Gelsight-Welds

Paired tactile (GelSight Mini) and top-down RGB (Intel RealSense D555) captures of weld beads on a flat sample plate. Each sample (session) is scanned with a 10×10 grid of GelSight touches; the D555 takes one overhead shot per session, which we crop into per-tile RGB images aligned with each touch.

Loading

from datasets import load_dataset

ds = load_dataset("Gelsight-Welds", split="all")  # uses config "v2" by default; no train/val/test split is imposed
print(ds[0]["rgb_image"].size, ds[0]["gelsight_image"].size)

Versions

• v1 — Initial release, 8 sessions / 800 rows. Crop bounds tuned by eye against gelsight_stitched.png; build parameters snapshotted at data/v1/build_config.json. Single shard data/v1/all-00000-of-00001.parquet.
• v2 — Superset of v1 plus a new collection batch (default config). Same camera / crop parameters as v1; sessions are sharded one parquet per session at data/v2/<session_id>.parquet so adding a batch is append-only. data/v2/build_config.json records the per-version crop params and the session list under each source version.

Per-session previews (v2)

GelSight stitched panorama (left) and D555 overhead with crop-grid overlay (right), one composite per session. The block below is auto-generated from previews/v2/*.png by scripts/v2/compile.py; do not hand-edit between the sentinels.

2026-06-12_16-42-16

2026-06-12_16-52-55

2026-06-12_17-03-10

2026-06-12_17-14-29

2026-06-12_17-28-42

2026-06-12_17-38-46

2026-06-12_17-51-05

2026-06-12_18-02-13

2026-06-16_14-45-30

2026-06-16_14-55-09

2026-06-16_15-22-43

2026-06-16_15-32-14

2026-06-16_15-42-43

2026-06-16_15-53-55

2026-06-16_16-03-36

2026-06-16_16-13-21

2026-06-16_16-23-54

2026-06-16_16-33-55

2026-06-16_16-49-04

2026-06-16_17-03-24

2026-06-16_17-16-20

2026-06-16_17-27-07

2026-06-16_17-37-34

2026-06-16_17-48-31

2026-06-16_17-59-22

2026-06-16_18-10-38

2026-06-16_18-20-47

2026-06-16_18-31-11

Row schema (configs v1, v2 — identical)

Column	Type	Notes
session_id	string	timestamp of the collection session, e.g. 2026-06-12_16-42-16
capture_index	int32	1..100 within the session
tile_row	int32	0..9, row in the 10×10 grid (matches stitch.py)
tile_col	int32	0..9, col in the 10×10 grid
robot_x_m	float32	plate-frame x of this tile's center, meters
robot_y_m	float32	plate-frame y of this tile's center, meters
rgb_image	Image	per-tile crop from the D555 overhead, 44×33 landscape
gelsight_image	Image	raw GelSight saved at touch, 160×120 landscape
gelsight_depth	Array2D (240, 320) float32	raw Poisson-integration output; multiply by 0.0634 for mm
d555_overview	Image	full D555 overhead capture, 1280×720, duplicated across this session's 100 rows
gelsight_baseline	Image	no-contact gel reference for this session
rgb_stitched	Image	the D555 crop region as a portrait panorama (330×440)
gelsight_stitched	Image	the 10×10 GelSight panorama (1200×1600 portrait)
defective	bool	per-tile defect label; defaults to False, overrides live in source/annotations.csv
notes	string	free-form per-tile annotation; defaults to "", same override path

Hand-edit source/annotations.csv (columns: session_id, capture_index, defective, notes) to relabel tiles and rerun 02_build_dataset.py. Tiles not listed inherit defaults. The file lives in source/ rather than data/<version>/ because defects describe physical capture events that don't change between versions.

Geometry

• Each GelSight tile covers ~10 mm × 7.5 mm of the plate (after the 25% border crop, keep_fraction = 0.5).
• The 10×10 grid spans ~100 mm × 75 mm of the plate.
• The data collection script puts the robot origin (plate frame (0, 0)) at the center of cell (row=5, col=5); the grid is asymmetric, extending 5 cells on the +x / +y side and 4 cells on the −x / −y side.
• gelsight_depth is in raw Poisson-integration units. The gsrobotics-documented lateral scale is PX_TO_MM = 0.0634 at the 320×240 depth resolution.
• D555 depth was found to be unreliable for this setup and is not included.

Build

Scripts to reproduce from source/ live in scripts/<version>/:

# v1 (frozen snapshot of the initial 8 sessions)
uv run python scripts/v1/01_crop_d555.py     # writes rgb_*.png + rgb_stitched.png into source/v1/<session>/
uv run python scripts/v1/02_build_dataset.py # writes data/v1/all-*.parquet and data/v1/build_config.json
uv run python scripts/v1/03_make_previews.py # writes previews/v1/<session>.png for the README

# v2 (default; v1 sessions + new batches under source/v2/)
uv run python scripts/v2/01_crop_d555.py     # crops only source/v2/<session>/ (v1 keeps its v1 crop)
uv run python scripts/v2/02_build_dataset.py # writes data/v2/<session>.parquet per session + build_config.json
uv run python scripts/v2/03_make_previews.py # writes previews/v2/<session>.png for all v1 + v2 sessions

# v2 one-shot: runs the three scripts above then rewrites the README v2 gallery
# from previews/v2/*.png between the <!-- v2-gallery-{start,end} --> sentinels.
uv run python scripts/v2/compile.py