Neural-Network-Based Tactile Perception System Using Ultrahigh-Resolution Tactile Sensor.

Area of Science:

• Materials Science
• Robotics
• Sensory Science

Background:

• Human tactile perception is complex and crucial for product evaluation.
• Existing methods for tactile evaluation lack quantitative precision.
• Microelectromechanical systems (MEMS) offer potential for high-resolution sensing.

Purpose of the Study:

• To develop the first tactile perception system for sensory evaluation using a high-resolution MEMS tactile sensor.
• To quantitatively assess fabric tactile properties and enable comparison with human sensory evaluation.
• To utilize machine learning for predicting tactile sensations and informing product design.

Main Methods:

• Developed a tactile perception system utilizing a MEMS tactile sensor with micrometer resolution.
• Conducted sensory evaluation on 17 fabrics using a semantic differential method.
• Acquired tactile signals at 1 µm spatial resolution and trained a convolutional neural network (CNN) regression model.
• Validated system performance on unseen fabric data.

Main Results:

• Achieved a mean squared error (MSE) of 0.27 for tactile perception at 300 mm data length.
• Successfully predicted 89.2% of sensory evaluation words for unknown fabrics.
• Visualized fabric tactile sensation variations using heatmaps, identifying regions affecting perception.

Conclusions:

• The developed system enables quantitative comparison of tactile sensations across fabrics.
• The system's high accuracy in predicting sensory evaluations supports its application in product design.
• Heatmap visualization provides insights for achieving desired tactile product qualities.