Fault- and saturation-aware quadrotor tracking via MRAC with an RL-inspired adaptive reference governor
This paper presents a hybrid, constraint-aware tracking architecture for quadrotor UAVs that explicitly targets the failure mode arising from the combination of actuator saturation and loss-of-effectiveness (LOE) faults. While model-based controllers augmented with adaptive terms can compensate for matched uncertainties locally, their performance can still collapse when the commanded trajectory becomes infeasible under post-fault authority reduction, leading to prolonged saturation and rapidly growing tracking errors. To address this issue, an policy-based adaptive reference governor with optional offline parameter tuning is introduced that shapes the commanded reference in real time, without modifying the stabilizing control law. The governor is implemented as a bounded, rate-limited first-order reference filter whose channel-wise time constants and rate limits are generated by a compact parametric policy driven by interpretable features: output-error norm, output-rate proxy, saturation ratio, and a fault indicator. The policy parameters may optionally be refined offline via a toolbox-free CEM-like random search, yielding a lightweight and reproducible parameter-tuning procedure that is reproducible and safety-oriented by construction. Nonlinear simulations under two regimes–(A) nominal operation without saturation/faults and (B) simultaneous rotor-speed saturation with an LOE fault–demonstrate the effectiveness of policy-based adaptive reference shaping in improving practical feasibility under constrained operation. In nominal conditions, the proposed governor improves practical performance by substantially reducing peak control magnitude, control energy, and control-rate activity while maintaining bounded tracking. In the stress regime, baseline configurations exhibit sustained saturation and divergence, whereas the RL-inspired governor maintains bounded tracking, eliminates time spent at saturation, and reduces error and effort metrics by orders of magnitude. These results indicate that embedding a low-dimensional parametric policy at the reference level provides an effective and safety-compatible mechanism for fault- and constraint-aware quadrotor tracking.