yxma/gelsight-mini-pretrain

GelSight Mini Pretrain

A unified, parquet-native collection of ~700 K GelSight Mini tactile RGB frames for self-supervised representation learning. Eleven public datasets are aggregated under one schema, each filtered through a unified area+intensity contact rule + per-capture phash dedupe:

• ~503 K real-world frames from 9 captures (FoTA labeled+unlabeled, 3DCal, FEATS, GelSLAM, TactileTracking, Real Tactile MNIST, FeelAnyForce, TacQuad-mini)
• 400 K simulated frames from 2 Mini-calibrated Taxim renders (Sim Tactile MNIST, Sim Starstruck)
• Plus a companion CC-BY-NC extension with another ~15 K real frames from Meta Sparsh / FAF force-estimation

Every row carries a domain column ("real" or "sim") and a markered flag (gel has tracking dots?) so you can mix or filter freely.

➡️ For a full per-subset breakdown (intro, paper, license, processing recipe, sample grids, stats) see SOURCES.md.

📦 Need more data and OK with non-commercial use? A companion extension at yxma/gelsight-mini-pretrain-nc adds GelSight Mini sources whose licenses (e.g. CC-BY-NC-4.0) aren't compatible with this commercial-friendly repo. Same schema, same pipeline, strictly additive. Load both and concatenate for the largest pool.

TL;DR

from datasets import load_dataset

# Largest subset, markerless, for VAE / MAE / contrastive pretraining
ds = load_dataset("yxma/gelsight-mini-pretrain", "fota_unlabeled", split="train")

# Pose-labelled subset for supervised fine-tuning
ds = load_dataset("yxma/gelsight-mini-pretrain", "fota_labeled",   split="train")

# Calibration sweep on a sphere indenter
ds = load_dataset("yxma/gelsight-mini-pretrain", "threedcal",      split="train")

# Markered (dotted) gel with force labels
ds = load_dataset("yxma/gelsight-mini-pretrain", "feats",          split="train")

img = ds[0]["image"]                       # PIL.Image (decoded on the fly)
meta = {k: ds[0][k] for k in ds.features if k != "image"}

Why this exists

Public GelSight Mini data is scattered. Authors release each dataset with its own folder structure, file format (.jpg / .png / .npy pickled dicts / WebDataset tars), naming convention, and metadata layout. For most pretraining pipelines you end up writing four bespoke dataloaders. This release does the boring part once:

1. Unifies the format: every frame is a JPEG byte string in a parquet row, alongside a single shared metadata schema (poses, indenter labels, forces — all optional, all null where N/A).
2. Re-encodes uniformly: all images at JPEG quality 92, so file sizes are comparable across sources.
3. Tags the gel variant: a single markered bool tells you whether the gel has tracking dots — useful because dotted gels look visually very different from smooth ones, and mixing them confuses self-supervised objectives.
4. Splits by source, not just train/val: each upstream dataset is its own HF "config", so you can train on combinations of your choosing.

How this dataset was built

The pipeline below was applied to four public GelSight Mini datasets (FoTA, py3DCal, FEATS, and — pending license — FeelAnyForce) to produce a single homogeneous parquet collection. Each step is a separate script in this repository's source tree; the four steps below are what changed between the raw upstream releases and what you load_dataset here.

_{A 3-row overview across the three currently-released subsets. Each row is a different upstream dataset; every frame is a real GelSight Mini capture.}

1. Source verification — "is this really GelSight Mini?" Each upstream release was checked against the Mini's known native modes (640×480 RGB, occasionally 320×240) and the dotted/dotless gel variants documented by GelSight Inc. We cross-referenced every dataset's metadata.json, paper, and folder naming convention to confirm the sensor model before inclusion. FAF is held back until licensing is clarified.

2. Format unification — one schema, one image codec Upstream layouts are wildly different: FoTA ships WebDataset tar shards, py3DCal ships loose PNGs in pose-named folders, FEATS ships .npy pickled dicts containing the image plus a 32×24 depth grid plus forces. Each was decoded to a numpy array, re-encoded as JPEG quality 92 for fair file-size comparison, and packed into the unified schema below as a parquet row. Shards target ~2 GB each so they stream efficiently from HF Hub.

3. Marker detection — fixing FoTA's mixed-gel surprise FoTA does not ship a markered/markerless label, but on visual inspection the captures are clearly mixed (some have ~80 visible tracking dots, others are smooth). We classify automatically: for each FoTA capture (one continuous press of an object), average ~50 evenly-spaced frames into a single mean image, then run dark-blob detection on it. Markered gels yield ≥10 well-sized dark dots in the mean; markerless gels yield <10 scattered noise blobs. Result: 36 of 124 captures are markered (all on the right finger), 88 are markerless. The per-row markered column reflects this.

Markered captures (top half)	Markerless captures (bottom half)

4. Empty-frame removal — cleaning FEATS FEATS' raw frames include captures where the indenter was hovering off the gel (no contact at all). These would teach a pretraining model nothing useful. We filtered them with |f_z| < 0.5 N (using FEATS's force-sensor ground truth), removing 5,302 frames out of 22,013. We also added a gel_variant column distinguishing the two physical sensor setups used in FEATS (black_dot for the main markered gel, different for the second sensor used in test_diff_sensor_new_gel).

_{FEATS samples after empty-frame removal — every frame now shows a real contact.}

5. Tier-2 expansion: adding four more markerless Mini sources After the initial release a second pipeline pass added GelSLAM, TactileTracking, Real Tactile MNIST, and FeelAnyForce — four more public CC-BY/MIT-licensed GelSight Mini datasets. Each was processed through a shared backbone:

• Adaptive subsampling per source, capped at 200K kept frames so no single source dominates.
• Validity filter for video sources (GelSLAM, TactileTracking): a per-capture baseline is computed from the median of the first 10 frames; each subsequent frame is kept only if its central deformation from baseline exceeds ~4 grey levels. Up to 3% of kept frames are allowed below this threshold (variance).
• Filter disabled for already-curated sources (FeelAnyForce indentation stills, Real Tactile MNIST middle-of-touch frames) where every frame is by construction in-contact.
• Perceptual-hash dedupe within each capture (Hamming ≤ 4 on 8×8 DCT low-frequency hash) to drop near-identical adjacent frames from slow indentation videos.
• Per-source frame budget per dataset (see "Composition" above for kept counts; raw → kept summary in §How this dataset was built).

Final result: ~865K frames across 8 subsets, ~24 GB on disk, one schema, one image codec, one markered flag.

The entire conversion script lives in this repo at scripts/make_parquet_v2.py. It implements the per-source decoders, the validity filter, the perceptual-hash dedupe, the 200K-per-source budget, and the parquet sharding. The CLI:

python make_parquet_v2.py probe   <subset>   # measure dynamism + empty fraction
python make_parquet_v2.py process <subset>   # full pipeline + write parquet shards
python make_parquet_v2.py stats              # row counts across all subsets

<subset> is one of gelslam, tactile_tracking, real_tactile_mnist, feelanyforce.

Statistics at a glance

Composition across the 8 subsets, coloured by gel variant:

Image resolution — the dataset retains both GelSight Mini native modes. Filter by height/width if you need a uniform-resolution pool:

FEATS normal-force distribution and indenter-shape mix (the only subset with force labels):

py3DCal probe-position coverage — confirms the dense calibration grid:

Real Tactile MNIST · digit-class balance (used as a sanity check that the upstream touch sampling was uniform across digits 0–9):

Composition

Subset	Source dataset	Frames	Gel	Has labels
fota_labeled	FoTA — panda_warped still captures	29,494 (66% markerless, 34% markered)	mixed¹	end-effector x,y,z + quaternion
fota_unlabeled	FoTA — same captures, video frames	66,761 train-only (stride-subsampled from 516K — see ³ below)	mixed¹	object name only
threedcal	py3DCal sphere indentation grid	36,270	markerless	probe x, y, penetration depth (mm)
feats	FEATS indentation with force grids	16,711	markered (two gel variants — see below)	indenter shape/size + contact forces
gelslam	GelSLAM tactile SLAM tracking + reconstruction	89,612 (28K tracking + 61K reconstruction)	markerless	episode + object name
tactile_tracking	TactileTracking (NormalFlow) 6DoF pose tracking	1,605	markerless	object + trial id
real_tactile_mnist	Real Tactile MNIST 3D-printed digit touches	16,961 (14K train + 2.8K test)	markerless	digit class (0–9) + round id
feelanyforce	FeelAnyForce force-controlled indentations	50,997	markerless²	object name
sim_tactile_mnist	SIM · Taxim-rendered Mini imagery of digit touches	200,000	markerless	digit class + episode
sim_starstruck	SIM · Taxim-rendered Mini imagery of star objects	200,000	markerless	episode

¹ FoTA used different gels on the two gripper fingers for many of its captures. Approximately 36 of 124 captures use a markered gel on the right finger and a markerless gel on the left; the remaining 88 captures use markerless gels on both. The per-row markered column was set by averaging ~50 frames per capture and counting visible dark dots in the mean image (threshold ≥10 dots). Use it to filter:

² FeelAnyForce is colloquially described as a "markered Mini" dataset in some references, but visual inspection of the released tactile images confirms the gel surface is smooth (markerless). The markered column reflects what is actually observed in the data.

³ fota_unlabeled was uniformly stride-subsampled from 516,523 → 66,761 frames (~12.9 % retention) and consolidated as train-only on 2026-05-18. Reason: only 11 objects across 60 captures means within-capture frames are near-duplicate gripper-trajectory snapshots, and the raw subset was ~70 % of the aggregated pretraining pool. Because this subset has only obj_name as a label (no pose / force / ground-truth contact) there is nothing to benchmark, so the prior train/val partition served no purpose and has been dropped. Every (object, init_pose, side) combination is still represented; the subset's share of the aggregated pool drops to ~7 %.

ds = load_dataset("yxma/gelsight-mini-pretrain", "fota_labeled", split="train")
markerless = ds.filter(lambda r: not r["markered"])
markered   = ds.filter(lambda r: r["markered"])

Sample images

fota_labeled  ·  29,494 frames  ·  mixed gel  ·  +6DoF pose

One labeled still captured at every recorded end-effector pose along a Franka Panda trajectory pressing one of 13 household objects into the gel. The arc-shaped imprints are tactile signatures of objects (here, plier handles, clamps, knives, etc.). FoTA used both markered and markerless gels — use the markered column to filter.

Random sample (mixed):

Per gel variant:

markerless (66% of the data)	markered (34% of the data)

threedcal  ·  36,270 frames  ·  markerless  ·  +xyz pose

A motorised sphere indenter is pressed into the gel at 1,209 different (x, y) positions at a fixed 3 mm depth. The bright spot moves as the probe walks across the sensor surface — useful for learning a calibrated position→appearance mapping.

feats  ·  16,711 frames  ·  markered  ·  +force

Six indenter shapes (sphere, cuboid, cylinder, pyramid, cross, plus one "unknown" set of held-out probes), each pressed into a markered (dotted) GelSight Mini gel. Provides f_x, f_y, f_z forces and a 32×24 depth grid per image. Forces span from ~0 N (light touch) to −73 N (heavy normal compression).

⚠️ FEATS has two physical gel variants. A new column gel_variant distinguishes them:

• "black_dot" — the standard dotted Mini gel used for train, val, test, test_unknown_indenters, test_diff_sensor_old_gel.
• "different" — a second Mini sensor with a different gel (smaller / dimmer / differently-coloured markers) used only in the test_diff_sensor_new_gel split. This split was designed by the FEATS authors to test cross-gel generalisation. Visually, marker dots are less prominent here.

No-contact frames removed. The upstream FEATS dataset included ~5,300 frames where the indenter was hovering off the gel (|f_z| < 0.5 N). These have been filtered out here to give a cleaner pretraining set. Original counts (with the empty frames) were 22,013; current counts are 16,711.

See assets/samples_feats_by_split.png for one example per (split, indenter) pair.

fota_unlabeled  ·  66,761 frames  ·  train-only

Visually identical to fota_labeled — same sensor, same objects, same captures — except these are the dense video frames between the labelled stills, with object name only (no pose). This is the bulk of the data and the primary target for self-supervised pretraining.

gelslam  ·  60,982 frames  ·  markerless  ·  +per-episode 6DoF

The GelSLAM dataset of Huang et al. 2025 — markerless GelSight Mini videos of an object being pressed into the gel and slid around. Two splits:

• train (27,763): the tracking dataset — 140 short episodes across 20 objects, each ~21 s long. Suitable for SLAM, pose-from-touch, and continuous pretraining.
• recon (33,219): the reconstruction dataset — 15 longer videos (1–30 min) of tactile scans across 15 objects (food, rocks, tool handles).

Per-frame frame_idx and per-row episode columns are populated. After empty-frame and dedupe filtering, kept 21% of raw frames (the majority of GelSLAM frames are pre/post-contact approach motion, which the validity filter removes).

tactile_tracking  ·  1,143 frames  ·  markerless  ·  +per-trial 6DoF

The TactileTracking benchmark from the NormalFlow paper (Huang et al. 2024) — 84 tracking trials across 12 objects. Aggressively filtered (empty-frame + perceptual-hash dedupe) because the trials are short and most consecutive frames are near-identical. Kept rate after filtering: 15%.

real_tactile_mnist  ·  153,600 frames  ·  markerless  ·  +digit class

From the Real Tactile MNIST benchmark (Schneider et al. 2025) — 600 3D-printed MNIST digits, each touched 256 times by a robot arm, giving 153,600 touches in total. The upstream release ships each touch as a short video clip; here we keep one middle-frame per touch video (near peak contact). The digit_class column gives the digit 0–9; the episode column gives the print id (which of the 600 physical digits was touched).

feelanyforce  ·  50,997 frames  ·  markerless  ·  +per-indentation object

The FeelAnyForce dataset (Sharei et al. 2024) of robotically-controlled indentations against 42 unique objects (cylinders, cubes, spheres, fruits, household items). Aggressively dedupe-filtered (50% of raw frames dropped) because each indentation is a slow press-and-hold, so adjacent frames are visually near-identical. The schema retains the obj_name and episode columns; the upstream force labels (in TacForce_train_set.csv etc.) are not currently joined in — see the upstream release for forces.

Useful statistics

fota_labeled — pose coverage

• 13 distinct contact objects · 5 initial poses · 15,148 unique (object, pose, side) captures
• Frames split equally between left / right gripper finger: 14,747 / 14,747
• End-effector range: x ∈ [−25.7, 129.1] mm · y ∈ [−137.3, 137.3] mm · z ∈ [−38.6, 38.6] mm
• Top objects by frame count:

object	frames
tapemeasure	4,800
whiteclamped	3,600
blackclamp	3,016
blackclampclosed	2,772
key1	2,400
plierhandle	2,400
foldingknife	2,156
wrench	2,134
blackclampedclosed	1,922
printedstrawberry	1,800

threedcal — calibration grid

• All 36,270 frames are sphere indentations at z = 3 mm penetration
• x ∈ [0, 19] mm · y ∈ [0, 15] mm · 1,209 unique (x,y) grid positions
• ~30 repeated frames per position (lighting noise / repeat trials), useful for variance estimation

feats — indenter & force coverage

Shape	Sizes (mm)	Total frames
cuboid	2, 7, 10, 12, 15, 20	4,683
sphere	10, 15, 20	2,883
cylinder	8, 10	1,291
cross	15	894
pyramid	10	854
(unannotated)	—	11,408

• Normal force range: f_z ∈ [−73.3, 0.0] N (mean −9.30 N, std 10.66 N)
• Shear force range: f_x ∈ [−4.86, 4.86] N, f_y ∈ [−5.89, 5.87] N
• Six labeled splits including out-of-distribution tests:
  • train (15,670), val (921), test (1,845) — in-distribution
  • test_diff_sensor_new_gel (395) — new physical gel
  • test_diff_sensor_old_gel (393) — different sensor unit
  • test_unknown_indenters (2,789) — held-out probe shapes

fota_unlabeled — characterisation

• 11 distinct objects (subset of fota_labeled's 13)
• 60 captures (11 objects × ~5 init-poses × 2 gripper sides)
• Roughly balanced left / right (~50/50)
• ~1000 frames per capture after stride-subsampling³

Unified schema

Every row, regardless of source, has the same columns. Optional fields are null when not applicable.

Column	Type	Description
image	image (binary)	tactile RGB frame, JPEG bytes; auto-decoded by datasets to PIL.Image
image_format	string	always "jpeg"
source	string	one of "fota_labeled" / "fota_unlabeled" / "3dcal" / "feats"
markered	bool	does the gel have tracking dots?
capture	string	per-source capture / scene / probe identifier
split	string	dataset split (e.g. train, val, test_unknown_indenters, …)
height, width	int32	image dimensions in pixels
obj_name	string	(FoTA) which object was pressed
init_pose	int32	(FoTA) initial-pose index for this capture
side	string	(FoTA) "left" or "right" gripper finger
x_mm, y_mm, z_mm	float32	probe / end-effector position
quat_x..w	float32	(FoTA labeled) end-effector orientation
indenter	string	(FEATS) probe shape, e.g. "sphere"
indenter_param	string	(FEATS) probe size in mm
f_x, f_y, f_z	float32	(FEATS) total force on probe (N)
grid_z_max, grid_z_mean	float32	(FEATS) summary of per-pixel depth grid
gel_variant	string	(FEATS only) "black_dot" (standard markered gel) or "different" (the second sensor / different gel used in test_diff_sensor_new_gel)
domain	string	"real" for real-world captures or "sim" for Taxim-rendered Mini imagery

Real vs simulated data

Every row carries a domain column. The 758 K real-world frames span 8 upstream datasets capturing physical robot–object contact on a real GelSight Mini sensor; the 400 K simulated frames come from the Mini-calibrated Taxim renderer of Schneider et al.

# Real-only markerless pool (largest pretraining target)
real_markerless = concatenate_datasets([
    load_dataset("yxma/gelsight-mini-pretrain", c, split="train"
                ).filter(lambda r: r["domain"] == "real" and not r["markered"])
    for c in ["fota_unlabeled", "threedcal", "gelslam", "tactile_tracking",
              "real_tactile_mnist", "feelanyforce"]
])

# Sim pool — useful for sim-to-real transfer or as augmentation
sim_pool = concatenate_datasets([
    load_dataset("yxma/gelsight-mini-pretrain", c, split="train")
    for c in ["sim_tactile_mnist", "sim_starstruck"]
])

The two sim subsets are from the same authors as real_tactile_mnist (Schneider et al. 2025) and use the same Mini sensor model.

Recommended uses

• Self-supervised pretraining (VAE / MAE / SimCLR / DINO): use fota_unlabeled + fota_labeled + threedcal (markerless, ~133K frames). Hold feats out since its dotted-gel appearance will dominate the reconstruction objective if mixed in.
• Force / shear regression: fine-tune on feats with its f_{x,y,z} labels.
• Pose estimation: fine-tune on fota_labeled (xyz + quat) or threedcal (xy + depth).
• Marker invariance studies: train markerless ↔ test on feats, or use a marker-mask augmentation.

Provenance and citation

This dataset is a re-packaging of existing public datasets — please cite the upstream sources if you use the data:

• FoTA — Foundation Tactile, Zhao et al., 2024. HF dataset, GitHub, arXiv:2406.13640 · MIT License.
• 3D Cal (py3DCal) — Kota, Shah, Colgate, Reardon (2025). Zenodo 18462608 · CC-BY-4.0.
• FEATS — Helmut (2025). HF dataset · MIT License.
• GelSLAM — Huang et al., 2025. HF dataset, GitHub, arXiv:2508.15990 · MIT License.
• TactileTracking / NormalFlow — Huang, Kaess, Yuan (2024). HF dataset, GitHub, IEEE RA-L 2024 · MIT License.
• Real Tactile MNIST — Schneider et al., 2025. HF dataset family, GitHub, arXiv:2506.06361 · CC-BY-2.0.
• FeelAnyForce — Sharei et al., 2024. HF dataset · CC-BY-4.0.
• Sim Tactile MNIST / Sim Starstruck — Schneider et al., 2025 (same authors as Real Tactile MNIST). tactile-mnist-touch-syn-single-t32-320x240, tactile-mnist-touch-starstruck-syn-single-t32-320x240. Mini-calibrated Taxim renderer. CC-BY-2.0.
• Taxim simulator (used by the above sim sources) — Si & Yuan, 2022. GitHub, arXiv:2109.04027.

Conversion details:

• All images are re-encoded to JPEG at quality 92. Original PNGs in threedcal are decoded losslessly and re-encoded; FEATS .npy dicts have their gs_img array extracted and saved as JPEG; FoTA WebDataset shards are unpacked and re-encoded. Metadata is preserved verbatim.
• FeelAnyForce (Sharei et al., 2024) ~200K markerless frames may be added in a future revision pending license verification with the upstream authors.

License

This aggregated release is released under CC-BY-4.0 (the most restrictive license among the included sources). Cite the component datasets above.

Acknowledgments

Thanks to the FoTA, py3DCal, and FEATS authors for releasing their data publicly.