# Sources

Per-source details for [`yxma/tactile-video-pretrain`](https://huggingface.co/datasets/yxma/tactile-video-pretrain).

For the full roadmap (priority phasing, license gates), see **[SHORTLIST.md](SHORTLIST.md)**.

---

## Included

### FreeTacMan ✅

- **Upstream:** [OpenDriveLab/FreeTacMan](https://huggingface.co/datasets/OpenDriveLab/FreeTacMan) on HuggingFace.
- **Paper:** *FreeTacMan: Robot-free Visuo-Tactile Data Collection System for Contact-rich Manipulation*, Wu et al., ICRA 2026 ([arXiv:2506.01941](https://arxiv.org/abs/2506.01941)).
- **Project page:** https://opendrivelab.com/freetacman
- **Sensor:** LED-based visuo-tactile sensor with multi-color illumination + dot-grid markers. Functionally equivalent to GelSight Mini / DIGIT family (McTac-derived).
- **Hardware guide:** [Google Doc](https://docs.google.com/document/d/1Hhi2stn_goXUHdYi7461w10AJbzQDC0fdYaSxMdMVXM/edit) (the rig is robot-free — humans hold the sensor by hand).
- **What we pulled:**
  - 47 task folders (one is empty after CSV-validity filtering ⇒ 46 with rows).
  - 6,338 demonstrations, 11,967 tactile-only rows, 11,965 tactile_rgb rows.
  - 29.71 h tactile, 15.70 h wrist scene RGB.
- **Stream-to-modality mapping:**
  - 3-cam tasks: `camera1 → tactile_left`, `camera2 → tactile_right`, `camera3 → scene`.
  - 2-cam tasks (`FragileCup`, `Stamp`, `TextureClassify`, `Write`): `Camera1 → tactile_left`, `Camera2 → scene`.
- **Verification:** mid-frames from `camera1` show the classic GelSight signature — multi-color rainbow gradient + dot-grid markers — confirming this is a true vision-tactile sensor, not a wrist RGB camera.
- **License:** MIT.

---

## Shortlist (planned acquisitions)

See [SHORTLIST.md](SHORTLIST.md) for the master table and phase plan.

### 🟡 Phase 1 — next push

| # | Dataset | Sensor | Modality | License | Why now |
|---|---|---|---|---|---|
| 1 | **VisGel** | GelSight (markered) | tactile + RGB | unspec — needs audit before push | Biggest old GelSight corpus (3 M frames); best-documented cross-modal training set |
| 2 | **Touch and Go** | GelSight + ego RGB | tactile + RGB | CC-BY-4.0 | Best in-the-wild diversity; material labels + touch onsets |
| 3 | **GelSLAM Tracking** | GelSight Mini (markerless) | tactile-only | MIT | ~73k frames with 6-DoF + contact mask + gradient maps for geom-aware SSL |
| 4 | **GelSLAM Reconstruction** | GelSight Mini (markerless) | tactile-only | MIT | Long-horizon (1–30 min) scans for surface-coverage consistency |

### 🟢 Phase 2 — geometry-supervised adjuncts

| # | Dataset | Sensor | Modality | License | Why |
|---|---|---|---|---|---|
| 5 | **TactileTracking** (NormalFlow) | GelSight Mini (markerless) | tactile + aux webcam | MIT | Dense per-frame gradient + mask supervision |
| 6 | **YCB-Slide (real)** | DIGIT | tactile + RGB | MIT | Sliding dynamics + GT poses, well-documented |
| 7 | **YCB-Slide (sim)** | DIGIT (Taxim) | tactile (sim) | MIT | Heightmaps + contact masks for sim2real |
| 8 | **YCB-Sight** | GelSight + Azure Kinect | tactile + RGB-D | data CC-BY-SA-4.0; code MIT | Object-centric paired set. ⚠ **SA propagation** — may need `-sa` sibling repo. |

### 🔵 Phase 3 — auxiliary

| # | Dataset | Sensor | Modality | License | Why |
|---|---|---|---|---|---|
| 9 | **Tactile MNIST (real sequences)** | GelSight Mini | tactile-only | CC-BY-2.0 | Controlled benchmark for invariances / active touch curricula |

### 🟠 Phase 4 — non-commercial sibling

Will populate a separate `yxma/tactile-video-pretrain-nc` repo (mirrors the `yxma/gelsight-mini-pretrain-nc` pattern):

| # | Dataset | Sensor | Modality | License | Why isolated |
|---|---|---|---|---|---|
| 10 | **Touch-Slide** | DIGIT (markerless) | tactile-only | **CC-BY-NC-4.0** | Largest unlabeled DIGIT corpus (180k frames) but NC license cannot mix with the MIT main repo. |

---

## Investigated but not included

### AgiBot World (Alpha, Beta, 2026, Challenge-2025/202509/2026)

- **Upstream:** [`agibot-world/*`](https://huggingface.co/agibot-world) on HuggingFace (Alpha + Beta are gated; the rest are open).
- **Marketing claim:** Alpha / Beta READMEs advertise *"Cutting-edge hardware: visual tactile sensors / 6-DoF dexterous hand / mobile dual-arm robots"*.
- **Spreadsheet flag:** A 7-task subset of Beta (`666 Close the pen cap`, `675 Insert the straw`, `676 Unplug the charger`, `677 Insert the plug`, `694 Twist the bottle cap`, `737 Peel the skin`, `774 Insert the key and open the door`) is labeled `grippers with tactile sensor` in the [public task spreadsheet](https://docs.google.com/spreadsheets/d/1GWMFHYo3UJADS7kkScoJ5ObbQfAFasPuaeC7TJUr1Cc/).
- **Why excluded:** the released data does not actually contain a tactile video stream. Verified through:
  1. **Full file-tree probe of all 7 tactile-flagged tasks:** every observation tarball contains exactly the same 8 streams — `back_left/right_fisheye_color`, `hand_left/right_color`, `head_center/left/right_fisheye_color`, `head_color` — plus `depth/`. No tactile-named camera, no tactile sub-folder.
  2. **Visual inspection of every stream's mid-frame:** `hand_left_color` and `hand_right_color` are wrist-mounted RGB cameras looking *outward* (with the gripper visible from the outside, not from inside the finger), 640×480 — not GelSight-style multi-color illumination of a gel pad. `head_*` and `back_*` are scene-view fisheye cameras.
  3. **`parameters/866961-940165.tar` (special tar covering the tactile-task episode range):** only the same 8 RGB-camera intrinsics — no tactile camera entry.
  4. **`proprio_stats/866961-940165.tar` H5 schema, scanned at scale:** all 773 H5 files covering the 7 tactile tasks have `state/effector/force` with shape `(0,)` — empty. No extra keys related to tactile / gel / finger / sensor.
- **Conclusion:** the tactile-equipped grippers were used *during data collection*, but the tactile sensor signal was never exported in the public release. The dataset has **0 hours of tactile video** in its current form.

### Other exclusions (sensor / format / scope mismatch)

| Dataset | Reason |
|---|---|
| **PoseIt** | 1,840 grasp datapoints over hold-phase transitions, not a tactile-video corpus |
| **The Feeling of Success** (Calandra) | 9,269 grasp outcomes, no modern tactile-video sequences exposed |
| **ObjectFolder Real** | Tactile component released as readings, not sequences; included in image-level `yxma/gelsight-mini-pretrain` instead |
| **OmniViTac** | Mixes GelSight Mini + Tac3D + Xense — not gel-only without manual sensor filtering |
| **TacQuad / VTV150K** | Same mix problem — Tac3D split is not photometric-stereo tactile; would need to drop those rows |
| **Touch-Vision-Language (TVL)** | Text-paired variant, not the SSL-pretrain target |
| **Sim-only tactile (Taxim / TACTO / DiffTactile)** | Out of scope for this real-data video repo; would be `sim_tactile_video_pretrain` |
| **Eagle Shoal** | Piezoresistive (non-optical) tactile — wrong sensor family |

### Companion repos

- [`yxma/gelsight-mini-pretrain`](https://huggingface.co/datasets/yxma/gelsight-mini-pretrain) — image-level GelSight Mini SSL corpus (~853K frames, 12 sources). The video-level repo here is meant to complement it.
- [`yxma/gelsight-mini-pretrain-nc`](https://huggingface.co/datasets/yxma/gelsight-mini-pretrain-nc) — NC sibling for Sparsh-licensed data.
- Planned: `yxma/tactile-video-pretrain-nc` — NC sibling for Touch-Slide once Phase 4 lands.

---

## How sources combine in the parquet

Every row has a `source` string column. When future sources are added, the schema is preserved and a `source = "freetacman" | "touch_and_go" | "visgel" | …` filter is all you need:

```python
ds = load_dataset("yxma/tactile-video-pretrain", "tactile_only", split="train")
freetacman = ds.filter(lambda r: r["source"] == "freetacman")
```