tac-roboprompt

This repository is part of the Open-X-Tactile Dataset. For detailed information about the project and dataset contributions, please visit the Open-X-Tactile (OXT) official website.

Overview

tac-roboprompt is a tabletop bimanual manipulation dataset collected through teleoperation on an i2rt dual-arm platform.

Each of the two grippers is equipped with one Xsense Photon Tactile Sensor to capture tactile images. The visual setup consists of three cameras: one wrist camera on each arm and one fixed camera providing a first-person view.

Directory Structure

The dataset is organized by task. Each task folder contains a ZIP archive that extracts to a Zarr dataset with the corresponding task name.

Tac-Roboprompt/
├── README.md
└── tac-roboprompt/
    ├── bottle/
    │   └── bottle.zip
    ├── lego/
    │   └── lego.zip
    └── plug/
        └── plug.zip

Dataset Statistics

Task	Episodes	Frames
bottle	50	44,742
lego	50	106,044
plug	21	35,932
Total	121	186,718

Data Format

The data is stored in Zarr format and includes camera images, left and right gripper tactile images, robot joint states, timestamps, and episode boundaries.

• Left and right tactile images are stored in data/left_tactile_data_gripper and data/right_tactile_data_gripper, respectively, with shape (T, 1, 700, 400, 3) and dtype uint8.
• The three camera streams are stored in data/camera_ego_rgb, data/left_wrist_camera_rgb, and data/right_wrist_camera_rgb. The camera_ego_rgb stream corresponds to the fixed camera providing a first-person view.
• meta/episode_ends stores the cumulative, exclusive end indices of the episodes.
• The sensor identifier stored in the existing data is XenseOG. The hardware name used in this README, Xsense Photon Tactile Sensor, follows the information provided by the data collector.