TacVerse/opendata

Language: English (current) · 中文

Representative frames from TacVerse's bimanual demonstrations.

Collected with XTac-UMI-G1 grippers, released as LeRobot datasets.

TacVerse Open Data

Collection of 122 LeRobot v3.0 task datasets — 17,690 episodes, 370.2 hours, 40.0M frames, ~145 GB. Each subfolder is a standalone LeRobot dataset (meta/info.json, data/, videos/). Collection timestamps have been removed from titles and metadata.

Every frame carries six synchronized video streams — two wrist cameras and four Xense Neutron visuotactile channels (two per gripper) — at 30 fps, alongside per-hand 6-DoF pose and continuous jaw aperture.

from huggingface_hub import snapshot_download
from lerobot.datasets.lerobot_dataset import LeRobotDataset

root = snapshot_download("TacVerse/opendata", repo_type="dataset")
ds = LeRobotDataset("taccap-g1-wipe-mirror", root=f"{root}/taccap-g1-wipe-mirror")

At a glance

	This collection	Planned total
Task datasets	122	350+
Scene families	3	3
Episodes	17,690	~480,000
Hours	370.2	10,000+
Frames	39,977,972	~1.1B
Frame rate	30 fps	30 fps
Visual streams	2 wrist (640×480)	—
Tactile streams	4 (700×400)	—
Robot type	bi_taccap_gripper	—
Format	LeRobot v3.0	LeRobot v3.0
Total size	~145 GB	~13 TB
Size per hour	0.39 GB	—

This collection is measured from this repository. Planned total is the TacVerse programme's target — this release is one openly published slice of it; see the landing page for programme-wide figures. Planned episodes, frames and size are all extrapolated from this repository's own density (48 episodes and 0.39 GB per hour); the planned size follows the programme's own 1.3 GB per hour, not this repository's re-encoded density.

Scenes

The collection spans three scene families.

Scene	Datasets	Episodes	Hours
Office	44	3,028	55.5
Home	70	13,124	286.9
Workbench	8	1,538	27.7
Total	122	17,690	370.2

Task families

Scene	Task family	Datasets	Episodes	Hours	Largest tasks
Office	Desk tidying	5	138	2.7	arrange-desk-items, remove-item-from-display-cover, put-item-in-display-cover
	Document handling	3	51	1.5	flip-bound-document, pull-single-sheet, organize-folders
	Stationery	2	31	1.5	paper-punching, arrange-pens-sticky-notes
	Print area	3	231	9.3	organize-printing-paper-stack, organize-printing-papers-stack, remove-and-place-paper
	Meeting room	4	65	2.6	organize-meeting-table-items, prepare-whiteboard-marker, align-remote-and-laser-pointer
	Pantry	7	194	3.9	tear-open-the-coffee-bag, scoop-coffee-granules, grasp-the-water-cup-and-twist-off-the-cup-lid
	Cable management	12	1,824	28.7	plug-the-charger-into-the-power-strip, connect-the-charger-to-the-type-c-data-cable, locate-cover-grip-clip-into-slot
	Connectors and ports	8	494	5.2	unplug-usb-cable-from-port, insert-usb-connector-into-port, put-earphone-into-case
Home	Kitchen and dining	15	1,256	39.3	place-food-with-tongs, rotate-seasoning-bottle-to-target, place-egg-into-egg-tray
	Drinks and food service	6	329	8.5	place-paper-cup, pick-up-bread-slices, assemble-and-place-cup
	Food packaging	10	925	13.1	open-resealable-bag, arrange-food-packaging-bags, pick-up-marshmallows
	Household items	13	2,778	51.9	arrange-tabletop-items, press-remote-buttons, place-chess-pieces
	Cabinets and storage	6	3,267	66.2	operate-shoe-box, snap-plastic-buckle, insert-hook-assembly
	Cleaning	4	1,131	16.8	wipe-mirror, clean-desk, mask-extraction-and-stacking
	Clothing and textiles	5	2,389	54.6	fold-garment, open-folded-towel, fold-towel
	Personal care	7	996	34.5	arrange-skincare-products, cosmetics-are-scanned-and-boxed, press-pump-bottle
	Medical and lab	4	53	2.0	liquid-medicine-bottles-cartoning, arrange-medicine-boxes, identify-and-place-reagent-bottle
Workbench	Precision assembly	3	1,061	19.2	place-washer-onto-stud, insert-divider, insert-jst-connector
	Inspection and sorting	5	477	8.6	compare-elastomer-hardness, sort-resistors, sort-screws-nuts-washers-into-bins

Sensor key map

Tactile keys map to fingertip sensors; wrist keys to gripper cameras.

This collection carries the two wrist and four tactile keys only; it has no head streams.

Provenance

These datasets were assembled from the TacVerse capture programme and re-exported for this release. They do not carry meta/sessions.json, so the per-day capture sessions behind each task are not recoverable from the files themselves.

End-effector poses were recomputed where the capture stack had not yet applied the measured tracker→TCP mount transform (added upstream on 2026-08-01). Only the 9 pose dimensions per arm were touched — jaw aperture, wrist video and tactile video are the captured values throughout.

License

• License: cc-by-sa-4.0
• Homepage: TacVerse/TacVerse

Citation

BibTeX:

@misc{tacverse2026,
  title        = {TacVerse: A Large-Scale Bimanual Visuo-Tactile Manipulation Dataset},
  author       = {Xense Robotics},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/TacVerse/TacVerse}},
  note         = {Collected with the XTac-UMI-G1 handheld gripper and Xense Neutron visuotactile sensors}
}