---
pretty_name: DeformableObjectsGrasping
---

# DeformableObjectsGrasping — OXT Conversion

This directory contains an Open-X-Tactile / FTP-1-style conversion of the
DeformableObjectsGrasping fruit interaction dataset.

For the Open-X-Tactile data format and project information, see
https://open-x-tactile.github.io/.

## Hosted file layout

The repository stores the two Zarr datasets as archives under the required
inner dataset directory:

```text
DeformableObjectsGrasping/
├── pinch.zarr.zip
└── slide.zarr.zip
```

Extract an archive before opening it with Zarr. Each archive contains one
top-level directory named `pinch.zarr` or `slide.zarr`.

## Source paper

**Learning Generalizable Vision-Tactile Robotic Grasping Strategy for
Deformable Objects via Transformer**

- Paper: https://arxiv.org/abs/2112.06374
- Authors: Yunhai Han*, Kelin Yu*, Rahul Batra, Nathan Boyd, Chaitanya Mehta,
  Tuo Zhao, Yu She, Seth Hutchinson, Ye Zhao

This release corresponds to the exploratory-action formulation in Section
III-A and the fruit-grasping experiments in Section IV-B (Figs. 4–5) of the
source paper. It contains the locally available five-fruit subset described
below.

## Contents

- `pinch.zarr`: 633 pinch episodes, 30 synchronized frames per episode.
- `slide.zarr`: 633 slide episodes, 30 synchronized frames per episode.
- Each frame contains a RealSense RGB image and one GelSight tactile image.
- The dataset covers apple, lemon, orange, plum, and tomato interactions.

## Custom gripper and joint-state scope

The source platform uses a custom gripper. No synchronized gripper aperture,
joint-state, wrist-pose, or action trajectory is available for the exported
image frames. Therefore:

- no `*_hand_joints` field is exported;
- no `*_hand_joints_idx` field applies;
- FAAS gripper slot 28 is not assigned because there is no exported joint
  channel;
- no missing gripper state or action is synthesized.

`hand_joints_idx.png` records this intentional N/A state for manual review.
The dataset should be treated as synchronized vision-tactile observation data,
not as a complete action-conditioned policy trajectory dataset.

## Tactile topology

This OXT release exports one GelSight tactile stream from a sensor pad on one
gripper jaw. No second tactile channel is included. The sole exported channel
is represented as:

- `right_tactile_data_gripper`: `(T, 1, 150, 200, 3)`;
- `right_tactile_area_gripper`: `(T, 1)`, constant functional-area index `0`;
- `right_tactile_sensor_gripper`: constant value `GelSight`;
- `right_tactile_type_gripper`: constant value `image`.

The physical pad location and index are illustrated in
`tactile_func_area_idx.png`.

## Intentionally omitted source fields

The source `Position.npy` and `Force.npy` streams are not aligned one-to-one
with the 30 synchronized RGB/GelSight frames and are not exported. The original
per-trial outcome labels and thresholds are also outside the current FTP-1
observation schema and are not exported. These omissions are documented rather
than replaced with inferred values.

## Quality-control artifacts

`oxt_check_output/` contains the metadata/schema precheck and VLM review
artifacts. The metadata precheck retains a non-fatal warning for absent
`hand_joints`; this is expected given the documented source-data scope above.