We release two tactile perception datasets collected using the GelSight Mini (without markers) sensor:
- Tactile-based Long-Horizon Object Tracking Dataset
- Tactile-based Object Reconstruction Dataset
For real-time long-horizon 6DoF object pose estimation and high-fidelity object reconstruction using only touch, please refer to our work [GelSLAM](https://github.com/rpl-cmu/gelslam). For short-horizon object tracking using only touch, see [NormalFlow](https://github.com/rpl-cmu/normalflow) and the corresponding [NormalFlow Dataset](https://huggingface.co/datasets/joehjhuang/TactileTracking).
---
## Long-Horizon Object Tracking Dataset
This benchmark dataset is designed to evaluate the performance of tactile-based 6DoF object tracking algorithms over long horizons.
### Collection Setup
The dataset includes 20 objects and 140 tracking episodes, with 7 episodes per object corresponding to seven distinct initial contact locations. The 20 objects consist of 14 everyday objects, 3 small textured objects, and 3 geometric shapes. The objects and their associated initial contact locations are shown below.
In each episode, the tactile sensor moves against a fixed object on the workbench under continuous contact, without contact breaks. A motion capture system records the sensor pose during contact, providing precise 6DoF ground-truth poses for every tactile frame. The data collection setup is illustrated in the figure below.
### Dataset Structure
The tracking dataset is located in `dataset/tracking_dataset/`. Each episode directory contains:
- **gelsight.avi**: Tactile video with N frames.
- **true_start_T_currs.npy**: An (N, 4, 4) array representing the sensor’s 6DoF pose for each tactile frame in `gelsight.avi`, formatted as homogeneous transformation matrices.
- **contact_masks.npy**: An (N, H, W) array of contact masks computed from the tactile images in `gelsight.avi`.
- **gradient_maps.npy**: An (N, H, W, 2) array of gradient maps computed from the tactile images in `gelsight.avi`.
- **background.png**: Reference tactile image without contact.
Contact masks and gradient maps are generated from the tactile video using `gs_sdk`, with the calibration file located in `dataset/gelsight_calibrations/gelsight3`. All episodes share the same calibration file.
### Dataset Statistics
On average, each episode lasts approximately 21 seconds and contains 523 frames. The dataset exhibits substantial accumulated 6DoF motion per episode, enabling rigorous evaluation of long-horizon tracking performance. The average accumulated 6DoF motion for all episodes is shown below.
---
## Object Reconstruction Dataset
This dataset is designed to qualitatively evaluate the performance of tactile-based object reconstruction algorithms.
---
### Data Collection Setup
The dataset contains 15 objects, including 3 tool handles, 7 food items, 4 rocks and fossils, and 1 textured object. Ground-truth meshes are not provided. Data is collected in an in-the-wild setup shown below where both the object and the GelSight Mini sensor are manually held during scanning. Contact breaks and re-initializations are frequent and may exceed 100 occurrences for certain objects. Scanning trajectories are guided by visualing the real-time reconstruction results using [GelSLAM](https://github.com/rpl-cmu/gelslam), allowing the operator to adaptively explore the object surface for full coverage.
### Dataset Structure
The reconstruction dataset is located in `dataset/reconstruction_dataset/`. For each object, the dataset contains a single tactile scanning episode consisting of:
- **gelsight.avi**: Full tactile video.
- **config.yaml**: Configuration file specifying sensor calibration and device information.
The calibration file path and device information are specified under `device_config` in `config.yaml`. To evaluate your own reconstruction algorithm, only the tactile video, calibration file, and device information are required. Reconstruction results obtained using GelSLAM are provided in `dataset/gelslam_reconstruction_results/`. You may run the [GelSLAM](https://github.com/rpl-cmu/gelslam) code to reproduce these results. They are provided for qualitative comparison with your own method.
### Dataset Statistics
The smallest object in the dataset is the Seed (8 × 8 × 8 mm), while the largest is the Avocado (85 × 61 × 58 mm). The tactile video duration in our dataset range from 1 to 30 minutes. Detailed statistics for each object, along with object images and the corresponding GelSLAM reconstruction results, are shown in the figure below.
---
## Cite Us
If you find this dataset useful, please consider citing our arXiv paper:
```
@misc{huang2025gelslam,
author={Hung-Jui Huang and Mohammad Amin Mirzaee and Michael Kaess and Wenzhen Yuan},
title={GelSLAM: A Real-time, High-Fidelity, and Robust 3D Tactile SLAM System},
year={2025},
eprint={2508.15990},
archivePrefix={arXiv},
primaryClass={cs.RO},
url={https://arxiv.org/abs/2508.15990},
}
```
