Nhat is currently a first-year PhD student at
National University of Singapore
and is a Graduate Researcher at
A*STAR Singapore
, where he is advised by
Prof. Ivor Tsang,
Prof. Jin-Song Dong,
and Prof. Qing Guo
in making robots safer whenever they
share space with humans, particularly through
learning semantic-action representations and stress-testing.
| Feb 2026 |
Paper accepted
LIBERO-Mem, our work on stress-testing robots to remember object-level interaction histories for manipulation, was accepted to AAAI 2026.
|
|---|---|
| Jan 2026 | |
| Aug 2025 |
Started as a PhD student at NUS's School of Computing and Graduate Researcher at A*STAR's IAIC, CFAR.
|
| Jul 2025 |
Paper accepted
BiMa, our work on reducing visual and linguistic biases in text-video retrieval, was accepted to ACM Multimedia 2025.
|
He earned a B.Eng. and a M.Eng. from International University, Vietnam National University HCMC, graduating top of his class. During his studies, he was mentored by Dr. Tuan-Anh Vu in competitive programming and computer vision research. He later worked as an AI Engineer at FPT Software, where he was mentored by Prof. Ngan Le, Prof. Tung Kieu, and Prof. Anh Nguyen in robotics and multimodal systems.
Nhat's technical background spans embodied AI, trustworthy systems, and visual perception.
Nhat develops semantic representations that help robots connect perception, memory, and action. SlotVLA models objects and their relations for manipulation, while LIBERO-Mem maintains object-centric task states over long horizons, giving embodied agents structured knowledge for deciding what to do next.
Nhat studies whether semantic representations remain dependable against ambiguities and adversaries. OBEYED-VLA grounds actions in task-relevant objects and geometry, while DepthVanish, MAGIC, and typographic attacks expose how physical and linguistic perturbations can distort understanding and downstream decisions.
Nhat's visual perception research turns complex scenes into semantic concepts that embodied agents can reason about. Open-vocabulary camouflage segmentation identifies concealed instances, BgSubNet separates foreground dynamics, and UNO represents interactions as scene relations when objects, motion, and context are difficult to interpret.
Nhat's research now asks how semantic representations can support decisions that remain meaningful, grounded, and robust in the open world. The long-term goal is to build contextually safe robots that understand what matters in a scene, recognize when their understanding may be unreliable, and act accordingly around people.
