---
license: apache-2.0
task_categories:
  - robotics
tags:
  - manipulation
  - tactile
  - visuo-tactile
  - deformable-objects
  - soft-body
  - libero
  - isaac-lab
pretty_name: SoftVTBench
size_categories:
  - 1K<n<10K
---

# SoftVTBench

Visuo-tactile manipulation demonstrations 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)

Earlier releases (evaluation USD assets and the first partial data drops) were
moved to `Arthur12137/SoftVTBench-archive` on both hubs.

## Contents

Four subsets, each **10 tasks × 100 successful demonstrations = 1000 demos**
and **4000 videos**.

| Subset | Scene / goal | Object | Video res. | Episode len | hdf5 | Videos | Total |
|---|---|---|---|---|---|---|---|
| `spatial-rigid` | kitchen table → plate | rigid pastry | 512×512 | 141–181 | 1250 MB | 560 MB | 1.8 GB |
| `object-rigid` | floor → basket | rigid pastry | 512×512 | 107–130 | 475 MB | 482 MB | 975 MB |
| `object-soft` | floor → basket | FEM soft body | 1024×1024 | 112–137 | 919 MB | 1074 MB | 2.0 GB |
| `spatial-soft` | kitchen table → plate | FEM soft body | 1024×1024 | 120–144 | 1390 MB | 1138 MB | 2.5 GB |

Every episode carries `success = True`; failures were filtered out at assembly
time (the skipped records are kept in `failure.jsonl`).

## Layout

```
<subset>/
├── manifest.jsonl              # one line per demo: task_id, demo_id, language, num_samples,
│                               # success, paths to the hdf5 and the 4 videos
├── assemble_summary.json       # per-task counts produced by the assembler
├── duplicate.jsonl             # empty in all four subsets
├── failure.jsonl               # demos dropped during assembly (soft subsets only)
└── libero_{spatial,object}/
    └── libero_{spatial,object}_task{0..9}/
        ├── replayed_demos/<task>_<language>_replayed_demo.hdf5     # 100 demos
        └── video_datasets/<task>/
            ├── videos/demo_<i>_agentview_rgb.mp4
            ├── videos/demo_<i>_eye_in_hand_rgb.mp4
            └── tactile_outputs/demo_<i>_gsmini_{left,right}_markers_rgb.mp4
```

## HDF5 schema

```
data/demo_<i>/
├── actions            (T, 13) float32     # full action record
├── actions_binary     (T, 7)  float32     # 6-DoF delta + binary gripper
├── obs/                                   # policy-facing observations
├── soft_extras/                           # raw / extra channels used for analysis
├── states/                                # per-step full simulator state
└── initial_state/                         # reset state (unique per demo)
```

`obs/` fields, all four subsets:

`actions`, `applied_torque`, `arm_joint_pos`, `computed_torque`,
`eef_axis_angle`, `eef_pose`, `fem_bbox_dims`, `finger_force`,
`gripper_binary`, `gripper_close_norm`, `gripper_marker_motion`,
`gripper_net_force`, `gripper_pos`, `gripper_width`

The two soft subsets add four deformation channels:
`fem_deformation_max`, `fem_deformation_rms`,
`fem_kabsch_max_pct`, `fem_kabsch_rms_pct`.

`gripper_marker_motion` is the tactile signal: `(T, 2, 2, 99, 2)` — two GelSight
Mini pads × (reference, current) × 99 markers × (x, y) displacement.

Per-demo attributes include `asset_name`, `demo_id`, `language`, `num_samples`,
`success`, `soft_task_kind`, `task_id`, `task_suite`, and a `metadata_json`
blob with the sampled gripper closure.

## Verified integrity

Checked over the full release, not sampled unless noted:

- demo count = manifest lines = `success=True` count = 1000 per subset
- 4 videos per demo, no zero-byte or truncated files
- every path referenced by `manifest.jsonl` exists
- video frame count equals hdf5 episode length (360 pairs sampled)
- no NaN/Inf in any float field; all `obs/` fields aligned with `actions`
- 1000/1000 unique action trajectories and 1000/1000 unique `initial_state`
  per subset — the per-demo randomisation (object XY, goal XY, robot joints,
  gripper closing force) is genuinely distinct

## Known caveats

- **Video resolution differs**: rigid subsets are 512×512, soft subsets 1024×1024.
- **`demo_id` conventions differ** across subsets (9-digit global ids in
  `spatial-rigid`/`object-rigid`, 5–6-digit in `object-soft`, 0–350 in
  `spatial-soft`). In `spatial-soft` the id is only unique *within* a task.
  Index by `(subset, task_id, demo)` when merging subsets.
- **HDF5 compression is not uniform**: the two `object-*` subsets were written
  with gzip, the two `spatial-*` subsets uncompressed. This is why
  `object-rigid` is much smaller on disk despite holding the same amount of data.
- **`gripper_marker_motion` is stored twice**, identically, under both `obs/`
  and `soft_extras/`; it accounts for ~40% of each file.
- **Task 6 language is a placeholder** (`"golden pastry"`) in `object-rigid`
  and `object-soft`.
- Paths inside `manifest.jsonl` are relative to the original collection root
  (`outputs/...`), not to this repository.

## Download

```bash
pip install -U huggingface_hub
hf download Arthur12137/SoftVTBench --repo-type dataset --local-dir ./SoftVTBench
```

```bash
pip install -U modelscope
modelscope download --dataset Arthur12137/SoftVTBench --local_dir ./SoftVTBench
```

## License

Apache-2.0.