Past seminars

All past seminars

• Candidature review

  13 May 2026 · 13:00–14:00

  Candidate:

  Chuhan Wang

  Supervisor:

  Prof. Olaf Maennel and Prof. Ruidong Chang

  Date and Time:

  Wednesday 13 May 2026, 13:00–14:00 (Adelaide)

  Title:

  Towards Reliable Digital Twin State: Definition, Maintenance, and Use under Imperfect Sensing

  Abstract:

  Digital Twin (DT) systems for the building environment promise continuous, model-based insight into the condition of buildings and infrastructure. However, the sensing data that drives these systems is noisy, intermittent, and heterogeneous and most existing frameworks respond by treating sensor observations directly as system state in practice. This leads to various instability: transient objects corrupt geometric models, sensor failures leave the state undefined, and residual uncertainty propagates unchecked to downstream monitoring and maintenance decisions etc. This research proposes a state-centric perspective for DT systems, in which system state is an explicitly defined, continuously maintained, and reliably construct from but not equate with sensor observations. A unified three-module framework is developed to address the full state lifecycle under imperfect sensing: (1) a geometric state definition module that infers persistent structural state from multi-temporal LiDAR point clouds and images, distinguishing permanent features from transient elements; (2) a state maintenance module that coordinates heterogeneous IoT sensor streams, identify sensor reliability, and resolves inter-sensor conflicts to produce a trusted Digital Twin state; (3) a decision module that formulates resource-constrained maintenance and sensing strategies, accounting for residual state uncertainty. The framework is validated using data from an ARC-funded large-scale IoT deployment across residential buildings, providing real-world conditions of sensing noise, heterogeneity, and long-term operation. Expected contributions include a conceptual framework for state-centric Digital Twins, and an empirically validated system demonstrating practical deployability. This research addresses a cross-domain gap identified across geometric modeling, multi-sensor integration, and reliability-aware decision-making literature.

  Location:

  Ingkarni Wardli 4.63 (IW 4.63) — also available via Teams

  Microsoft Teams meeting:

  Join Teams meeting

• Candidature review

  13 May 2026 · 10:00–11:00

  Candidate:

  Ziyang Ye

  Supervisor:

  Prof Olaf Maennel

  Date and Time:

  Wednesday 13 May 2026, 10:00–11:00 (Adelaide)

  Title:

  Towards Secure and Robust Visual Intelligence: Adversarial Robustness in Vision Foundation Models

  Abstract:

  Vision Foundation Models are increasingly becoming the backbone of modern visual intelligence, enabling a wide range of systems to perceive, interpret, and reason about visual information. Their growing reuse across different applications brings significant benefits in generalisation and transferability, but it also introduces a new security concern: adversarial vulnerabilities may no longer be confined to individual models or isolated tasks. Instead, weaknesses in shared visual representations may transfer across model families and propagate through downstream systems, creating broader risks for the visual intelligence ecosystem. This research investigates the adversarial robustness of vision foundation models as a core AI security problem. It aims to understand how adversarial vulnerabilities emerge, how they transfer across different models and learning paradigms, and how they affect systems built upon shared visual backbones. It also examines whether current defence mechanisms are sufficient under stronger and more systematic evaluation, and explores the trade-off between improving robustness and preserving the generalisation capabilities that make foundation models valuable. Through a comparative and system-oriented methodology, this study will evaluate robustness at multiple levels, from foundation representations to downstream system behaviour, using a range of adversarial settings and defence strategies. The expected outcome is a more unified understanding of vulnerability propagation, defence limitations, and robustness–generalisation trade-offs, contributing to the development of more secure and reliable visual intelligence systems.

  Microsoft Teams meeting:

  Join Teams meeting

• Candidature review

  24 Apr 2026 · 15:00

  Candidate:

  Simranjeet Singh Dahia

  Supervisor:

  Prof Claudia Szabo, Dr Ruslan Puscasu, Dr Azhar Iqbal

  Date and Time:

  Friday 24 April 2026, 15:00 (Adelaide)

  Title:

  Quantum multi-agent reinforcement learning for decision-making in complex systems

  Abstract:

  Multi-agent reinforcement learning is a natural framework for decentralised decision-making in complex systems, but it faces persistent challenges under partial observability, strong inter-agent coupling and increasing coordination complexity. Quantum multi-agent reinforcement learning (QMARL) has recently emerged as a possible way to enhance policy and value representations in such settings. However, the field remains methodologically fragmented, is dominated by hybrid simulation-based studies and lacks both a consistent operational definition and disciplined evaluation standards. Our systematic review found that most existing work concentrates on encoding and policy approximation, while learning, coordination and evaluation remain underdeveloped. This PhD investigates QMARL as a representational framework rather than a computational speedup claim. We established a centralized-training, decentralized-execution QMARL pipeline, implemented matched quantum-classical benchmarking and introduced a Dec-POMDP calibration benchmark to test whether the framework can resolve non-classical correlation structure. Building on this foundation, ongoing experiments in cooperative MARL environments examine when quantum-parameterized actors and critics remain competitive and when stronger coordination regimes begin to expose representation-dependent behaviour. These findings motivate the next stage of the project: a complex phase space representation designed to explicitly encode coordination-relevant dynamics and a conflict-aware stabilisation mechanism for more expressive modelling of complex multi-agent dynamics.

  Microsoft Teams meeting:

  Join Teams meeting

• Special session

  10 Mar 2026 · 14:00

  Speaker:

  Panel (CSIT HDR)

  Date and Time:

  Tuesday 10 March 2026, 14:00 (Adelaide)

  Title:

  Industry collaboration and IP — panel discussion

  Location:

  Ingkarni Wardli, Level 4 seminar room

  Abstract:

  Special session with industry partners on collaboration models, intellectual property, and commercialisation pathways for HDR research.

Candidature review

• Candidature review

  13 May 2026 · 13:00–14:00

  Candidate:

  Chuhan Wang

  Supervisor:

  Prof. Olaf Maennel and Prof. Ruidong Chang

  Date and Time:

  Wednesday 13 May 2026, 13:00–14:00 (Adelaide)

  Title:

  Towards Reliable Digital Twin State: Definition, Maintenance, and Use under Imperfect Sensing

  Abstract:

  Digital Twin (DT) systems for the building environment promise continuous, model-based insight into the condition of buildings and infrastructure. However, the sensing data that drives these systems is noisy, intermittent, and heterogeneous and most existing frameworks respond by treating sensor observations directly as system state in practice. This leads to various instability: transient objects corrupt geometric models, sensor failures leave the state undefined, and residual uncertainty propagates unchecked to downstream monitoring and maintenance decisions etc. This research proposes a state-centric perspective for DT systems, in which system state is an explicitly defined, continuously maintained, and reliably construct from but not equate with sensor observations. A unified three-module framework is developed to address the full state lifecycle under imperfect sensing: (1) a geometric state definition module that infers persistent structural state from multi-temporal LiDAR point clouds and images, distinguishing permanent features from transient elements; (2) a state maintenance module that coordinates heterogeneous IoT sensor streams, identify sensor reliability, and resolves inter-sensor conflicts to produce a trusted Digital Twin state; (3) a decision module that formulates resource-constrained maintenance and sensing strategies, accounting for residual state uncertainty. The framework is validated using data from an ARC-funded large-scale IoT deployment across residential buildings, providing real-world conditions of sensing noise, heterogeneity, and long-term operation. Expected contributions include a conceptual framework for state-centric Digital Twins, and an empirically validated system demonstrating practical deployability. This research addresses a cross-domain gap identified across geometric modeling, multi-sensor integration, and reliability-aware decision-making literature.

  Location:

  Ingkarni Wardli 4.63 (IW 4.63) — also available via Teams

  Microsoft Teams meeting:

  Join Teams meeting

• Candidature review

  13 May 2026 · 10:00–11:00

  Candidate:

  Ziyang Ye

  Supervisor:

  Prof Olaf Maennel

  Date and Time:

  Wednesday 13 May 2026, 10:00–11:00 (Adelaide)

  Title:

  Towards Secure and Robust Visual Intelligence: Adversarial Robustness in Vision Foundation Models

  Abstract:

  Vision Foundation Models are increasingly becoming the backbone of modern visual intelligence, enabling a wide range of systems to perceive, interpret, and reason about visual information. Their growing reuse across different applications brings significant benefits in generalisation and transferability, but it also introduces a new security concern: adversarial vulnerabilities may no longer be confined to individual models or isolated tasks. Instead, weaknesses in shared visual representations may transfer across model families and propagate through downstream systems, creating broader risks for the visual intelligence ecosystem. This research investigates the adversarial robustness of vision foundation models as a core AI security problem. It aims to understand how adversarial vulnerabilities emerge, how they transfer across different models and learning paradigms, and how they affect systems built upon shared visual backbones. It also examines whether current defence mechanisms are sufficient under stronger and more systematic evaluation, and explores the trade-off between improving robustness and preserving the generalisation capabilities that make foundation models valuable. Through a comparative and system-oriented methodology, this study will evaluate robustness at multiple levels, from foundation representations to downstream system behaviour, using a range of adversarial settings and defence strategies. The expected outcome is a more unified understanding of vulnerability propagation, defence limitations, and robustness–generalisation trade-offs, contributing to the development of more secure and reliable visual intelligence systems.

  Microsoft Teams meeting:

  Join Teams meeting

• Candidature review

  24 Apr 2026 · 15:00

  Candidate:

  Simranjeet Singh Dahia

  Supervisor:

  Prof Claudia Szabo, Dr Ruslan Puscasu, Dr Azhar Iqbal

  Date and Time:

  Friday 24 April 2026, 15:00 (Adelaide)

  Title:

  Quantum multi-agent reinforcement learning for decision-making in complex systems

  Abstract:

  Multi-agent reinforcement learning is a natural framework for decentralised decision-making in complex systems, but it faces persistent challenges under partial observability, strong inter-agent coupling and increasing coordination complexity. Quantum multi-agent reinforcement learning (QMARL) has recently emerged as a possible way to enhance policy and value representations in such settings. However, the field remains methodologically fragmented, is dominated by hybrid simulation-based studies and lacks both a consistent operational definition and disciplined evaluation standards. Our systematic review found that most existing work concentrates on encoding and policy approximation, while learning, coordination and evaluation remain underdeveloped. This PhD investigates QMARL as a representational framework rather than a computational speedup claim. We established a centralized-training, decentralized-execution QMARL pipeline, implemented matched quantum-classical benchmarking and introduced a Dec-POMDP calibration benchmark to test whether the framework can resolve non-classical correlation structure. Building on this foundation, ongoing experiments in cooperative MARL environments examine when quantum-parameterized actors and critics remain competitive and when stronger coordination regimes begin to expose representation-dependent behaviour. These findings motivate the next stage of the project: a complex phase space representation designed to explicitly encode coordination-relevant dynamics and a conflict-aware stabilisation mechanism for more expressive modelling of complex multi-agent dynamics.

  Microsoft Teams meeting:

  Join Teams meeting

Thesis defence (public seminar)

No seminars in this section.

HDR seminar

No seminars in this section.

Special session

• Special session

  10 Mar 2026 · 14:00

  Speaker:

  Panel (CSIT HDR)

  Date and Time:

  Tuesday 10 March 2026, 14:00 (Adelaide)

  Title:

  Industry collaboration and IP — panel discussion

  Location:

  Ingkarni Wardli, Level 4 seminar room

  Abstract:

  Special session with industry partners on collaboration models, intellectual property, and commercialisation pathways for HDR research.