Wearable thumb sleeves enabled by self-supervised learning with few stretchable sensors and few-shot data for...

Area of Science:

• Human-Computer Interaction
• Wearable Technology
• Machine Learning

Background:

• Human fingers offer high dexterity for natural human-machine interaction.
• Conventional methods often require multiple devices per finger and extensive labeled data, limiting user and task adaptability.
• Existing systems struggle with user independence and data efficiency for diverse finger-related tasks.

Purpose of the Study:

• To develop a user-independent and data-efficient wearable system for recognizing various finger-related tasks.
• To enable rapid adaptation to new users and tasks using minimal labeled data.
• To explore the potential of a novel thumb sleeve as a replacement for traditional input devices.

Main Methods:

• A wearable thumb sleeve equipped with two stretchable sensors at the thumb joints was developed.
• Self-supervised learning was employed to learn latent features from unlabeled random thumb movement data.
• Fine-tuning with minimal (five-shot) labeled data enabled rapid adaptation to new users and tasks.

Main Results:

• The system demonstrated user independence and high data efficiency.
• It successfully recognized eight directional commands and 10 knuckle key inputs.
• The model allowed seamless task switching without retraining, showcasing adaptability.

Conclusions:

• The proposed wearable thumb sleeve with self-supervised learning offers a promising solution for natural human-machine interaction.
• This approach significantly reduces the need for labeled data and extensive retraining, enhancing user and task flexibility.
• The system has strong potential for real-world applications, including replacing mouse and keyboard functionalities for tasks like online shopping.