Robustness margins for neural-augmented agile flight control: quantifying admissible control-dependent model discrepancy under Lipschitz-constrained neural feedback

Pries and Ryll (IROS 2025) introduced the Neural Geometric Tracking Controller (NGTC), a quadrotor tracking controller that augments a geometric base controller with a contracting recurrent equilibrium network $Q$ acting on the one-step model-prediction residual. Stability by contraction is proven only for the nominal case where the controller's internal model matches the plant; under a control-dependent model discrepancy $\Delta G = G-\hat{G}$ (for example, unmodeled aerodynamic drag), the residual becomes control-dependent and closes a feedback loop through $Q$ that can destabilize the closed loop. The paper asserts that Lipschitz bounds on $Q$ must be enforced to preserve closed-loop passivity, but gives no explicit bound, no tuning methodology, and no robustness margin quantifying the admissible $\Delta G$. The open problem is to establish that margin: an explicit, computable sufficient condition relating a bound on the incremental gain of the model discrepancy over the agile flight envelope to the Lipschitz constant enforced on $Q$, under which the closed loop retains its stability guarantee, together with a constructive tuning procedure instantiated on the quadrotor geometric-control loop (including its constraint-dependent control allocation) and a coverage analysis showing that realistic aerodynamic discrepancies lie within the margin. Such a margin would in particular convert the paper's empirical observation — that policies trained in a drag-free simulator transfer to drag-affected platforms without domain randomization or fine-tuning — into a certified deployment condition.

Background

High-speed quadrotor flight stresses trajectory-tracking controllers at the platform's operational limits: actuator saturation, aerodynamic effects, and unknown disturbances, all under hard real-time constraints. Three control paradigms dominate. Nonlinear model predictive control (NMPC) repeatedly solves an optimal-control problem online; it handles constraints but costs milliseconds per update and lacks rigorous closed-loop stability proofs. Differential-flatness-based control (DFBC, a form of geometric control) exploits the quadrotor's differential flatness — the property that its states and inputs are exact functions of a flat output such as the position trajectory and a finite number of its derivatives — to build a fast cascade of position and attitude controllers; it is efficient but degrades under large disturbances and dynamically infeasible references. Learned (neural) policies achieve strong empirical performance but typically rely on domain randomization — randomizing simulator physics during training so the policy tolerates the sim-to-real gap — and carry no stability certificates.

The Neural Geometric Tracking Controller (NGTC) of Pries and Ryll (2025) unifies these strengths. A geometric base controller $k(x)$ (following Sun et al. 2022, with a tilt-prioritized attitude cascade and quadratic-programming-based control allocation mapping desired wrench to rotor thrusts) is augmented by a neural feedback $Q$ that acts on the one-step prediction residual $\tilde{x}_t = x_t - \hat{x}_t$, where $\hat{x}_t = \hat{G}(x_{t-1},u_{t-1})$ is the prediction of a nominal model $\hat{G}$ running inside the controller: $u_t = k(x_t) + Q(\tilde{x}_t) + r_t$, where $r_t$ is a bounded reference feedforward input. This residual-injection structure is an instance of the (nonlinear) Youla parameterization, which parameterizes stabilizing controllers by a stable filter — the Youla parameter — driven by the estimation innovation; here $Q$ is realized as a recurrent equilibrium network (REN), a class of implicit recurrent neural models whose parameterization guarantees, by construction and without constrained optimization, that every model in the class is contracting (a strong incremental stability property: distance between any two trajectories decays exponentially in a suitable metric) and satisfies any prescribed incremental integral quadratic constraint (IQC) — an inequality constraining a system's incremental input-output behavior, generalizing small-gain and passivity conditions — including a user-chosen Lipschitz bound and incremental passivity. Training uses analytic policy gradients through the (provably stable) closed-loop simulation.

In the nominal case $G=\hat{G}$, the residual reduces to the exogenous disturbance $\tilde{x}=d$, so the output of the contracting $Q$ acts as a bounded exogenous input and closed-loop contraction is preserved. Under model discrepancy $\Delta G = G-\hat{G}$, however, the residual $\tilde{x} = \Delta G(x,u)+d$ depends on the control input, so the signal passes through $Q$ back into $u$: a feedback loop that can destabilize the system. Pries and Ryll state that "certain Lipschitz bounds of $Q$ need to be enforced to ensure passivity of the closed-loop", citing van der Schaft's $L_2$-gain and passivity techniques, and defer a comprehensive robustness discussion to the Youla-REN literature (Wang and Manchester 2022). Empirically the controller is strikingly robust — trained in simulation without any aerodynamic drag model (for comparability with baselines) and without domain randomization or fine-tuning, it nonetheless tracks agile trajectories under +50% drag, mass and motor-time-constant mismatch, external forces, and real wind — but the paper provides no theory predicting when such transfer succeeds.

The general margin-shaped theory now exists at framework level. The Youla-REN line (Wang and Manchester 2022; Barbara, Wang, and Manchester 2023) proves built-in contraction and user-tunable Lipschitz guarantees for Youla-parameterized neural policies, including the partially-observed case with observers. Most directly, Barbara, Wang, Megretski, and Manchester (2025, revised 2026) prove that when the innovation channel is coupled through disturbances or model parameter uncertainty, closed-loop contraction and Lipschitzness follow from an incremental-IQC loop-shaping condition — reducing, for Lipschitz-type IQCs, to the incremental small-gain condition $\mu\gamma < 1$ on the product of the Lipschitz bounds of the weighted uncertainty loop and of the neural parameter — and demonstrate a tuning methodology (Youla-$\gamma$REN: choosing the enforced Lipschitz bound via small-gain analysis with shaping filters) on a linearized cart-pole with pole-mass uncertainty. The same work shows the residual (non-Youla) architecture — exactly NGTC's shape — is certifiable via small-gain arguments but generally conservatively, and that with nonlinearity plus partial observation plus disturbances, full incremental stability can be lost. None of this has been instantiated for aerial vehicles: no published work computes such a margin for a quadrotor geometric-control loop, treats the aerodynamic drag discrepancy class at agile-flight speeds, analyzes the loop through the constraint-dependent QP control allocation, or connects the margin to drag-free training transfer.

Problem Statement

For the neural-augmented tracking architecture of Pries and Ryll (2025) — full-state feedback, geometric base controller $k(x)$ with tilt-prioritized attitude cascade and QP-based control allocation, and a contracting recurrent equilibrium network $Q$ acting on the one-step model-prediction residual $\tilde{x}_t = x_t - \hat{G}(x_{t-1},u_{t-1})$ so that $u_t = k(x_t) + Q(\tilde{x}_t) + r_t$ — establish a formal robustness margin against control-dependent model discrepancy $\Delta G = G - \hat{G}$. Specifically: determine an explicit, computable sufficient condition relating (i) a bound on the incremental gain (Lipschitz constant) of $\Delta G$ over the agile flight envelope — covering, at minimum, a quantified class of unmodeled aerodynamic drag — to (ii) the Lipschitz constant enforced on $Q$ (and any loop-shaping weights), under which the closed loop retains its stability guarantee, whether in the form of contraction or of the passivity property asserted in the source. Provide a constructive tuning methodology — how to select and enforce the bound on $Q$ so the margin holds while preserving tracking performance — instantiate the analysis on the quadrotor geometric-control loop, including its constraint-dependent (QP) control-allocation map, and quantify whether realistic aerodynamic discrepancies at agile-flight speeds lie within the resulting margin. As a direct corollary, this yields sufficient conditions under which a policy trained on the drag-free nominal model (as in the source: no domain randomization, no fine-tuning) is guaranteed to deploy safely on the drag-affected physical platform.

References

1. Lukas Pries and Markus Ryll, 'Learning Robust Agile Flight Control with Stability Guarantees', 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2025, DOI 10.1109/iros60139.2025.11246103; arXiv:2510.12611, https://arxiv.org/abs/2510.12611
2. Max Revay, Ruigang Wang, and Ian R. Manchester, 'Recurrent Equilibrium Networks: Flexible Dynamic Models with Guaranteed Stability and Robustness', IEEE Transactions on Automatic Control, vol. 69, no. 5, pp. 2855-2870, 2024, DOI 10.1109/TAC.2023.3294101, https://doi.org/10.1109/TAC.2023.3294101; arXiv:2104.05942, https://arxiv.org/abs/2104.05942
3. Ruigang Wang and Ian R. Manchester, 'Youla-REN: Learning Nonlinear Feedback Policies with Robust Stability Guarantees', 2022 American Control Conference (ACC), 2022, DOI 10.23919/ACC53348.2022.9867842; arXiv:2112.01253, https://arxiv.org/abs/2112.01253
4. Nicholas H. Barbara, Ruigang Wang, and Ian R. Manchester, 'Learning Over Contracting and Lipschitz Closed-Loops for Partially-Observed Nonlinear Systems', 62nd IEEE Conference on Decision and Control (CDC), 2023, DOI 10.1109/cdc49753.2023.10383269; arXiv:2304.06193, https://arxiv.org/abs/2304.06193
5. Nicholas H. Barbara, Ruigang Wang, Alexandre Megretski, and Ian R. Manchester, 'React to Surprises: Stable-by-Design Neural Feedback Control and the Youla-REN', arXiv:2506.01226 (v1 June 2025, v3 June 2026), https://arxiv.org/abs/2506.01226
6. Glen Chou, Necmiye Ozay, and Dmitry Berenson, 'Model Error Propagation via Learned Contraction Metrics for Safe Feedback Motion Planning of Unknown Systems', 60th IEEE Conference on Decision and Control (CDC), 2021, DOI 10.1109/CDC45484.2021.9683354, https://doi.org/10.1109/CDC45484.2021.9683354; arXiv:2104.08695, https://arxiv.org/abs/2104.08695
7. Akash Harapanahalli, Samuel Coogan, and Alexander Davydov, 'Learning Certified Neural Network Controllers Using Contraction and Interval Analysis', 2026, arXiv:2603.28011, https://arxiv.org/abs/2603.28011
8. Dalim Wahby, Lorenzo Schenk, and Guillaume Ducard, 'SOS-based Stability Verification for Saturated INDI Control of Hybrid-VTOL Aircraft Pitch Rate Dynamics', 2026, arXiv:2606.14198, https://arxiv.org/abs/2606.14198
9. Arjan van der Schaft, 'L2-Gain and Passivity Techniques in Nonlinear Control', Springer, London, 2000, DOI 10.1007/978-1-4471-0507-7, https://doi.org/10.1007/978-1-4471-0507-7
10. Sihao Sun, Angel Romero, Philipp Foehn, Elia Kaufmann, and Davide Scaramuzza, 'A Comparative Study of Nonlinear MPC and Differential-Flatness-Based Control for Quadrotor Agile Flight', IEEE Transactions on Robotics 38(6), 2022, DOI 10.1109/TRO.2022.3177279, https://doi.org/10.1109/TRO.2022.3177279