The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
shape: list<item: int64>
child 0, item: int64
data_type: string
chunk_grid: struct<name: string, configuration: struct<chunk_shape: list<item: int64>>>
child 0, name: string
child 1, configuration: struct<chunk_shape: list<item: int64>>
child 0, chunk_shape: list<item: int64>
child 0, item: int64
chunk_key_encoding: struct<name: string, configuration: struct<separator: string>>
child 0, name: string
child 1, configuration: struct<separator: string>
child 0, separator: string
fill_value: double
codecs: list<item: struct<name: string, configuration: struct<endian: string, level: int64, checksum: bool>> (... 1 chars omitted)
child 0, item: struct<name: string, configuration: struct<endian: string, level: int64, checksum: bool>>
child 0, name: string
child 1, configuration: struct<endian: string, level: int64, checksum: bool>
child 0, endian: string
child 1, level: int64
child 2, checksum: bool
attributes: struct<>
zarr_format: int64
node_type: string
storage_transformers: list<item: null>
child 0, item: null
to
{'attributes': {}, 'zarr_format': Value('int64'), 'node_type': Value('string')}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2951, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2461, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2486, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 547, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 430, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
shape: list<item: int64>
child 0, item: int64
data_type: string
chunk_grid: struct<name: string, configuration: struct<chunk_shape: list<item: int64>>>
child 0, name: string
child 1, configuration: struct<chunk_shape: list<item: int64>>
child 0, chunk_shape: list<item: int64>
child 0, item: int64
chunk_key_encoding: struct<name: string, configuration: struct<separator: string>>
child 0, name: string
child 1, configuration: struct<separator: string>
child 0, separator: string
fill_value: double
codecs: list<item: struct<name: string, configuration: struct<endian: string, level: int64, checksum: bool>> (... 1 chars omitted)
child 0, item: struct<name: string, configuration: struct<endian: string, level: int64, checksum: bool>>
child 0, name: string
child 1, configuration: struct<endian: string, level: int64, checksum: bool>
child 0, endian: string
child 1, level: int64
child 2, checksum: bool
attributes: struct<>
zarr_format: int64
node_type: string
storage_transformers: list<item: null>
child 0, item: null
to
{'attributes': {}, 'zarr_format': Value('int64'), 'node_type': Value('string')}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
TacVerse
TacVerse is a large-scale bimanual visuotactile manipulation dataset from Xense Robotics. Human operators perform everyday tasks with a pair of XTac-UMI-G1 handheld self-centering grippers. Every frame carries synchronized wrist-camera views and four Xense Neutron visuotactile streams (one fingertip sensor per finger, two per gripper) at 30 fps, alongside per-hand 6-DoF end-effector pose and continuous jaw aperture. Part of the data was additionally recorded with a head-mounted stereo camera and head tracking, and carries an egocentric head view and head pose; the rest has wrist views only (see Subsets). In total: 177 tasks, 72,886 trajectories, 157 M frames (≈1,451 hours).
This repository converts the full TacVerse recordings (merged per task, with the hand-tracker poses repaired) into the FTP-1 intermediate Zarr contract used by Open-X-Tactile. For the data format, key naming, pose convention and dataset discovery rules, refer to Open-X-Tactile and the FTP-1 data-processing documentation.
Files
TacVerse-OXT/
├── TacVerse-OXT/ # one zip per task
│ ├── taccap-g1-<task>.zip # -> taccap-g1-<task>.zarr (wrist-view, 155 tasks)
│ ├── xtac-umi-g1-<task>.zip # -> xtac-umi-g1-<task>.zarr (head-view, 22 tasks)
│ └── ...
├── h264_codec_pkg/ # reader support for the video arrays (install this first)
├── supplementary_tactile_note.json
└── README.md
Each *.zip holds a single Zarr store at its top level, so unpacking side by
side gives a flat directory of .zarr folders:
hf download Vertax/TacVerse-OXT --repo-type dataset --include 'TacVerse-OXT/*.zip' --local-dir TacVerse-OXT
cd TacVerse-OXT && unzip 'TacVerse-OXT/*.zip' -d data/
Before reading: the video arrays use a custom zarr codec
xense.h264that stores the original H.264 bitstream verbatim (bit-exact to the recordings, no re-encode). Install the bundled package once — no reader code changes needed. It must be installed by the reading side too; without it Zarr raisesUnknownCodecErrorfor every*_rgb/*_data_gripperarray:# pip cannot install a sub-folder of a dataset repo directly; fetch it first hf download Vertax/TacVerse-OXT --repo-type dataset --include 'h264_codec_pkg/*' --local-dir . pip install ./h264_codec_pkgAll other arrays use standard
bytes+zstd.
Subsets
All tasks share the same grippers, tactile sensors and wrist cameras. Some tasks were additionally recorded with a head-mounted stereo camera and head tracking; those carry an egocentric head view and head pose.
| Subset | Source | Streams | Tasks | Trajectories | Frames | Hours |
|---|---|---|---|---|---|---|
| Wrist-view | TacVerse recordings, merged per task (taccap-g1-*) |
2 wrist RGB + 4 tactile | 155 | 62,835 | 140,578,423 | 1,301.7 |
| Head-view | TacVerse recordings, merged per task (xtac-umi-g1-*) |
2 wrist RGB + 1 head RGB + head pose + 4 tactile | 12 | 9,977 | 16,035,912 | 148.5 |
| Head-view samples | XTac-UMI-G1 Samples (xtac-umi-g1-*) |
2 wrist RGB + 1 head RGB + head pose + 4 tactile | 10 | 74 | 108,353 | 1.0 |
| Total | 177 | 72,886 | 156,722,688 | 1,451.1 |
The two head-view groups are separate recording sessions with no overlapping trajectories (checked frame-by-frame on the gripper signal).
Tasks
taccap-g1-operate-shoe-box-collector is the same task as taccap-g1-operate-shoe-box
but recorded with a different capture stack (the portable TacCap-Collector rig instead
of the desktop recorder: different video encoder and tactile rectification). It is
kept as a separate task on purpose; the stack-specific calibration files are not part
of the FTP-1 format and are not included.
Tactile quality control: episodes whose tactile streams had accidentally been recorded as difference images (grey background, markers only) were removed, and isolated tactile frames corrupted in transmission (colour streaks / blocks, including partial-frame and fully grey frames) were replaced by interpolating the nearest good frames of the same episode (≈13 k frames, <0.01 % of tactile frames). Only the affected GOPs were re-encoded.
Trajectory cleaning (2026-10-14): leading and trailing idle segments (both grippers slower than 2 cm/s with no gripper motion for at least 3 s) were trimmed, keeping 0.5 s of margin; episodes with a mis-oriented world frame or swapped left/right poses were re-aligned with a world-frame rotation (with the measured tracker-to-TCP mounts exchanged where the trackers were swapped); and episodes that were static, unusually slow, had a tracker position freeze of 5 s or more, or could not be re-aligned were removed (≈1.5 k episodes).
Robot and sensor
- Robot: XTac-UMI-G1 handheld gripper, bimanual (left + right), human-operated (UMI-style)
- End effector: self-centering gripper
- Tactile sensor: XenseNeutron (marker-based visuotactile RGB), 2 per gripper, 4 in total
- Wrist camera: 1 per gripper, 640×480 RGB
- Head camera (
xtac-umi-g1-*only): right eye of the head-mounted stereo camera, 640×480 RGB - Gripper state: normalized continuous jaw aperture in
[0, 1], copied from the sourcegripper.pos - Gripper FAAS/UAS index:
28 - Tactile MTTS functional-area index:
0= the finger on the operator's left,1= the finger on the operator's right
Modalities
Every task Zarr contains, for each hand <side> ∈ {left, right}:
<side>_wrist_camera_rgb: wrist RGB, shape(T, 480, 640, 3)<side>_wrist_pose: TCP pose[x, y, z, rx, ry, rz](position + rotation vector), converted from the source 6D rotation representation. Axis convention (OXT UMI frame): right gripperzforward /xdown /yleft; left gripperzforward /xup /yright (on both handsypoints from the back of the hand to the palm)<side>_hand_joints: 1-D gripper jaw aperture- XenseNeutron tactile RGB, shape
(T, 2, 400, 700, 3)— two finger pads per gripper - Per-step natural-language instruction
timestamps(seconds, restarting at 0 for each trajectory) andmeta/episode_ends
Head-view tasks (xtac-umi-g1-*) additionally contain:
camera_ego_rgb: head RGB (right eye of the head-mounted stereo camera), shape(T, 480, 640, 3)camera_ego_pose: head camera pose[x, y, z, rx, ry, rz], same representation as the wrist poses; camera axeszforward /xright /ydown
The tactile channels are raw sensor images (no force field is derived in this
release). See supplementary_tactile_note.json.
FTP-1 / Open-X-Tactile format
Each .zarr (Zarr v3) follows:
task.zarr/
├── data/
│ ├── timestamps
│ ├── left_wrist_camera_rgb / right_wrist_camera_rgb
│ ├── left_wrist_pose / right_wrist_pose
│ ├── left_hand_joints / right_hand_joints
│ ├── left_hand_joints_idx / right_hand_joints_idx
│ ├── left_tactile_data_gripper / right_tactile_data_gripper
│ ├── left_tactile_area_gripper / right_tactile_area_gripper
│ ├── left_tactile_sensor_gripper / right_tactile_sensor_gripper
│ ├── left_tactile_type_gripper / right_tactile_type_gripper
│ ├── camera_ego_rgb (head-view tasks only)
│ ├── camera_ego_pose (head-view tasks only)
│ └── sub_task_instruction
└── meta/
└── episode_ends
Video arrays (*_wrist_camera_rgb, *_tactile_data_gripper, camera_ego_rgb)
use the xense.h264 codec; all other arrays use standard bytes + zstd.
For format definitions and dataset discovery, refer to Open-X-Tactile and the
FTP-1 data-processing documentation.
Source
Converted from TacVerse LeRobot v3.0 datasets: the full recordings merged per task with
hand-tracker poses repaired (wrist-view and head-view), plus the public XTac-UMI-G1 samples.
The source action field is not carried over (FTP-1 has no action channel), and
observation.tracker_confidence is dropped. For head-view tasks the left eye of the stereo head camera is also dropped (only the right eye is kept as camera_ego_rgb).
License
cc-by-sa-4.0, inherited from the source release.
Reference
- Dataset: https://huggingface.co/datasets/TacVerse/TacVerse
- Homepage: https://huggingface.co/datasets/Vertax/TacVerse-OXT
- Public subset of the source data: https://huggingface.co/datasets/TacVerse/opendata, https://huggingface.co/collections/TacVerse/xtac-umi-g1-samples-6a97cf6d157fe67e799a1bc7
- OXT project: https://open-x-tactile.github.io/
- Authors: Xense Robotics
- Contact: dylan@xenserobotics.com
- Downloads last month
- 211