README.md

# H-Tac Release Sample InternDesk

This package is a sample release of H-Tac tactile manipulation data. It uses a LeRobot v2.1-style episode layout with H-Tac metadata, and contains complete episode groups from two dataset families:

| directory | subset | episodes | frames | videos |
| --- | --- | ---: | ---: | ---: |
| `DeskTaskTac/RawHandPip` | human hand desk tasks with tactile data | 10 | 10,974 | 10 |
| `DeskTaskTac/AprilTagPip` | AprilTag-assisted desk tasks with tactile data | 4 | 11,104 | 4 |
| `InternDataTac/curated__genie1_mano_tactile` | robot MANO tactile data | 49 | 7,553 | 49 |
| `InternDataTac/curated__lift2_mano_tactile` | robot MANO tactile data | 35 | 6,351 | 35 |

Total sample size: 98 complete episodes, 35,982 frames, and 98 included top-view videos.

## Release Documents

- `README.md`: release overview, layout, shared fields, and loading notes.
- `SCHEMA.md`: cross-family schema reference for loader implementation.
- `InternDataTac/README.md`: InternDataTac-specific fields, masks, tactile normalization, and video availability.
- `DeskTaskTac/README.md`: DeskTaskTac-specific fields, pipeline differences, timing notes, and subset summaries.
- `LICENSE`: license terms for this sample package.
- `CITATION.md`: citation text for this sample package.

## Scope

Each episode is complete within the package: parquet frames, episode metadata, task records, per-episode records, and available top-view videos are aligned by release-local episode indices.

## Dataset Families

The sample has two related but not identical data families:

- `InternDataTac`: robot manipulation data with MANO-based 351-D tactile pressure fields. Its two subsets share the same per-frame schema.
- `DeskTaskTac`: human-hand desktop manipulation data with TacGlove-style 137-D tactile fields. Its two subsets share the same per-frame schema.

Both families use the same LeRobot v2.1-style directory layout. They share timing, task, pose, MANO, video, and action/state fields; tactile and hand-control fields are family-specific.

## Layout

Each subset follows this layout:

- `data/episode_*.parquet`: per-frame episode data.
- `meta/episodes.parquet`: episode-level metadata.
- `meta/tasks.jsonl`: task language and task indices.
- `meta/features.json`: machine-readable feature schema.
- `meta/dataset_info.json`: dataset-level metadata.
- `meta/episode_records/*.json`: per-episode metadata records.
- `videos/video.top_view/episode_*.mp4`: top-view videos when available.

`meta/features.json` is the machine-readable frame schema. `meta/dataset_info.json` records dataset-level counts, modality metadata, fps, tactile notes, and coordinate systems.

## Access Notes

This release does not include dataset loading code. Use a parquet-capable reader that supports list columns, and treat the metadata files as the interface contract:

- Use `meta/features.json` as the authoritative per-frame schema for each subset.
- Use `meta/dataset_info.json` for dataset-level counts, modality names, fps, tactile normalization notes, and coordinate-system definitions.
- Use `meta/tasks.jsonl` to map `task_index` to text.
- Use `meta/episodes.parquet` for episode length, task, relative parquet path, and relative video path.
- Resolve `video.top_view` relative to the subset root.
- Treat the common fields below as the shared interface, then branch by family for tactile and hand-control fields.

## Schema Compatibility

Within this package, the subset schemas have the following properties:

- All episode parquet files in each subset have column names matching the subset's `meta/features.json`.
- All vector fields have the declared lengths.
- `frame_index` is contiguous from `0` to `length - 1` inside each episode.
- `episode_done` is `true` only on the last frame of each episode.
- Frame totals in `meta/dataset_info.json` match the actual parquet row counts.

Per-frame schema compatibility:

| comparison | result |
| --- | --- |
| `InternDataTac` subsets compared with each other | same 40-column per-frame schema |
| `DeskTaskTac` subsets compared with each other | same 38-column per-frame schema |
| `InternDataTac` compared with `DeskTaskTac` | not identical; 30 common columns plus family-specific fields |

Notable family-specific differences:

| family | family-specific fields |
| --- | --- |
| `InternDataTac` | gripper fields, active-arm masks, `tactile.*_mano_tactile` and `tactile_pred.*_mano_tactile` |
| `DeskTaskTac` | dexhand fields, `tactile.*_tacglove` and `tactile_pred.*_tacglove` |

## Common Per-Frame Fields

The following fields are common to both families and 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 from `0`. |
| `timestamp` | `float32` scalar | Episode-relative timestamp in seconds. |
| `timestamp_ms` | `float64` scalar | Millisecond timestamp. In `InternDataTac` this is episode-relative; in `DeskTaskTac` this preserves the captured timeline timestamp. |
| `task_index` | `int64` scalar | Task id, resolved through `meta/tasks.jsonl`. |
| `task` | `string` scalar | Task language or task name copied onto each frame. |
| `session_name` | `string` scalar | Release-local episode/session identifier. |
| `session_dir` | `string` scalar | Release-local session path or identifier. |
| `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 final frame is `true`. |
| `state.eef_position` | `float32[3]` | Right-side end-effector or wrist position in `first_person_camera_opencv`. |
| `state.eef_rotation` | `float32[3]` | Right-side end-effector or wrist rotation as axis-angle in `first_person_camera_opencv`. |
| `state.left_eef_position` | `float32[3]` | Left-side end-effector or wrist position in `first_person_camera_opencv`. |
| `state.left_eef_rotation` | `float32[3]` | Left-side end-effector or wrist rotation as axis-angle 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 rotation as axis-angle in `world_scene`. |
| `video.top_view` | `video` / stored as string path | Relative path to the top-view MP4 for this episode when present. |
| `video.top_view_frame_index` | `int64` scalar | Frame index in the top-view video. In this sample it matches `frame_index`. |
| `action.eef_position` | `float32[3]` | Right-side action target or delta in `first_person_camera_opencv`; declared with `absolute=false`. |
| `action.eef_rotation` | `float32[3]` | Right-side action rotation as axis-angle; declared with `absolute=false`. |
| `action.left_eef_position` | `float32[3]` | Left-side action target or delta in `first_person_camera_opencv`; declared with `absolute=false`. |
| `action.left_eef_rotation` | `float32[3]` | Left-side action rotation as axis-angle; declared with `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. |

## Coordinate Systems

`first_person_camera_opencv` is the current-frame egocentric camera frame following the OpenCV convention:

- `x`: image right
- `y`: image down
- `z`: camera forward

`world_scene` is the scene/world frame used to express the first-person camera pose. `InternDataTac` defines it as the SimBox camera-to-environment scene frame. `DeskTaskTac` defines it as the scene-board world frame anchored to the reference scene AprilTag.

## Notes

`meta/episodes.parquet` contains family-specific auxiliary columns. Use the common fields required by your loader, and branch by dataset family for additional metadata.

Top-view videos are part of the released modalities. Desk-task videos show hands, forearms, tools, and tabletop objects; robot-scene videos show the robot workspace and manipulated objects.

## License And Citation

License terms are provided in `LICENSE`. Citation text is provided in `CITATION.md`.

Family-specific schema details are documented in:

- `InternDataTac/README.md`
- `DeskTaskTac/README.md`