Learning-based control & sim-to-real

A controller is designed from a model, but the plant it acts on is only partly known. Unmodelled dynamics, time-varying parameters, load disturbances, actuator hysteresis and, in traffic, the variable reaction delay of human drivers all separate the model from the real system. The program studies control laws that remain stable and accurate despite that gap, for wheeled robots, quadcopters, magnetic levitation systems, electric motors and grippers driven by shape memory alloy wires, together with plant and driver models identified from measured data. A second thread studies multi-objective optimisation under uncertainty for routing, scheduling and supply-network design.