README.md

---
license: mit
tags:
- robotics
- tactile-sensing
- force-torque
- multimodal
pretty_name: TaF-Dataset
viewer: false
task_categories:
- robotics
- feature-extraction
- other
---


# TaF-Dataset

TaF-Dataset is a large-scale, synchronized tactile-force dataset introduced in
**TaF-VLA: Tactile-Force Alignment in Vision-Language-Action Models for
Force-aware Manipulation**. It is designed for learning force-aware tactile
representations and studying contact-rich robotic manipulation.

During data collection, an **ATI six-axis force/torque sensor** records the
interaction wrench (`Fx`, `Fy`, `Fz`, `Tx`, `Ty`, and `Tz`). A piezoelectric
sensing array provides a spatial pressure map that identifies the contact
region, while a vision-based tactile sensor records the corresponding tactile
image. These signals are temporally synchronized, enabling models to associate
local tactile deformation and contact location with the physical interaction
force.

## Dataset overview

- **10,053,265** synchronized frames
- **3,592** episodes
- **408** object/sensor sequences
- Tactile images, `12 x 12` pressure maps, and six-axis force/torque signals

The dataset covers six tactile-sensor configurations:

| Configuration | Parent directory | Sequences |
| --- | --- | ---: |
| GelSight Mini with markers | `taf_dataset/gs_mini/` | 124 |
| GelSight Mini without markers | `taf_dataset/gs_mini/` | 84 |
| Custom-designed sensor without markers | `taf_dataset/custom_designed/` | 84 |
| Custom-designed sensor with `6 x 6` markers | `taf_dataset/custom_designed/` | 74 |
| Custom-designed sensor with `7 x 7` markers | `taf_dataset/custom_designed/` | 10 |
| Custom-designed sensor with `8 x 8` markers | `taf_dataset/custom_designed/` | 32 |

## Data modalities

Each frame contains the following synchronized observations:

| Field | Shape | Description |
| --- | --- | --- |
| `observation.image` | `240 x 320 x 3` | RGB tactile image |
| `observation.pressure_matrix` | `12 x 12` | Piezoelectric pressure map indicating the contact region and spatial pressure distribution |
| `observation.force_torque` | `6` | ATI force/torque measurement ordered as `Fx, Fy, Fz, Tx, Ty, Tz` |
| `timestamp` | `1` | Frame timestamp in seconds |
| `frame_index` | `1` | Frame index within the episode |
| `episode_index` | `1` | Episode identifier |

The normal-force channel `Fz` provides the clearest indication of loading along
the primary contact direction. The complete six-dimensional wrench is retained
to capture tangential forces and rotational interaction dynamics.

## Data organization

The dataset follows the LeRobot v3 layout. The repository is organized as
follows:

```text
TaF-Dataset/
├── README.md
└── taf_dataset/
    ├── gs_mini/
    │   ├── gs_mini_obj1/
    │   ├── ...
    │   └── gs_mini_wo_marker_obj84/
    └── custom_designed/
        ├── custom_designed_no_mark_obj1/
        ├── custom_designed_6*6_mark_obj1/
        ├── custom_designed_7*7_mark_obj1/
        ├── custom_designed_8*8_mark_obj1/
        └── ...
```

The `gs_mini/` and `custom_designed/` directories contain 208 and 200 recorded
sequences, respectively. Every sequence uses the same LeRobot layout:

```text
<sensor_configuration>_obj<ID>/
├── data/chunk-000/file-000.parquet
├── videos/observation.image/chunk-000/file-000.mp4
└── meta/
    ├── episodes/chunk-000/file-000.parquet
    ├── collection_config.json
    ├── info.json
    ├── stats.json
    └── tasks.parquet
```

All sequence-directory names use a numeric ID within their corresponding
tactile-sensor configuration:

| Relative path pattern | ID range |
| --- | ---: |
| `taf_dataset/gs_mini/gs_mini_obj<ID>` | 1–124 |
| `taf_dataset/gs_mini/gs_mini_wo_marker_obj<ID>` | 1–84 |
| `taf_dataset/custom_designed/custom_designed_no_mark_obj<ID>` | 1–84 |
| `taf_dataset/custom_designed/custom_designed_6*6_mark_obj<ID>` | 1–74 |
| `taf_dataset/custom_designed/custom_designed_7*7_mark_obj<ID>` | 1–10 |
| `taf_dataset/custom_designed/custom_designed_8*8_mark_obj<ID>` | 1–32 |

The numeric IDs are local to each sensor configuration: directories with the
same ID but different prefixes represent separate recorded sequences and
should not be assumed to contain the same physical object.


## Citation

If this dataset is useful for your research, please cite:

```bibtex
@article{huang2026tafvla,
  title   = {TaF-VLA: Tactile-Force Alignment in Vision-Language-Action Models for Force-aware Manipulation},
  author  = {Huang, Yuzhe and Lin, Pei and Li, Wanlin and Li, Daohan and Li, Jiajun and Jiang, Jiaming and Xiao, Chenxi and Jiao, Ziyuan},
  journal = {arXiv preprint arXiv:2601.20321},
  year    = {2026}
}
```