Reliability-Aware Supervisory Flight Control of Quadrotors Using Causal Condition Estimation

Abstract Quadrotor flight control under time-varying uncertainty is challenging because no single controller remains uniformly superior across nominal operation, payload variation, wind disturbance, and rotor degradation. This study presents a reliability-aware supervisory flight-control architecture that selects between adaptive proportional–integral–derivative control and linear quadratic regulation using causal condition estimation and requirement-conditioned reliability information. The framework combines controller-conditioned estimation of payload, wind magnitude, rotor effectiveness, and affected-rotor identity with reliability-based decision logic, asymmetric fault latching, hysteresis, dwell time, decision persistence, and bumpless command transfer. The estimator operates causally using a 0.50 s trailing window updated every 0.25 s. Once rotor degradation is confirmed, the fault state is permanently latched and post-fault reversal from linear quadratic regulation to adaptive PID is prohibited. The final architecture was evaluated exclusively in nonlinear simulation using 500 independent stratified missions comprising payload/wind-dominant, rotor-degradation-dominant, sequential mixed, and boundary scenarios, resulting in 2,000 paired simulations. The practical supervisor achieved 71.2% mission reliability, compared with 35.0% for fixed adaptive PID and 25.4% for fixed LQR, while matching the perfect-information oracle in all 500 missions. The improvement over both fixed controllers was statistically significant, and the practical architecture satisfied a predeclared 5 percentage-point noninferiority margin relative to the oracle. Successful-mission median tracking error increased by only 1.65% relative to the oracle. The estimator achieved 91.31% rotor-fault recall, a 0.326% false-positive rate, 0.393 s median causal detection delay, and 100% rotor-identity accuracy. These results demonstrate that causal condition estimation and asymmetric reliability-aware supervision can recover oracle-level mission reliability without requiring perfect operating-condition knowledge within the modeled operating envelope.