Research

Research vision: people and machines learning together

My long-term goal is to develop intelligent machines that understand human intention, adapt to individual needs, and expand what people can do in everyday life. I connect biosignal and behavior understanding, personalization and continual learning, safe physical interaction, and human-centered design.

Prosthetic hands provide a concrete setting for studying these questions: the user learns to operate the device while the device must adapt to the user and changing conditions. I also investigate how virtual evaluation and real-world evidence can guide the design and training of assistive robotic systems. These are research goals; their value must be demonstrated through reliable operation, reduced burden, and useful capabilities for the person.

Connected research directions

• Intention and behavior: biosignal decoding, time-series data segmentation, and multimodal understanding of actions and outcomes.
• Learning through use: efficient personalization, continual adaptation, and the relationship between machine adaptation and human skill development.
• Physical interaction: sensing, compliant mechanisms, and shared control for safe grasping and robot manipulation.
• Evaluation and design: simulation and immersive environments linked to physical testing and user feedback, to improve devices and training approaches.

My current work in the Nakamura Laboratory connects time-series segmentation and robot manipulation to these broader questions.

Previous research: myoelectric prosthetic hands

The following sections describe my research background in biosignal processing and prosthetic hands.

My research focuses on human-centered myoelectric prosthetic hands: systems that decode a user’s intention, adapt to changing biosignals, interact safely with objects, and remain mechanically forgiving in real tasks.

1. Continuous multi-DOF myoelectric control

Most pattern-recognition prostheses select one discrete grasp at a time. I study regression-based control that maps surface electromyography (sEMG) directly to multiple finger-joint angles, enabling smoother transitions and more biomimetic motion.

Our TF2AngleNet framework combines raw time-domain sEMG with multidimensional time–frequency representations to estimate six finger-joint angles. The work also uses non-contact motion measurement to align hand kinematics with high-rate sEMG.

Key questions

• Which signal representations preserve movement information while remaining suitable for real-time control?
• How should multi-joint motion be modeled when several fingers are activated together?
• How can inference and communication be implemented reliably on an embedded prosthetic-hand system?

2. Personalization with limited calibration

Myoelectric signals vary substantially across people, days, fatigue levels, electrode positions, and donning conditions. A model that performs well once is not enough; it must be easy to personalize and maintain.

I investigate few-shot fine-tuning, cross-user transfer, robustness to sensor displacement, and online adaptation. A central design constraint is calibration burden: adaptation should use a small number of representative movements and fit naturally into a user’s routine.

Current directions include:

• selecting compact calibration sets that improve both trained and untrained movements;
• adapting across days and users without erasing previously learned behavior;
• learning from reliable, successful interactions rather than every incoming sample;
• moving personalization from a workstation toward edge and on-device deployment.

3. Prosthetic finger mechanics and compliant grasping

Control intelligence cannot compensate for every uncertainty in contact. I therefore also design finger mechanisms that use passive mechanics to improve grasping.

This work explores passive hyperextension at distal finger joints, compliant fingertips, underactuated transmission, and stiffness selection. The aim is to let the hand conform to objects and tolerate positioning errors without requiring more motors or more complicated control.

Evaluation combines:

• standardized prosthetic-hand assessment tasks;
• benchtop force–angle characterization;
• healthy-participant and amputee experiments;
• comparison of joint constraints, spring settings, and fingertip compliance.

4. Multimodal grasp-state understanding and safe control

A practical prosthesis should recognize not only what the user intends, but also what is happening to the object. I study the integration of wrist-camera images, fingertip pressure, and inertial sensing to estimate grasp stage and stability.

This supports shared-control strategies that can suppress accidental opening during transport while preserving intentional release. It also creates a route toward self-improvement: successful interactions can be identified and reused as high-quality adaptation data.

Research principles

Methods and platforms

My work combines deep learning, signal processing, embedded inference, Bluetooth communication, motion capture, mechanical prototyping, force testing, standardized prosthetic-hand assessments, and virtual/physical task environments. Experimental systems include multi-DOF myoelectric hands such as the BIT-UEC HAND.

For papers and publication links, see Publications. For collaboration inquiries, see Contact.