Learning-Based Hybrid Approach to Online Human-Pilot Behavior Characterization

This work presents a learning-based hybrid framework for the online characterization of human-pilot behavior. The approach focuses on estimating time-invariant parameters of a modified Tustin--McRuer pilot model in a pilot-in-the-loop setting, where key internal compensatory states are not directly measurable. To address this limitation, a Temporal Convolution Network augments the model's observability, enabling a discrete Unscented Kalman Filter to perform real-time estimation. Simulation results using a Lift+Cruise VTOL aircraft model demonstrate that the proposed architecture can accurately recover both the internal pilot state and time-invariant parameters while predicting pilot control-effort across varying flight conditions. These findings are further supported by experimental testing, where the architecture successfully reconstructed pilot control signals with high-fidelity online, showcasing the framework's applicability for real-world pilot-in-the-loop systems.