CoolFace
Datasetpublic

Mickey27tsing/mocap-infra-0828-trial22223

XINGYING RGB-D Tactile Cylinder Trial 0828 This public dataset contains a recovered 201-second multimodal robotics capture recorded with XINGYING/NOKOV motion capture, an Intel RealSense D435i, and one left DM-Tac tactile sensor. The original XINGYING CAP and companion directory are retained beside inspectable Parquet, Zarr, MP4, CSV, and HDF5 derivatives. Important quality status This is a useful recovered capture, not a clean benchmark episode.… See the full description on the dataset page: https://huggingface.co/datasets/Mickey27tsing/mocap-infra-0828-trial22223.

sourceHugging Faceupdated 1mo agoView on Hugging Face
0likes498downloads
Dataset Card

XINGYING RGB-D Tactile Cylinder Trial 0828

This public dataset contains a recovered 201-second multimodal robotics capture recorded with XINGYING/NOKOV motion capture, an Intel RealSense D435i, and one left DM-Tac tactile sensor. The original XINGYING CAP and companion directory are retained beside inspectable Parquet, Zarr, MP4, CSV, and HDF5 derivatives.

Important quality status

This is a useful recovered capture, not a clean benchmark episode.

  • —export_complete=true: the published bundle was exported successfully and all files declared by the manifest are present.
  • —source_complete=false: the original recorder stopped when an older sequence gate treated a reused depth frame (6245 -> 6245) as fatal.
  • —The 6,031 stored RGB-D pairs and 12,072 NOKOV frames preceding that event are readable and have continuous stored frame numbers.
  • —The target rigid body cylinder has a finite pose in 6,316 / 12,072 frames (52.32%). There are 10 invalid runs; the longest is about 32.30 seconds.
  • —Only the left tactile sensor (M26190152) is present. There is no right tactile stream in this trial.
  • —Tactile force mode is deferred_completed, but finger geometry status is unknown. Wrench values are readable relative outputs, not certified absolute force ground truth.

Do not filter these limitations out when reporting results derived from this dataset.

Contents

ModalityCount / status
NOKOV frames12,072
RGB-D pairs6,031
Left DM-Tac frames6,038
NOKOV duration201.166 s
RGB-D to NOKOV absolute time deltamedian 4.287 ms, max 9.267 ms
Left tactile to NOKOV absolute deltamedian 1.388 ms, max 50.233 ms
Left tactile to RGB-D absolute deltamedian 8.293 ms, max 57.933 ms

The larger tactile maxima are confined to the short tactile tail retained by the recovery export.

Data layout

text
.
├── session.json
├── calibration.json
├── color.mp4
├── preview.mp4
├── depth.zarr/depth_raw
├── tactile/left_M26190152.h5
├── tables/
│   ├── mocap_frames.parquet
│   ├── markers.parquet
│   ├── rigid_bodies.parquet
│   ├── rigid_kinematics.parquet
│   ├── sync.parquet
│   ├── tactile_sync.parquet
│   └── six_points.csv
└── xingying/
    ├── 0828_trial222231.cap
    └── 0828_trial222231/

tables/sync.parquet and tables/tactile_sync.parquet are the authoritative time-pairing tables. MP4 timing is for playback only. Cross-device association uses nearest host monotonic timestamps and is software synchronization, not hardware triggering or temporal resampling.

Positions are stored in metres. Rigid-body quaternions use xyzw order. Depth is raw uint16; convert it with depth_scale_m from session.json.

Publication notes

Local filesystem paths and NOKOV network addresses were redacted from the published session.json. Hardware serials were retained because they identify the calibration and tactile source used for this trial. The color and preview videos may reveal the laboratory scene.

No reuse license is granted by this initial publication. Contact the repository owner before redistribution or incorporation into another released dataset.