Tactile Data Recording System for Clothing with Motion-Controlled Robotic Sliding

AI summary

Overview

Research area: Robotics and haptics — robotic tactile sensing for clothing and textiles, combined with machine-learning-based material identification.

Technical level: Intermediate. The paper is a short SIGGRAPH Asia 2025 poster (SA Posters '25) built on accessible hardware and a comparative CNN classification study; the concepts are clear, but familiarity with tactile sensing and classification metrics helps.

Scope: The paper presents a robotic-arm system that non-destructively strokes intact garments with a simulated fingertip under controlled motion (8 directions, 5 speeds, 2 kinds of force) to build motion-labeled multimodal tactile databases, and evaluates whether motion labels improve clothing identification.

What This Paper Is About

The feel of clothing matters for comfort, but existing clothing tactile databases were collected with uncontrolled finger-mounted sensors and lack systematic recordings of motion variables such as sliding speed and direction. Recent work shows that stroking with systematically varied speeds and directions encodes material properties more accurately, but those methods measure cut fabric swatches rather than whole garments, which does not scale to large clothing databases where non-destructive measurement is essential. The goal of this work is to recording tactile data directly from intact clothing using a motion-controlled robotic arm.

Key Contributions

1. A robotic arm-based system whose sensor head — two omnidirectional microphones (also used for noise cancelling), one triaxial accelerometer, and one load cell, with a rubber (urethane) simulated fingertip — slides across fabric surfaces with controlled movements.
2. Non-destructive, automated recording: when the user places clothing on the attachment, the system records multimodal tactile data (frictional audio signals, surface images, and acceleration) under preset conditions of 8 directions, 5 speeds, and 2 kinds of force.
3. A machine-learning evaluation on 23 distinct clothing items showing that adding motion-related parameters (velocity and direction) improves clothing identification accuracy.
4. A stated foundation for objectively quantifying clothing tactile sensation, oriented toward quality investigation, classification, and tactile reproduction applications.

Main Findings

• Motion labels help audio-based identification: clothing identification accuracy with audio alone was 87.09% without motion-related parameters and 93.75% with them (Table 1).
• Best overall single result: the highest accuracy reported in Table 1 is 93.75%, achieved with audio plus motion-related parameters.
• Acceleration alone is weak: accuracy with acceleration alone was 44.57% with motion and 43.34% without motion — the lowest numbers in the table.
• Combined modalities: audio and acceleration together reached 93.48% with motion and 92.80% without motion.
• Acceleration is complementary, then potentially redundant: when motion labels are absent, adding acceleration improved accuracy (87.09% to 92.80%); when accurate motion labels already exist, acceleration may become redundant or act as noise (93.75% with audio only versus 93.48% with audio plus acceleration).
• Audio is information-rich: the authors conclude that audio contains a significant amount of information related to clothing tactile sensation, while acceleration mainly supplies motion information.
• Dataset diversity: evaluation used data from 23 distinct clothing items varying in material (such as cotton, polyester, and wool), weave, and thickness; the authors state that because the model classifies 23 clothing types with high accuracy, the recorded data contained diverse features readily distinguishable by CNN.
• Parallel to human perception: the authors note the result resembles human tactile perception, where accurate perception of motion increases the accuracy of texture perception.

Methodology in Plain English

A robotic arm (manipulator) carries a sensor head fitted with a rubber tip that acts as a simulated/pseudo-fingertip. The system presses this fingertip against clothing and slides it across the surface, with the arm controlling exactly how hard it presses, how fast it slides, and in which direction. The preset conditions used were 8 directions, 5 speeds, and 2 kinds of force. Clothing is placed on a cloth attachment, so nothing is cut or destroyed.

While sliding, the sensor head collects several signals at once: frictional audio from the microphones, acceleration from the triaxial accelerometer, and force from the load cell, along with surface images. This produces a multimodal, motion-labeled dataset.

For evaluation, the authors ran a clothing identification task using machine learning. They extracted features from audio and/or acceleration signals, and also tested adding the motion-related parameters (velocity and direction) as inputs. A CNN (as referenced in the discussion of results) then predicted the clothing label across 23 distinct items, and the authors compared accuracy across modality combinations with and without motion information.

Why This Matters

Impact on research: The work addresses a scalability gap — prior tactile clothing datasets used cut swatches and uncontrolled finger-mounted sensors, which limits their use for large clothing databases. By recording non-destructively from intact garments with labeled motion parameters, this system provides a way to build motion-labeled, multimodal tactile databases and supports objective quantification of clothing tactile sensation. It also echoes findings in human tactile perception about the role of motion in texture judgments.

Real-world applications:

• Clothing quality investigation and fabric inspection in apparel manufacturing or retail.
• Tactile reproduction of clothing, for example in haptic displays or virtual try-on experiences.
• Clothing classification and search — e.g., automatically characterizing garments by material, weave, or thickness from recorded tactile data.
• Informing comfort-oriented design decisions by linking recorded physical signals to garment properties.

Industry relevance: Apparel and e-commerce companies that handle large catalogs of garments stand to benefit, since non-destructive measurement is a prerequisite for catalog-scale data collection. The author affiliations include ZOZO Research alongside academic labs (University of Tsukuba, The University of Tokyo, Cluster Metaverse Lab), indicating direct industry interest in clothing-related tactile technology.

Future Directions

• Human sensory validation: The authors identify a lack of human sensory information and propose incorporating human validation to correlate objective tactile data with subjective ratings of clothing comfort and feel, bridging physical properties and qualitative human experience.
• Broader clothing variety: The dataset should include materials with very fine or complex textures, as well as extremely soft or deformable materials, to improve generalizability.
• Richer tactile properties: While the urethane rubber simulated finger captures roughness, future data acquisition should aim to capture stiffness, thickness, and warmth.
• Open question on multimodal redundancy: Since acceleration appears complementary when motion labels are absent but potentially redundant or noisy when they are present, how best to fuse or weight modalities in a motion-labeled setting remains open.

Target Audience

Researchers and practitioners in haptics, tactile sensing, and robotics who are interested in textile or material recognition; apparel and e-commerce engineers building clothing databases, quality-inspection pipelines, or virtual try-on and tactile-reproduction systems; and machine-learning practitioners working on multimodal (audio, acceleration, force) classification. The paper is a concise poster, so it is also well suited to readers who want a quick orientation to motion-controlled robotic tactile data collection rather than an in-depth methods treatment.

Note: The paper does not report the number of measurement samples per garment, recording duration, or dataset size in seconds or trials; it reports 23 clothing items, 8 directions, 5 speeds, and 2 kinds of force.

Authors’ abstract

The tactile sensation of clothing is critical to wearer comfort. To reveal physical properties that make clothing comfortable, systematic collection of tactile data during sliding motion is required. We propose a robotic arm-based system for collecting tactile data from intact garments. The system performs stroking measurements with a simulated fingertip while precisely controlling speed and direction, enabling creation of motion-labeled, multimodal tactile databases. Machine learning evaluation showed that including motion-related parameters improved identification accuracy for audio and acceleration data, demonstrating the efficacy of motion-related labels for characterizing clothing tactile sensation. This system provides a scalable, non-destructive method for capturing tactile data of clothing, contributing to future studies on fabric perception and reproduction.