README.md

---
license: mit
language:
- en
task_categories:
- zero-shot-image-classification
- zero-shot-classification
- feature-extraction
- image-feature-extraction
- tabular-classification
- tabular-regression
- depth-estimation
tags:
- tactile
- robotics
pretty_name: Sensor-Invariant Tactile Representation
size_categories:
- 1M<n<10M
viewer: false
---
# SITR Dataset & Weights

This repository hosts both the dataset and pre-trained model weights for the Sensor-Invariant Tactile Representation (SITR) paper. The dataset supports training and evaluating models for sensor-invariant tactile representations across simulated and real-world settings, while the pre-trained weights enable immediate deployment and fine-tuning for various tactile perception tasks.

The codebase implementing SITR is available on GitHub: [SITR Codebase](https://github.com/hgupt3/gsrl)

For more details on the underlying methods and experiments, please visit our [project website](https://hgupt3.github.io/sitr/) and read the [arXiv paper](https://arxiv.org/abs/2502.19638).

---

## Pre-trained Model Weights

The pre-trained model weights are available for immediate use in inference or fine-tuning. These weights were trained on our large-scale simulated dataset and have been validated across multiple real-world sensors.

### Downloading the Weights

```bash
wget https://huggingface.co/datasets/hgupt3/sitr_dataset/resolve/main/checkpoints.zip
unzip checkpoints.zip -d your_desired_directory
```

### Weights Directory Structure

The weights directory contains the following structure:

```
checkpoints/
├── SITR_B18.pth              # Base pre-trained model weights (371MB)
├── classification/           # Classification task weights
│   └── SITR_base/            # Base model with fine-tuned head for classification on 1 sensor
│       ├── sensor_0000.pth   # Weights for sensor 0
│       ├── sensor_0001.pth   # Weights for sensor 1
│       └── ...
└── pose_estimation/          # Pose estimation task weights
    └── SITR_base/            # Base model with fine-tuned head for classification on 1 sensor
        ├── sensor_0000.pth   # Weights for sensor 0
        ├── sensor_0001.pth   # Weights for sensor 1
        └── ...
```

You can use the SITR_B18.pth weight for:
1. **Zero-shot inference** on new tactile data
2. **Fine-tuning** for specific tasks
3. **Feature extraction** for downstream applications

For detailed usage instructions and examples, please refer to the [SITR Codebase](https://github.com/hgupt3/gsrl).

---

## Dataset Overview

The SITR dataset consists of three main parts:

1. **Simulated Tactile Dataset**  
   A large-scale synthetic dataset generated using physics-based rendering (PBR) in Blender. This dataset spans 100 unique simulated sensor configurations with tactile signals, calibration images, and corresponding surface normal maps. It includes 10K unique contact configurations generated using 50 high-resolution 3D meshes of common household objects, resulting in a pre-training dataset of 1M samples.

2. **Classification Tactile Dataset**  
   Data collected from 7 real sensors (including variations of GelSight Mini, GelSight Hex, GelSight Wedge, and DIGIT). For the classification task, 20 objects are pressed against each sensor at various poses and depths, accumulating 1K tactile images per object (140K images in total, with 20K per sensor). We used 16 objects for our classification experiments, as some items were deemed unsuitable (this was decided before experimentation). The dataset is provided as separate train (80%) and test sets (20%).

3. **Pose Estimation Tactile Dataset**  
   For pose estimation, tactile signals are recorded using a modified Ender-3 Pro 3D printer equipped with 3D-printed indenters. This setup provides accurate ground truth (x, y, z coordinates) for contact points, where all coordinates are specified in millimeters. Data were collected for 6 indenters across 4 sensors, resulting in 1K samples per indenter (24K images in total, 6K per sensor). This dataset is also organized into train (80%) and test sets (20%).

---

## Download and Setup

### Simulated Tactile Dataset

The simulated dataset is split into two parts due to its size:

- `renders_part_aa.zip`
- `renders_part_ab.zip`

Download both files using:

```bash
wget https://huggingface.co/datasets/hgupt3/sitr_dataset/resolve/main/renders_part_aa
wget https://huggingface.co/datasets/hgupt3/sitr_dataset/resolve/main/renders_part_ab
```

**To merge and unzip:**

1. **Merge the parts into a single zip file:**

```bash
cat renders_part_aa renders_part_ab > renders.zip
rm renders_part_aa renders_part_ab  # Remove the split files
```

2. **Unzip the merged file:**

```bash
unzip renders.zip -d your_desired_directory
rm renders.zip
```

### Real-World Datasets (Classification & Pose Estimation)

Download the classification dataset:

```bash
wget https://huggingface.co/datasets/hgupt3/sitr_dataset/resolve/main/classification_dataset.zip
unzip classification_dataset.zip -d your_desired_directory
rm classification_dataset.zip
```

Download the pose estimation dataset:

```bash
wget https://huggingface.co/datasets/hgupt3/sitr_dataset/resolve/main/pose_dataset.zip
unzip pose_dataset.zip -d your_desired_directory
rm pose_dataset.zip
```

Each dataset contains:
- `train_set/` (80% of the data)
- `test_set/` (20% of the data)

---

## File Structure

### 1. Simulated Tactile Dataset

```
data_root/
├── sensor_0000/
│   ├── calibration/          # Calibration images
│   │   ├── 0000.png          # Background image
│   │   ├── 0001.png
│   │   └── ...
│   ├── samples/              # Tactile sample images
│   │   ├── 0000.png
│   │   ├── 0001.png
│   │   └── ...
│   ├── dmaps/                # (Optional) Depth maps
│   │   ├── 0000.npy
│   │   └── ...
│   └── norms/                # (Optional) Surface normals
│       ├── 0000.npy
│       └── ...
├── sensor_0001/
└── ...
```

### 2. Classification Dataset

Each of the `train_set/` and `test_set/` directories follows this structure:

```
train_set/  (or test_set/)
├── sensor_0000/
│   ├── calibration/          # Calibration images
│   ├── samples/              # Organized by class
│   │   ├── class_0000/
│   │   │   ├── 0000.png
│   │   │   └── ...
│   │   ├── class_0001/
│   │   │   ├── 0000.png
│   │   │   └── ...
│   │   └── ...
├── sensor_0001/
└── ...
```

### 3. Pose Estimation Dataset

Each of the `train_set/` and `test_set/` directories is structured as follows:

```
train_set/  (or test_set/)
├── sensor_0000/
│   ├── calibration/          # Calibration images
│   ├── samples/              # Tactile sample images
│   │   ├── 0000.png
│   │   ├── 0001.png
│   │   └── ...
│   └── locations/            # Pose/Location data
│       ├── 0000.npy
│       ├── 0001.npy
│       └── ...
├── sensor_0001/
└── ...
```

---

## Citation

If you use this dataset or model weights in your research, please cite:

```bibtex
@inproceedings{
    gupta2025sensorinvariant,
    title={Sensor-Invariant Tactile Representation},
    author={Harsh Gupta and Yuchen Mo and Shengmiao Jin and Wenzhen Yuan},
    booktitle={The Thirteenth International Conference on Learning Representations},
    year={2025},
}
```

---

## License

This dataset and model weights are licensed under the MIT License. See the LICENSE file for details.

If you have any questions or need further clarification, please feel free to reach out.