-soft/spatial_pastry005_data_quality_check_20260625.md`.
- `spatial-soft/manifest.jsonl` covers tasks 0–4 only; tasks 5–9 are listed in
  `manifest_task5_9_copy_20260624.jsonl`. Derive demo counts from the HDF5 files,
  not from a manifest line count.

## Provenance & licensing

Released under the Apache License 2.0.

The rigid scene objects in `eval-assets/` originate from the
[LIBERO](https://github.com/Lifelong-Robot-Learning/LIBERO) benchmark (MIT),
converted to USD by [Tabero](https://github.com/NathanWu7/Tabero) (Apache-2.0).
The deformable assets and all trajectory data are contributed by this project.

If you use this dataset, please cite LIBERO and Tabero alongside SoftVTBench.