Dataset Viewer

The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.

LoopGlove

LoopGlove is an egocentric bimanual table-assembly dataset from Loopeva, recorded using Moxian fabric tactile gloves, VIVE trackers, and a static head camera. It is exported in the Open-X-Tactile (OXT) Zarr format.

Version: 1.0.0
Authors: Haoyu Wei, Xiuwei Xu, Wenxuan Guo, Yirui Wang
Contact: 1091537104w@gmail.com
Dataset: https://huggingface.co/datasets/WhyPeace/LoopGlove
License: Apache License 2.0

Contents

The release contains 2,388 source recordings distributed across 12 subsets, with up to 200 source recordings per subset. Synchronization filtering splits recordings at invalid time steps, producing 7,148 continuous segments and 1,771,492 frames, or approximately 16.40 hours at 30 Hz. Segment count is not the number of independent collection trials. Approximately 0.52% of candidate time steps were removed during synchronization filtering. Some resulting segments are shorter than one second; downstream training code should select segments long enough for its temporal window.

The task instruction is:

Hold the table with your left hand, place the white legs that are not properly installed on the gray table with your right hand, align the holes, and then put the screws into the holes. Then, use the wrench in the pen holder with your right hand to tighten the screws, and finally put the wrench back in place.

The release contains 63 independent ZIP64 archives (ZIP_STORED), totalling approximately 451.23 GiB. Each part contains up to roughly 8 GiB of already-compressed Zarr payload; ZIP metadata adds a small overhead. Archives are distributed under LoopGlove/table_0000/ through LoopGlove/table_0011/, rather than as individual Zarr chunk files. After extracting all archive parts, preserve the episode_batch_XXXX.zarr/ directory and all files inside it, including hidden Zarr metadata files. Each Zarr contains a time-major data/ group and meta/episode_ends, whose entries are exclusive segment end indices.

The repository layout is:

README.md
LICENSE
metadata.json
archive_manifest.json
LoopGlove/
  table_0000/
    table_0000.part-001-of-007.zip
    ...
  ...
  table_0011/
    ...

Each ZIP is an independent archive, not a binary split of a larger ZIP. Extract all parts for a selected subset into the same destination. Paths inside each archive already include table_XXXX/episode_batch_XXXX.zarr/; do not nest the archives in one another or concatenate ZIP files.

from pathlib import Path
from zipfile import ZipFile

# Example for one downloaded subset; repeat for other subsets as needed.
for archive_path in sorted(Path("LoopGlove/table_0000").glob("*.zip")):
    with ZipFile(archive_path) as archive:
        archive.extractall("extracted")
# Reader path: extracted/table_0000/episode_batch_0000.zarr

Data fields

Field Shape Description
timestamps (T,) Seconds relative to the original recording start; strictly increasing within each segment
camera_ego_rgb (T,720,1280,3) RGB uint8, JPEG quality 95 compression
camera_ego_pose (T,6) Zero pose: the static head camera optical frame is the global frame
{left,right}_wrist_pose (T,6) Position in metres and rotation vector in radians
{left,right}_hand_joints (T,20) Wuji Hand2.2 retargeted joint angles in radians
{left,right}_hand_joints_idx (T,20) OXT canonical joint slot indices
{left,right}_hand_pose (T,21,3) Observed human MediaPipe-order keypoint XYZ in metres, used by the official viewer
sub_task_instruction (T,) English task instruction
{left,right}_tactile_data_tip (T,5,4,4) Native fingertip ADC values
{left,right}_tactile_data_mid (T,5,2,4) Native finger-pad ADC values
{left,right}_tactile_data_palm (T,1,15,16) Native palm ADC values

Each tactile group includes the corresponding area, sensor, and type fields. Sensor name is MoxianFabricGlove, type is matrix, and values are uint8 ADC readings, not calibrated forces. Finger order is thumb, index, middle, ring, little. Tip area IDs are [0,1,2,3,4], finger-pad IDs [6,7,8,9,10], and palm ID [5]. Each hand retains 360 assigned tactile cells; unassigned cells are omitted. Left and right layouts are mapped independently according to the device protocol.

The 20 joint channels use thumb cmc_flex, cmc_abd, mcp, ip, followed by mcp_flex, mcp_abd, pip, dip for each remaining finger. Their slot indices are:

[1,26,2,3, 7,6,8,9, 12,11,13,14, 17,16,18,19, 22,21,23,24]

Coordinates and synchronization

Global coordinates are static head-camera optical coordinates: X right, Y down, Z forward. Wrist translations refer to the original VIVE tracker origins. Wrist orientations include fixed calibrated local-axis adjustments to the OXT hand convention. Human keypoints use the physical tracker-to-hand offset and calibrated scale, so the human wrist keypoint and VIVE origin need not coincide. Joint angles are retargeted robot states, rather than direct measurements of human joint angles.

Data is sampled onto a 30 Hz timeline using recorded monotonic timestamps. Poses use linear translation interpolation and rotation SLERP, with no interpolation across source gaps above 80 ms. Head RGB uses nearest-frame matching with a maximum error of 30 ms. Tactile uses the latest arrived sample no older than 30 ms, with no future samples. Some original head-camera recordings have frame rates below 30 Hz, so adjacent output frames may share an RGB source frame. Depth and tactile sample histories are not exported.

Reading

Use the common/replay_buffer.py implementation from OXT-QC, together with its dependencies. imagecodecs must be installed and its codecs registered to decode JPEG chunks; the official ReplayBuffer registers them on import. The stored format is Zarr v2, while the current official reader uses the Zarr3 Python API.

# Run with the OXT-QC directory on PYTHONPATH.
from common.replay_buffer import ReplayBuffer

buffer = ReplayBuffer.create_from_path(
    "/path/to/episode_batch_0000.zarr", mode="r"
)
episode = buffer.get_episode(0)
rgb = episode["camera_ego_rgb"]
left_angles = episode["left_hand_joints"]

Validation

All 12 subsets passed the official metadata/schema precheck with zero data errors and zero data warnings. Official VLM sampling and prompts were run on all 12 subsets using gpt-6.1-sol through a Responses transport adapter: dataset-level results are 0 pass, 8 warning, and 4 fail. These automated semantic flags remain unresolved pending collection-owner review; they are not an API failure and are not presented as an overall QC pass. The official VLM selects camera_ego_rgb and left_tactile_data_mid, rather than all six tactile groups. Sparse snapshots cannot establish full task completion or precise synchronization. The underlying ADC values are not calibrated forces. All data ZIPs have passed full member CRC validation; sample Zarr metadata and JPEG frames were also read back from every subset. The official Viser browser scene has been recorded from two observation viewpoints for manual coordinate/trajectory review. No final OXT acceptance is claimed.

Downloads last month
74