Datasets:
Schema Reference
This document summarizes the schema of H-Tac_Release_Sample_InternDesk. The authoritative machine-readable schemas are the per-subset meta/features.json and meta/dataset_info.json files.
Compatibility
| scope | per-frame schema |
|---|---|
InternDataTac/curated__genie1_mano_tactile |
40 columns |
InternDataTac/curated__lift2_mano_tactile |
same 40 columns |
DeskTaskTac/RawHandPip |
38 columns |
DeskTaskTac/AprilTagPip |
same 38 columns |
| InternDataTac vs DeskTaskTac | 30 shared columns plus family-specific fields |
InternDataTac subsets share one per-frame schema. DeskTaskTac subsets share one per-frame schema. Tactile and hand-control fields differ by family.
Common Frame Columns
These columns appear in every data/episode_*.parquet file.
| field | dtype / shape | meaning |
|---|---|---|
episode_index |
int64 scalar |
Episode id within the release subset. |
frame_index |
int64 scalar |
Frame id inside the episode, starting at 0. |
timestamp |
float32 scalar |
Episode-relative timestamp in seconds. |
timestamp_ms |
float64 scalar |
Millisecond timestamp. InternDataTac uses episode-relative milliseconds; DeskTaskTac preserves captured timeline milliseconds. |
task_index |
int64 scalar |
Task id resolved through meta/tasks.jsonl. |
task |
string scalar |
Task instruction or task name. |
session_name |
string scalar |
Release-local episode/session id. |
session_dir |
string scalar |
Release-local session path or id. |
manus_frame_index |
int64 scalar |
Source Manus/MoCap frame index when available; -1 means unavailable. |
episode_done |
bool scalar |
End-of-episode marker; only the last frame is true. |
state.eef_position |
float32[3] |
Right-side EEF or wrist position in first_person_camera_opencv. |
state.eef_rotation |
float32[3] |
Right-side EEF or wrist axis-angle rotation in first_person_camera_opencv. |
state.left_eef_position |
float32[3] |
Left-side EEF or wrist position in first_person_camera_opencv. |
state.left_eef_rotation |
float32[3] |
Left-side EEF or wrist axis-angle rotation in first_person_camera_opencv. |
state.right_beta |
float32[10] |
Right MANO shape parameter. |
state.left_beta |
float32[10] |
Left MANO shape parameter. |
state.right_theta |
float32[45] |
Right MANO pose parameter. |
state.left_theta |
float32[45] |
Left MANO pose parameter. |
state.first_person_camera_position_world |
float32[3] |
First-person camera position in world_scene. |
state.first_person_camera_rotation_world |
float32[3] |
First-person camera axis-angle rotation in world_scene. |
video.top_view |
stored as string path | Relative path to the top-view video for this episode when present. |
video.top_view_frame_index |
int64 scalar |
Top-view video frame index. In this sample it matches frame_index. |
action.eef_position |
float32[3] |
Right-side action target or delta in first_person_camera_opencv; declared absolute=false. |
action.eef_rotation |
float32[3] |
Right-side action axis-angle rotation; declared absolute=false. |
action.left_eef_position |
float32[3] |
Left-side action target or delta in first_person_camera_opencv; declared absolute=false. |
action.left_eef_rotation |
float32[3] |
Left-side action axis-angle rotation; declared absolute=false. |
action.right_beta |
float32[10] |
Right MANO action shape parameter. |
action.left_beta |
float32[10] |
Left MANO action shape parameter. |
action.right_theta |
float32[45] |
Right MANO action pose parameter. |
action.left_theta |
float32[45] |
Left MANO action pose parameter. |
InternDataTac-Specific Frame Columns
| field | dtype / shape | meaning |
|---|---|---|
state.right_gripper |
float32[1] |
Right gripper openness/state in [0, 1]; -1 means inactive/masked. |
state.left_gripper |
float32[1] |
Left gripper openness/state in [0, 1]; -1 means inactive/masked. |
action.right_gripper |
float32[1] |
Right gripper action target. |
action.left_gripper |
float32[1] |
Left gripper action target. |
mask.right_arm |
bool scalar |
Right side inactive/masked flag. |
mask.left_arm |
bool scalar |
Left side inactive/masked flag. |
tactile.right_mano_tactile |
float32[351] |
Right MANO compact tactile pressure vector. |
tactile.left_mano_tactile |
float32[351] |
Left MANO compact tactile pressure vector. |
tactile_pred.right_mano_tactile |
float32[351] |
Right predicted or processed MANO tactile vector. |
tactile_pred.left_mano_tactile |
float32[351] |
Left predicted or processed MANO tactile vector. |
InternDataTac tactile values are normalized by a subset-specific dataset_robot_max scale recorded in meta/dataset_info.json. Nonnegative values are divided by that scale; masked inactive-arm values remain -1.
DeskTaskTac-Specific Frame Columns
| 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 TacGlove tactile vector. |
tactile.left_tacglove |
float32[137] |
Left TacGlove tactile vector. |
tactile_pred.right_tacglove |
float32[137] |
Right predicted or processed TacGlove tactile vector. |
tactile_pred.left_tacglove |
float32[137] |
Left predicted or processed TacGlove tactile vector. |
For DeskTaskTac, state.eef_* and state.left_eef_* are hand wrist or anchor poses in the first-person OpenCV camera frame, not robot EEF poses.
Coordinate Systems
first_person_camera_opencv:
x: image righty: image downz: camera forward
world_scene:
- InternDataTac: scene world frame used by SimBox camera-to-environment poses.
- DeskTaskTac: scene-board world frame anchored to the reference scene AprilTag.
Rotations are stored as axis-angle vectors where declared by meta/features.json.
Metadata Files
| path | purpose |
|---|---|
meta/features.json |
Authoritative frame-level feature schema. |
meta/dataset_info.json |
Dataset-level counts, modalities, coordinate systems, and tactile notes. |
meta/tasks.jsonl |
Task id to task text mapping. |
meta/episodes.parquet |
Episode-level index with lengths, task ids, relative parquet paths, and video paths. |
meta/episode_records/*.json |
Detailed per-episode metadata records. |
meta/episodes.parquet has family-specific auxiliary columns. InternDataTac includes robot_type, right_arm_mask, and left_arm_mask. DeskTaskTac includes camera, timeline, processing, and pipeline metadata.
Missing And Masked Values
- InternDataTac inactive sides use masks and commonly fill tactile, gripper, pose, or MANO fields with
-1. - Treat
-1on masked InternDataTac sides as missing data, not physical zero contact. manus_frame_index = -1means no source Manus/MoCap frame was available for that frame.- Resolve
video.top_viewpath strings relative to the subset root.