Search in the TU Delft Repository Collections
Recently Added Records
Analysis of tunnel segment connection selection under differential settlement using a multi-criteria decision-making approach
A Case Study of the Oosterweelknoop Project in Antwerp
Tunnel segment connections in cut-and-cover tunnels founded on shallow foundations are subjected to differential settlements during their service life. This research develops a systematic methodology for the assessment and selection of tunnel segment connections using a combination of geotechnical modeling, structural verification, and Multi-Criteria Decision Analysis (MCDA). The Oosterweelknoop project in Antwerp was used as a case study, considering four connection concepts: a shear key, collar construction, dowel connection, and free-settlement joint. The results show that the dowel connection provides the most suitable balance between robustness, feasibility, durability, sustainability, and cost. The developed methodology supports transparent and reproducible decision-making for similar tunnel projects.
...
Tunnel segment connections in cut-and-cover tunnels founded on shallow foundations are subjected to differential settlements during their service life. This research develops a systematic methodology for the assessment and selection of tunnel segment connections using a combination of geotechnical modeling, structural verification, and Multi-Criteria Decision Analysis (MCDA). The Oosterweelknoop project in Antwerp was used as a case study, considering four connection concepts: a shear key, collar construction, dowel connection, and free-settlement joint. The results show that the dowel connection provides the most suitable balance between robustness, feasibility, durability, sustainability, and cost. The developed methodology supports transparent and reproducible decision-making for similar tunnel projects.
Coordinated Infrastructure Improvements for European Military Mobilisation
A Network Analysis approach to optimise corridor‑level infrastructure upgrades for European military mobilisation
The ability of armed forces to rapidly deploy personnel, equipment, and supplies across Europe has become increasingly important in response to the deteriorating international security environment, following Russia's large-scale invasion of Ukraine in 2022. Despite renewed political attention and investment, large-scale military mobility through Europe remains constrained by a combination of fragmented governance and inadequate transport infrastructure. Rather than evaluating individual infrastructure objects in isolation, this thesis develops a network-based approach to identify and prioritise infrastructure upgrades according to their contribution to overall network efficiency, under military mobility constraints. The results show that improving military mobility is not about spending more, but about making smarter choices. With only a limited number of targeted interventions, Europe can greatly strengthen its military mobility, and with it its deterrence and security. In the end however, the challenge lies not in reinforcing the infrastructure itself, but in prioritising investments in such a way that strategic necessity algins with political reality.
...
The ability of armed forces to rapidly deploy personnel, equipment, and supplies across Europe has become increasingly important in response to the deteriorating international security environment, following Russia's large-scale invasion of Ukraine in 2022. Despite renewed political attention and investment, large-scale military mobility through Europe remains constrained by a combination of fragmented governance and inadequate transport infrastructure. Rather than evaluating individual infrastructure objects in isolation, this thesis develops a network-based approach to identify and prioritise infrastructure upgrades according to their contribution to overall network efficiency, under military mobility constraints. The results show that improving military mobility is not about spending more, but about making smarter choices. With only a limited number of targeted interventions, Europe can greatly strengthen its military mobility, and with it its deterrence and security. In the end however, the challenge lies not in reinforcing the infrastructure itself, but in prioritising investments in such a way that strategic necessity algins with political reality.
Context-Aware Teleimpedance Through hierarchical LLM–VLM Reasoning
For contact-rich tasks
Master thesis
(2026)
-
H.S. Sathyanarayanan, L. Peternel, Jasper Schol, Cock Heemskerk, Carlos Hernández
Contact-rich robotic teleoperation requires continuous impedance regulation for safe physical interaction, but manually configuring multi-dimensional stiffness parameters places a substantial burden on operators. To address this, this paper
presents a zero-shot, context-aware hierarchical LLM-VLM reasoning framework that estimates task-dependent Cartesian stiffness parameters directly from natural-language instructions and visual scene context. The system integrates a Vision-Language Model (VLM) and a dual Large Language Model (LLM) pipeline
to generate six-degree-of-freedom stiffness profiles, alongside a human-in-the-loop web interface for safe oversight. Evaluation through a controlled user study (N = 30) demonstrated that AI assistance reduced median task completion time (15 s vs. 32 s) and parameter adjustments (3 vs. 10), while decreasing cognitive workload (p < 0.001).A real-world deployment on a PAL TIAGo Pro mobile manipulator validated the framework’s ability to modulate stiffness proportionally to physical demands across four cleaning tasks Scrubbing Dried Coffee (6.64 N) > Erasing Pencil Marks (5.54 N) > Wiping Spilled Water (2.47
N) > Sweeping Loose Dust (1.97 N) without requiring task specific training data. These findings demonstrate that semantic reasoning can effectively bridge high-level intent and low-level robot compliance. ...
presents a zero-shot, context-aware hierarchical LLM-VLM reasoning framework that estimates task-dependent Cartesian stiffness parameters directly from natural-language instructions and visual scene context. The system integrates a Vision-Language Model (VLM) and a dual Large Language Model (LLM) pipeline
to generate six-degree-of-freedom stiffness profiles, alongside a human-in-the-loop web interface for safe oversight. Evaluation through a controlled user study (N = 30) demonstrated that AI assistance reduced median task completion time (15 s vs. 32 s) and parameter adjustments (3 vs. 10), while decreasing cognitive workload (p < 0.001).A real-world deployment on a PAL TIAGo Pro mobile manipulator validated the framework’s ability to modulate stiffness proportionally to physical demands across four cleaning tasks Scrubbing Dried Coffee (6.64 N) > Erasing Pencil Marks (5.54 N) > Wiping Spilled Water (2.47
N) > Sweeping Loose Dust (1.97 N) without requiring task specific training data. These findings demonstrate that semantic reasoning can effectively bridge high-level intent and low-level robot compliance. ...
Contact-rich robotic teleoperation requires continuous impedance regulation for safe physical interaction, but manually configuring multi-dimensional stiffness parameters places a substantial burden on operators. To address this, this paper
presents a zero-shot, context-aware hierarchical LLM-VLM reasoning framework that estimates task-dependent Cartesian stiffness parameters directly from natural-language instructions and visual scene context. The system integrates a Vision-Language Model (VLM) and a dual Large Language Model (LLM) pipeline
to generate six-degree-of-freedom stiffness profiles, alongside a human-in-the-loop web interface for safe oversight. Evaluation through a controlled user study (N = 30) demonstrated that AI assistance reduced median task completion time (15 s vs. 32 s) and parameter adjustments (3 vs. 10), while decreasing cognitive workload (p < 0.001).A real-world deployment on a PAL TIAGo Pro mobile manipulator validated the framework’s ability to modulate stiffness proportionally to physical demands across four cleaning tasks Scrubbing Dried Coffee (6.64 N) > Erasing Pencil Marks (5.54 N) > Wiping Spilled Water (2.47
N) > Sweeping Loose Dust (1.97 N) without requiring task specific training data. These findings demonstrate that semantic reasoning can effectively bridge high-level intent and low-level robot compliance.
presents a zero-shot, context-aware hierarchical LLM-VLM reasoning framework that estimates task-dependent Cartesian stiffness parameters directly from natural-language instructions and visual scene context. The system integrates a Vision-Language Model (VLM) and a dual Large Language Model (LLM) pipeline
to generate six-degree-of-freedom stiffness profiles, alongside a human-in-the-loop web interface for safe oversight. Evaluation through a controlled user study (N = 30) demonstrated that AI assistance reduced median task completion time (15 s vs. 32 s) and parameter adjustments (3 vs. 10), while decreasing cognitive workload (p < 0.001).A real-world deployment on a PAL TIAGo Pro mobile manipulator validated the framework’s ability to modulate stiffness proportionally to physical demands across four cleaning tasks Scrubbing Dried Coffee (6.64 N) > Erasing Pencil Marks (5.54 N) > Wiping Spilled Water (2.47
N) > Sweeping Loose Dust (1.97 N) without requiring task specific training data. These findings demonstrate that semantic reasoning can effectively bridge high-level intent and low-level robot compliance.
Selective antimicrobial peptides
Machine learning driven design and discovery
Antibiotic resistance has become a major health risk globally. Consequently,
antimicrobial peptides (AMPs) have been recognised as potential alternatives to
antibiotics. Much computational work has been done on the classification of peptides
into AMPs or non-AMPs, however, sparsity remains in the methods which address
selectivity. The selectivity is described as the ability of a peptide to act against
a target species while not inhibiting off-target species. This research explores a
machine learning pipeline for the discovery and optimisation of selective AMPs.
To achieve this, a standardised database was constructed, called sAMPle, giving a
total of 38,978 peptide-species combinations with corresponding minimum inhibitory
concentrations (MIC). Peptide features were represented using ESM-2 embeddings,
a well-known protein language model, while bacteria were represented by name-
based and taxonomy-based features. Two overlapping modelling approaches were
employed, a joint interaction model, which was trained on all bacterial species
simultaneously, and species-specific regression models. The best performing model
was a species-specific multilayer perceptron, achieving a root mean square error of
0.55–0.57 depending on the species, which is comparable to other MIC prediction
methods found in literature. Subsequently, the MIC predictions were used to define
a selectivity score, applied to the most prominent bacterial species in the sAMPle
dataset, namely E. coli and S. aureus, which are also clinically relevant species. The
selectivity scoring provided a ranked dataset in which the most selective peptides
were the known peptides apidaecin-1A and nosiheptide, effective against E. coli and
S. aureus respectively. These high performers were used as seeds for a beam search
method which improved selectivity. Additionally, a toxicity penalty was incorporated
into the optimisation to try and generate non-toxic and selective peptides. The
results indicate that the computational framework can optimise and identify selective
peptides. However, experimental validation is required to confirm the MIC values,
and subsequently the selectivity predictions. ...
antimicrobial peptides (AMPs) have been recognised as potential alternatives to
antibiotics. Much computational work has been done on the classification of peptides
into AMPs or non-AMPs, however, sparsity remains in the methods which address
selectivity. The selectivity is described as the ability of a peptide to act against
a target species while not inhibiting off-target species. This research explores a
machine learning pipeline for the discovery and optimisation of selective AMPs.
To achieve this, a standardised database was constructed, called sAMPle, giving a
total of 38,978 peptide-species combinations with corresponding minimum inhibitory
concentrations (MIC). Peptide features were represented using ESM-2 embeddings,
a well-known protein language model, while bacteria were represented by name-
based and taxonomy-based features. Two overlapping modelling approaches were
employed, a joint interaction model, which was trained on all bacterial species
simultaneously, and species-specific regression models. The best performing model
was a species-specific multilayer perceptron, achieving a root mean square error of
0.55–0.57 depending on the species, which is comparable to other MIC prediction
methods found in literature. Subsequently, the MIC predictions were used to define
a selectivity score, applied to the most prominent bacterial species in the sAMPle
dataset, namely E. coli and S. aureus, which are also clinically relevant species. The
selectivity scoring provided a ranked dataset in which the most selective peptides
were the known peptides apidaecin-1A and nosiheptide, effective against E. coli and
S. aureus respectively. These high performers were used as seeds for a beam search
method which improved selectivity. Additionally, a toxicity penalty was incorporated
into the optimisation to try and generate non-toxic and selective peptides. The
results indicate that the computational framework can optimise and identify selective
peptides. However, experimental validation is required to confirm the MIC values,
and subsequently the selectivity predictions. ...
Antibiotic resistance has become a major health risk globally. Consequently,
antimicrobial peptides (AMPs) have been recognised as potential alternatives to
antibiotics. Much computational work has been done on the classification of peptides
into AMPs or non-AMPs, however, sparsity remains in the methods which address
selectivity. The selectivity is described as the ability of a peptide to act against
a target species while not inhibiting off-target species. This research explores a
machine learning pipeline for the discovery and optimisation of selective AMPs.
To achieve this, a standardised database was constructed, called sAMPle, giving a
total of 38,978 peptide-species combinations with corresponding minimum inhibitory
concentrations (MIC). Peptide features were represented using ESM-2 embeddings,
a well-known protein language model, while bacteria were represented by name-
based and taxonomy-based features. Two overlapping modelling approaches were
employed, a joint interaction model, which was trained on all bacterial species
simultaneously, and species-specific regression models. The best performing model
was a species-specific multilayer perceptron, achieving a root mean square error of
0.55–0.57 depending on the species, which is comparable to other MIC prediction
methods found in literature. Subsequently, the MIC predictions were used to define
a selectivity score, applied to the most prominent bacterial species in the sAMPle
dataset, namely E. coli and S. aureus, which are also clinically relevant species. The
selectivity scoring provided a ranked dataset in which the most selective peptides
were the known peptides apidaecin-1A and nosiheptide, effective against E. coli and
S. aureus respectively. These high performers were used as seeds for a beam search
method which improved selectivity. Additionally, a toxicity penalty was incorporated
into the optimisation to try and generate non-toxic and selective peptides. The
results indicate that the computational framework can optimise and identify selective
peptides. However, experimental validation is required to confirm the MIC values,
and subsequently the selectivity predictions.
antimicrobial peptides (AMPs) have been recognised as potential alternatives to
antibiotics. Much computational work has been done on the classification of peptides
into AMPs or non-AMPs, however, sparsity remains in the methods which address
selectivity. The selectivity is described as the ability of a peptide to act against
a target species while not inhibiting off-target species. This research explores a
machine learning pipeline for the discovery and optimisation of selective AMPs.
To achieve this, a standardised database was constructed, called sAMPle, giving a
total of 38,978 peptide-species combinations with corresponding minimum inhibitory
concentrations (MIC). Peptide features were represented using ESM-2 embeddings,
a well-known protein language model, while bacteria were represented by name-
based and taxonomy-based features. Two overlapping modelling approaches were
employed, a joint interaction model, which was trained on all bacterial species
simultaneously, and species-specific regression models. The best performing model
was a species-specific multilayer perceptron, achieving a root mean square error of
0.55–0.57 depending on the species, which is comparable to other MIC prediction
methods found in literature. Subsequently, the MIC predictions were used to define
a selectivity score, applied to the most prominent bacterial species in the sAMPle
dataset, namely E. coli and S. aureus, which are also clinically relevant species. The
selectivity scoring provided a ranked dataset in which the most selective peptides
were the known peptides apidaecin-1A and nosiheptide, effective against E. coli and
S. aureus respectively. These high performers were used as seeds for a beam search
method which improved selectivity. Additionally, a toxicity penalty was incorporated
into the optimisation to try and generate non-toxic and selective peptides. The
results indicate that the computational framework can optimise and identify selective
peptides. However, experimental validation is required to confirm the MIC values,
and subsequently the selectivity predictions.
Integral Project Management in the Future of the Infrastructure Sector
An exploratory research on the implementation of IPM in the context of asset management
The infrastructure sector is characterised by the complexity in its projects. Integral Project Management (IPM), created and used by Rijkswaterstaat, provides a project management team structure and way of working to make optimal decisions in these complex projects. It divides the team into the disciplines of project management, project control management, stakeholder management, technical management and contract management. These disciplines integrally make the decisions in projects. The infrastructure sector is shifting from a project to asset management context. This shift requires a different focus, making the application of IPM in this new context uncertain. The research focuses on the possibility to implement IPM in the asset management context. It investigates the current applications of IPM in projects and then adjusts it to the new focus. The research follows multiple methods: a literature review to understand the current field of research, a desk study on the origin of IPM, semi-structured interviews with project managers and an expert session with experts in asset management to evaluate the final advice.
The findings of the research uncovered that projects apply IPM differently in three dimensions. Projects adjust these three dimensions to fit IPM in the assignment and specifics of the project. Together the dimensions form a IPM cube that represents the area of possible IPM variations. The first dimension is the “structure to control the project” dimension and focuses on the hard team structure. Projects adjust the core team to bring in additional expertise and roles that are needed for the project. The second dimension is the “interconnectedness of the team” dimension and focuses on the way of working within the team. Teams have strict or loose application of IPM, meaning that some divide decisions while others strictly make all decision together. The third dimension is the “personal preference” dimension and focuses on the soft aspects of project management. Personalities and preferences determine if people can work together or which adjustments in the other dimensions are made.
There are possibilities for implementing this IPM cube in the asset management context. The research created an advice. Minor adjustments to the dimensions are necessary to implement the different focus, but the five disciplines also fit as a team in asset management. This means that the asset management team mirrors the project management team. The asset management team focuses on the long-term and creates the project for the project management team to realise. An information manager between the two teams ensures communication takes place and that information is shared correctly. Continuously evaluating and adjusting the three IPM dimensions ensures that IPM is optimally applied in asset management.
The research concludes with an advice that shows the possibilities of IPM in the asset management context. Further, it contributes to existing literature by providing an insight into the current practice of IPM. The research could be used as a foundation for future research on IPM.
...
The findings of the research uncovered that projects apply IPM differently in three dimensions. Projects adjust these three dimensions to fit IPM in the assignment and specifics of the project. Together the dimensions form a IPM cube that represents the area of possible IPM variations. The first dimension is the “structure to control the project” dimension and focuses on the hard team structure. Projects adjust the core team to bring in additional expertise and roles that are needed for the project. The second dimension is the “interconnectedness of the team” dimension and focuses on the way of working within the team. Teams have strict or loose application of IPM, meaning that some divide decisions while others strictly make all decision together. The third dimension is the “personal preference” dimension and focuses on the soft aspects of project management. Personalities and preferences determine if people can work together or which adjustments in the other dimensions are made.
There are possibilities for implementing this IPM cube in the asset management context. The research created an advice. Minor adjustments to the dimensions are necessary to implement the different focus, but the five disciplines also fit as a team in asset management. This means that the asset management team mirrors the project management team. The asset management team focuses on the long-term and creates the project for the project management team to realise. An information manager between the two teams ensures communication takes place and that information is shared correctly. Continuously evaluating and adjusting the three IPM dimensions ensures that IPM is optimally applied in asset management.
The research concludes with an advice that shows the possibilities of IPM in the asset management context. Further, it contributes to existing literature by providing an insight into the current practice of IPM. The research could be used as a foundation for future research on IPM.
...
The infrastructure sector is characterised by the complexity in its projects. Integral Project Management (IPM), created and used by Rijkswaterstaat, provides a project management team structure and way of working to make optimal decisions in these complex projects. It divides the team into the disciplines of project management, project control management, stakeholder management, technical management and contract management. These disciplines integrally make the decisions in projects. The infrastructure sector is shifting from a project to asset management context. This shift requires a different focus, making the application of IPM in this new context uncertain. The research focuses on the possibility to implement IPM in the asset management context. It investigates the current applications of IPM in projects and then adjusts it to the new focus. The research follows multiple methods: a literature review to understand the current field of research, a desk study on the origin of IPM, semi-structured interviews with project managers and an expert session with experts in asset management to evaluate the final advice.
The findings of the research uncovered that projects apply IPM differently in three dimensions. Projects adjust these three dimensions to fit IPM in the assignment and specifics of the project. Together the dimensions form a IPM cube that represents the area of possible IPM variations. The first dimension is the “structure to control the project” dimension and focuses on the hard team structure. Projects adjust the core team to bring in additional expertise and roles that are needed for the project. The second dimension is the “interconnectedness of the team” dimension and focuses on the way of working within the team. Teams have strict or loose application of IPM, meaning that some divide decisions while others strictly make all decision together. The third dimension is the “personal preference” dimension and focuses on the soft aspects of project management. Personalities and preferences determine if people can work together or which adjustments in the other dimensions are made.
There are possibilities for implementing this IPM cube in the asset management context. The research created an advice. Minor adjustments to the dimensions are necessary to implement the different focus, but the five disciplines also fit as a team in asset management. This means that the asset management team mirrors the project management team. The asset management team focuses on the long-term and creates the project for the project management team to realise. An information manager between the two teams ensures communication takes place and that information is shared correctly. Continuously evaluating and adjusting the three IPM dimensions ensures that IPM is optimally applied in asset management.
The research concludes with an advice that shows the possibilities of IPM in the asset management context. Further, it contributes to existing literature by providing an insight into the current practice of IPM. The research could be used as a foundation for future research on IPM.
The findings of the research uncovered that projects apply IPM differently in three dimensions. Projects adjust these three dimensions to fit IPM in the assignment and specifics of the project. Together the dimensions form a IPM cube that represents the area of possible IPM variations. The first dimension is the “structure to control the project” dimension and focuses on the hard team structure. Projects adjust the core team to bring in additional expertise and roles that are needed for the project. The second dimension is the “interconnectedness of the team” dimension and focuses on the way of working within the team. Teams have strict or loose application of IPM, meaning that some divide decisions while others strictly make all decision together. The third dimension is the “personal preference” dimension and focuses on the soft aspects of project management. Personalities and preferences determine if people can work together or which adjustments in the other dimensions are made.
There are possibilities for implementing this IPM cube in the asset management context. The research created an advice. Minor adjustments to the dimensions are necessary to implement the different focus, but the five disciplines also fit as a team in asset management. This means that the asset management team mirrors the project management team. The asset management team focuses on the long-term and creates the project for the project management team to realise. An information manager between the two teams ensures communication takes place and that information is shared correctly. Continuously evaluating and adjusting the three IPM dimensions ensures that IPM is optimally applied in asset management.
The research concludes with an advice that shows the possibilities of IPM in the asset management context. Further, it contributes to existing literature by providing an insight into the current practice of IPM. The research could be used as a foundation for future research on IPM.