Initial upload of synthetic grasp dataset
Browse files
README.md
ADDED
|
@@ -0,0 +1,50 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language: en
|
| 3 |
+
license: mit
|
| 4 |
+
task_categories:
|
| 5 |
+
- robotics
|
| 6 |
+
- computer-vision
|
| 7 |
+
tags:
|
| 8 |
+
- 3d
|
| 9 |
+
- grasp-synthesis
|
| 10 |
+
- tactile-sensing
|
| 11 |
+
- objaverse
|
| 12 |
+
pretty_name: Synthetic Grasp Dataset (Curated)
|
| 13 |
+
---
|
| 14 |
+
|
| 15 |
+
# 🖐️ Synthetic Grasp Dataset (Objaverse-LVIS Curated)
|
| 16 |
+
|
| 17 |
+
This dataset contains high-quality synthetic grasp data generated for robotic manipulation research.
|
| 18 |
+
It focuses on the fusion of **vision** and **tactile** sensing by providing visibility and occlusion analysis for each contact point.
|
| 19 |
+
|
| 20 |
+
## 📊 Dataset Statistics
|
| 21 |
+
- **Number of objects:** 10
|
| 22 |
+
- **Source:** Curated objects from [Objaverse-LVIS](https://objaverse.allenai.org/) (categories: cup, bottle, hammer, screwdriver, wrench)
|
| 23 |
+
- **Grasp Strategies:** front_back
|
| 24 |
+
- **Camera Resolution:** 640x480
|
| 25 |
+
|
| 26 |
+
## 🛠️ Data Format
|
| 27 |
+
|
| 28 |
+
Each object folder contains:
|
| 29 |
+
- `rgb.png`: Monocular RGB render.
|
| 30 |
+
- `grasp_<strategy>.json`: Contact points with position, normal, tangent, and visibility status.
|
| 31 |
+
- `grasp_<strategy>.npz`: NumPy version of the contact points.
|
| 32 |
+
- `grasp_<strategy>_overlay.png`: Visual overlay of the grasp on the object.
|
| 33 |
+
- `metadata.json`: Object-specific metadata (surface visibility, bounding box, complexity).
|
| 34 |
+
|
| 35 |
+
## 🔍 Visibility Classification
|
| 36 |
+
Every contact point is classified based on camera occlusion:
|
| 37 |
+
- **VISIBLE**: Point is directly seen by the camera.
|
| 38 |
+
- **SILHOUETTE**: Point is on the visual horizon (critical for tactile تکمیل).
|
| 39 |
+
- **OCCLUDED**: Point is hidden by the object itself (back side or self-occlusion).
|
| 40 |
+
|
| 41 |
+
## 📜 How to use
|
| 42 |
+
This dataset is designed to train models that predict contact stability from visual data or to simulate-to-real transfer for tactile controllers.
|
| 43 |
+
|
| 44 |
+
```python
|
| 45 |
+
from huggingface_hub import snapshot_download
|
| 46 |
+
path = snapshot_download("jack635/grasp-dataset-curated", repo_type="dataset")
|
| 47 |
+
```
|
| 48 |
+
|
| 49 |
+
---
|
| 50 |
+
*Generated using the [Grasp Dataset Generator](https://github.com/635jack/grasp-dataset-gen) pipeline.*
|