FORMAT.md

# HTT dataset

Paired multi-sensor tactile data for the Heterogeneous Tactile Transformer
(HTT). Four sensors — two vision-based (GelSight Mini, 9DTact) and two taxel
arrays (Xela uSkin, TAC-02) — organized into four task splits.

| sensor | type | raw format |
|---|---|---|
| `gsmini` | vision (GelSight Mini) | JPEG `[224, 224, 3]` |
| `9dtact` | vision (9DTact) | JPEG `[224, 224, 3]` |
| `xela` | taxel array (Xela uSkin) | float `[T, 72]` |
| `tac02` | taxel array (TAC-02) | float `[T, 66]` |

```
├── pretrain/        1. self-supervised pretraining (paired episodes, no labels)
├── classification/  2. 20-object classification (paired episodes, labeled)
├── force/           3. 6D force estimation (per-sensor static probe episodes)
├── slip/            4. slip-stage detection (per-sensor sliding episodes)
├── bg_data/         background/reference frames per sensor
├── STATS.json       measured per-split statistics (machine-readable)
└── MANIFEST.sha256  checksums for every file
```

## Dataset statistics — 1.59M frames

One *paired sample* = 1 optical frame (224×224 RGB) + 1 synchronized chunk of
10 consecutive taxel frames. The npz splits are counted as force-aligned
timesteps (taxel sensors at native rate, vision sensors at camera rate).

| corpus | episodes | frames (native rate) |
|---|---:|---:|
| pretrain (paired, both pairs) | 1,789 | 710,633 (64,603 optical + 646,030 array) |
| classification (paired, both pairs) | 599 | 239,184 (21,744 optical + 217,440 array) |
| force task: static + sliding modes | 401 + 405 | 423,284 (static 209,869 + sliding 213,415) |
| slip task: sliding modes | 405 | 213,415 |
| **frame–task instances** | | **1,586,516 ≈ 1.59M** |
| distinct physical frames | | 1,373,101 |

Sliding-mode episodes carry both 6D force and slip labels, so they serve two
tasks; the frame–task total counts them twice, the distinct total once.

**Note on the force task protocol:** the sliding episodes under `slip/` also
contain time-aligned 6D force and can be used for force training
(`mode_filter=None` in the loaders reads both modes from a merged root). All
force results reported in the HTT paper use the **static mode only**
(`mode_filter: static`), i.e. exactly the `force/` split; slip results use
exactly the `slip/` split.

Per-sensor force-aligned frame counts:

| | xela | tac02 | 9dtact | gsmini |
|---|---:|---:|---:|---:|
| force (static) | 105,691 | 54,718 | 24,829 | 24,631 |
| slip (sliding) | 108,649 | 54,717 | 24,900 | 25,149 |

## 1. `pretrain/` — paired episodes, label-free

WebDataset tar shards for two sensor pairs, collected with the sensor pair
touching the same object simultaneously:

- `pretrain/xela_9dtact/pretrain_{train,val,test}_{0000..0003}.tar`
- `pretrain/tacniq_gsmini/pretrain_{train,val,test}_{0000..0003}.tar`
  (`tacniq` is the TAC-02 sensor)

Each episode contributes, with a shared basename prefix:

```
episode_N.bg.jpg              background frame (vision sensor)
episode_N_XXXXXX.<vis>.jpg    vision frame XXXXXX
episode_N_XXXXXX.<tax>.npy    taxel chunk XXXXXX  [frames_per_chunk, D]
episode_N_meta.json           structural metadata only
_manifest.json                per-shard episode list + shard info
```

**This split is fully unlabeled.** All object and action annotations have
been removed from both `episode_N_meta.json` and `_manifest.json`; the
metadata retains only structural fields (episode id, frame counts, chunking).
The removal is audited programmatically: no label key and no label vocabulary
string appears anywhere in these shards, and the pretrain episode ids are
disjoint from the classification episode ids, so object identity cannot be
recovered by cross-referencing the labeled split.

## 2. `classification/` — 20-object supervised episodes

Same episode format and pairs as `pretrain/`
(`classification/<pair>/supervised_{train,val,test}_{0000..0003}.tar`), with
`object` (20 classes) and `action` (press / slide / twist) kept in
`episode_N_meta.json` and `_manifest.json`.

## 3. `force/` — 6D force estimation

`force/<sensor>/processed/p{1..4}_static/*.npz` — static pressing episodes for
4 probe tips, 50 episodes each, with time-aligned ATI force-torque readings.

Vision npz keys: `ref_frame [224,224,3]`, `ref_force [6]`,
`tactile_img [T,224,224,3] uint8`, `6d_force [T,6]`, `probe`, `mode`.
Taxel npz keys: `ref_tactile [D]` and `tactile [T,D]` instead of images.

## 4. `slip/` — slip-stage detection

`slip/<sensor>/processed/p{1..4}_sliding/*.npz` — sliding episodes in the same
npz format as `force/` — plus
`slip/<sensor>/sliding_labeled/p{1..4}_sliding/*.labeled.npz` with per-frame
3-class labels (`sliding_labels_bracket`: 0 static / 1 incipient / 2 gross)
and friction diagnostics (`mus`, `fz`, `c_plus`, `c_minus`). A label file maps
to its episode by path convention:
`processed/pK_sliding/foo.npz` ↔ `sliding_labeled/pK_sliding/foo.labeled.npz`.

## Loading

The HTT code release ships dataloaders for every split; point them at this
directory:

- pretrain / classification: `data/xela_9dtact_dataloader_webdataset.py`,
  `data/tacniq_gsmini_dataloader_webdataset.py` (`data_root` = the pair dir,
  `dataset_split_type` = `pretrain` | `supervised`)
- force / slip: `data/taxel_force_4probe_50each_dataloader.py`,
  `data/gsmini_force_4probe_50each_dataloader.py` (`data_root` =
  `<split>/<sensor>/processed`, `mode_filter` = `static` | `sliding`)

## Integrity

```bash
sha256sum -c MANIFEST.sha256
```