Trust-Driven Hybrid Control Integrating MPC and LSTM for VTOL Flight Transitions

Tite: Trust-Driven Hybrid Control Integrating MPC and LSTM for VTOL Flight Transitions (DOI)
​​​​​​​Authors:  Navid Mohammadi, Morteza Tayefi.
Journal:  IEEE Transactions on Automation Science and Engineering
​​​​​​​  ​​​​​​​
Abstract: This paper introduces a hybrid control strategy that combines Model Predictive Control (MPC) and Long Short-Term Memory (LSTM) networks. The method is designed to improve stability and adaptability during Vertical Take-off and Landing (VTOL) transitions between multi-copter and fixed-wing modes. A trust-based switching mechanism based on an exponential moving average dynamically allocates control between MPC and LSTM based on their recent performance. This ensures reliable operation under a wide range of flight conditions. The framework balances stability and computational efficiency by using MPC to handle constraints and LSTM to approximate control behavior in steady-state conditions. A VTOL transition scenario is used as a case study to test the controller’s performance under model uncertainties and external disturbances. Simulations show that the trust mechanism enables smooth switching between MPC and LSTM under various noise levels. A Monte Carlo sensitivity analysis indicates that a trust threshold of around 0.7 provides the best trade-off between accuracy and control effort. The hybrid controller achieves almost identical performance to a standalone robust MPC while reducing average computation time by about 40%. Eigenvalue and Lyapunov analysis further confirm closed-loop stability, validating the robustness of the hybrid control design. Overall, the proposed framework offers a scalable, reliable control solution for autonomous systems, enabling adaptive selection between model-based and learning-based strategies. Note to Practitioners—Engineers working on aerial robotics and autonomous systems face uncertainty, noisy sensors, tight actuator limits, and strict online constraints during transitions between flight modes. We present a straightforward hybrid controller that uses MPC to handle safety and constraints, while an LSTM provides fast commands when the flight condition is stable. A single trust score, computed from recent tracking errors and proximity to state and input limits, determines which controller takes the lead: MPC steps in near limits or under disturbances, and the LSTM runs in calm segments to improve responsiveness and reduce compute. To deploy it, you can keep your current state estimation and actuator limits, train the LSTM to mimic the MPC on representative flights, and tune two simple parameters, the trust threshold and a short memory factor, during routine tests. The proposed approach smooths the transition phase and reduces average computation time, which is important for implementing MPC in real-time embedded systems using microprocessors. It does require data that cover the operating range, and major configuration changes will need retraining.