DR-NTU (Digital Repository of NTU)
Not a member yet
116018 research outputs found
Sort by
Maritime navigation by rule-based optimization for global path planning
No full textIn this project, a hybrid autonomous navigation system combining Rule-based
algorithms and Deep Reinforcement Learning (DRL) is developed to address the
challenges of path planning in complex and dynamic environments. Traditional
Rule-based methods, such as A*, Dijkstra, RRT*, DWA, and Potential Field, are
effective in static environments but often struggle when dealing with unpredictable
obstacles or dynamic scenarios. On the other hand, DRL methods, including PPO,
DQN, and SAC, provide strong adaptability but require significant training resources
and may lack global path optimality.
To combine the strengths of both approaches, this project integrates Rule-based
planners with DRL agents within a flexible architecture. Multiple hybrid approaches
are designed and implemented, including A* + PPO, Dijkstra + DQN, RRT* + SAC,
DWA + PPO, and Potential Field + PPO. Each hybrid approach is evaluated in
simulated environments with varying obstacle densities and dynamic elements.
Experimental results demonstrate that hybrid methods significantly outperform
standalone Rule-based or DRL approaches in terms of success rate, path efficiency,
and adaptability in dynamic environments. This project provides valuable insights
into the design of reliable and intelligent navigation systems, with potential
applications in autonomous ships, marine robots, and other mobile navigation fields.Bachelor's degre
Ode to ayah...
No full textAdulthood is a unique developmental phase, marked by the ongoing growth of the prefrontal cortex. During this time, emerging adults often seek independence, while still relying on parental guidance and feeling uncertain about their self-sufficiency. The experience of loss and death during this period can be particularly impactful, bringing unexpected and complex emotions. Though grief is challenging, it can also be a catalyst for personal growth. I aim to shift the perspective on grief by highlighting the importance of relationships and fostering empathy. Instead, I want to demonstrate how loss may make us more appreciative and compassionate of the people in our lives.
Ode to Ayah… is a personal illustrated documentation presented as a series of three diary zines, visually narrating my journey with grief as an emerging adult. Illustrating over more than 100 days of personal reflections, conversations and moments since the loss of my Ayah (Dad), the zines merge both my external observations and internal emotions to portray this visual storytelling. Rather than defining grief, the series delves into how it feels, thus offering readers a reflective space to engage with my grief journey. This project aims to encourage emerging adults to approach grief with resilience while fostering meaningful conversations for future generations.Bachelor's degre
Power splitting in nanoplasmonic waveguide N-junctions I
No full textDue to global health crisis, there has been an increased in demands for biosensors. These
biosensors potentially contain photonics or plasmonic networks and the X/Y/T-Junction arrays
in both photonics and plasmonic networks have been widely studied, but the angles utilised
tends to be idealised value which may not be geometrically realistic and most project uses
Finite-Difference Time-Domain (FDTD) for these simulations. This project uses COMSOL
Multiphysics software which uses Finite Element Method (FEM) to investigate the power
splitting properties of light when angles have been varied in photonics and plasmonic Yjunction waveguides. The results found may contribute to the ongoing work in developing
robust and sensitive biosensors.Bachelor's degre
The third UN ocean conference in Nice: a global summit to safeguard the ocean
No full textThe third United Nations Ocean Conference, held in the French city of Nice from 9 to 12 June 2025, constitutes a significant milestone in the evolving international framework for ocean governance. Despite the adoption of several tangible commitments, the conference also exposed persistent divisions among states on key issues, reflecting the considerable challenges that remain in achieving comprehensive, effective, and sustainable protection of the ocean commons.Published versio
Realistic grasping with deep reinforcement learning in Isaac Sim
No full textSince AlphaGo's landmark victories over Fan Hui (2015), Lee Sedol (2016), and Ke Jie (2017), Deep Reinforcement Learning (DRL) has emerged as a key research area in AI, with applications spanning autonomous driving, robotic control, financial trading, game AI (e.g., Dota 2, StarCraft), and drug discovery. Despite significant theoretical and practical progress, real-world deployment of DRL—particularly in robotics—remains challenging due to the Sim-to-Real Gap. This gap is especially evident in robotic grasping, where traditional rule-based methods lack adaptability to dynamic environments. DRL offers the ability to learn robust, generalizable grasping strategies, but training such models in the physical world is resource-intensive and risky. Consequently, simulations are increasingly used for training. However, many simulators prioritize accurate physics over realistic visuals, limiting their effectiveness for vision-based DRL models that rely on RGB or depth inputs. This visual-physical mismatch, along with real-world complexities like lighting changes and noise, hinders model transferability. To bridge this gap, this paper explores training and transferring the Visual Pushing and Grasping (VPG) method within NVIDIA Isaac Sim. VPG enhances manipulation by integrating pushing and grasping actions, while Isaac Sim provides high-fidelity physics and photorealistic rendering. Its focus on aligning visual and physical realism improves the transferability of learned policies to real-world robotic tasks.Bachelor's degre
Light emission from free electron-driven crystalline materials
No full textThe free-electron induced light emission phenomenon in crystalline materials has drawn significant attention for potential applications in optoelectronics, nanophotonics, and high-definition imaging. Among all the possible materials, van der Waals heterostructures present a new platform for investigating free-electron-induced radiation due to their high-tunability in electronic and photonic characteristics. A systematic investigation into free-electron driven light emission in van der Waals materials is conducted in this dissertation, focusing on parametric X-ray radiation and coherent bremsstrahlung. These mechanisms arise when high-energy electron beams interact with the periodic lattice structures of the materials. These radiation processes present new opportunities in the creation of compact, tunable, and high-brightness X-ray sources.
In order to describe scattering dynamics and the dissipation of energy in a systematic way, Monte Carlo simulations are used. These allow for a complete assessment of the effects of both electronic and structural parameters on X-ray emission. It is reported that variations in the thicknesses and types of the van der Waals materials result in significant changes in the intensity and spectral tuneability of the emitted X-rays. The capability of van der Waals materials to generate multicolor X-rays through engineered design is investigated, demonstrating their potential for highly tunable X-ray sources.
This dissertation deepens the understanding of X-ray radiation production mechanism driven by free electrons from van der Waals materials and points the way toward developing smaller, next-generation X-ray emitters. These results have wide-reaching implications for biomedical imaging, nanoscale spectroscopies and lab-on-a-chip devices, linking free-electron physics with practical X-ray technologies.Master's degre
Explainable machine learning for 2D material layer group prediction with automated descriptor selection
No full textCrystal symmetry is a fundamental aspect of material properties and plays a pivotal role in the discovery and design of new materials. Existing approaches for predicting the symmetries of two-dimensional (2D) materials typically focus on space groups, which often overlook the distinct in-plane and out-of-plane symmetries inherent to 2D systems. To address this, we present Au2LaP (Automated Descriptor Selection Enhanced 2D Material Layer Group Predictor), the first machine learning framework designed to predict layer groups of 2D materials directly from their chemical composition. Au2LaP integrates Light Gradient Boosting Machine (LightGBM) algorithm with SHapley Additive exPlanations (SHAP) for automated descriptor selection, optimizing both predictive accuracy and model interpretability. The explainability of Au2LaP ensures transparency by highlighting the most significant chemical descriptors contributing to layer group classification, thereby enhancing its utility for material discovery and design. Au2LaP outperforms seven state-of-the-art models based on chemical composition, achieving a top-1 accuracy of 0.8102 and a top-3 accuracy of 0.9048. Remarkably, it delivers superior performance using only 20 key descriptors, outperforming models trained with descriptor sets as large as 546. Furthermore, we demonstrate that Au2LaP can effectively predict polymorph structures by identifying multiple possible layer groups for given compositions, further advancing the predictive modelling of new material phases. This work sets a new benchmark for 2D material symmetry prediction, paving the way for more efficient crystal structure prediction, polymorph studies, and material design.National Research Foundation (NRF)AI SingaporeThis work acknowledges the support from National Research Foundation, Singapore, under its Competitive Research Programme (CRP) (NRF-CRP22-2019-0007). This research is also supported by the National Research Foundation Singapore and DSO National Laboratories under the AI Singapore Programme (AISG Award No: AISG2-GC-2023-009)
Leveraging enterprise resource planning system for effective project budget and cost management
No full textIn today’s dynamic and increasingly complex project environments, organizations face significant challenges in maintaining financial control due to fragmented data, evolving project scopes, and cost overruns. Traditional cost management approaches often lack real-time visibility, leading to inefficiencies and financial risks. To address these issues, many company are adopting Enterprise Resource Planning (ERP) systems to enhance project budgeting and cost management. ERP systems provide a centralized platform that integrates project-related processes, automates financial workflows, and ensures seamless access to consolidated data, enabling better decision-making and resource allocation. However, despite their advantages, ERP implementations often encounter challenges such as high costs, limited human resource support, customization difficulties, process misalignment, and resistance to change.
This research examines how ERP systems enhance operational efficiency by streamlining project processes, improving data accuracy, and facilitating cost control. Particularly, this survey explores the function of ERP system templates and best practices in aligning business processes with project-specific requirements, ensuring organizations can maximize the benefits of ERP adoption while mitigating implementation risks. The research using a mixed-method approach, include qualitative interviews with key stakeholders involved in project execution, primarily within the construction industry and a questionnaire survey conducted with 50 participants who have up to one year of ERP system experience. Additionally, a literature review provides further context on ERP adoption trends and challenges.
By evaluating the effectiveness of ERP systems in improving project financial management, this study offers valuable insights into both their advantages and the obstacles faced during implementation. The findings highlight that cost constraints and resistance to change are the primary obstacles to successful ERP adoption. Survey results and interview feedback indicate that securing sustainable financial support is a critical challenge, as organizations often struggle with the high budget required for ERP implementation and maintenance cost. As a result, insufficient budgets often lead to limited system functionality, inadequate user training, and delays in integrating essential modules. Additionally, resistance from ERP users who stemming from unfamiliarity with the system, lack of training, and reluctance to adapt to new processes and further complicates adoption. This resistance leads to low user engagement, frequent operational errors, and inefficient system use, ultimately undermining productivity, delaying full implementation, and reducing the overall return on investment.
Through a structured analysis of ERP-driven project budgeting strategies, this research provides the development of best practices for organizations seeking to optimize financial performance in complex project environments. A key recommendation for successful ERP implementation is cost optimization through a phased adoption approach. This strategy involves breaking down the implementation process into manageable phases, enabling periodic reviews of costs, resource allocation, and system performance by management and key users.Master's degre
Improvement of optical coherence tomography images from patients with diabetic retinopathy using generative deep learning
No full textDiabetic retinopathy (DR) is a common complication of diabetes and is the leading cause
of vision loss in adults. Optical Coherence Tomography (OCT) offers a way to diagnose
DR and monitor the disease progression, but its widespread use is hindered by high cost
and limited availability of high-quality OCT images. Even though Spectral-domain OCT
(SDOCT) is most commonly used in clinics due to its low cost, it lacks penetration depth
and can hinder clinical assessment for DR. On the other hand, swept-source OCT (OCT)
offers higher penetration depth, but is not widely available due to higher cost. As such,
there is a gap for a low cost but accurate method of DR assessment using OCT images.
This project explores the potential of using a previously developed generative deep
learning model, based on the Pix2Pix GAN architecture, to synthetically enhance the lower
cost SDOCT images to achieve image quality that is comparable to SSOCT images for DR
analysis.
The Diabetes Study in Nephropathy and other Microvascular Complications (DYNAMO)
dataset which contained B-scan volumes of eyes with varying DR severity was used to
evaluate the model’s performance for this project. Three clinically relevant measures were
selected for this analysis: choroidal thickness, choroidal area and choroidal vascularity
index. Results show that there is strong correlation in all three measures, which showcases
the model’s ability to learn and reconstruct choroidal structures accurately.
Throughout the project, significant improvements were made on the pre-processing
workflow, which included a semi-automatic retinal pigment epithelium (RPE) flattening
algorithm and image registration to produce good quality inputs for the generative deep
learning model. Despite the improvements, the RPE flattening process still requires manual
intervention. A fully automated segmentation solution is needed to further streamline the
workflow and scale the approach.Bachelor's degre
Text-guided city street images semantic segmentation with large-scale vision language model
No full textThis dissertation proposes a text-guided open-set semantic segmentation method in
city street images based on a large-scale vision-language model. Traditional semantic
segmentation models often struggle with unknown classes in city street images. To
address this challenge, we integrate large-scale vision-language models with
text-guided city street images segmentation, leveraging cross-modal alignment to
enhance recognition of unknown classes. By using natural language descriptions, the
method achieves more fine-grained object recognition and segmentation.
Experimental results demonstrate that our approach outperforms existing methods
across multiple public data sets, showcasing superior generalizability and scalability.Master's degre