Skip to main content
Back to timeline
researchsquare.comSource publication:

Morphing-aircraft attitude control drops the analytic adaptation law: TD3-MRAC cuts roll IAE by 46% across 500 perturbation scenarios

Synopsis

As a hypersonic morphing vehicle sweeps its wings, its aerodynamic and inertial characteristics keep changing, which makes an analytic adaptation law impractical to write configuration by configuration. TD3-MRAC uses a fixed-wing model as reference, backstepping as the baseline, a linear extended state observer to estimate the lumped disturbance, and a TD3 agent to generate a bounded compensation increment in place of the hand-designed adaptation law. Across 500 multi-source perturbation scenarios the same trained policy, with no retraining, reduced mean roll IAE by 46% and yaw by 13% relative to a robust adaptive controller, while pitch was slightly worse.

AI-generated editorial illustration: Model-Reference-Based Intelligent Active Disturbance Rejection Attitude Control for Morphing Aircraft

Interpretation

Attitude control for a morphing vehicle no longer depends on an analytic adaptation law written for each configuration; an offline-trained bounded compensation policy absorbs the model mismatch that morphing introduces. Conventional model reference adaptive control writes the uncertain dynamics as a parameter vector times a regressor and assumes constant parameters; for a morphing vehicle the aerodynamic and inertial maps depend on configuration, so the true parameters move and a constant-parameter law persistently chases a moving target. Across 500 scenarios built from positive, negative and zero biases on six disturbance parameters, the same trained policy was evaluated without retraining or fine-tuning: mean roll IAE was 20.894 deg·s against 38.719 deg·s for the robust adaptive controller.

The control input is a superposition of three parts: a backstepping baseline for basic stabilization, an observer feedforward that cancels the estimated lumped disturbance, and a TD3 policy that outputs only a bounded compensation increment. A linear extended state observer treats the total disturbance as an extra state, so low-dimensional output measurements such as attitude angles suffice to reconstruct the system states and estimate the combined effect of unmodeled dynamics and external disturbances. The simulation runs 70 seconds, with the sweep angle following a 24-second trapezoidal schedule between 10 and 50 degrees and attitude commands forming a compound banked climb-and-recovery maneuver across roll, pitch and yaw.

Perspective

The conclusions rest on the nonlinear hypersonic morphing-vehicle dynamics derived in Section 3 and are evaluated in 70-second closed-loop numerical simulation, with perturbations drawn from combinations of initial attitude offsets and three-axis moment-coefficient biases; hardware-in-the-loop and flight tests are not covered.

The stability results require only that the learned action be bounded and do not rely on TD3 convergence, which remains an open problem under nonlinear function approximation; in pitch the robust adaptive controller has 34% lower mean IAE, narrowing to a 5% median gap, so TD3-MRAC's pitch performance degrades mainly in a subset of outlier scenarios.