pretty_name: VT-MUSE Real-World Visuotactile Manipulation Dataset
task_categories:
- robotics
language:
- en
tags:
- robotics
- tactile-sensing
- visuotactile
- manipulation
- open-x-tactile
VT-MUSE Real-World Visuotactile Manipulation Dataset
This repository hosts the real-world tactile manipulation data contributed with VT-MUSE to the Open-X-Tactile (OXT) dataset.
The release follows the FTP-1-compatible Zarr organization required by OXT. Each task is distributed as an independently downloadable ZIP archive containing one standardized Zarr dataset.
Dataset contents
| Sub-dataset | Trajectories |
|---|---|
insert_tube |
250 |
press_toaster |
100 |
pull_out_drawer |
100 |
wipe_board |
201 |
| Total | 651 |
The standardized release contains 192,225 temporally aligned frames. It includes the 200 trajectories used for the final VT-MUSE real-world experiments (50 per task) and 451 additional valid real-world trajectories. The authoritative episode-level split is recorded in the OXT submission manifest.
Repository layout
insert_tube/
insert_tube.zarr.zip
press_toaster/
press_toaster.zarr.zip
pull_out_drawer/
pull_out_drawer.zarr.zip
wipe_board/
wipe_board.zarr.zip
README.md
supplementary_tactile_note.json
Sensors and observations
- Fixed third-person RGB camera, exported as
camera_main_rgb. - Robot state/action: seven arm joints and one gripper degree of freedom.
- Two GelSight Mini tactile pads, with aligned tactile RGB observations.
- Supplemental per-pad tactile depth maps are documented in
supplementary_tactile_note.json.
Format and quality control
The archives use the OXT/FTP-1-compatible Zarr schema. Metadata and structure validation was run with the official OXT-QC toolkit. The four task-level Zarr datasets passed the metadata precheck with no reported errors or warnings.
For the data-format specification and the broader collection, visit the Open-X-Tactile website.
Institution and contact
Shanghai Jiao Tong University
Contact: acondaway@sjtu.edu.cn