Projects

A camera observation image and a text language instruction feed a VLA model that produces robot actions.

Physical AI Safety

Trustworthy Vision-Language-Action models and embodied agents

JSPS KAKENHI · 26K25580

I study how embodied AI systems can perceive, reason, and act safely in changing environments. My research addresses robustness to environmental variation and adversarial inputs, alongside evaluation frameworks for trustworthy Vision-Language-Action models.

  • Robustness under deployment shifts
  • Adversarial robustness
  • Safety evaluation for trustworthy VLAs

Awesome Physical AI Safety
An LLM-assisted bibliography on safety, security, robustness, and trustworthy deployment across VLA, robotics, autonomous systems, and embodied agents.

View Collapse · Preprint
Overview of asymmetric representation regularization for adversarial training.

Adversarial Robustness & Generalization

Understanding failures and learning robust representations

I investigate why models fail under adversarial perturbations and perceptual ambiguity, and how to make learning and predictions more reliable. My work connects the mechanisms of these failures with robust training, defense, calibration, and generalization across tasks and modalities.

  • Understanding and analyzing failure mechanisms
  • Robust training
  • Calibration and generalization
Brain Decoding with CLIP · PreprintDifferential Attention · ICLR 2026Allegory of the Cave · APSIPA ASC 2025AR-AT · ICLR 2025Infrared Patch Defense · ICIP 2024DAC · ICML 2023Adversarial Transferability · WACV 2023
Overview of multimodal adversarial training with one-to-many image-text relationships.

Reliable Vision-Language Models

Robust multimodal understanding and evaluation

I develop training methods and benchmarks for vision-language models that understand visual and textual information reliably. My work brings together multimodal adversarial defense, language supervision, image-text alignment, and evaluation of when models should read or ignore text.

  • Robust multimodal learning and adversarial defense
  • Language supervision and image-text alignment
  • Safety evaluation, typographic robustness, and text recognition
RIO-Bench · ECCV 2026Text-Printed Image · CVPR 2026MAT · WACV 2026QT-AFT · ACM MM 2025
Semantic Color Naturalness Breaker adds a perturbation to a grayscale image, causing unauthorized AI colorization to produce unnatural colors.

Model Ownership & Content Protection

Protecting models and visual content from unauthorized use

I develop methods to establish model ownership and protect visual content against unauthorized AI processing. My work explores fingerprints resilient to model merging, traceable watermarks for image editing, and content-aware perturbations that prevent unwanted colorization.

  • Ownership verification after model merging
  • Traceable watermarking for image editing
  • Perception- and content-aware protection against colorization
Semantic Color Naturalness Breaker · SMC 2026AngelFingerprint · PreprintUncolorable Examples · APSIPA ASC 2025MergePrint · ACL 2025

Research Funding

JSPS KAKENHI — Research Activity Start-up

科研費・研究活動スタート支援

信頼できるVLAの実現に向けた安全評価基盤の構築

Building a Safety Evaluation Framework for Trustworthy Vision-Language-Action Models

Project number
26K25580
Period
July 31, 2026 – March 31, 2028
Role
Principal Investigator
Total funding
JPY 2,600,000Direct costs: JPY 2,000,000 · Indirect costs: JPY 600,000
View project on KAKEN ↗

Past Research Funding

JSPS Research Fellowships for Young Scientists (DC2)

日本学術振興会 特別研究員（DC2）

Period
April 2024 – March 2026
Role
JSPS Research Fellow (DC2)
View fellowship on JSPS ↗

JST AIP Challenge Program

JST AIPチャレンジプログラム

Period
April 2023 – March 2024
Research grant
JPY 1,000,000
View program on JST ↗