---
license: cc-by-4.0
pretty_name: FusionX Multimodal Sample Dataset (V3)
tags:
  - robotics
  - tactile
  - multimodal
  - stereo-vision
  - imu
  - manipulation
size_categories:
  - 10K<n<100K
---

# FusionX Multimodal Sample Dataset (V3)

This repository contains a multimodal dataset capturing synchronized stereo vision, RGB, IMU, and tactile glove data across 11 distinct tasks. It is intended for research in multimodal perception, manipulation learning, and tactile-vision fusion.

## Dataset Overview

Each task is captured with the following synchronized modalities:

- **Mono Stereo Vision** — Left and right monochrome camera streams stored as raw `.png` files at **640×400** resolution.
- **RGB** — Color camera stream stored as `.jpg` files at **720p**.
- **IMU** — Inertial measurements from the on-board OAK IMU, stored in `oak_imu.parquet`.
- **Tactile** — Per-frame glove sensor data aligned with vision, stored in `frames.parquet`.
- **Calibration** — Intrinsic and extrinsic camera parameters in `calib.json`.

## Coordinate Systems & Alignment

To facilitate 3D reconstruction and multimodal learning, alignment is defined as follows:

- **Monochrome Stereo** — Left and right monochrome cameras operate in their own independent coordinate frames. Extrinsics between them are provided in `calib.json`.
- **Tactile ↔ Vision** — Per-frame alignment between glove readings and image frames is provided via the index in `frames.parquet`.

**Resolution note:** RGB is 720p while the monochrome streams are 640×400.

## Repository Structure

The dataset is organized by task. Each task folder follows the same layout:

```
Dataset-V3/
├── task1/
│   ├── calib.json
│   ├── frames.parquet
│   ├── oak_imu.parquet
│   ├── mono_left/
│   │   ├── 000001.png
│   │   ├── 000002.png
│   │   └── ...
│   ├── mono_right/
│   │   ├── 000001.png
│   │   ├── 000002.png
│   │   └── ...
│   └── rgb/
│       ├── 000001.jpg
│       ├── 000002.jpg
│       └── ...
├── task2/
│   └── ...
├── ...
└── task11/
    └── ...
```


## How to Use

### Download with `huggingface_hub`

Download a single task:

```python
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="touchtronix/FusionX-Multimodal-Sample-Data-V3",
    repo_type="dataset",
    allow_patterns="strap1/*",
    local_dir="./fusionx_v3",
)
```

Download the full dataset:

```python
snapshot_download(
    repo_id="touchtronix/FusionX-Multimodal-Sample-Data-V3",
    repo_type="dataset",
    local_dir="./fusionx_v3",
)
```

### Load the per-frame index

```python
import pandas as pd

frames = pd.read_parquet("fusionx_v3/strap1/frames.parquet")
imu    = pd.read_parquet("fusionx_v3/strap1/oak_imu.parquet")
```

### Read calibration

```python
import json

with open("fusionx_v3/strap1/calib.json") as f:
    calib = json.load(f)
```

## License

**Creative Commons Attribution 4.0 (CC BY 4.0)**

You are free to:

- **Use** — Incorporate the data into your own projects.
- **Share** — Copy and redistribute the material in any medium or format.
- **Modify** — Remix, transform, and build upon the material.
- **Commercial Use** — Use the data for commercial purposes.

…as long as appropriate attribution is provided to **Touchtronix Robotics**.

## Citation

```bibtex
@misc{touchtronix_fusionx_v3_2026,
  title  = {FusionX Multimodal Sample Dataset (V3)},
  author = {Touchtronix Robotics},
  year   = {2026},
  url    = {https://huggingface.co/datasets/touchtronix/FusionX-Multimodal-Sample-Data-V3},
}
```