---
pretty_name: VisuoTactile-GS
tags:
  - robotics
  - tactile
  - open-x-tactile
  - ftp-1
  - densetact
task_categories:
  - robotics
---

# VisuoTactile-GS

Visuo-tactile contact data with DenseTact optical tactile images paired to RGB
views, packaged in the [Open-X-Tactile](https://open-x-tactile.github.io/) /
FTP-1 zarr format.

The data format, key naming, pose convention and field semantics are defined by
Open-X-Tactile. Please refer to the
[OXT website](https://open-x-tactile.github.io/) and the
[FTP-1 data processing documentation](https://github.com/michaelyuancb/ftp1-policy/blob/main/data_processing/README.md)
for how to read and interpret these files.

## Contents

| Sub-dataset | Frames | Scene |
| --- | --- | --- |
| `block.zarr` | 166 | textured block |
| `bunny_real.zarr` | 134 | bunny figurine |
| `mirror.zarr` | 113 | mirror (specular) |

Total: 3 tasks, 3 trajectories, 413 frames. Images are 224x224. The tactile
stream is a single DenseTact pad (`right_tactile_area_gripper = 0`, type
`image`); the hand stream is a single gripper aperture channel
(`right_hand_joints_idx = 28`).

## Loading

```python
import zarr

root = zarr.open_group("bunny_real.zarr", mode="r")
data = root["data"]
print(list(data.keys()))
print(data["right_tactile_data_gripper"].shape)  # (T, 1, 224, 224, 3)
```

## A note on how the pairs were formed

The source data captured RGB views and tactile contacts as independent sweeps,
so there is no shared timestamp between the two streams. Each tactile
contact is paired with an RGB view by **pose**: the view is chosen so that the
contact point projects inside the image, is in front of the camera, and the
camera's optical axis is well aligned with the finger's approach direction. The
RGB frame is then cropped to a square window centred on the projected contact.

Consequently a frame index is a contact, not a timestep, and consecutive frames
are not a continuous trajectory. Treat this as a paired visuo-tactile contact
set rather than a manipulation demonstration.