Coupling learned aerodynamic dynamics with saturation-limited flight control: cascaded attitude-loop and actuation-constrained MPC interaction and extrapolation limits for Neural-Fly-style agile flight in strong winds

Neural-Fly (O'Connell et al., Science Robotics 2022) achieved rapid wind adaptation for agile quadrotor flight by combining a meta-learned nonlinear aerodynamic representation with fast online adaptation, and proved a closed-loop tracking error bound that scales with the learned representation error $d(t)$ and the rate of change $\|\dot{a}(t)\|$ of the wind-dependent coefficients, but analyzed only a nominal unsaturated outer-loop vehicle model. The paper explicitly leaves open (i) how the learned dynamics block interacts with cascaded attitude inner loops operating near actuation limits and with actuation-constrained planning-and-control structures such as model predictive control (MPC), including the effect of learned-model-induced command changes on inner-loop stability, and (ii) how to reduce and quantify representation error when extrapolating beyond approximately uniform, static training winds to fast-changing, spatially nonuniform fields. Follow-up work (FALCON, hierarchical meta-learning, meta-learning-augmented MPC, Proto-MPC, DATT, sliced learning, online-uncertainty comparative studies) addresses adjacent ingredients but provides neither a stability or performance theory for the learned-block/cascaded-saturation coupling with flight validation nor a principled extrapolation or safe-envelope characterization for the learned aerodynamic representation. The problem remains open.

Background

A quadrotor's aerodynamic forces in strong wind depend nonlinearly on airspeed, angle of attack, and rotor interactions, so first-principles models degrade exactly when agile flight is hardest. Neural-Fly (O'Connell, Shi, Shi, Azizzadenesheli, Anandkumar, Yue, and Chung; Science Robotics 7:eabm6597, 2022) addressed this by learning part of the dynamics: a four-layer neural network feature map $\phi(\cdot)$ is trained on wind-tunnel data so that the residual aerodynamic force on the vehicle is modeled as $f=\phi(s)\,a+d(t)$, where $s$ collects the vehicle state and input, $a$ is a small vector of linear coefficients adapted online by a composite adaptation law (an adaptive-control parameter update that combines tracking error and prediction error), and $d(t)$ is the representation error — the part of the true dynamics the learned features cannot express. Only the final linear layer (the coefficients $a$) is adapted online. Meta-learning (learning-to-learn: training the shared representation $\phi$ so that adaptation to a new wind condition from few samples is fast) yields residual-force estimates with sub-second lag after wind changes, and a domain-adversarial variant (DAIML, Domain-Adversarially Invariant Meta-Learning, which trains $\phi$ to be invariant across wind domains) enabled transfer between two different vehicles. On this structure the authors proved a Lyapunov-based closed-loop tracking error bound: the tracking error exponentially converges to an error ball whose size is proportional to $\|d(t)+\epsilon\|$, where $\epsilon$ is measurement noise, and to $\|\dot{a}(t)\|$, the rate of change of the wind-dependent coefficients (in the paper's words, how fast the wind condition changes). In Caltech's Real Weather Wind Tunnel (a wall of 1296 individually controllable fans), figure-eight trajectories were tracked in constant winds up to 43.6 km/h (12.1 m/s) — about twice the 22 km/h maximum of the static, approximately uniform winds in the training data — as well as in sinusoidally time-varying wind, and outdoor flight was demonstrated in gentle breeze (about 17 km/h). Two structural limitations frame the open problem. First, real flight stacks are cascaded: an outer loop (here, the learned position/velocity dynamics plus adaptive controller) commands attitude and thrust to an inner attitude loop, and both the inner loop and any planning layer enforce actuation limits. The published guarantee analyzes a nominal unsaturated model, so it does not describe what happens when the learned block's commands push the inner loop near or into saturation, nor how the guarantee composes with an actuation-constrained planner such as model predictive control (MPC — receding-horizon optimization of the control input subject to explicit input/state constraints). The paper states verbatim that further work is needed on the interaction of the learned dynamics with the cascaded control design when implementing a tracking attitude controller, and that further research is required on the coupling between planning and learning-based control near actuation limits, suggesting a combined planning-and-control structure such as MPC. Second, the learned model was trained on static and approximately uniform wind data; when extrapolating to the strongest wind speeds tested (roughly twice the training maximum) some model mismatch appears, and since the tracking bound scales with $d(t)$, performance in fast-changing, spatially nonuniform wind (large $\|\dot{a}(t)\|$) is limited by representation error. Enriching $\phi$ reduces $d(t)$ but typically slows or destabilizes online adaptation, a trade-off the paper leaves unresolved. Related threads exist but do not close these gaps: FALCON (Lale et al., npj Robotics 2024) puts a learned Fourier basis inside MPC with stability and robustness guarantees under extreme turbulence, but on a stationary airfoil testbed rather than a free-flying vehicle; hierarchical meta-learning (Xie et al., ICRA 2024) extends meta-learned adaptation to latent (unmeasured) disturbances; meta-learning-augmented MPC (Lapandić et al., IEEE L-CSS 2024), Proto-MPC (Gu, Cheng, Hovakimyan, arXiv:2401.15508), and comparative online-uncertainty-learning studies (Gu, Zhao, Rizzo, JIRS 2025) each address one ingredient in simulation or partial hardware, without a saturation-aware cascaded-interaction theory or an extrapolation-bound characterization; DATT (Huang et al., arXiv:2310.09053) handles large disturbances and actuation limits empirically with learned feedforward plus $\mathcal{L}_1$ adaptive augmentation (a robust adaptive-control architecture that filters the adaptive control signal to decouple adaptation rate from robustness), without guarantees for the learned block; sliced learning (Gao et al., arXiv:2508.14422) proves exponential convergence for learned disturbance identification at the attitude level on SO(3) (the group of 3D rotations), but leaves the coupling of learned inner and outer dynamics near saturation and extrapolation beyond training winds untouched.

Problem Statement

For Neural-Fly-style flight control, in which a meta-learned aerodynamic representation $\phi$ with rapidly adapted linear coefficients captures wind-induced residual dynamics inside an outer tracking loop whose certified error ball scales with the representation error $d(t)$ (the part of the true dynamics the learned features cannot express) and with the rate of change $\|\dot{a}(t)\|$ of the wind-dependent coefficients (how fast the wind condition changes): (1) establish a stability and performance analysis of the closed loop in which the learned dynamics block is coupled to a cascaded attitude inner loop operating near or at actuation saturation and composed with an actuation-constrained planning-and-control layer such as model predictive control, characterizing how learned-model-induced command changes propagate to inner-loop stability; and (2) give a principled characterization of the representation error $d(t)$ when operating beyond the training distribution of static, approximately uniform wind — fast-changing, spatially nonuniform wind fields and higher wind speeds — yielding quantified safe-extrapolation envelopes or bounds, and clarifying how the learned representation can be enriched to reduce $d(t)$ across broader aerodynamic regimes in light of the trade-off between representation richness and online adaptation speed. These two components share one scientific objective: extending the certified operating envelope of learned aerodynamic flight control from the nominal unsaturated outer-loop regime demonstrated in the source paper to the saturation-limited, distribution-shifted regime required for agile flight in the strongest real winds.

References

1. Michael O'Connell, Guanya Shi, Xichen Shi, Kamyar Azizzadenesheli, Anima Anandkumar, Yisong Yue, Soon-Jo Chung, 'Neural-Fly enables rapid learning for agile flight in strong winds', Science Robotics 7, eabm6597 (2022). DOI: 10.1126/scirobotics.abm6597, https://www.science.org/doi/10.1126/scirobotics.abm6597 (preprint: arXiv:2205.06908, https://arxiv.org/abs/2205.06908)
2. Sahin Lale, Peter I. Renn, Kamyar Azizzadenesheli, Babak Hassibi, Morteza Gharib, Anima Anandkumar, 'FALCON: Fourier Adaptive Learning and Control for Disturbance Rejection Under Extreme Turbulence', npj Robotics 2 (2024). DOI: 10.1038/s44182-024-00013-0, https://doi.org/10.1038/s44182-024-00013-0
3. Fengze Xie, Guanya Shi, Michael O'Connell, Yisong Yue, Soon-Jo Chung, 'Hierarchical Meta-learning-based Adaptive Controller', ICRA 2024. DOI: 10.1109/ICRA57147.2024.10611562, https://doi.org/10.1109/ICRA57147.2024.10611562 (preprint: arXiv:2311.12367, https://arxiv.org/abs/2311.12367)
4. Dženan Lapandić, Fengze Xie, Christos K. Verginis, Soon-Jo Chung, Dimos V. Dimarogonas, Bo Wahlberg, 'Meta-Learning Augmented MPC for Disturbance-Aware Motion Planning and Control of Quadrotors', IEEE Control Systems Letters 8 (2024). DOI: 10.1109/LCSYS.2024.3520023, https://doi.org/10.1109/LCSYS.2024.3520023 (preprint: arXiv:2410.06325, https://arxiv.org/abs/2410.06325)
5. Yuliang Gu, Sheng Cheng, Naira Hovakimyan, 'Proto-MPC: An Encoder-Prototype-Decoder Approach for Quadrotor Control in Challenging Winds', arXiv:2401.15508 (2024), https://arxiv.org/abs/2401.15508
6. Kevin Huang, Rwik Rana, Alexander Spitzer, Guanya Shi, Byron Boots, 'DATT: Deep Adaptive Trajectory Tracking for Quadrotor Control', arXiv:2310.09053 (2023), https://arxiv.org/abs/2310.09053
7. Tianhua Gao, Masashi Izumita, Kohji Tomita, Akiya Kamimura, 'A Sliced Learning Framework for Online Disturbance Identification in Quadrotor SO(3) Attitude Control', arXiv:2508.14422 (2025), https://arxiv.org/abs/2508.14422
8. Weibin Gu, Jiance Zhao, Alessandro Rizzo, 'Learning uncertainties online for quadrotor flight control: A comparative study', Journal of Intelligent & Robotic Systems 111 (2025). DOI: 10.1007/s10846-025-02305-5, https://doi.org/10.1007/s10846-025-02305-5
9. Mingxin Wei, Min Li, Bingzhi Zou, Lanxiang Zheng, 'Online Learning-based Model Predictive Control for Quadrotors Under Uncertain Wind Disturbances', Proceedings of the 2024 4th International Conference on Control and Intelligent Robotics (ICARM) (2024). DOI: 10.1145/3687488.3687518, https://doi.org/10.1145/3687488.3687518