datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
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… See the full description on the dataset page: https://huggingface.co/datasets/yxma/gelsight-mini-pretrain.Gelsight-Welds
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… See the full description on the dataset page: https://huggingface.co/datasets/ppak10/Gelsight-Welds.gelsight-mini-pretrain-nc
GelSight Mini Pretrain · Non-Commercial Extension
⚠️ Non-commercial use only. This repository is licensed
CC-BY-NC-4.0 because it includes upstream sources whose licenses
restrict commercial use. For commercial-friendly Mini tactile data,
see the main yxma/gelsight-mini-pretrain repo
(CC-BY-4.0).
This dataset is the CC-BY-NC extension of yxma/gelsight-mini-pretrain.
It contains only the GelSight Mini sources whose upstream licenses are
not compatible with CC-BY-4.0 aggregation.… See the full description on the dataset page: https://huggingface.co/datasets/yxma/gelsight-mini-pretrain-nc.gelsight-mini-pretrain-video
GelSight Mini Pretrain · Video / Sequence Subset
🎬 Companion to yxma/gelsight-mini-pretrain.
Where the main repo treats every kept frame as an independent image, this repo
preserves temporal sequences — one row per frame, ordered, with explicit
sequence-id + position metadata, for video tactile pretraining.
Why this repo
The main repo's pipeline applies perceptual-hash dedupe within each capture
to drop near-identical adjacent frames. That's great for image-level… See the full description on the dataset page: https://huggingface.co/datasets/yxma/gelsight-mini-pretrain-video.