yxma/gelsight-mini-pretrain

GelSight Mini Pretrain

~853K GelSight Mini tactile RGB frames, 12 public sources, one parquet schema. Built for self-supervised representation learning (VAE / MAE / SimCLR / DINO) — every frame contact-filtered, channel-normalized, and re-encoded as JPEG q92.

	Frames	Sources
Real	536K	FoTA (labeled+unlabeled), 3DCal, FEATS, GelSLAM, TactileTracking, RTM, FeelAnyForce, UniT, TacQuad
Sim	317K	sim_tactile_mnist, sim_starstruck (Taxim-rendered, Mini-calibrated)
NC extension (repo)	+66K	Sparsh (CC-BY-NC)

Quick start

from datasets import load_dataset, concatenate_datasets

# Single source
ds = load_dataset("yxma/gelsight-mini-pretrain", "fota_unlabeled", split="train")
img = ds[0]["image"]                  # PIL.Image (auto-decoded from JPEG bytes)

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

Composition

Subset	Frames	Splits	Gel	Labels
fota_unlabeled	66,761	train	mixed¹	object name
gelslam	114,019	train + recon	markerless	episode + object
feelanyforce	48,197	train	markerless	42 unique objects
real_tactile_mnist	30,956	train + test	markerless	digit + print id
fota_labeled	26,394	train + val	mixed¹	6-DoF pose + object
feats	16,969	6-split OOD bench	markered	indenter + 3-axis force
tacquad	12,195	indoor/outdoor/fine	markerless	181 objects
threedcal	6,924	train	markerless	(x, y) sphere grid
tactile_tracking	2,408	train	markerless	trial + object
unit	387	train	markered	3D-pose target
sim_starstruck	166,104	train + test	markerless	episode (sim)
sim_tactile_mnist	150,601	train + test	markerless	digit + episode (sim)

¹ FoTA mixes markered + markerless gels per finger; the per-row markered column was auto-detected from dot density and is correct.

Pipeline (applied to every source)

1. Unified contact filter — area ≥ 40 px ∧ intensity ≥ I_min on the central-50% greyscale diff vs per-source baseline (I_min = 12 real, 10 sim); 1.5 % background-diversity keep rate.
2. Channel-order normalization — Mini's at-rest gel has B > R (3 colored LEDs); subsets where the upstream stored BGR are auto-detected (per-image R-B sign) and swapped to RGB. After this, every frame is guaranteed RGB.
3. JPEG q=92 re-encode + chunked-binary parquet writes (handles >2 GB shards safely).
4. Object diversity preserved — ~8,500 unique object instances across 13 physical sensor configurations.

Schema (30 columns, every row identical)

image (JPEG bytes), source, domain (real/sim), markered (bool), gel_variant (markered/markerless), capture, split, height, width, obj_name, episode, frame_idx, pose fields (x_mm, y_mm, z_mm, quat_*), FEATS fields (indenter, indenter_param, f_x, f_y, f_z, grid_z_*), digit_class, etc. — all optional fields are null when not applicable.

For per-subset details (paper, license, processing recipe, sample grids, stats), see SOURCES.md.

Sample images

fota_labeled (markerless)	fota_labeled (markered)
gelslam	feats (markered + force)
real_tactile_mnist	tacquad (181 household objects)
sim_tactile_mnist	sim_starstruck

Recommended uses

• Self-supervised pretraining (VAE / MAE / SimCLR / DINO) — concat all markerless real subsets (~472K frames), then fine-tune.
• Pose / force regression — fine-tune on fota_labeled (6-DoF), threedcal (xy + depth), or feats (3-axis force).
• Sim-to-real transfer — pretrain on sim_*, fine-tune on real.
• Marker-invariance studies — train markerless ↔ test on feats (markered).

Citations

Please cite both this aggregation and the upstream sources you use:

• FoTA (HF, arXiv:2406.13640) · MIT
• py3DCal (Zenodo) · CC-BY-4.0
• FEATS (HF) · MIT
• GelSLAM (HF, arXiv:2508.15990) · MIT
• TactileTracking / NormalFlow (HF, RA-L 2024) · MIT
• Real Tactile MNIST (HF family, arXiv:2506.06361) · CC-BY-2.0
• FeelAnyForce (HF) · CC-BY-4.0
• UniT (GitHub) · BSD-3-Clause-style
• TacQuad / AnyTouch (HF) · CC-BY-4.0
• Taxim (sim renderer, GitHub, arXiv:2109.04027)

Investigated but not included

Touch-and-Go, TVL (Touch-Vision-Language), facebook/gelsight-force-estimation, YCB-Sight, TACTO/MidasTouch/DiffTactile — see SOURCES.md for reasons (wrong sensor, license, or not Mini-calibrated).

License

CC-BY-4.0 for this aggregation. Cite the component datasets above. The companion yxma/gelsight-mini-pretrain-nc repo adds Sparsh (CC-BY-NC-4.0) for non-commercial use.