UKRI AI CDT in Dependable and Deployable AI for Robotics (CDT-D2AIR)

Constance Frachon

Project
Hybrid Solutions to Non-stationarity in Underwater Navigation

Summary
Reinforcement Learning (RL) methods often struggle in non-stationary settings, where environmental dynamics evolve over time and violate stationarity assumptions in standard RL theory. This challenge is particularly pronounced in underwater robotics, where turbulence and time-varying disturbances complicate robust navigation and path planning. In contrast, traditional control and robotic design offer well-established approaches to disturbance rejection but typically lack adaptability and generalisability. This project aims to explore hybrid approaches that combine learning-based adaptability with model-based robustness, towards resilient and generalisable solutions.