Research Projects
Innovative AI research initiatives bridging academic excellence with real-world applications
Active Projects
Arm Robot for Bin-Picking in Unstructured Environments
The project aims to develop an intelligent bin-picking system that enables robots to detect, localize, and grasp objects in cluttered, unstructured environments. By combining advanced vision techniques such as instance segmentation, graph-transformer reasoning, depth refinement, shape completion, and 6D pose estimation, the project seeks to overcome challenges like occlusion and sensor noise.
Key Technologies: Instance Segmentation, Graph Transformers, 6D Pose Estimation
Status: In Progress
IntelliScrape: An AI-assisted Web Scraping Module Generator
The project aims to develop an AI-assisted module development system that partially automated COOCON’s scraping module creation process. Leveraging artificial intelligence, network traffic analysis, and browser automation technologies, the proposed system aims to reduce manual effort, minimize human error, and accelerate the overall development lifecycle of COOCON developer.
Key Technologies: LLMs, Browser Automation, Network Traffic Analysis
Status: In Progress
AI-Driven Administration (AIDA)
MPTC AI-Driven Administration is a collaborative initiative between the Ministry of Post and Telecommunications of Cambodia (MPTC) and CBNU's AICLab. The project applies LLM technologies to address operational inefficiencies in government administration. Its goal is to develop AI-driven tools that streamline workflows, improve information processing, and enhance organizational efficiency, ultimately supporting faster and more consistent public service delivery.
Key Technologies: Large Language Models, Administrative Automation, Decision Support Systems
Status: In Progress
Research Interests
Beyond our active projects, AICLab pursues several long-term research directions shaped by the collective expertise of our team. These directions are methodologically broad and increasingly intersect with data-intensive scientific domains, including computational biology and biomedical research, where we welcome interdisciplinary collaboration.
NL2SQL: Natural Language to SQL
This research develops an AI system that converts natural-language questions into accurate SQL queries, removing SQL expertise as a barrier to data access. The research focuses on schema understanding, query validation, hallucination reduction, and automated dataset curation using multiple AI agents. The lab's MRS-Agent submission reached rank #21 on the BIRD benchmark, a large-scale cross-domain Text-to-SQL evaluation, achieving 75.13% execution accuracy on the test set. The same methods apply naturally to scientific data access, where large research databases remain difficult to query without specialized training.
Key Technologies: Multi-Agent Systems, Schema Understanding, Query Validation, Hallucination Reduction, Text-to-SQL
Related Publication: QA-SQL: Query-Augmented SQL generation using few-shot prompting with data augmentation, PeerJ Computer Science (2025). View Paper
Quantum Computing: Large-Scale Optimization
This research explores quantum annealing and hybrid quantum-classical methods for tackling highly complex, large-scale optimization problems that are difficult for conventional computing approaches. The work aims to apply these methods to domain-specific challenges such as synchrotron accelerator scheduling, parameter optimization, resource allocation, and operational control. Because these optimization methods generalize across domains, the lab is also interested in computationally hard problems in molecular modeling and structural analysis relevant to the life sciences.
Key Technologies: Quantum Annealing, Hybrid Quantum-Classical Algorithms, Combinatorial Optimization, Quantum Error Mitigation
Related Publication: Strategies for Noise-Resilient Quantum Approximate Optimization Algorithms: A Review and Classification of Error Mitigation, IEEE Access (2025). View Paper
AI Scientist: Automating Scientific Discovery
This research develops an AI system that supports scientific research by generating ideas, designing experiments, analyzing results, and checking whether the findings are reliable. Multiple AI agents work together to reduce errors and improve the quality and reproducibility of the research. The direction draws on the team's combined experience in backend systems engineering, data analytics, and multimodal AI, and is especially suited to literature-dense fields such as the life sciences, where automated synthesis of published findings can accelerate hypothesis generation.
Key Technologies: Multi-Agent Systems, Large Language Models, Experimental Design, Reproducibility Validation
Related Publication: AI-Scientist Web-Based System for Streamlined Scientific Paper Generation, BIGDAS Conference (2024). View Conference
Past Projects
Fall Detection Systems
The project aims to develop a real-time fall detection system using YOLO for high-speed human pose estimation to extract skeletal keypoints from video streams. Through feature engineering, it computes a dozen biomechanical angles capturing body orientation and velocity, which are then fed into a classification-based model to accurately distinguish intentional movements from sudden, unintentional falls.
Key Technologies: YOLO, Pose Estimation, Classification Models
Status: Completed
Automated Motorcycle Helmet Violation Detection
The project aims to enhance motorcycle rider safety and aid law enforcement in Vietnam by automatically detecting riders without helmets using surveillance camera footage. The proposed method uses the YOLOv5 object detection algorithm, a state-of-the-art model custom-trained on a local dataset to detect three specific classes: "Helmet", "No-helmet", and "Rider". A critical post-processing stage is then applied to analyze the relationships between these detected objects.
Key Technologies: YOLOv5, Object Detection, Post-Processing
Status: Completed
Real-Time Abandoned Object Detection System
The project aims to develop an automated surveillance system to enhance public safety by detecting abandoned objects (specifically luggage and paper boxes) in real-time. The system is designed to minimize security risks and false alarms by distinguishing between attended and unattended items using CCTV footage. The core of the system is a custom-trained YOLOv8 model, which performs initial object detection to identify "persons", "luggage", and "paper boxes". This is integrated with the ByteTrack algorithm to assign unique IDs and track the movement of these objects across consecutive frames.
Key Technologies: YOLOv8, ByteTrack, Object Tracking
Status: Completed