Hugging Face's logo

Datasets:
BeingBeyond
/
H-Tac_Sample

Modalities:
Tabular
Text
Time-series
Formats:
parquet
Size:
10K - 100K
Libraries:
Datasets
Dask
Polars
Dataset card Data Studio Files
xet
 Community
1
H-Tac_Sample / DeskTaskTac /README.md
zawnpn's picture
zawnpn
Upload folder using huggingface_hub
dfdf7ab verified
preview code
|
Raw
History Blame Contribute Delete
5.85 kB

DeskTaskTac Sample

DeskTaskTac contains two complete-episode desk-task samples:

subset episodes frames videos notes
RawHandPip 10 10,974 10 raw-hand desk-task episodes
AprilTagPip 4 11,104 4 AprilTag-assisted desk-task episodes

Both subsets follow the release layout described in the top-level README and include top-view MP4 videos. Episode ids, task ids, parquet paths, and video paths are local to this release package.

Schema Consistency

RawHandPip and AprilTagPip share the same 38-column per-frame schema. Common frame fields are documented in the top-level README.md. This file lists DeskTaskTac-specific fields and metadata.

The two DeskTaskTac subsets differ in collection/reconstruction pipeline and episode-level metadata:

  • RawHandPip: raw-hand desk-task episodes. The raw-hand pipeline recovers hand state from image keypoint detection and multi-view triangulation, with tactile glove readings.
  • AprilTagPip: AprilTag-assisted desk-task episodes. The AprilTag pipeline uses scene AprilTags and hand reference structures to recover bimanual poses, with tactile and MoCap glove information.

Loading Notes

  • RawHandPip and AprilTagPip share one per-frame schema.
  • Read meta/features.json before selecting parquet columns.
  • meta/episodes.parquet contains subset-specific auxiliary columns.
  • state.eef_* is the hand wrist/anchor pose in first_person_camera_opencv.
  • Resolve video.top_view relative to the subset root.
  • timestamp is episode-relative seconds; timestamp_ms preserves the captured millisecond timeline.

DeskTaskTac-Specific Per-Frame Fields

DeskTaskTac adds dexhand and TacGlove tactile fields on top of the common schema.

field dtype / shape meaning
state.dexhand_position float32[6] Right-side dexhand/hand-position state vector.
state.left_dexhand_position float32[6] Left-side dexhand/hand-position state vector.
action.dexhand_position float32[6] Right-side dexhand/hand-position action vector.
action.left_dexhand_position float32[6] Left-side dexhand/hand-position action vector.
tactile.right_tacglove float32[137] Right-hand TacGlove tactile vector.
tactile.left_tacglove float32[137] Left-hand TacGlove tactile vector.
tactile_pred.right_tacglove float32[137] Right-hand predicted or processed TacGlove tactile vector.
tactile_pred.left_tacglove float32[137] Left-hand predicted or processed TacGlove tactile vector.

Pose And Timing Notes

  • state.eef_position and state.eef_rotation are right-hand wrist/anchor pose fields in first_person_camera_opencv, not robot EEF fields.
  • state.left_eef_position and state.left_eef_rotation are the left-hand equivalents.
  • action.eef_* and action.left_eef_* are declared with absolute=false; treat them as action targets or deltas according to the original conversion policy.
  • timestamp is episode-relative seconds on the captured first-person video timeline.
  • timestamp_ms preserves the captured millisecond timeline timestamp. It is not simply frame_index * 1000 / 30.
  • video.top_view_frame_index matches frame_index in this sample.
  • manus_frame_index is meaningful when a Manus/MoCap source was available. In RawHand episodes it is commonly -1; in AprilTag episodes it may contain source frame indices.

Tactile Representation

DeskTaskTac tactile fields are 137-D TacGlove vectors for the right and left hands. The paired tactile_pred.* fields keep the same shape and represent predicted or processed tactile values for the corresponding hand.

DeskTaskTac combines first-person camera timeline alignment, wrist pose, MANO theta/beta, tactile glove signals, top-view video, and task labels.

Episode Metadata

The per-frame parquet schema is the same for RawHandPip and AprilTagPip, but meta/episodes.parquet is not identical between the two subsets.

Common DeskTaskTac episode-level fields include:

field meaning
episode_index Episode id within the release subset.
session_name, session_dir Release-local episode/session identifiers.
length Number of frames in the episode.
task_index, task Task id and task name.
fps Nominal frame rate, 30.0.
camera_serials Camera selection string used by the subset metadata.
multiview_min_hand_tags Minimum required hand tags/views in the source pipeline.
rts_smooth_wrist_pose Whether wrist pose smoothing was enabled.
first_person_camera_serial First-person camera id, usually camera_0.
video.top_view Relative top-view video path.
parquet_path Relative path to the episode parquet file.
mano_device Device used for MANO processing, such as cpu.
duration_seconds Episode duration in seconds.
timeline_mode, timeline_camera_serial Timeline alignment mode and reference camera.
pad_frozen_data, manus_frozen_left, manus_frozen_right Flags for padded or frozen source data.

RawHandPip also includes extra raw-hand processing fields such as pipeline_mode, hand_2d_backend, interpolation_space, and trusted_3d_*. AprilTagPip has a smaller episode-level metadata table.

Task Names

meta/tasks.jsonl maps task_index to desk-task names. Examples include assemble_lego, operate_desk_lamp, fold_paper, work_with_cylinder, operate_calculator, and play_tangram.

Release Files

Each subset contains the standard release layout: data/, meta/, meta/episode_records/, and videos/video.top_view/.

Subset counts are summarized above and in the top-level README.md. Field definitions shared by both subsets are listed in SCHEMA.md.