Dayanıklı insan-otonomi iş birliği: Öğrenmeye dayalı yardımcı sistemler ve güvenli paylaşımlı kontrol

This thesis presents novel control architectures for shared control systems, where human operator and the automation share the control responsibility. Proposed architectures are formally verified by mathematical methods, and shown to improve performance/reduce workload in experiments and simulations for flight control scenarios. Firstly, with use of an adaptive human operator model and Long Short-Term Memory (LSTM) network, a neural network based pilot assistance system is proposed and shown to improve tracking performance and reduce workload in simulations. Then, this architecture is improved with a more advanced pilot model and a modulation mechanism, such that it does not violate pilot autonomy in case of a malfunction or a disagreement between the pilot and the assistance signal. Lastly, a general, automation-agnostic architecture for automation assistance regulation for shared control systems is proposed. This architecture employs control barrier functions (CBFs) to modulate the automation assistance, where shared control metrics are specified in terms of barrier functions. Finally, results are verified with simulations and human-in-the-loop experiments.