Key information:
| Student | Carl Richardson |
| Academic Supervisors | |
| Cohort | 3 |
| Pure Link | Active Project |
Next-generation control systems for robotics, aircrafts, and autonomous systems will inevitably feature elements of adaptation. They will be required to work in uncertain dynamic environments, needing fast online learning and adaptation to achieve optimal performance. Examples include sophisticated UAVs, operating in unknown wind conditions or compensating for in-flight damage.
Conventional adaptive control algorithms have limited potential to offer this level of adaptability: they consist of fixed structure control laws supplemented with rudimentary algorithms which update these structures based on the current environment. These update algorithms are mainly based on stability considerations required to guarantee some level of safe operation, but are unsophisticated, with limited, inflexible learning potential. Their key advantage is the ease in which they can be implemented on modest hardware.
Techniques from machine learning offer, potentially, an appealing alternative to the existing rudimentary algorithms for the update of adaptive control laws. Utilising the measured data, these techniques offer more flexibility, higher performance and may improve the other well-known limitation in adaptive control: robustness. Their key deficiencies are their difficulty in implementation on embedded devices and the difficulties in guaranteeing stability (the minimum requirement for control systems).
This research aims to develop safe and robust learning-based control for edge computing, thereby enabling the implementation of such algorithms on mobile, network-deprived autonomous systems. The focus of the research will be on enhancing robust control strategies with the ability to learn and adapt in order to optimise performance whilst always guaranteeing safe operation.