Human-robot interaction (HRI) enables intuitive and intelligent collaboration between humans and robots in real-world environments. This paper introduces a humanoid robot prototype designed as a flexible testbed for developing and integrating artificial intelligence (AI) modules in HRI tasks. The system features a 12 degree-of-freedom (DOFs) dual-arm mechanism and a 2 DOFs head with an expressive LCD screen to express facial emotions. All hardware components are controlled by a custom-designed controller board with real-time AI processing supported by an onboard Jetson module. The system incorporates three AI modules: (1) gesture recognition using MediaPipe Pose and an LSTM classifier, (2) object detection with YOLO and 3D localization, and (3) voice-command processing through speech recognition and large language model(LLM)-based semantic parsing. The platform is validated through experiments on positioning accuracy, with results showing average manipulation errors of approximately 1.83 cm. To demonstrate its versatility, experimental results show over 90
Sydney Brenner, Nobel Laureate in Physiology, once said, “Progress in science depends on new techniques, new discoveries, and new ideas, probably in that order.” A paradigm-shifting technology of our time is artificial intelligence (AI), which offers new ways to accelerate scientific discovery. This talk will explore how AI is transforming science by automating discovery workflows—representing and leveraging scientific artefacts and knowledge, predicting properties and phenomena, optimising and generating new designs, and reasoning about competing hypotheses. We will also discuss the emerging concept of AI co-scientists: autonomous agents that emulate the scientific discovery process by combining creative exploration with systematic reasoning. These agents collectively propose hypotheses, design experiments, analyse results, and communicate findings while collaborating with human scientists. Specific examples will be drawn from molecular space exploration, drug–protein interactions, chemical synthesis, crystal structure prediction, and property prediction for advanced materials such as superconductors and alloys.
The advent of 6G networks promises unprecedented advancements in connectivity, low latency, and massive machine-type communication, setting the stage for a future where artificial intelligence (AI) will play a pivotal role in network management, automation, and optimization. However, the integration of AI into 6G networks also introduces a wide range of security challenges that need to be addressed to ensure the safe and reliable operation of these next-generation networks. This talk will explore the evolving security landscape of AI-driven 6G networks, focusing on potential threats, vulnerabilities, and mitigation strategies. Key topics include the risks of AI-powered attacks, such as adversarial machine learning, data poisoning, and model inference attacks, as well as vulnerabilities in network slicing, edge computing, and Internet of Things (IoT) devices that are expected to proliferate in 6G environments. The seminar will also discuss the importance of trust, privacy, and security in AI-based decision-making processes.
The integration of digital twins into smart manufacturing has emerged as a transformative approach, revolutionizing the way factories operate, monitor, and optimize performance. In Malaysia's push toward Industry 4.0 and its upcoming transition to Industry 5.0, digital twinenhanced communication architectures play a critical role in ensuring real-time data synchronization, predictive analytics, and intelligent automation. This paper presents a comprehensive investigation into the design, implementation, and optimization of digital twin-augmented communication infrastructures tailored to Malaysia's smart manufacturing context. Leveraging edge computing, industrial IoT, 5G, and semantic interoperability standards, we propose a multilayered communication framework that addresses latency, data reliability, and system scalability challenges. Our research combines simulation modeling, industrial testbeds, and stakeholder interviews across key industrial corridors in Selangor and Penang. Findings indicate that digital twinenabled communication systems improve real-time process visibility by 58%, reduce downtime by 35%, and enhance predictive maintenance accuracy by 62% compared to conventional setups. Through architectural modeling, latency analysis, and digital twin synchronization protocols, the study demonstrates the viability of the proposed framework in small-to-medium enterprises (SMEs) and multinational manufacturing plants alike. The results underscore the necessity of standardized digital twin interfaces (e.g., OPC UA, MQTT, and ISO 23247), edge-driven latency mitigation strategies, and cloud-twin data fusion mechanisms to ensure full-loop control and scalability. This work contributes a technically validated, policy-aligned, and future-proof architecture that supports Malaysia's Smart Manufacturing Blueprint (2021-2030) and aligns with the MyDigital roadmap.
This paper presents Parameter-Space Distillation (PSD), a dataset distillation approach that compresses datasets by embedding their knowledge into the parameters of a pretrained generative model, rather than optimizing synthetic pixels as in traditional methods. This parameter-centric design improves flexibility across distillation ratios and network architectures while enabling faster and more stable training. Instead of storing fixed synthetic images, PSD dynamically generates diverse training samples and minimizes prediction discrepancies between real and generated data across multiple architectures. Experiments on CIFAR10, SVHN ($32 \times 32$), and the higher-resolution Imagenette dataset ($128 \times 128$) show that PSD achieves higher accuracy with fewer distilled samples and generalizes better across networks. These results highlight the effectiveness of learning dataset information directly in the parameter space of generative models. Full experimental details are provided in Section 4.
Labeling large-scale datasets in domains such as traffic analysis remains a labor-intensive and error-prone process, with mislabeling often leading to suboptimal deep learning performance. To address this challenge, we introduce a semiautomatic labeling system that combines the object detection capabilities of YOLOv11 with the semantic verification power of CLIP (ViT-B/32). In our framework, YOLOv11 first generates object labels, which are then verified and calibrated by CLIP. Labels exceeding a confidence threshold are retained, while uncertain cases are corrected using CLIP and subsequently validated by the end user. This pipeline significantly reduces manual labeling effort while ensuring consistent quality. The performance in terms of confusion representation, loss, precision, recall, mean average precision, and confidence results is evaluated. Experimental results demonstrate that the proposed system achieves an accuracy of over 90 %, effectively leveraging the complementary strengths of YOLOv11 in image-based detection and CLIP in multimodal semantic understanding. The approach offers a practical solution for large-scale traffic dataset annotation, striking a balance between efficiency and reliability.
Traditional keyword search fails to unlock knowledge from unstructured enterprise data. This paper introduces an AI agent that improves document lookup using a hybrid Retrieval-Augmented Generation (RAG) pipeline, which combines sparse, dense, and reranking techniques for high accuracy. The agent is made extensible with the Model Context Protocol (MCP) for external tool use, such as web searches. Our approach achieves a state-of-the-art nDCG@5 of 0.91, significantly outperforming keyword search (0.62) and standard RAG (0.76), while also enabling conversational queries, multidocument synthesis, and external knowledge integration.
The aerodynamic stability derivatives of a flying object are crucial parameters for evaluating its aerodynamic quality, control system performance, and overall flight characteristics. In recent years, the use of Computational Fluid Dynamics (CFD) to determine these stability derivatives has gained widespread adoption due to its high reliability. Among different approaches, the forced oscillation method derived from dynamic tests in wind tunnel has been widely used thanks to their physical similarity. In this study, numerical simulations of an Army-Navy Basic Finner projectile undergoing forced sinusoidal pitching oscillation are carried out to predict longitudinal dynamic stability derivatives. This study utilizes the open-source CFD code SU2 to perform the transient harmonic pitching simulations in a wide range of angles of attack from 0 to 85° and extract aerodynamic derivatives related to pitching moment coefficient. Computational results demonstrate a good agreement with experimental wind tunnel data. The results suggest the potential of the SU2 framework to predict dynamic stability derivatives for complex, newly designed flying bodies or those experiencing unsteady aerodynamics.
In real-world settings, facial recognition systems frequently suffer reduced accuracy because of changes in lighting, the use of glasses, and obstructions such as protective masks. This study proposes a lightweight, interpretable framework that leverages geometric facial keypoints and support vector machine classification to enhance recognition robustness in such challenging conditions. Two specialized pipelines were developed: one optimized for clear, lowlight, and glasses scenarios, and another designed for mask-induced occlusion using upper-face landmarks. Evaluation on a six-participant dataset revealed that both models achieved accuracy rates approaching 90 %, confirming the practicality of SVM landmark techniques for robust face recognition across unconstrained and masked settings.
Table detection and structure recognition are fundamental tasks in document image analysis, as they play a critical role in information extraction from digital and scanned documents. Yet, existing methods often suffer from imprecise localization and noisy boundary predictions. Building on the success of PyramidTabNet (PTN), we propose PyramidTabNet++ (PTN++), which integrates a Pyramid Vision Transformer backbone, a Cascade R-CNN detector, and an enhanced boundary line correction module for table detection. This combination leverages the strengths of transformer-based feature extraction, multi-stage object detection, and geometry-aware refinement to achieve more precise and reliable table localization. Extensive experiments on public benchmarks, including ICDAR-2017 POD, ICDAR-2019, and TableBank, demonstrate that PTN++ achieves accurate and robust performance across diverse datasets.
The increasing prevalence of digital image manipulation and AI-generated content has posed significant challenges to the image authenticity verification. These challenges, in turn, facilitate the spread of misinformation in the media and forensic applications. We propose a novel hybrid image forensics model to classify images into three different categories: authentic, spliced and generated by diffusion modeling. Existing methods lack a unified framework for three-class classification of authentic, spliced and diffusion-generated images. To address this gap, the proposed model integrates traditional forensic techniques, specifically Error Level Analysis (ELA) for splicing detection and frequency analysis using Fast Fourier Transform (FFT) for identifying artifacts from diffusion-based generation, with a lightweight Convolutional Neural Network (CNN) for robust feature extraction. This research advances digital image verification by introducing a unified framework for detecting various forms of image manipulation, thereby supporting broader efforts to combat misinformation in synthetic media.
Plug and Play (PnP) technology has demonstrated superior efficacy in modern computing systems by enabling automatic device detection and configuration without manual intervention. Building upon this principle, this research proposes LitePnP, a lightweight Plug-and-Play model specifically designed for resource-constrained IoT devices, which typically lack comprehensive operating systems and are limited in memory capacity, processing power, and storage capabilities. The model focuses on three primary components: (1) a sensor identification mechanism based on active and passive scanning through RS485/Modbus RTU protocols; (2) lightweight JSON-based drivers that enable sensor operation abstraction while reducing memory overhead; and (3) a minimalist PnP process that faciliates seamless driver installation and service deployment. The significant contribution of this research lies not only in proposing the lightweight PnP model but also in comparing it with typical PnP models such as MicroPnP, IEEE 1451, and TEDS. The comparison demonstrates that while current models provide comprehensive functionality or high standardization, they often require specialized hardware or substantial resources. In contrast, LitePnP model offers minimalism and ease of integration, making it particularly suitable for low-cost IoT devices and opening up a viable approach for large-scale applications in smart agriculture, environmental monitoring, and industrial automation.
In this paper, we investigate cheating behaviors in computer-based examinations and propose an effective detection approach based on candidate behavioral analysis. The dataset was collected through simulated exam sessions and subsequently annotated frame by frame to support model training. To extract relevant features of the head, eyes, and mouth, we employed the Face Mesh framework from Mediapipe. We further introduce AnoResLSTM, a hybrid model combining a ResNet module with an AttLSTM network, designed to enhance the representation and analysis of sequential behavioral patterns. Experimental results demonstrate that our proposed method achieves high accuracy in detecting cheating while maintaining efficient processing time.
Wireless Sensor Networks (WSNs) have permeated many aspects of our lives, attracting significant interest from researchers and industry professionals. This paper analyzes latency and energy performance in a WSN assisted by Unmanned Aerial Vehiclen (UAV) that employs backscatter communication. We model a scenario where sensor nodes offload data to a UAV serving as a hybrid access point and a fixed base station. The UAV is equipped with a backscatter reader and a lightweight Mobile Edge Computing (MEC) server, capable of migrating computing tasks to the more powerful base station server. We derive closed-form expressions for the successful computation probability (SCP) and energy outage probability (EOP) of the proposed system. We investigate the impact of key factors such as transmit power, UAV altitude, number of users, and offloading task allocation ratio on system performance. MonteCarlo simulations are provided to validate our analysis and offer deeper insights into the system's behavior.
We present a deployment-ready, real-time vehicle re-identification (Re-ID) pipeline designed for motorcycle-dominated, non-lane urban traffic. The system integrates Apache Kafka and Spark Structured Streaming with modern detection, tracking, and embedding models, enabling city-scale streaming operation with temporalgated FAISS retrieval and persistent metadata for analysis. To support evaluation under realistic conditions, we release a new multi-camera dataset from Ho Chi Minh City that captures dense occlusions and irregular flows often missing in existing benchmarks. Our work contributes (i) a scalable big-data Re-ID pipeline, (ii) a motorcycle-centric dataset with reproducible evaluation, and (iii) a foundation for practical intelligent transportation and urban security applications in developing cities. The source code and dataset are publicly available at: https://github.com/HaPXV/A-Streaming-Big-Data-pipeline-for-Real-Time-Vehicle-Re-Identification-in-Chaotic-Urban-Traffic.
Permissioned blockchains using Proof-ofAuthority (PoA) deliver high throughput but face issues of predictability and centralization, while token-weighted governance risks plutocracy that undermines fairness. This paper proposes Proof-of-Merit (PoM), a consensus and governance framework that integrates PoA with Verifiable Random Functions (VRFs) and a dual-token model. PoM selects validators through a weighted combination of transferable stake (UIT-Coin) and non-transferable academic reputation (UIT-Rep), earned via verifiable onchain learning activities. Governance follows the same principle, anchoring voting rights in Sybil-resistant merit rather than pure capital. To ensure sustainability, PoM introduces reputation decay, preventing long-term power concentration and promoting continuous participation. We implement PoM on Hyperledger Besu and evaluate it with Hyperledger Caliper. Results show PoM achieves strong performance while significantly improving fairness, with a much lower Gini coefficient and higher Nakamoto coefficient compared to IBFT 2.0. Sensitivity analysis further highlights the need for dynamic reputation mechanisms to avoid saturation. These contributions establish PoM as a scalable, equitable, and sustainable foundation for Learn-to-Earn ecosystems, where influence derives from ongoing educational engagement instead of wealth accumulation.
This research addresses the difficulty of precisely identifying coconut maturity, which is essential for the Philippine agricultural sector and currently relies on unreliable techniques such as tapping guided by skilled judgment. To provide a more dependable solution, this study employs convolutional neural networks (CNNs) particularly transfer learning models for categorizing coconuts into three groups: mature, pre-mature, and young. A collection of 3,000 coconut images was collected and augmented, with an equal number at each of the three stages of ripeness and then enhanced to increase variation. Before training, the images were preprocessed by resizing and normalizing before applying three pre-trained models (EfficientNet, MobileNetV2, and VGG16). The assessment involved measurements such as accuracy, rate of errors, sensitivity, specificity, false positive rate, precision, and F1-score. Results showed that the VGG16 model achieved the highest accuracy of 96.17 %, followed by EfficientNet at 80.17 %, and MobileNetV2 at 65.33 %. This demonstrates how CNNs are more reliable than traditional methods for categorizing coconut ripeness. The device was made by incorporating the high-performing VGG16 model into the Jetson Nano microprocessor. This offers a feasible and scalable answer for enhancing farming practices and production methods within the coconut industry. This development could greatly enhance efficiency and standards in coconut farming, which would help farmers and the overall agricultural industry, ultimately promoting economic growth and sustainability.
Command-line interfaces like Bash are essential tools for developers and system administrators, enabling efficient interaction with operating systems. However, existing support tools, such as auto-completion, still rely on simple solutions based on history lookups or pre-defined, static rules, failing to learn from individual user workflows and adapt to personalized patterns. This paper presents a lightweight, data-driven approach for Bash command recommendation using a statistical n-gram model designed to run efficiently on typical hardware without NPUs/GPUs or cloud-based services, thereby ensuring data privacy and low latency. This paper also introduces a novel sessionized dataset constructed from GitHub user histories and Q&A forums. Our model handles two key tasks, which are are intra-command completion and inter-command prediction. Experiments show high effectiveness and efficiency: key results include $\mathbf{7 0 \%}$ Top-1 accuracy for intra-command completion, over $\mathbf{7 1 \%}$ Top-3 accuracy for inter-command, and latencies under 50 ms on a mid-range CPU. This work confirms that a lightweight approach can provide a more personalized and practical command line experience, outperforming rule-based tools like bash-completion in adaptability while avoiding the computational overhead and data privacy risks of cloud-based LLMs.
With the rapid advancement of information technology, the integration of digital solutions such as Quick Response codes and facial recognition into event attendance management has become an essential trend to enhance efficiency, security, and user experience. The combined system of Quick Response codes and facial recognition streamlines check-in procedures, minimizes fraud risks, optimizes resource utilization, and elevates the prestige and reputation of events. This research aims to develop a modern, practical system that fosters the application of cutting-edge technology in event organization and customer information management. In this paper, a design of a mobile system leveraging computer vision, incorporating key components such as a camera, a Quick Response code scanner, a Raspberry Pi embedded computer, an LCD, and a loudspeaker, is proposed. The system was tested in various real-world environments, achieving an average check-in time of 20 seconds per user and a false positive rate of approximately 7 %, based on over 88 experiments involving 20 real customers. These results demonstrate the promising potential of the proposed design for future development and implementation.
While modern phishing detection models achieve high accuracy, they often remain vulnerable to sophisticated adversarial attacks. This paper argues that enhancing adversarial robustness is a critical challenge that requires augmenting models with external contextual knowledge. We introduce a novel framework centered on Retrieval-Augmented Generation (RAG) as a core defensive mechanism. To effectively integrate this external knowledge with multimodal inputs (URLs, HTML, structured features), we propose a Selective Cross-Attention Fusion mechanism. This fusion technique is specifically designed to maximize the benefits of RAG by selectively applying context only to semantically rich modalities, thereby improving signal alignment and minimizing statistical noise. Through systematic experiments against 4,000 multi-layered adversarial samples, we demonstrate that our RAG-integrated framework significantly reduces the Attack Success Rate (ASR) compared to a strong baseline. The results confirm that our dual-pronged approach-leveraging RAG for robustness and a specialized fusion mechanism for effective integration-establishes a more resilient and interpretable defense against modern phishing threats.
