SOURCES.md

# Source Datasets

This document describes the **10 public datasets** aggregated into
`yxma/gelsight-mini-pretrain` — **8 real-world** captures and **2 simulated**
(Mini-calibrated Taxim) sources. The `domain` column on every row tags each
frame as `"real"` or `"sim"` so users can filter or mix freely.

For each source you'll find:

- a one-paragraph intro
- the paper / project page / canonical download
- the upstream license
- the upstream format and how we transformed it
- the resulting parquet stats
- a 40-image sample grid (4×10, central crop to square, ~144 px thumbs)

A high-level summary is at the bottom in [§ Aggregate statistics](#aggregate-statistics).

The full schema and load-time examples are in the main [README](README.md);
the conversion script is at [`scripts/make_parquet_v2.py`](scripts/make_parquet_v2.py).

![overview](assets/combined_overview.png)

---

## 1 · FoTA — Foundation Tactile (`fota_labeled` + `fota_unlabeled`)

**Intro.** FoTA is a large multi-sensor tactile-foundation dataset released
with the *T3* tactile-foundation-model paper. Its `panda_warped` subset
records a Franka Panda robot grasping 13 household objects against a
GelSight Mini in both fingers, with end-effector poses logged at every
recorded "still" and dense video frames in between.

**Source release.**
- 📄 Paper · *Towards a Tactile Foundation Model* — Zhao et al., 2024.
  [arXiv:2406.13640](https://arxiv.org/abs/2406.13640)
- 🤗 [`alanz-mit/FoundationTactile`](https://huggingface.co/datasets/alanz-mit/FoundationTactile)
- 🐙 [github.com/alanzjl/t3](https://github.com/alanzjl/t3)
- 📜 License: **MIT**

**Original format.** WebDataset (`.tar`) shards mixing per-sensor frames
with JSON metadata. The `panda_warped` subset of FoTA is the only piece
we use here; other sensors in T3 (DIGIT, GelSight Wedge, Soft Bubble…)
are excluded.

**How we processed it.**
1. Unpacked WebDataset tars and kept only frames recorded with a
   *GelSight Mini* sensor (other sensors discarded).
2. Re-encoded each frame as JPEG quality 92.
3. **Marker classification.** FoTA does not ship a markered/markerless
   label, but visual inspection of the captures showed a mix of dotted
   and smooth gels (the two gripper fingers used different gels for some
   recordings). We averaged ~50 frames per capture and ran dark-blob
   detection on the mean image; thresholding at ≥10 well-sized dots gave
   the per-row `markered` boolean. Of 124 captures, **36 are markered
   (all on the right finger), 88 are markerless**.
4. Split into two HF subsets: `fota_labeled` (the recorded stills, with
   x,y,z + quaternion poses) and `fota_unlabeled` (the dense video
   frames between stills, with object name only).

**Stats after processing.**

| Subset           | Frames  | Resolution | Markered / Markerless | Splits           |
|------------------|--------:|------------|----------------------:|------------------|
| `fota_labeled`   | 26,394  | 640 × 480  | 10,025 / 16,369       | train 21,139 · val 5,255    |
| `fota_unlabeled` | 66,761  | 640 × 480  | 36,592 / 30,169       | train 66,761 (train-only)    |

13 distinct contact objects, 5 initial-pose indices, 15,148 unique
(object, pose, side) captures. End-effector range
`x ∈ [−25.7, 129.1] mm`, `y ∈ [−137.3, 137.3] mm`,
`z ∈ [−38.6, 38.6] mm`.

**40 random samples** (`fota_labeled`, mixed gel):

![fota_labeled](assets/samples_40_fota_labeled.png)

**40 markered + 40 markerless** (`fota_labeled`):

| markered (right finger) | markerless (left finger) |
|:---:|:---:|
| ![fota markered](assets/samples_40_fota_labeled_markered.png) | ![fota markerless](assets/samples_40_fota_labeled_markerless.png) |

**40 random samples** (`fota_unlabeled`):

![fota_unlabeled](assets/samples_40_fota_unlabeled.png)

---

## 2 · py3DCal — sphere-indentation calibration grid (`threedcal`)

**Intro.** A motorised sphere indenter is pressed into a markerless
GelSight Mini gel at a regular **(x, y) grid of 1,209 positions**, each
at a fixed 3 mm penetration depth, with ~30 repeated frames per
position. Intended as a calibration / photometric-stereo training set
for the Mini.

**Source release.**
- Kota, Shah, Colgate, Reardon (2025).
- 💾 [Zenodo 18462608](https://zenodo.org/records/18462608)
- 📜 License: **CC-BY-4.0**

**Original format.** Loose PNGs in pose-named folders (e.g.
`x_010_y_006_z_3/0001.png`). PNG resolution is the GelSight Mini
low-resolution 320 × 240 mode.

**How we processed it.**
1. Decoded PNGs losslessly, re-encoded to JPEG quality 92 (large file-size
   reduction with no perceptible difference for tactile imagery).
2. Mapped the folder-encoded pose to `x_mm`, `y_mm`, `z_mm` (the z is the
   constant 3 mm penetration).
3. Tagged `markered=False` (gel is smooth).
4. Single `train` split.

**Stats after processing.**

| Subset      | Frames | Resolution | Markered | Probe coverage |
|-------------|-------:|------------|---------:|----------------|
| `threedcal` |  6,924 | 320 × 240  | 0        | 1,209 grid positions |

`x ∈ [0, 19] mm`, `y ∈ [0, 15] mm`, fixed `z = 3 mm`, ~30 frames per
position.

**40 random samples:**

![threedcal](assets/samples_40_threedcal.png)

**Probe coverage heatmap:**

![threedcal coverage](assets/threedcal_coverage.png)

---

## 3 · FEATS — Force Estimation for Tactile Sensors (`feats`)

**Intro.** A robotically-controlled indentation dataset designed for
*force* and *depth* regression from tactile RGB. Six indenter shapes
(sphere, cuboid, cylinder, cross, pyramid + held-out "unknown" probes)
are pressed into a markered (dotted) GelSight Mini gel with a 6-axis
F/T sensor logging f_x, f_y, f_z plus a 32×24 ground-truth depth grid
per image.

**Source release.**
- 📄 Author: Erik Helmut (2025).
- 🤗 [`erikhelmut/FEATS`](https://huggingface.co/datasets/erikhelmut/FEATS)
- 📜 License: **MIT**

**Original format.** `.npy` pickled dicts, one per frame, containing
`gs_img` (the RGB image), `f_x/y/z`, and `grid_z` (depth grid). Six
splits including out-of-distribution test sets.

**How we processed it.**
1. Loaded each `.npy`, extracted `gs_img`, re-encoded as JPEG q=92.
2. Recorded `f_x`, `f_y`, `f_z`, `grid_z_max`, `grid_z_mean` per row.
3. Parsed filename stem (e.g. `113_cuboid_12`) into
   `indenter="cuboid"`, `indenter_param="12"`.
4. **Added a `gel_variant` column** to distinguish the two physical
   sensor setups used in FEATS: `"black_dot"` (standard dotted Mini for
   `train`/`val`/`test`/`test_unknown_indenters`/`test_diff_sensor_old_gel`)
   vs `"different"` (a second Mini sensor with a differently-styled gel,
   used only in `test_diff_sensor_new_gel`).
5. **Removed empty frames.** The raw release includes ~5,300 frames where
   the indenter was hovering off the gel (`|f_z| < 0.5 N`). These are
   filtered out here; original 22,013 rows → kept 16,711.

**Stats after processing.**

| Split                          | Frames |
|--------------------------------|-------:|
| `train`                        | 11,415 |
| `test_unknown_indenters`       | 2,581  |
| `test`                         | 1,342  |
| `val`                          | 693    |
| `test_diff_sensor_new_gel`     | 341    |
| `test_diff_sensor_old_gel`     | 339    |
| **Total**                      | **16,969** |

Normal-force range `f_z ∈ [−73.3, 0.0] N` (mean −9.30 N, std 10.66 N);
shear `f_x ∈ [−4.86, 4.86] N`, `f_y ∈ [−5.89, 5.87] N`.

**40 random samples:**

![feats](assets/samples_40_feats.png)

**Force distribution and indenter mix:**

![feats force](assets/force_distribution.png)

---

## 4 · GelSLAM — tactile SLAM tracking + reconstruction (`gelslam`)

**Intro.** A real-time tactile SLAM dataset from CMU's RPL. Markerless
GelSight Mini videos of an object being pressed into the gel and
*slid* across the sensor surface; the data is annotated with per-frame
6DoF sensor pose, contact masks, and surface-gradient maps. Two splits:
**tracking** (140 short episodes, 20 objects) and **reconstruction**
(15 longer scans, 1–30 minutes each).

**Source release.**
- 📄 Paper · *GelSLAM: Real-time, High-Fidelity, Robust 3D Tactile SLAM* —
  Huang et al., 2025. [arXiv:2508.15990](https://arxiv.org/abs/2508.15990)
- 🤗 [`joehjhuang/GelSLAM_dataset`](https://huggingface.co/datasets/joehjhuang/GelSLAM_dataset)
- 🐙 [github.com/rpl-cmu/gelslam](https://github.com/rpl-cmu/gelslam)
- 📜 License: **MIT**

**Original format.** Single `dataset.zip` (~73 GB). Each episode is a
folder containing `gelsight.avi` (the RGB tactile video, ~25 FPS),
`true_start_T_currs.npy` (per-frame 4×4 pose), `contact_masks.npy`,
`gradient_maps.npy`. Reconstruction objects similarly have
`gelsight.avi` + `config.yaml`.

**How we processed it (v2 — area+intensity validity).**
1. Decoded all 155 `gelsight.avi` files frame-by-frame with OpenCV.
2. **Validity filter** — per-episode baseline = median of the first 10
   frames (typically the no-contact prologue); on every subsequent frame
   compute `pixel_diff = |frame_center − baseline|`, `mask = pixel_diff
   > 10` (sensor noise floor), `area = mask.sum()`, `intensity =
   pixel_diff[mask].mean()`. Keep iff **area ≥ 200 px AND intensity ≥ 12
   grey-levels**. This drops pre/post-contact frames cleanly without
   throwing out softer-but-real contacts (which a single mean-deform
   scalar would conflate).
3. **Perceptual-hash dedupe** within each episode (Hamming ≤ 4 on 8×8
   DCT-low-frequency hash) to drop near-identical consecutive frames.
4. Split-tagged `train` (the tracking dataset) vs `recon` (the
   reconstruction dataset). Per-row `episode` and `frame_idx` columns
   are populated.

Kept rate: **~380K raw → 114,019** (~30%). The validity rule keeps
more legitimate contact frames than the old mean-deform≥4 scalar (which
dropped softer contacts due to background-pixel dilution).

**Stats after processing.**

| Split   | Frames | Description |
|---------|-------:|-------------|
| `train` | 28,264 | 140 tracking episodes across 20 objects |
| `recon` | 61,348 | 15 long-form reconstruction scans |

**40 random samples:**

![gelslam](assets/samples_40_gelslam.png)

---

## 5 · TactileTracking / NormalFlow — 6DoF pose tracking (`tactile_tracking`)

**Intro.** The benchmark dataset for the *NormalFlow* paper — markerless
GelSight Mini videos of 12 different objects pressed against the gel
across 84 short tracking trials, recorded simultaneously with a webcam
for ground-truth visualisation. Used to evaluate contact-based 6DoF
pose tracking.

**Source release.**
- 📄 Paper · *NormalFlow: Fast, Robust, and Accurate Contact-based Object
  6DoF Pose Tracking with Vision-based Tactile Sensors* —
  Huang, Kaess, Yuan, IEEE RA-L 2024.
  [DOI 10.1109/LRA.2024.3505815](https://doi.org/10.1109/LRA.2024.3505815)
- 🤗 [`joehjhuang/TactileTracking`](https://huggingface.co/datasets/joehjhuang/TactileTracking)
- 🐙 [github.com/rpl-cmu/normalflow](https://github.com/rpl-cmu/normalflow)
- 📜 License: **MIT**

**Original format.** Single `dataset.zip` (~7 GB) with 84 trial folders
(e.g. `corner3`, `hammer1`, …). Each contains `gelsight.avi`,
`webcam.avi`, `true_start_T_currs.npy`, `contact_masks.npy`,
`gradient_maps.npy`. We use only the GelSight video.

**How we processed it.**
1. Parsed the trial folder name (e.g. `corner3` → object `corner`, trial
   `3`) and decoded the GelSight video.
2. Same area+intensity validity filter as GelSLAM (`A ≥ 200, I ≥ 12`,
   PIXEL_THRESH = 10), and perceptual-hash dedupe.
3. The trials are short (~10 s each at ~25 FPS) and contain a lot of
   slow contact motion, so the dedupe pass is *very* aggressive
   (~50% of valid frames are near-duplicates of the previous one).

Kept rate: **~10K raw → 2,408** (~24%) — the smallest subset, but
visually distinct from the others (varied real-world contact objects).

**Stats after processing.**

| Subset            | Frames | Resolution | Unique objects |
|-------------------|-------:|------------|---------------:|
| `tactile_tracking`|  2,408 | 320 × 240  | 12             |

**40 random samples:**

![tactile_tracking](assets/samples_40_tactile_tracking.png)

---

## 6 · Real Tactile MNIST — 3D-printed digit touches (`real_tactile_mnist`)

**Intro.** A large benchmark for *active tactile perception*: 600 3D-
printed MNIST digits, each touched 256 times by a robot-arm-mounted
GelSight Mini, producing 153,600 unique touches. Each touch is a short
video clip; we keep one representative frame per touch.

**Source release.**
- 📄 Paper · *Tactile MNIST: A Benchmark for Active Tactile Perception* —
  Schneider, de Farias, Calandra, Chen, Peters, 2025.
  [arXiv:2506.06361](https://arxiv.org/abs/2506.06361)
- 🤗 [`TimSchneider42/tactile-mnist-touch-real-seq-t256-320x240`](https://huggingface.co/datasets/TimSchneider42/tactile-mnist-touch-real-seq-t256-320x240)
- 🐙 [github.com/TimSchneider42/tactile-mnist](https://github.com/TimSchneider42/tactile-mnist)
- 🌐 [Project page](https://sites.google.com/robot-learning.de/tactile-mnist/)
- 📜 License: **CC-BY-2.0** (code is MIT)

**Original format.** Parquet "rounds": one row = one digit object,
containing a list of 256 short video clips (`sensor_video[].bytes`,
MP4-encoded), plus per-touch position, gel-frame pose, and timestamps.

**How we processed it (v3 — area+intensity rule).**
1. For each touch video clip, decode all ~60–73 frames.
2. Compute a per-clip baseline = median of the first 5 frames (no-contact prologue).
3. Use upstream `touch_start_time_rel` / `touch_end_time_rel` timestamps to
   identify the actual contact window inside the clip (typically only ~6
   frames near the end).
4. Within that window, pick the frame with **maximum mean |frame − baseline|**
   — the true peak-contact frame.
5. On the picked frame, compute on the central 50% crop:
   ```
   pixel_diff = |frame - baseline|
   mask       = pixel_diff > 10              # sensor-noise floor
   area       = mask.sum()                   # # of "lit" pixels
   intensity  = pixel_diff[mask].mean()      # avg deformation in lit pixels
   ```
   **Keep the touch iff area ≥ 40 px AND intensity ≥ 15 grey-levels.** Drops
   ~89 % of touches that produced no real imprint (3D-printed digits are
   small and many touches are glancing); the surviving ~16,961 frames all
   visibly show a digit-edge streak.

   *Earlier iterations:* the first version picked the middle of the video
   (~30 frames before any contact); the second picked the middle of the
   touch window (still a press *transition*, not peak compression). Both
   gave samples that looked like bare gel. The area+intensity rule with
   peak-deformation frame selection is what finally produced clean samples.
6. No dedupe — each touch is a unique digit at a unique position.
7. Per-row metadata: `digit_class` (0–9), `episode` (which of the 600
   physical digit objects), `obj_name="digit_<N>"`, `frame_idx` (the
   selected frame index within the source video).

Kept rate: **153,600 raw → 153,600** (100%).

**Stats after processing.**

| Split    | Frames  | Resolution |
|----------|--------:|------------|
| `train`  | 128,000 | 320 × 240  |
| `test`   | 25,600  | 320 × 240  |

**40 random samples:**

![rtm](assets/samples_40_real_tactile_mnist.png)

**Digit-class balance:**

![rtm digits](assets/rtm_digit_distribution.png)

### Why we don't use the authors' `touch-real-single-t256-320x240` directly

TimSchneider42 also publishes a "single-frame-per-touch" variant
(`tactile-mnist-touch-real-single-t256-320x240`) where they pre-extract one
image per touch. We deliberately use the **`-seq-`** variant (full videos)
and do our own peak-frame + area+intensity extraction. Reason:

- The authors' `single` release **ships every touch** (153,600 frames),
  with **no validity filter** — they include barely-perceptible "glancing"
  touches alongside real imprints.
- Probing 5,120 random touches from their `single` release with the same
  metric we use everywhere else (cross-touch median baseline,
  `A ≥ 40 px above 10 grey-levels`, `I ≥ 15 grey-levels`): only
  **11.0 % of authors' touches pass** — meaning ~89 % visually look like
  bare gel.
- Our extraction yields **~17 K curated frames** where every kept image
  visibly shows a digit-edge imprint.

Side-by-side comparison (top: authors' raw release; bottom: our filtered
extraction):

![rtm authors vs ours](assets/rtm_authors_vs_ours.png)

The authors' raw release covers the same 600 physical digits × 256 touches
but ships everything; we apply a peak-frame picker (using their upstream
`touch_start_time_rel`/`touch_end_time_rel` annotations) and then a
visibility filter on top. So in principle their `single` release is "our
extraction minus the validity filter".

---

## 7 · FeelAnyForce — force-controlled indentations (`feelanyforce`)

**Intro.** A robotic-indentation dataset originally collected to learn
contact-force estimation from tactile RGB. 42 distinct objects (geometric
primitives + lemons / fruit / household items) pressed into a GelSight
Mini gel under controlled force trajectories.

**Source release.**
- 📄 Sharei et al., 2024 (paper title: *FeelAnyForce*).
- 🤗 [`amirsh1376/FeelAnyForce`](https://huggingface.co/datasets/amirsh1376/FeelAnyForce)
- 📜 License: **CC-BY-4.0**

**Note on gel variant.** Some references describe FeelAnyForce as a
"markered Mini" dataset, but visual inspection of the released tactile
images confirms the gel surface is **smooth (markerless)** — there is no
visible dot grid in any sample. We label it `markered=False` to reflect
what is actually in the data.

**Original format.** Multi-part zip archive (`dataset.zip` + `.z01` +
`.z02` + `.z03` + `dataset_part_a/b/c`, reassembled to ~82 GB extracted).
Each of 42 objects has subfolders `tactile/`, `tactile_nobg/` (background
subtracted), `depth/`. We use only `tactile/` (raw RGB). Plus three CSVs
(`TacForce_train/val/test_set.csv`) with per-frame force labels — *not
joined in here*; see upstream for force regression work.

**How we processed it.**
1. Iterated all `tactile/*.png` files (320 × 240 PNG).
2. Re-encoded each as JPEG q=92.
3. **Validity filter disabled** — the data is already curated, every
   frame is a real indentation moment, and the per-capture-median
   baseline approach would yield false positives for "empty" since the
   median itself sits in a contact frame.
4. **Perceptual-hash dedupe active** — slow indentation press-and-hold
   means many adjacent frames are visually near-identical; dedupe
   drops ~50%.

Kept rate: **~100K raw → 48,197** (~48%). Drops are all dedupe; zero
empty drops.

**Stats after processing.**

| Subset         | Frames  | Resolution | Markered | Unique objects |
|----------------|--------:|------------|---------:|---------------:|
| `feelanyforce` | 48,197  | 320 × 240  | 0        | 42             |

**40 random samples:**

![feelanyforce](assets/samples_40_feelanyforce.png)

---

---

## 8 · Sim Tactile MNIST — Mini-calibrated Taxim render of RTM (`sim_tactile_mnist`)

**Intro.** A simulation companion to `real_tactile_mnist`, from the same
authors (Schneider et al.). Renders the same digit-touch geometry through
**Taxim** — a tactile simulator re-calibrated to the GelSight Mini —
producing visually plausible synthetic Mini RGB. Useful as additional
markerless pretraining data, or for sim-to-real transfer studies.

**Source release.**
- 🤗 [`TimSchneider42/tactile-mnist-touch-syn-single-t32-320x240`](https://huggingface.co/datasets/TimSchneider42/tactile-mnist-touch-syn-single-t32-320x240)
- 🐙 [github.com/TimSchneider42/tactile-mnist](https://github.com/TimSchneider42/tactile-mnist)
- 📜 License: **CC-BY-2.0**

**Original format.** Parquet "rounds" — one row = one digit object, with
`sensor_image` as a list of 32 already-rendered JPEG byte structs (one image
per touch, already at peak contact — no video decode needed).

**How we processed it.**
1. Iterated parquet rows; for each row's 32 touches, decoded `sensor_image[i].bytes`
   directly as a JPEG.
2. **No validity filter** — every sim frame is by construction at peak contact
   (it's the rendered output, not a video).
3. No dedupe — each sim touch is at a unique gel-frame pose.
4. **Capped at 200 K kept frames** to balance against real sources.
5. Per-row metadata: `domain="sim"`, `digit_class` (0–9), `episode` (object id),
   `obj_name="digit_<N>"`, `frame_idx` (touch index within the round).

**Stats after processing.**

| Subset                | Frames  | Resolution | Domain |
|-----------------------|--------:|------------|--------|
| `sim_tactile_mnist`   | 150,601 | 320 × 240  | sim    |

**40 random samples:**

![sim_tactile_mnist](assets/samples_40_sim_tactile_mnist.png)

---

## 9 · Sim Starstruck — Mini-calibrated Taxim render of star objects (`sim_starstruck`)

**Intro.** Sim companion to the "Starstruck" star-shape search benchmark in
the Tactile MNIST family. Same Taxim Mini-calibrated renderer, but the
indenter geometry is a star instead of a digit — angular edges produce
characteristic radial imprints.

**Source release.**
- 🤗 [`TimSchneider42/tactile-mnist-touch-starstruck-syn-single-t32-320x240`](https://huggingface.co/datasets/TimSchneider42/tactile-mnist-touch-starstruck-syn-single-t32-320x240)
- 📜 License: **CC-BY-2.0**

**Original format.** Same as `sim_tactile_mnist`: parquet rows with `sensor_image`
list of 32 JPEG byte structs per row.

**How we processed it.** Identical recipe to `sim_tactile_mnist` — no validity
filter, no dedupe, capped at 200 K kept frames.

**Stats after processing.**

| Subset           | Split | Frames  | Resolution | Domain |
|------------------|-------|--------:|------------|--------|
| `sim_starstruck` | train | 150,000 | 320 × 240  | sim    |
| `sim_starstruck` | test  |  16,104 | 320 × 240  | sim    |
| **Total**        |       | **200,000** |        |        |

**40 random samples:**

![sim_starstruck](assets/samples_40_sim_starstruck.png)

---

## 10 · UniT — continuous 3D-pose tracking (`unit`)

**Intro.** UniT (Yu et al., 2024) is a self-supervised pretraining
dataset designed for tactile pose estimation. It consists of a long
recording of a known calibration object being slid around against the
GelSight Mini gel, with paired 3D pose targets (x, y, z, yaw). The data
is delivered as a single zarr replay-buffer with **continuous contact**
throughout (no approach / release phases), so we skip the area+intensity
filter and keep every frame.

**Source release.**
- 🔗 [UniT repo](https://github.com/ZeyuYong/UniT) (zarr replay-buffer
  format)
- 📜 License: BSD-3-Clause-style permissive (per upstream repo metadata)

**Original format.** A `replay_buffer.zarr` with two arrays:
- `data/tactile_image` — `(11340, 240, 320, 3)` uint8, zstd-compressed,
  9-frame chunks
- `data/3Dpose` — `(11340, 4)` float32 — (x_mm, y_mm, z_mm, yaw)

**How we processed it.** Read each frame from zarr, re-encoded as JPEG
quality 92, filled the unified schema with pose metadata mapped into
the `x_mm`/`y_mm`/`z_mm`/`quat_z` columns (yaw stored in `quat_z`).
Wrote `unit/train-00000-of-00001.parquet`. **No contact filter** — UniT
is a continuous-contact dataset by design, so every frame is contact-
bearing.

**Stats after processing.**

| Subset | Frames | Resolution | Markered / Markerless | Splits |
|--------|-------:|------------|----------------------:|--------|
| `unit` |    387 | 320 × 240  | 387 / 0               | train 387 |

After re-ingesting with the contact filter (initially we kept all 11,340
frames assuming continuous contact, but 86% turned out to be near-black
LED-off frames), only **387 frames** show actual gel contact. The visible
frames clearly show **markered gel** (regular dot pattern), so this
subset goes in the markered pool.

**40 random samples:**

![unit](assets/samples_40_unit.png)

---

## 11 · TacQuad — quad-sensor benchmark, Mini stream (`tacquad`)

**Intro.** TacQuad (Feng et al., 2025) is a 4-sensor synchronized
benchmark — every contact was captured simultaneously with **4 different
tactile sensors** (GelSight Mini, DIGIT, DuraGel, Tac3D). We ingest
**only the GelSight Mini stream**, contributing the highest object-
diversity component of the aggregation: **181 unique household, outdoor,
and "fine-grain" objects** captured under controlled indenter pressure.

**Source release.**
- 🔗 [TacQuad / AnyTouch project](https://github.com/anytouch-project)
- 📜 License: CC-BY-4.0

**Original format.** Three folders (one per environment):
`data_indoor/` (101 objects), `data_outdoor/` (50 objects), `data_fine/`
(30 objects). Each object has subfolders for each sensor's stream; we
read only the `gelsight/` subfolder (PNG files numbered 0.png, 1.png…).

**How we processed it.** Walked each environment, read all `gelsight/*.png`
frames per object, applied the unified area+intensity filter (I_min=12,
A_min=40, 1.5 % bg-diversity). Each environment becomes one split. Per-
row metadata: `obj_name`, `episode=obj_name`, `frame_idx` = the PNG
sequence number, `domain="real"`, `gel_variant="markerless"`.

**Stats after processing.**

| Subset    | Split        | Frames | Resolution | Domain |
|-----------|--------------|-------:|------------|--------|
| `tacquad` | data_indoor  |  5,363 | 320 × 240  | real   |
| `tacquad` | data_outdoor |  3,934 | 320 × 240  | real   |
| `tacquad` | data_fine    |  2,898 | 320 × 240  | real   |
| **Total** |              | **12,195** |        |        |

24,866 raw Mini frames in the upstream zips, retention ~49 % after the
contact filter. Distinguishing feature: largest object diversity (181
real-world objects) in the entire dataset.

**40 random samples:**

![tacquad](assets/samples_40_tacquad.png)

---


## Investigated but not included

- **`facebook/gelsight-force-estimation`** — CC-BY-NC-4.0 license is
  incompatible with this CC-BY-4.0 repo. (Has been moved to our
  companion NC repo
  [`yxma/gelsight-mini-pretrain-nc`](https://huggingface.co/datasets/yxma/gelsight-mini-pretrain-nc),
  where it is now superseded by `sparsh` — the same data at a larger
  upstream snapshot.)
- **TVL — Touch-Vision-Language** (Fu et al., ICML 2024,
  [project](https://tactile-vlm.github.io/),
  [dataset](https://huggingface.co/datasets/mlfu7/Touch-Vision-Language-Dataset))
  — **uses a DIGIT sensor, not GelSight Mini.** Quote from project page:
  *"Tactile data are collected using a DIGIT sensor: a compact, open-
  source tactile sensor that provides observations in the form of RGB
  images."* We initially ingested 209,795 frames before noticing the
  sensor mismatch. Parquet preserved locally for potential future
  `digit-pretrain` repo.
- **Touch-and-Go** (Yang et al., NeurIPS 2022, [GitHub](https://github.com/fredfyyang/Touch-and-Go)) —
  **uses a different GelSight model, not the Mini.** The dataset's
  `gelsight.mp4` clips show 640×480 frames with white-light illumination
  (channel means R/G/B ≈ 137/135/140, flat ~grey) — completely
  different from the Mini's 3-colored-LED signature (Mini at-rest has
  B > G > R with ~30 grey-level spread). We initially ingested 2,066
  frames before noticing the channel-signature mismatch and verifying
  against the paper. Removed from this aggregate; could go to a
  separate "gelsight-large-pretrain" repo if needed.
- **YCB-Sight** — uses the **original (larger) GelSight sensor**, not the Mini, so the gel surface and illumination geometry differ. Also CC-BY-SA-4.0 (viral copyleft) and the sim is not Mini-calibrated. Not included.
- **TACTO**, **MidasTouch**, **DiffTactile**, generic GelSight sims —
  either DIGIT-specific, depth-only, or not Mini-calibrated. Not
  included.

---

## Channel-order normalization

Some upstream sources stored their RGB frames as **BGR**, which would
have produced color-inverted tactile patterns when loaded as RGB. We
fixed this by computing the mean per-channel (R, G, B) for every
subset, and unconditionally swapping R↔B for those where R > B at rest
(GelSight Mini's at-rest gel illumination has B > R consistently).

The affected subsets (now corrected to RGB) were:
- `fota_unlabeled` — globally BGR-stored upstream
- `unit` — globally BGR-stored upstream
- (NC repo) `faf_force_estimation` — globally BGR-stored
- (NC repo) `sparsh` — **mixed** (some files RGB, some BGR);
  per-image conditional swap applied based on per-image R > B

The diagnostic is published as `assets/channel_order_diagnosis.json`
(both repos). After normalization, every image in both repos has
B > R at rest, matching the Mini's reference illumination geometry.

---

## Aggregate statistics

![composition](assets/composition.png)

| Subset                | Domain | Frames  | Bytes (GB) | Resolution | Markered | Markerless |
|-----------------------|--------|--------:|-----------:|------------|---------:|-----------:|
| `fota_labeled`        | real   | 26,394  | 0.46       | 640 × 480  | 10,025   | 16,369     |
| `fota_unlabeled`      | real   | 66,761  | 3.02       | 640 × 480  | 36,592   | 30,169     |
| `threedcal`           | real   |  6,924  | 0.05       | 320 × 240  | 0        |  6,924     |
| `feats`               | real   | 16,969  | 0.28       | 320 × 240  | 16,969   | 0          |
| `gelslam`             | real   | 114,019 | 1.03       | 320 × 240  | 0        | 114,019    |
| `tactile_tracking`    | real   |  2,408  | 0.02       | 320 × 240  | 0        |  2,408     |
| `real_tactile_mnist`  | real   | 30,956  | 0.16       | 320 × 240  | 0        | 30,956     |
| `feelanyforce`        | real   | 48,197  | 0.45       | 320 × 240  | 0        | 48,197     |
| `unit`                | real   | 11,340  | 0.01       | 320 × 240  | 0        | 11,340     |
| `tacquad`             | real   | 12,195  | 0.11       | 320 × 240  | 0        | 12,195     |
| `sim_tactile_mnist`   | sim    | 150,601 | 1.28       | 320 × 240  | 0        | 150,601    |
| `sim_starstruck`      | sim    | 166,104 | 1.39       | 320 × 240  | 0        | 166,104    |
| `tvl`                 | real   | 209,795 | 7.90       | mixed¹     | 0        | 209,795    |
| **Real total**        |        | **745,958** | **~13.3** |        | **63,586** | **682,372** |
| **Sim total**         |        | **316,705** | **~2.67** |        | **0**     | **316,705** |
| **Grand total**       |        | **1,062,663** | **~16.0** |       | **63,586** | **999,077** |

![summary pies](assets/summary_pies.png)

**Resolution distribution:**

![resolution](assets/resolution_distribution.png)

**Gel-variant pool sizes after aggregation:**

- **Markerless pool** (all `_*[markerless]` and pure-markerless subsets — `fota_*[markerless]` + `threedcal` + `gelslam` + `tactile_tracking` + `real_tactile_mnist` + `feelanyforce` + `unit` + `tacquad`): **472,577 real markerless frames** (789,282 if including sims)
- **Markered pool** (`feats` + `fota_labeled[markered]` + `fota_unlabeled[markered]`): **63,586 frames**

Filter examples:

```python
from datasets import load_dataset, concatenate_datasets

# Big markerless pool for VAE / MAE / contrastive pretraining
sources = ["fota_unlabeled", "threedcal", "gelslam", "tactile_tracking",
           "real_tactile_mnist", "feelanyforce"]
markerless = concatenate_datasets([
    load_dataset("yxma/gelsight-mini-pretrain", s, split="train"
                ).filter(lambda r: not r["markered"])
    for s in sources
])

# Markered pool
markered = concatenate_datasets([
    load_dataset("yxma/gelsight-mini-pretrain", "feats",      split="train"),
    load_dataset("yxma/gelsight-mini-pretrain", "fota_labeled", split="train"
                ).filter(lambda r: r["markered"]),
])
```

---

## Citation chain

If you use this aggregated release, please cite the **upstream sources**
for the subsets you use. See the per-subset sections above for paper /
DOI references; the [main README](README.md) consolidates the BibTeX-
ready citation list.