add README

README.md

@@ -0,0 +1,146 @@
+---
+license: cc-by-4.0
+task_categories:
+  - image-classification
+  - feature-extraction
+tags:
+  - tactile
+  - gelsight
+  - gelsight-mini
+  - robotics
+  - tactile-sensing
+  - pretraining
+size_categories:
+  - 100K<n<1M
+configs:
+  - config_name: fota_labeled
+    data_files:
+      - split: train
+        path: fota_labeled/train-*.parquet
+      - split: val
+        path: fota_labeled/val-*.parquet
+  - config_name: fota_unlabeled
+    data_files:
+      - split: train
+        path: fota_unlabeled/train-*.parquet
+      - split: val
+        path: fota_unlabeled/val-*.parquet
+  - config_name: threedcal
+    data_files:
+      - split: train
+        path: threedcal/train-*.parquet
+  - config_name: feats
+    data_files:
+      - split: train
+        path: feats/train-*.parquet
+      - split: val
+        path: feats/val-*.parquet
+      - split: test
+        path: feats/test-*.parquet
+      - split: test_diff_sensor_new_gel
+        path: feats/test_diff_sensor_new_gel-*.parquet
+      - split: test_diff_sensor_old_gel
+        path: feats/test_diff_sensor_old_gel-*.parquet
+      - split: test_unknown_indenters
+        path: feats/test_unknown_indenters-*.parquet
+---
+
+# GelSight Mini Pretrain
+
+A unified collection of public **GelSight Mini** tactile images for self-supervised
+pretraining (VAE, MAE, contrastive, etc.). Aggregates several existing datasets
+into a single parquet-based, HuggingFace-native format with a common schema and
+marker / markerless split.
+
+## Quick start
+
+```python
+from datasets import load_dataset
+
+# Default subset: largest, markerless, unlabeled video frames
+ds = load_dataset("yxma/gelsight-mini-pretrain", "fota_unlabeled", split="train")
+
+# Or fetch a specific subset:
+ds = load_dataset("yxma/gelsight-mini-pretrain", "fota_labeled", split="train")
+ds = load_dataset("yxma/gelsight-mini-pretrain", "threedcal", split="train")
+ds = load_dataset("yxma/gelsight-mini-pretrain", "feats", split="train")
+
+print(ds[0]["image"])  # PIL.Image
+print({k: ds[0][k] for k in ds.features if k != "image"})
+```
+
+## What's inside
+
+All images are 240×320 or 320×240 RGB JPEG (re-encoded at quality 92 for
+storage efficiency), captured with the **GelSight Mini** tactile sensor.
+
+| Config | Source | Frames | Gel | Has labels | Notes |
+|---|---|---:|---|---|---|
+| `fota_labeled`     | Foundation Tactile (FoTA) — panda_warped subset, sampled stills paired with end-effector pose | 29,494  | markerless | end-effector x,y,z + quat | One labeled still per pose tick along a robot trajectory. |
+| `fota_unlabeled`   | FoTA — continuous video frames from the same captures | 516,523 | markerless | obj_name only | Bulk of the data; intended for self-supervised pretraining. |
+| `threedcal`        | py3DCal (Kota, Shah, Colgate, Reardon 2025) — sphere indentation grid for calibration | 36,270  | markerless | x_mm, y_mm, penetration_depth_mm | Indenter probe at varying (x, y, z). PNG images, 320×240. |
+| `feats`            | FEATS (Helmut 2025) — sphere / cuboid / cylinder / etc. indentation with force | 22,014  | **markered** (dotted gel) | indenter shape + size, f_x, f_y, f_z | Marker-grid gel; useful for force/shear learning. Multiple test splits include unseen indenters and a second sensor. |
+
+*A `feelanyforce` config containing ~200K additional markerless frames may be added later, pending license verification with the upstream authors.*
+
+**Combined markerless: 582,287 frames** &nbsp; · &nbsp; **markered: 22,014 frames**
+
+## Schema
+
+Every row, regardless of source, has the same columns. Optional fields are `null`
+where not applicable to a given source.
+
+| Column           | Type    | Description |
+|---|---|---|
+| `image`          | image   | The tactile RGB frame (JPEG bytes; auto-decoded by `datasets`) |
+| `image_format`   | string  | Always `"jpeg"` |
+| `source`         | string  | `"fota_labeled"` / `"fota_unlabeled"` / `"3dcal"` / `"feats"` |
+| `markered`       | bool    | `True` if the gel has tracking dots |
+| `capture`        | string  | Capture/scene/probe identifier from the source dataset |
+| `split`          | string  | `train` / `val` / `test` / `test_diff_sensor_new_gel` / etc. |
+| `height`, `width`| int32   | Image dimensions in pixels |
+| `obj_name`       | string  | (FoTA) object pressed against the gel |
+| `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 | (FoTA labeled, 3DCal) 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 (mm) |
+| `f_x`, `f_y`, `f_z` | float32 | (FEATS) total force on probe |
+| `grid_z_max`, `grid_z_mean` | float32 | (FEATS) per-pixel depth grid statistics |
+
+## Recommended use
+
+- **Self-supervised pretraining (VAE / MAE / SimCLR / DINO):** train on
+  `fota_unlabeled + fota_labeled + threedcal` (markerless, ~582K frames).
+  Keep `feats` separate since its dotted-gel appearance is very different.
+- **Force / shear regression:** fine-tune on `feats` with its `f_x/y/z` labels.
+- **Pose estimation:** fine-tune on `fota_labeled` with its `xyz + quat` labels.
+- **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. If you use it, please
+cite the upstream sources:
+
+- **FoTA** — *Foundation Tactile*, Zhao et al. 2024. [Dataset on
+  HF](https://huggingface.co/datasets/alanz-mit/FoundationTactile), [GitHub](https://github.com/alanzjl/t3), [arXiv:2406.13640](https://arxiv.org/abs/2406.13640).
+  MIT License.
+- **3D Cal (py3DCal)** — Kota, Shah, Colgate, Reardon (2025). Zenodo record
+  [18462608](https://zenodo.org/records/18462608). CC-BY-4.0.
+- **FEATS** — Helmut (2025). [Dataset on
+  HF](https://huggingface.co/datasets/erikhelmut/FEATS). MIT License.
+
+All conversions performed here are lossy only by JPEG re-encoding at quality 92.
+Pose / force / metadata are preserved verbatim.
+
+## License
+
+This aggregated release is **CC-BY-4.0** (the most restrictive license among the
+included sources). Component datasets retain their original licenses for
+non-image content.
+
+## Acknowledgments
+
+Thanks to the FoTA, py3DCal, and FEATS authors for releasing their data publicly.