tamago117/cluster-haptic-texture-dataset

Cluster Haptic Texture Dataset

Sliding-contact recordings of 118 real material surfaces: contact sound, 3-axis acceleration, normal force and probe position, captured under a controlled sweep of scan direction, velocity and pressing force.

Paper | GitHub | figshare

Want to render these textures in VR? HaptoFlow is a Flow Matching model trained on this dataset that generates vibrotactile feedback in real time. Check it out!

Directory structure

.
├── images/
│   ├── images_scan_area/   (images of the 100x100 mm scan area)
│   └── crop_images/        (for evaluation of texture classification)
├── sensor_data/
│   ├── audio/              (after noise cancelling)
│   ├── raw_audio/
│   ├── accel/
│   ├── force/
│   └── position/
├── texture_list.xlsx       (texture names and their friction coefficients)
└── README.md

images_scan_area file naming rule

• img(texture_id).jpg
• example: img0.jpg

sensor_data file naming rule

(texture_id)_(direction)_(velocity)_(force)_(repeat count).parquet or .flac

texture_id

See texture_list.xlsx.

direction

Clockwise angle (degree) — 0, 45, 90, 135, 180, 225, 270, 315.

velocity

Scan speed (mm/s) — 20, 30, 40, 50, 60.

force

Push force (mN) — 500 or 1000.

repeat count

Number of scan — 0 or 1.

example: 0_135_40_500_0.parquet

• texture id: 0
• direction: 135 degree
• velocity: 40 mm/s
• force: 500 mN
• repeat count: 0

audio

Sound source after noise cancelling.

• file format: FLAC (mono, 16-bit)
• sampling rate: 44.1 kHz

raw_audio

Raw sound source.

• file format: FLAC (stereo, 16-bit)
  • channel 0: main mic signal (texture + machine noise signal)
  • channel 1: sub mic signal (machine noise signal)
• sampling rate: 44.1 kHz

accel

• file format: Parquet
• time_ns (int64, nanoseconds)
• 3-axis acceleration X, Y, Z (float32, g)
• sampling rate: 6 kHz

time_ns   X       Y       Z
999053    0.014   0.067   -0.960
1209053   0.034   0.048   -0.946
1339053   0.029   0.062   -0.952

force

• file format: Parquet
• time_ns (int64, nanoseconds)
• force (float32, N)
• sampling rate: 6 kHz

time_ns   force
999053    1.009
1209053   1.009
1339053   1.009

position

• file format: Parquet
• time_ns (int64, nanoseconds)
• 2D position X, Y (float32, mm)
• sampling rate: 100 Hz

time_ns   X     Y
613000    0.0   45.0
10477000  0.0   45.0
19940000  0.0   45.0

Reading the sensor tables

import pandas as pd

df = pd.read_parquet("sensor_data/accel/0/0_135_40_500_0.parquet")
df.insert(0, "time", df.pop("time_ns") / 1e9)   # seconds

Citation

@misc{eguchi2025clusterhaptictexturedataset,
      title={Cluster Haptic Texture Dataset: Haptic Texture Dataset with Varied Velocity-Direction Sliding Contacts},
      author={Michikuni Eguchi and Tomohiro Hayase and Yuichi Hiroi and Takefumi Hiraki},
      year={2025},
      eprint={2407.16206},
      archivePrefix={arXiv},
      primaryClass={cs.HC},
      url={https://arxiv.org/abs/2407.16206},
}

@inproceedings{eguchi2025whc_wip,
  title={Cluster Haptic Texture Database: Haptic Texture Database with Controlled Sliding-Contact Interactions},
  author={Michikuni Eguchi and Yuichi Hiroi and Tomohiro Hayase and Takefumi Hiraki},
  booktitle={2025 IEEE World Haptics Conference Work in Progress (WHC WiP)},
  year={2025},
  organization={IEEE}
}