# Roadmap & Source Shortlist

Public roadmap for [`yxma/tactile-video-pretrain`](https://huggingface.co/datasets/yxma/tactile-video-pretrain). Tracks every public GelSight-family video dataset we've audited and the order in which we'll fold them into this repo.

## Status legend

- ✅ included
- 🟡 next-priority add (Phase 1)
- 🟢 medium-priority add (Phase 2)
- 🔵 low-priority / auxiliary (Phase 3)
- 🟠 license restricts → goes to companion **NC** repo (Phase 4)
- ❌ investigated, will not include (sensor/format/scope mismatch)

## Master table

| # | Dataset | Sensor | Modality | Scale | FPS | License | Status |
|---|---|---|---|---|---|---|---|
| 1 | [**FreeTacMan**](https://huggingface.co/datasets/OpenDriveLab/FreeTacMan) | McTac-derived (GelSight-style, dual finger) | tactile + wrist RGB | 50 tasks / 10k+ traj / 3M+ pairs | 30 | MIT | ✅ |
| 2 | [**VisGel**](https://visgel.csail.mit.edu/) | GelSight (markered) | tactile + RGB | 195 obj / 12k touches / 3M frames | – | unspec (audit) | 🟡 |
| 3 | [**Touch and Go**](https://touch-and-go.github.io/) | GelSight + ego RGB | tactile + RGB | "in-the-wild", global N unspec | – | CC-BY-4.0 | 🟡 |
| 4 | [**GelSLAM Tracking**](https://huggingface.co/datasets/joehjhuang/GelSLAM_dataset) | GS Mini (markerless) | tactile-only | 20 obj / 140 ep / ~73k frames | 25 | MIT | 🟡 |
| 5 | [**GelSLAM Reconstruction**](https://huggingface.co/datasets/joehjhuang/GelSLAM_dataset) | GS Mini (markerless) | tactile-only | 15 obj / 1 scan each / 1–30 min | 25 | MIT | 🟡 |
| 6 | [**TactileTracking**](https://huggingface.co/datasets/joehjhuang/TactileTracking) (NormalFlow) | GS Mini (markerless) | tactile + aux webcam | 12 obj / 84 trials / ~21k frames | 25 | MIT | 🟢 |
| 7 | [**YCB-Slide (real)**](https://github.com/rpl-cmu/YCB-Slide) | DIGIT | tactile + RGB | 10 YCB / 50 logs / ~180k frames | 30 | MIT | 🟢 |
| 8 | [**YCB-Slide (sim)**](https://github.com/rpl-cmu/YCB-Slide) | DIGIT (Taxim) | tactile (sim) | 10 YCB / 50 traj | – | MIT | 🟢 |
| 9 | [**YCB-Sight**](https://github.com/Robo-Touch/YCB-Sight) | GelSight + Kinect | tactile + RGB-D | 30 sim + 6 real YCB | – | data CC-BY-SA-4.0, code MIT | 🟢 *propagation flag* |
| 10 | [**Touch-Slide**](https://github.com/facebookresearch/sparsh) | DIGIT (markerless) | tactile-only | 9 obj × 5 traj / 180k frames | – | **CC-BY-NC-4.0** | 🟠 → NC sibling repo |
| 11 | [**Tactile MNIST (real seq)**](https://huggingface.co/TimSchneider42) | GS Mini | tactile-only | 600 digits / 153.6k samples | – | CC-BY-2.0 | 🔵 |

## Phased acquisition plan

| Phase | Add | New hours (est) | License gate |
|---|---|---|---|
| **0** (done) | FreeTacMan | 29.7 h tactile + 15.7 h RGB | MIT ✅ |
| **1** | VisGel · Touch and Go · GelSLAM (track + recon) | +significant tactile-RGB; +GS Mini long sequences | VisGel license audit; CC-BY-4.0 (TaG); MIT (GelSLAM) |
| **2** | YCB-Slide (real + sim) · YCB-Sight · TactileTracking | small but geom-supervised | MIT; YCB-Sight CC-BY-SA-4.0 needs propagation → may emit `-sa` sibling |
| **3** | Tactile MNIST | controlled benchmark | CC-BY-2.0 |
| **4** | Touch-Slide | ~180k DIGIT frames (NC) | **CC-BY-NC-4.0** → `yxma/tactile-video-pretrain-nc` |

After each phase we'll regenerate `assets/` (composition, samples, pies) so the dataset card always reflects the current mix.

## Out of scope (with reason)

| Dataset | Reason |
|---|---|
| **PoseIt** | 1,840 grasp datapoints, not tactile video |
| **The Feeling of Success** | 9,269 grasp outcomes, no modern tactile video |
| **ObjectFolder Real** | "tactile readings" not tactile video sequences |
| **OmniViTac** | mixes GS Mini + Tac3D + Xense — not gel-only without manual filtering |
| **AgiBotWorld (all releases)** | **Audited:** 0 h of tactile video in public release; all 8 video streams are scene RGB; `state/effector/force` is empty `(0,)` across all 773 tactile-flagged episodes. See [SOURCES.md](SOURCES.md#agibot-world). |
| **Eagle Shoal** | piezoresistive (non-optical) tactile |

## Recommended training recipe

```
phase 1   tactile-only SSL          : Touch-Slide + GelSLAM (track+recon) + TactileTracking + Tactile MNIST
phase 2   cross-modal alignment     : VisGel + Touch and Go + YCB-Slide (real)
phase 3   geometry/pose supervision : YCB-Slide (sim heightmap+mask) + YCB-Sight + GelSLAM/TT pose-gradient heads
phase 4   manipulation finetuning   : FreeTacMan
phase 5   evaluation                : held-out tasks/objects + sensor-shift (markered ↔ markerless)
```

Frugal alternative: **Touch-Slide → VisGel → FreeTacMan** captures ~80 % of the value with much less plumbing.