Fifteen years from classical robust control into artificial intelligence, and the two have not come apart: an agent that acts in the world is a control problem wearing new clothes.

Research

My work sits where artificial intelligence meets control theory. AI decides what to do; control theory is the century of mathematics about whether a thing that decides and acts will actually stay stable, recover from error, and behave the same way tomorrow. Most of what I build needs both.

AI for healthcare

Diagnostic models where being right is not enough on its own, because a clinician has to be able to see why. The work covers stroke segmentation and rapid stroke diagnosis, breast cancer detection from mammography, ultrasound and MRI, and interpretable multimodal architectures that combine imaging with other clinical signals.

Two threads run through all of it. The first is explainability: a model that cannot show its reasoning cannot be trusted with a diagnosis, whatever its accuracy on a benchmark. The second is robustness: performance has to survive a different scanner, a different hospital and a different population, which is a question about domain shift rather than about architecture.

Current project. Multimodal AI for breast cancer detection, 2024 to present: deep models that integrate ultrasound, mammography and MRI with explainable AI, aimed at earlier detection that a clinician can actually act on.

Recently completed. Transformer based deep learning for rapid stroke diagnosis, 2023 to 2025, run as an international network with knowledge exchange and training for early career researchers.

Agentic AI

The newest line, and the one I expect to occupy most of the next few years.

An agent is a model that does not simply answer. It holds a goal across many steps, chooses actions, calls tools, reads what came back, and decides what to do next. That loop is a dynamical system, and the questions worth asking about it are the questions control theory already knows how to ask. Does it converge or oscillate? Is it stable when the environment is uncertain? What does it do when a tool fails or returns something adversarial? Where do you put a bound so that a bad step cannot become a bad hundred steps?

Three strands:

• Agents with guarantees. Treating the perceive, decide, act loop as a feedback system, so that limits on cost, safety and behaviour are properties of the design rather than instructions in a prompt.
• Agents that teach. The tutor on the course platform is an agent under supervision: it knows the topic, knows what the student has already got wrong, gives the next step rather than the answer, and works against a problem bank whose every answer has been re-derived by a computer algebra system before a student meets it.
• Agents that build. Using the same tools on the research and teaching work itself, and treating what breaks as evidence rather than annoyance. The platform this lab runs on was built this way.

Robust and optimal control

Negative imaginary systems theory, robust stability analysis and feedback design. The doctoral work behind this line, New Results on Negative Imaginary Systems Theory with Application to Flexible Structures and Nano-Positioning, with Professor Ian Petersen at UNSW, extended the theory to systems with free body dynamics and to interconnections mixing negative imaginary and passivity properties.

The applications that keep pulling on it are physical: flexible structures that ring, nano-positioning stages that must settle without overshoot, and interconnected systems where each part is well behaved and the whole is not. More recent work brings the same conditions to learned models, including Koopman operator approximation under negative imaginary constraints and controllability analysis for vision state space models, which is where this line and the AI line meet.

Robotics and autonomy

Perception, autonomous navigation and online learning for UAVs and ground robots, including work at the KAUST Robotics Lab and the Mohamed Bin Zayed International Robotics Challenge, where the team placed second runner-up in

1. Recent work covers search missions in adversarial environments and

UAV assisted networks under wind.

Dynamics and game theory

Stability analysis, nonlinear dynamics and strategic interaction: passivity of replicator dynamics and higher order evolutionary dynamics, and the thread that connects the control work to models of how agents, rather than machines, settle into equilibria.

AI in mathematics education

The teaching platform at learn.mabroklab.com is a research interest as much as a teaching tool: what a course looks like when every answer is machine checkable, when a tutor knows what a student has already got wrong, and when the record of a term is a dataset rather than a gradebook. It is built, funded and run personally, and used with real classes before any part of it is made general.

Earlier work: quantum systems

Quantum control and quantum computing, including the QQatar initiative founded at Qatar University. Not an active line now, though the systems questions carried over into the control and agent work above.

The group

Four researchers at present, working on AI for healthcare, foundation models in diagnostics, stroke source identification, and machine learning for additive manufacturing. Several of the papers listed under Publications are first-authored by students in the group.

Who they are, and what they are working on →

Funding

Lead PI on grants totalling over one million dollars.