Generative AI

Long-Chain Reasoning of LLM for Diagnostic Knowledge Tracing and Image-Text Fusion

Long-Chain Reasoning of LLM for Diagnostic Knowledge Tracing and Image-Text Fusion

Principal Investigators & Key Members:
Prof. Zengchang Qin
This project develops an AI-powered diagnostic system for gastrointestinal diseases, integrating multimodal data and clinical knowledge to enhance early cancer detection. By combining endoscopic imaging with real-time retrieval of medical guidelines and chain-of-thought reasoning, the system generates accurate, explainable, and structured diagnostic reports. It aims to reduce diagnostic errors, improve healthcare accessibility, and support clinicians with transparent decision-making tools. The research contributes to precision medicine and equitable healthcare by delivering scalable, trustworthy AI solutions for clinical environments.
Read More
Distributional Alignment and Diversity Control for Generative Models via Optimal Transport and Uncertainty Quantification

Distributional Alignment and Diversity Control for Generative Models via Optimal Transport and Uncertainty Quantification

Principal Investigators & Key Members:
Asst. Prof. Nguyen Tuan Binh
We will build methods that gently nudge a generative model’s behavior when it is generating words so its outputs match a desired domain or safety profile while staying useful and varied. The key tool is optimal transport, a mathematical way to align distributions. We pair it with calibrated confidence and safe refusal. We will test these ideas in Vietnamese healthcare and education, release open source tools and benchmarks, and collaborate with VinMec and VinSchool to increase the reliability and value of AI systems in real use.
Read More
On-Device Small Language Models: Toward Private, Efficient, and Agentic Edge Intelligence

On-Device Small Language Models: Toward Private, Efficient, and Agentic Edge Intelligence

Principal Investigators & Key Members:
Asst. Prof. Huynh Thanh Trung
This project aims to develop a new generation of intelligent recommendation systems based on Agentic Artificial Intelligence (AI) that can understand user needs, learn from feedback, and provide transparent explanations for their suggestions. Current recommender systems often operate as “black boxes,” limiting user trust and adaptability. By integrating large language models (LLMs) with reasoning, planning, and self-improvement capabilities, the project will enable more adaptive and explainable personalization. Applications include education, healthcare, and sustainable e-commerce. The anticipated outcomes are enhanced trust, efficiency, and fairness in digital recommendation technologies, contributing to the advancement of ethical and human-centered AI systems.
Read More
Mechanistic Understanding and Control of Multi-modal LLMs

Mechanistic Understanding and Control of Multi-modal LLMs

Principal Investigators & Key Members:
Asst. Prof. Khoa D. Doan
Multimodal LLMs that can solve complex visual, mathematical, programming, and discovery tasks still make simple mistakes - e.g., misjudging spatial relations, failing at genuine discovery, or lacking cultural awareness - limiting safe use in healthcare, finance, and low-resource/culturally diverse settings such as Vietnam. This project will build real-world benchmarks to reveal these weaknesses, explain why they occur using mechanistic analyses, and develop lightweight inference-time control and data-efficient fine-tuning methods to fix them without costly retraining. We will release open datasets, tools, and guidelines to enable safer, more inclusive AI, cut energy use through efficient training, and unlock trustworthy applications that support local innovation and economic growth.
Read More

Robotics VLA

AI-Enhanced Soft Robotic Endoscopy System for Precision Colonoscopy and Real-Time 3D Navigation

AI-Enhanced Soft Robotic Endoscopy System for Precision Colonoscopy and Real-Time 3D Navigation

Principal Investigators & Key Members:
Asst. Prof. Thai Mai Thanh
Gastrointestinal (GI) cancers remain major global health challenges, with early detection through colonoscopy crucial for survival. Conventional endoscopes are rigid and operator-dependent, causing discomfort and limiting precision. This project develops an AI-Enhanced Soft Robotic Endoscopy System that integrates AI-driven design automation and SLAM-based real-time 3D navigation to enhance safety, flexibility, and diagnostic accuracy. It advances soft continuum robotics, foundation model–based design, and visual–inertial SLAM for intelligent, adaptive navigation. The project aligns with healthcare innovation priorities, aiming to improve early GI cancer detection, reduce patient trauma, and accelerate the digital transformation of medical robotics through clinical and industrial collaboration.
Read More
Advancing Humanoid Robot Learning through Perception Modeling, World Model, and Biomimetic Machine Intelligence

Advancing Humanoid Robot Learning through Perception Modeling, World Model, and Biomimetic Machine Intelligence

Principal Investigators & Key Members:
Asst. Prof. Le Duy Dung
Current humanoid robots’ capabilities are limited: they excel in short-term tasks but struggle to thrive in long-horizon ones with high levels of uncertainty. Their safety protocols are usually not human-aware, and training them requires extensive expert demonstrations, incurring high budget, time, and effort. Our proposal addresses these challenges by introducing brain-inspired predictive perception modeling, enabling robots to learn internal representation of the world and imagine future outcomes before acting. This further enhances robots’ robustness in uncertain scenarios and improves their safety measures. Therefore, our research will lay the foundation for safer, more versatile, and more trustworthy humanoid robotics.
Read More

Biomorphic Intelligence

Physics-Guided Scientific Machine Learning for Trustworthy Mechanical and Physical Systems

Physics-Guided Scientific Machine Learning for Trustworthy Mechanical and Physical Systems

Principal Investigators & Key Members:
Asst. Prof. Nguyen Vu Linh
This project develops a new generation of Scientific Machine Learning models that combine the rigor of physics with the adaptability of artificial intelligence. By embedding physical laws into AI systems, the research aims to create models that are accurate, interpretable, and efficient for complex mechanical and physical systems such as robots, materials, and industrial machines. The outcomes will enhance predictive capability, safety, and energy efficiency while supporting Vietnam’s Industry 4.0 vision and advancing VinUni’s leadership in trustworthy, physics-guided AI research.
Read More
BioDroneX: AI-Enhanced Bio-Inspired Drone for Adaptive Multi-Environment Missions

BioDroneX: AI-Enhanced Bio-Inspired Drone for Adaptive Multi-Environment Missions

Principal Investigators & Key Members:
Prof. Nguyen Xuan Hung
BioDroneX is an AI-enhanced, bio-inspired drone with computer vision, designed to operate across air, land, and water using adaptive structures, advanced materials, and intelligent navigation. It enables safe access to hazardous or hard-to-reach areas for disaster response, environmental monitoring, smart agriculture, and industrial inspection. By combining biological principles, AI, and advanced 3D printing, the project aims to create autonomous cross-domain drones that reduce human risk and enhance societal resilience.
Read More

Human-Centric AI

3-D MRI-informed AI for Hand Bone 2-D X-ray Image Enhancement

3-D MRI-informed AI for Hand Bone 2-D X-ray Image Enhancement

Principal Investigators & Key Members:
Prof. Saeid Sanei
Carpal (wrist) bone fracture is very popular among ageing community, patients suffering from seizure, stroke, various degenerative brain diseases, and athletes due to impact or falling. The diagnosis procedure includes taking an X-ray and upon detecting the hand bone fracture, the treatment follows. In many cases the subject is sent home due to invisibility of fracture and in over 20% of the cases the patient returns to hospital with more pain and fracture symptoms within 2-3 weeks. At this stage, the doctor asks for taking 3-D structural MRI which provides approximately 100% diagnosis accuracy. However, x-ray is cheap and accessible while MRI is expensive and requires queueing. Therefore, in this project, the plan is to create an intelligent system which learns from pairs of X-ray-MRIs on how to enhance individual X-rays to have the same information visible in MRIs. Such a system alleviates the need for patient return, enhances the diagnosis and considerably reduces the cost for taking MRI. The application can be easily extended to other bone fractures.
Read More
Semantic EEG-to-Speech/text translation with visual feedback and large-vocabulary dictionary

Semantic EEG-to-Speech/text translation with visual feedback and large-vocabulary dictionary

Principal Investigators & Key Members:
Prof. Saeid Sanei
In this project the area of brainwave-to-speech translation, which is in its infancy, will be significantly developed to ensure error-free and semantic speech production. This is performed through visual feedback allowing the subject to decide among the most-likely responses suggested by the system and accept the correct word to be spoken. The major advantage is generating speech for those with full or partial speech disabilities (e.g. autistics, suffering stroke or brain degenerative diseases) covering more than 2.2% of Vietnam population.
Read More
Adaptive Reinforcement Learning for Scalable Traffic Control under Certainty

Adaptive Reinforcement Learning for Scalable Traffic Control under Certainty

Principal Investigators & Key Members:
Asst. Prof. Pham Duc Thinh
As urban traffic transportation faces increasing challenges related to traffic congestion, delays, and emissions. Efficient management of urban traffic networks is essential to enhancing operational efficiency, reducing travel time, and minimizing environmental impact. Despite advances in real-time and adaptive control algorithms, current solutions often fall short due to challenges of the real-world traffic network.This research seeks to develop a control policy that leverages advanced adaptive algorithms for rapidly adapting to environment changes and hierarchical multi-agent reinforcement learning for efficiently handling large-scale coordination complexity. The research supports Smart City goals to reduce congestion, emission and enhance mobility.
Read More