Human Strategies for Single Finger Texture Exploration

Data and software for studying how people perceive ten textures and vary their single-finger exploration. The supplied workbooks contain 40 participant columns, 2,000 perceptual ratings, and 400 reported force/speed pairs.

The analysis runs without a camera, force sensor, or downloaded videos. It reproduces summaries and explicitly defined analyses of the processed data. The original raw-force preprocessing and video-to-speed extraction are not supplied, so this does not establish a reproduction of those pipelines or of a paper's results.

The figure is generated by the analysis below. Each cell averages 40 participant ratings; coarseness retains the source direction (−10 fine, +10 coarse).

Get the repository

For analysis, use a partial clone excluding the large video directory (Git 2.37 or later):

git clone --filter=blob:none --sparse https://github.com/aoli0919/Human-Texture-Exploration-Experiment.git
cd Human-Texture-Exploration-Experiment
git sparse-checkout set --no-cone '/*' '!/data/raw/video/'

This includes both workbooks and all 400 raw-force CSVs. A full clone also downloads the videos and can be substantially larger. Run git sparse-checkout disable to materialize omitted files later.

Run the analysis

Use Python 3.11 or later. From the repository root:

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e .
python -m human_texture verify --data-dir data
python -m human_texture reproduce --data-dir data --output-dir outputs/reproduction

On Windows PowerShell, activate with .venv\Scripts\Activate.ps1. The default run checks the 400 referenced raw-force files as well as the workbooks. For a copy containing only data/processed/, use:

python -m human_texture reproduce --data-dir data --output-dir outputs/reproduction --skip-raw

Open outputs/reproduction/report.md for results and audit.json for validation details. Generated files include:

Output	Contents
observations.csv	400 participant–texture observations with source-file references
rating_summary.csv	Statistics for each texture and perceptual dimension
exploration_summary.csv	Statistics for supplied force and speed values
within_participant_correlations.csv	Spearman correlations across ten textures within each participant
association_summary.csv	Summary of participant-level correlation distributions
raw_force_qc.csv	Raw-force checks when raw checking is enabled
figures/	PNG and SVG ratings, force/speed distributions, and association figures

Rerunning replaces generated files in this output directory. A nonempty directory created outside this tool is rejected to protect existing work. To run the regression tests, use python -m unittest discover -s tests.

CI runs the tests and the complete analysis on Linux with Python 3.11 and 3.12, and uploads the results under each workflow run's Artifacts section. The supplied dataset passes 25 regression tests and full validation of 400 observations, 2,000 matched ratings, and 48,246 raw-force samples. Hardware operation is not part of these checks.

Correlations treat participants as repeated-measurement units and report their rho distributions. They do not treat all 400 observations as independent subjects or claim a paper's regression results. See methods and the data dictionary.

Data and layout

data/processed/       Original Excel workbooks
data/raw/force/       400 CSVs in 40 session folders
data/raw/video/       400 MP4s; optional for this analysis
data/raw/questionnaire/ Original export; not used by the analysis
src/human_texture/    Hardware-free analysis package
acquisition/          Configurable collection CLI
legacy/acquisition/   Original collection scripts
docs/                 Methods, data dictionary, acquisition guide, original PDFs
outputs/              Generated analysis and acquisition results

Original data and workbooks are preserved byte-for-byte; generated tables are separate. Ratings span −10 to +10 for traction, roughness, coarseness, hardness, and temperature sensation. The source calls the coarseness question Fineness, but its positive endpoint means coarse; the analysis preserves this sign.

For links to the old directory names, see the path migration table.

Reported force/speed units are not declared in the workbooks. The author confirms that speed processing used pixel-to-millimetre spatial calibration, but the coefficient and calibration files are no longer available; the final speed time unit is unconfirmed. Outputs preserve reported values with unspecified units. Workbook NF and SS are extracted/rounded values, not evidence of normalization.

Collect new data

Inspect a seeded trial plan without hardware dependencies or device access:

python -m acquisition --dry-run --seed 42

For a RobotiQ FT300-S sensor and compatible camera:

python -m pip install -e '.[acquisition]'
python -m acquisition --session-id pilot001 --output-dir outputs/acquisition --serial-port /dev/ttyUSB0 --camera-index 0 --seed 42

See the acquisition guide for controls and timing. Dry-run and software tests do not establish physical hardware operation.

Related work

These publications provide research context; their presence does not establish that this dataset reproduces their experiments or reported results.

• Velocity Modulation in Robotic Texture Exploration Based on Human Perceptual Dimensions, Humanoids 2025.
• Exploratory Movement Strategies for Texture Discrimination with a Neuromorphic Tactile Sensor, IROS 2025.
• Neuromorphic touch for robotics—a review, Neuromorphic Computing and Engineering, 2025.

Citation and contact

See CITATION.cff for repository citation metadata. Dataset questions: Ao Li or Benjamin Ward-Cherrier.