Datasets:

gemixin
/

touch-ex

language:
  - en
license: cc-by-4.0
pretty_name: Touch-Ex
task_categories:
  - image-classification
  - feature-extraction
tags:
  - tactile-sensing
  - visuo-tactile
  - robotics
  - digit-sensor
  - multimodal
  - tactile-language
  - tactile-perception
  - representation-learning
  - force-aware-perception

Touch-Ex: A Region-Level, Force-Annotated Visuo-Tactile Dataset

Developed by Gemma McLean and supervised by Dr Daniel Hao, University of Leicester.

Dataset Summary

Touch-Ex (Touch Exploration) is a region-level, force-annotated visuo-tactile dataset collected using a DIGIT vision-based tactile sensor and a collection of common UK household objects. The dataset was designed to support research in tactile representation learning, force-aware perception, object-region recognition, temporal tactile modelling, tactile-language alignment and in particular, active tactile exploration.

The dataset contains more than 135,000 visuo-tactile frames collected across 1,128 tactile interactions from 23 household objects spanning a wide range of materials, textures, geometries, and compliance levels. Data collection was designed to ensure balanced representation across objects, with each object in the training split contributing an equal number of tactile samples regardless of the number of annotated regions.

Touch-Ex is divided into two predefined splits:

Train: 14 objects used for model development, training, validation, and in-distribution evaluation.
Test Unseen: 9 objects reserved exclusively for object-level generalisation evaluation.

Each dataset sample corresponds to a single tactile frame and includes image data, force measurements, interaction metadata, semantic descriptions, and structured tactile attributes.

Example Samples

object	region	force_level	motion	material	hardness	description
hammer	handle	3	sliding	rubber	medium-hard	An elongated, solid rubber-coated gripping area featuring small, evenly spaced circular dimples that provide a repeating texture.
scissors	blades	2	rotation	metal	hard	A pair of elongated, flat metallic blades with polished surfaces that gradually taper to sharp outer edges and are joined at a raised pivot screw.
tv_remote	buttons	1	sliding	rubber	soft-medium	Small, compact raised rubber features, including rectangles, squares, and circles, that provide springy resistance when pressed.

Touch-Ex samples showing tactile images and their key annotations.

Dataset Modality

Following standard terminology for optical tactile sensors, "visuo-tactile" refers strictly to the internal image data captured by the DIGIT sensor's embedded camera. This dataset intentionally excludes external RGB camera views of the objects to prioritise compactness and pure tactile-language representation learning.

Why Touch-Ex?

Touch-Ex was designed to address several limitations of existing visuo-tactile datasets. While many tactile datasets focus primarily on object identity or contact localisation, Touch-Ex introduces region-level annotations, force-level annotations, and semantic tactile descriptions within a unified benchmark.

Touch-Ex combines:

Region-level object annotations
Force-level tactile annotations
Sequential tactile interactions
Controlled exploratory motions
Natural-language tactile descriptions
Structured semantic attributes

In addition to standard in-distribution evaluation, Touch-Ex provides a dedicated unseen object test split containing both alternative instances of known object categories and entirely new object categories. This enables systematic evaluation of object-level generalisation and the extent to which learned representations capture transferable tactile properties rather than object identity alone.

Data Collection

Touch-Ex was collected using a DIGIT vision-based tactile sensor operated manually by a human data collector, moving and orienting the DIGIT and the object as needed.

To enable force-aware data collection, a Grove Round Force Sensor (FSR402) was attached to the rear of the DIGIT sensor and connected to a Raspberry Pi Pico via a Grove Shield.

During data collection, the operator applied pressure to the rear of the sensor using their index finger (stabilising either side with their thumb and middle finger) while bringing the DIGIT into contact with the target object. The rear-mounted FSR measured this applied force, providing an estimate of how strongly the sensor was pressed against the object surface.

Data acquisition was performed using a custom collection tool developed specifically for Touch-Ex: https://github.com/gemixin/digit-fsr-gui.

The software simultaneously recorded tactile images and FSR measurements throughout data collection.

Each tactile interaction was performed using one of two exploratory motions:

Sliding: controlled movement across an object surface while maintaining contact.
Rotation: controlled rotational movement about a target contact point.

Every interaction consists of a sequence of 120 tactile frames collected from a labelled object region while maintaining a target force level. Every effort was made to ensure that contact remained within the target region throughout an interaction. However, for certain objects and regions, complete isolation was not always possible. For example, narrow football seams may include portions of adjacent panels due to their limited width.

Contact points were sampled within each labelled region to ensure diverse coverage of the object surface, with care taken to include any distinctive or unique features.

Force Annotation

To improve consistency across interactions, FSR measurements were used to categorise contact into three predefined force levels. These force levels were designed to capture meaningful differences in sensor deformation while remaining practical for consistent human application.

Force Level	Description
1	Contact Only
2	Light Press
3	Firm Press

During collection, a target force level was selected within the acquisition software. Real-time FSR measurements were used to monitor the applied force, while a connected LED bar provided visual feedback to the operator. The software additionally filtered tactile frames according to the selected target force level, ensuring that only samples falling within the predefined range for the chosen force category were retained.

FSR measurements were calibrated using a dedicated calibration framework: https://github.com/gemixin/digit-force-calibration.

The dataset includes both raw FSR voltage measurements and approximate calibrated force estimates derived from the rear-mounted force sensor. While these values provide additional information about relative contact intensity, the primary force annotation is the discrete force-level label. The calibrated force estimates should therefore be considered approximate rather than laboratory-grade force ground truth.

Semantic Annotations

The material labels, hardness labels, and natural-language descriptions were assigned by a human annotator and subsequently reviewed with the assistance of large language models to improve consistency.

Dataset Structure

The Hugging Face version of Touch-Ex uses a flattened tabular representation where each row corresponds to a single visuo-tactile frame.

Features

Feature	Type	Description
image	Image	DIGIT visuo-tactile image
object	string	Object label
region	string	Region label
object_region	string	Combined object-region label
force_level	string	Discrete force category (1, 2 or 3)
motion	string	Interaction motion type (sliding or rotation)
fsr_voltage	float32	Raw force-sensitive resistor voltage
force_n	float32	Approximate calibrated force estimate (Newtons)
hardness	string	Perceived hardness label
material	string	Primary material label
description	string	Natural-language tactile description
interaction_num	string	Interaction identifier within region and force level
frame_num	string	Frame number within interaction
interaction_id	string	Unique interaction identifier

Splits

Touch-Ex is divided into two predefined splits:

Split	Description
Train	14 objects used for model development, training, validation, and in-distribution evaluation
Test Unseen	9 objects reserved for object-level generalisation evaluation

The unseen test split contains both alternative instances of known object categories and entirely new object categories.

For evaluation, all frames originating from the same interaction should remain within the same dataset partition. Because interactions contain highly correlated frame sequences, frame-level train/test splitting is not recommended. The interaction_id field is provided to support leakage-free splitting.

Supplementary Files

The repository includes several supplementary files in addition to the main dataset:

File	Description
`baseline.jpg`	Reference DIGIT image captured without object contact. May be used for background subtraction and preprocessing.
`metadata.csv`	Interaction-level metadata containing object, region, force level, motion type, and interaction identifiers.
`descriptions.csv`	Region-level semantic annotations including material labels, hardness labels, and natural-language descriptions.

Dataset Contents

Training Objects

#	Object	Object Label	Region Labels	Specification
1	Toilet roll	toilet_roll	body, ends	New, full roll of unbranded, quilted white toilet paper
2	Pringles tube (empty)	tube_pringles	body, lid, base	Empty 185g Original Flavour Pringles tube with lid; freshness seal removed
3	Tin of beans (sealed)	tin_beans	body, lid, base	Sealed 415g tin of Heinz Baked Beans
4	Tennis ball	tennis_ball	body, seam	Standard yellow Penn tennis ball
5	Scissors	scissors	handle, blades	Large pair of general-purpose scissors
6	Hammer	hammer	head, neck, handle	Hammer with rubber-dimpled handle
7	Toothbrush	toothbrush	handle, head	Generic adult plastic toothbrush
8	Mug	mug	body, handle, rim	Plain white ceramic mug
9	Plastic bottle (empty)	plastic_bottle	body, cap, base	Empty 50cl Nestlé Pure Life still water bottle with lid; label removed
10	Sponge scourer	sponge_scourer	sponge_side, scourer_side	Generic sponge scourer
11	TV remote	tv_remote	body, buttons	Splinktech universal Samsung TV remote control
12	Wooden spoon	wooden_spoon	handle, head	30cm wooden spoon
13	Tea towel (folded)	tea_towel	surface, edge	Cotton tea towel with lattice-effect
14	Football	football	panels, seams	Mitre Impel L30P size 3 football

Unseen Test Objects

#	Object	Object Label	Region Labels	Specification
1	Tin of mushy peas (sealed)	tin_peas	body, lid, base	Sealed 300g tin of Batchelors Original Mushy Peas
2	Scissors	small_scissors	handle, blades	Small pair of general-purpose scissors
3	Mug	patterned_mug	body, handle, rim	Patterned white ceramic mug
4	Media remote	media_remote	body, buttons	Xbox media remote control
5	Table knife	table_knife	handle, blade	Standard stainless steel table knife
6	Microfibre cloth (folded)	microfibre_cloth	surface, edge	White ribbed microfibre cleaning cloth with yellow edging
7	Dish brush	dish_brush	handle, head	Spontex pink washing-up brush
8	Beaker	beaker	body, rim	Patterned plastic beaker
9	Rubber ball	rubber_ball	body, ridges	Chuckit! Ultra medium rubber ball

Usage

from datasets import load_dataset

dataset = load_dataset("gemixin/touch-ex")

train = dataset["train"]
test_unseen = dataset["test_unseen"]

sample = train[0] 
image = sample["image"]
print({key: value for key, value in sample.items() if key != "image"})

The image field is automatically loaded as a PIL image object by the Hugging Face Datasets library.

Supported Tasks

Touch-Ex supports a broad range of visuo-tactile learning tasks, including:

Object classification
Object-region classification
Force-level classification
Force-aware tactile perception
Sequential tactile modelling
Representation learning
Tactile-language alignment
Multimodal learning
Active tactile exploration

Reference implementations, baseline models, and future experiments will be available in the Touch-Ex repository: https://github.com/gemixin/touch-ex.

Dataset Statistics

Statistic	Value
Training objects	14
Training object regions	33
Unseen test objects	9
Unseen test object regions	20
Total objects	23
Total object regions	53
Force levels	3
Motion types	2
Frames per interaction	120
Total interactions	1,128
Total visuo-tactile frames	135,360

Related Resources

Raw Dataset

The original hierarchical version of Touch-Ex, including the complete folder structure, interaction-level metadata, semantic annotation files, and baseline image, is available here:

https://drive.google.com/drive/folders/1Jw3O8ewLCXLjAZWtHekXgwb4KDu-blX2?usp=sharing

Limitations

Data was collected using a single DIGIT tactile sensor.
Interactions were performed manually by a single operator.
Force measurements are intended primarily to support force-level categorisation and should not be interpreted as highly accurate force ground truth.
Objects are limited to household items collected in the United Kingdom.
The dataset does not include external RGB images or video streams of the objects.
Data was collected in a home office environment rather than a controlled laboratory setting.
Environmental conditions were not explicitly varied during collection.

Ethical Considerations

The dataset contains visuo-tactile observations of inanimate household objects and does not contain personal information, biometric data, human subjects, or sensitive content.

The dataset is intended for research in tactile perception, robotics, multimodal learning, and representation learning.

Citation

If you use Touch-Ex in your research, please cite:

@dataset{mclean2026touchex,
  author = {McLean, Gemma and Hao, Zhou Daniel},
  title = {Touch-Ex: A Region-Level, Force-Annotated Visuo-Tactile Dataset},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/gemixin/touch-ex}
}

Touch-Ex was collected using the DIGIT vision-based tactile sensor. If you use this dataset in your research, please additionally cite the original DIGIT paper:

@article{lambeta2020digit,
  title = {DIGIT: A Novel Design for a Low-Cost Compact High-Resolution Tactile Sensor with Application to In-Hand Manipulation},
  author = {Lambeta, Mike and Chou, Po-Wei and Tian, Stephen and Yang, Brian and Maloon, Benjamin and Most, Victoria Rose and Stroud, Dave and Santos, Raymond and Byagowi, Ahmad and Kammerer, Gregg and Jayaraman, Dinesh and Calandra, Roberto},
  journal = {IEEE Robotics and Automation Letters},
  volume = {5},
  number = {3},
  pages = {3838--3845},
  year = {2020},
  doi = {10.1109/LRA.2020.2977257}
}