Research area

AI and Hydrodynamics

Machine learning applied to maritime systems, where data-driven models are combined with the physics of ship and platform motion.

Classical hydrodynamic models are accurate but expensive. Purely data-driven models are fast but unconstrained by physics. We work in the space between the two — using machine learning to build fast, reliable models of marine vehicle behaviour while keeping the underlying physics in the loop.

Reinforcement learning for control

A major line of work is reinforcement learning for path following and collision avoidance of ships. Early controllers based on Deep Q-Networks outperformed traditional PID control in simulation and in lake experiments on free-running models. Later work explores continuous-action algorithms such as DDPG and PPO for path following and collision avoidance under the practical limits of underactuated displacement ships — asymmetric port–starboard dynamics, reduced rudder effectiveness at low speed, and slow steering gear.

Active heave compensation was an early bridge from classical control into learning-based methods. Comparisons of PID, sliding mode, model predictive control and linear quadratic regulation against RL controllers showed where learning-based approaches improve disturbance rejection and noise attenuation for crane loads in waves.

Data-driven models and applications

Beyond RL for vehicle control, the group develops data-driven and learning-based approaches to model predictive control, surrogate models for hydrodynamic response, and machine-learning methods for problems such as underwater acoustic signal classification. The aim is always the same: models that are useful for autonomy and design, not only accurate on the training set.