ViTaMIn-OXT

ViTaMIn-OXT is an Open X-Tactile (OXT)-formatted derivative of the public ViTaMIn dataset. It follows the FTP-1 data-processing schema and contains synchronized wrist RGB, dual-pad ViTaMIn tactile images, wrist pose, gripper opening width, and task instructions.

Release 1.0.1 corrects the wrist tool-frame orientation omitted in release 1.0.0. It applies R_new = R_orig @ Rz(+90deg) to every wrist pose and demonstration start/end pose. Position XYZ and the RGB, tactile, gripper, instruction, episode-boundary, and timestamp streams remain unchanged.

Dataset summary

Sub-dataset	Task	Trajectories	Frames
Articulated Object Manipulation	Manipulate the articulated object	75	22,848
Dynamic Peg Insertion	Insert the peg into the moving target	129	33,173
Orange Placement	Pick up the orange and place it on the plate	110	26,037
Scissor Hanging	Pick up the scissors and hang them on the hook	134	81,255
Sponge Insertion	Pick up the sponge and insert it into the cup	138	78,901
Test Tube Reorientation	Pick up the test tube and reorient it	161	119,125
Total	6 tasks	747	361,339

The data are single-hand, robot-free UMI-style demonstrations with one wrist camera and two image-based tactile pads. The 1.0.1 pose export uses the original data-bearing ViTaMIn ZIP archives; source filenames represented only by Git LFS pointers are not treated as datasets.

Repository layout

ViTaMIn-OXT/
├── ViTaMIn-Dataset/
│   ├── Articulated_Object_Manipulation_dataset.zarr.zip
│   ├── Dynamic_Peg_Insertion_dataset.zarr.zip
│   ├── Orange_Placement_dataset.zarr.zip
│   ├── Scissor_Hanging_dataset.zarr.zip
│   ├── Sponge_Insertion_dataset.zarr.zip
│   └── Test_Tube_Reorientation_dataset.zarr.zip
├── .gitattributes
├── README.md
├── metadata.json
└── SHA256SUMS

All six task archives are grouped in one ViTaMIn-Dataset/ directory. Each ZIP is a Zarr v2 store whose root contains data/ and meta/. Open the archive directly; do not add another directory level around the Zarr root when repacking.

Pose convention and correction

The original ViTaMIn pose rotations use camera-aligned UMI tool axes: x points right in the wrist-camera image, y points down, and z points forward. FTP-1's gripper/UMI tool convention uses x down, y left, and z forward.

The fixed local z rotation follows the official Touch in the Wild parser, pinned to commit 89fa681d6c014cce28300946b7526db808e0b1c1 for reproducibility.

For each original tool-to-world rotation matrix, release 1.0.1 computes:

R_new = R_orig @ Rz(+90deg)

Rz(+90deg) = [[ 0, -1,  0],
             [ 1,  0,  0],
             [ 0,  0,  1]]

This is a right multiplication in the local tool frame: x_new = y_orig, y_new = -x_orig, and z_new = z_orig. The world position is unchanged. Rotation vectors are converted to rotations, composed, and converted back; the correction is not added directly to a rotation-vector component.

The correction covers all frames of these three fields in every task:

• data/right_wrist_pose
• data/right_wrist_demo_start_pose
• data/right_wrist_demo_end_pose

All three use [x, y, z, rx, ry, rz], with position in metres and rotation vectors in radians. Dtypes remain float32 for right_wrist_pose and float64 for the demonstration start/end pose fields. Zarr root and pose-array attributes record the source/target conventions, correction formula, and local z rotation. Converter version is 0.3.0.

The demonstration start/end fields are source-provided anchors that remain constant within each episode. Some source anchors differ from that episode's first/last buffered wrist pose. Release 1.0.1 preserves those source values and their relation to the buffered trajectory while applying the same tool-frame rotation; it does not redefine the anchors from buffered endpoints. The re-review audit records the original discrepancies without attributing them to an unverified acquisition or processing cause.

Main fields

Key	Shape	Dtype	Meaning
data/timestamps	[T]	float64	Synthetic per-episode seconds at 30 FPS
data/right_wrist_camera_rgb	[T, 224, 224, 3]	uint8	Wrist RGB image
data/right_wrist_pose	[T, 6]	float32	Position XYZ in metres, then FTP-1 tool rotation vector in radians
data/right_wrist_demo_start_pose	[T, 6]	float64	Demonstration start pose in the same corrected convention
data/right_wrist_demo_end_pose	[T, 6]	float64	Demonstration end pose in the same corrected convention
data/right_hand_joints	[T, 1]	float32	Parallel-gripper opening width in metres; larger means more open
data/right_hand_joints_idx	[T, 1]	int32	Constant FTP-1 hand/gripper slot ID 28
data/right_tactile_data_gripper_left_pad	[T, 1, 224, 224, 3]	uint8	Left tactile-pad image
data/right_tactile_data_gripper_right_pad	[T, 1, 224, 224, 3]	uint8	Right tactile-pad image
data/right_tactile_area_gripper_left_pad	[T, 1]	int32	Constant functional-area ID 0
data/right_tactile_area_gripper_right_pad	[T, 1]	int32	Constant functional-area ID 1
data/right_tactile_sensor_gripper_{left,right}_pad	[T]	string	Constant sensor name ViTaMIn
data/right_tactile_type_gripper_{left,right}_pad	[T]	string	Constant tactile type image
data/sub_task_instruction	[T]	string	Natural-language task instruction
meta/episode_ends	[E]	int64	Exclusive cumulative end index of each trajectory

The source archives contain no timestamp stream. Timestamps were deterministically generated as frame_index_within_episode / 30.0; they reset to zero at every trajectory boundary. They are suitable for 30 FPS playback/alignment but are not hardware capture timestamps. This correction preserves these timestamps exactly.

right_hand_joints_idx = 28 is a schema slot identifier (zero-based slot 28 in the canonical 32-D FTP-1 hand/gripper space), not a physical joint number. The corresponding scalar is copied without rescaling from source robot0_gripper_width; its observed dataset range is approximately 0 to 0.0734371 m.

Loading an archive

Use Zarr 2.x and register the JPEG XL codec before opening image arrays:

pip install "zarr<3" imagecodecs

import zarr

try:
    from imagecodecs.numcodecs import register_codecs
except ImportError:
    from imagecodecs_numcodecs import register_codecs

register_codecs(verbose=False)

archive = "ViTaMIn-Dataset/Orange_Placement_dataset.zarr.zip"
store = zarr.ZipStore(archive, mode="r")
root = zarr.open_group(store=store, mode="r")

print(root["data/right_wrist_camera_rgb"].shape)
print(root["data/right_wrist_pose"].attrs["wrist_pose_rotation_formula"])
print(root["meta/episode_ends"][:])

store.close()

Verify downloaded files from the repository root with:

sha256sum -c SHA256SUMS

Conversion and quality checks

All six corrected working Zarr datasets passed the OXT metadata/schema precheck with no failures or warnings. Every frame of all three pose fields was checked against the original source with rotation matrices, yielding 1,084,017 checked pose records. Position components and non-pose numerical/string streams were compared exactly. Source and output RGB/tactile images were decoded and compared at each trajectory's first, middle, and last frame.

All six final ZIP archives also passed CRC, unique-member, Zarr-root-layout, full pose readback, and SHA-256 checks. All 1,084,017 encoded RGB/left-pad/right-pad image chunks were compared byte-for-byte with the original source archives and are unchanged; this release does not re-encode the stored images.

The maximum wrist rotation-matrix absolute error against R_orig @ Rz(+90deg) was 1.2858e-7, consistent with the retained float32 rotation vectors. The maximum demonstration-pose rotation-matrix absolute error was 9.5757e-16 with retained float64 rotation vectors. Regression tests also verify the local-axis mapping and distinguish right multiplication from an incorrect left multiplication.

An independent audit using the explicit rotation matrix and no converter imports also checked all 1,084,017 pose records. Its maximum rotation-matrix errors were 1.2858e-7 for wrist poses and 1.0825e-15 for demonstration anchors. It confirmed preservation of within-episode relative-motion magnitudes and source-anchor relationships; it did not assert that all source anchors equal buffered endpoints.

The re-review materials document the numerical pose audit, metadata precheck, archive checks, and revised visualization checks with their respective verification scope. A schema precheck or a visualization caption alone does not establish that a tool-frame conversion is correct; the full-frame matrix audit provides the numerical evidence for this correction.

Release history

• 1.0.0: Initial OXT export. Wrist and demonstration-pose rotations retained the original camera-aligned UMI axes; the local z-axis +90deg conversion was omitted.
• 1.0.1: Re-exported all three pose fields across all six tasks using R_new = R_orig @ Rz(+90deg). Positions, image/tactile streams, gripper widths and slot IDs, instructions, episode boundaries, and synthetic 30 FPS timestamps are preserved.

License and citation

This formatted derivative follows the MIT license declared by the source dataset. Please also consult and cite the original ViTaMIn work:

@article{liu2025vitamin,
  title   = {ViTaMIn: Learning Contact-Rich Tasks Through Robot-Free Visuo-Tactile Manipulation Interface},
  author  = {Liu, Fangchen and Li, Chuanyu and Qin, Yihua and Xu, Jing and Abbeel, Pieter and Chen, Rui},
  journal = {arXiv preprint arXiv:2504.06156},
  year    = {2025}
}

Project page: https://chuanyune.github.io/ViTaMIn_page/
OXT format and contribution information: https://open-x-tactile.github.io/
Contact: chuanyu.ne79@gmail.com