pith. sign in

arxiv: 2605.01556 · v1 · submitted 2026-05-02 · 📡 eess.SY · cs.SY

A Universal Optimal Control Strategy for a Tailsitter UAV

Animesh Kumar Shastry , Mangal Kothari This is my paper

Pith reviewed 2026-05-09 17:42 UTC · model grok-4.3

classification 📡 eess.SY cs.SY
keywords tailsitter UAVmodel predictive controltrajectory optimizationneural network approximationuniversal controllertransition maneuverquadrotor biplaneoptimal control
0
0 comments X

The pith

A single model predictive controller tracks neural network-generated optimal trajectories to handle hover, transition, and cruise in one tailsitter UAV without switching or gain scheduling.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a unified optimal control approach for a quadrotor biplane tailsitter UAV that must move between vertical hover, angled transition, and forward cruise. It first solves a nonlinear programming problem with direct collocation to produce safe, time-efficient transition trajectories that respect actuator limits and angle-of-attack bounds. These trajectories are then approximated by feedforward neural networks so they can be generated quickly for any initial cruise speed. The resulting state and input profiles serve as references for a model predictive controller that tracks the vehicle across all regimes. The result is one controller that works everywhere, shown in simulation to be more robust to model errors than a dynamic inversion alternative.

Core claim

The authors formulate a trajectory optimization scheme based on nonlinear programming with direct collocation that incorporates the vehicle's nonlinear dynamics, actuator saturation, and angle-of-attack constraints. Optimal cruise-to-hover trajectories are computed for a range of initial velocities and then learned by multilayer feedforward neural networks. These networks supply both feedforward inputs and reference state profiles that are tracked in real time by a model predictive controller. The MPC formulation eliminates the need for mode-specific controllers or gain scheduling, providing a single universal law that operates across hover, transition, and cruise. Numerical comparisons with

What carries the argument

The model predictive controller that uses neural network approximations of precomputed optimal trajectories as references while enforcing the full nonlinear dynamics and constraints in its prediction horizon.

If this is right

  • Optimal trajectories generated offline over a range of cruise speeds can be learned by neural networks for real-time constraint-satisfying reference generation.
  • The same MPC law tracks the vehicle in hover, transition, and cruise without any switching or retuning.
  • MPC achieves greater robustness to parameter uncertainties than a nonlinear dynamic inversion controller in simulation.
  • Two different numerical solvers for the MPC problem allow explicit trade-offs between computation time and closed-loop performance.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same neural-network-plus-MPC structure could be applied to other vehicles that perform large attitude changes, such as tilt-wing or tiltrotor aircraft, if comparable trajectory data sets are collected.
  • Hardware experiments would be needed to check whether unmodeled effects like wind gusts during transition still keep the angle-of-attack constraint satisfied.
  • Faster MPC solvers or embedded optimization hardware could reduce the execution-time gap between the two numerical schemes while preserving the reported robustness benefit.
  • Because the controller is formulated once for the entire envelope, certification or safety-case efforts could focus on a single set of stability and constraint proofs rather than separate ones for each mode.

Load-bearing premise

The nonlinear dynamics model used inside the optimizer and MPC accurately represents the vehicle's rapidly changing aerodynamics, and the neural network produces trajectories that remain feasible under real-world disturbances.

What would settle it

A hardware flight test in which measured aerodynamic forces during a high-speed transition cause the closed-loop system to violate the angle-of-attack limit or lose stability despite the MPC running at the reported rate.

Figures

Figures reproduced from arXiv: 2605.01556 by Animesh Kumar Shastry, Mangal Kothari.

Figure 2
Figure 2. Figure 2: Flight modes of a Biplane Quadrotor VTOL UAV. The vehicle is subjected to gravitational, aerodynamic, and propulsive forces, along with aerodynamic and propulsive moments. In the inertial Xi–Zi plane, the vehicle position is [x z] T and the velocity is [ ˙x z˙] T . The pitch angle about the Yi axis is denoted by θ, with pitch rate ˙θ. The total thrust generated by the rotors is represented by the control i… view at source ↗
Figure 3
Figure 3. Figure 3: Airfoil geometry and angle definitions. -200 -150 -100 -50 0 50 100 150 200 AOA (degrees) -3 -2 -1 0 1 2 3 Aerodynamic coeffients CL CD view at source ↗
Figure 5
Figure 5. Figure 5: Collocation point between two adjacent knots. view at source ↗
Figure 6
Figure 6. Figure 6: Nonlinear programming solution scheme. 6 view at source ↗
Figure 7
Figure 7. Figure 7: Planar view of optimized hover-to-cruise trajectories. Red arrows indicate thrust direction. view at source ↗
Figure 8
Figure 8. Figure 8: State trajectories for different target cruise speeds. view at source ↗
Figure 10
Figure 10. Figure 10: Angle of attack along the optimized trajectories. view at source ↗
Figure 11
Figure 11. Figure 11: State trajectories for the cruise-to-hover dataset. Different colors correspond to different initial view at source ↗
Figure 12
Figure 12. Figure 12: Control inputs for the trajectory dataset. 0 0.5 1 1.5 2 2.5 Time(s) -20 -10 0 10 20 α (deg) view at source ↗
Figure 14
Figure 14. Figure 14: Planar view of the cruise-to-hover trajec view at source ↗
Figure 16
Figure 16. Figure 16: Test MSE versus number of hidden neurons. Five neurons yield the best approxima￾tion accuracy. 5 Feedback Control Design To ensure that the UAV follows the optimal trajectories and rejects disturbances, a feedback control law is required. Such controllers use the desired state information supplied by the trajectory generator together with the current state estimates obtained from onboard sensors. The syst… view at source ↗
Figure 17
Figure 17. Figure 17: Block diagram of the Dynamic Inversion feedback controller. view at source ↗
read the original abstract

This work develops a unified optimal control framework for a Quadrotor Biplane tailsitter UAV capable of operating seamlessly across hover, transition, and cruise flight regimes. Although the tailsitter configuration enables mechanically simple mode switching, the transition maneuver remains challenging due to strong nonlinearities and rapidly varying aerodynamics. To address this, a trajectory optimization scheme based on nonlinear programming with direct collocation is formulated, incorporating nonlinear dynamics, actuator limits, and angle-of-attack constraints. The resulting optimal trajectories are safe, reliable, and time-efficient. For the cruise-to-hover maneuver, optimal trajectories are generated over a range of initial cruise velocities and subsequently learned using feedforward multilayer neural networks. The learned model generalizes across operating conditions and enables real-time generation of constraint-satisfying transition trajectories. These trajectories provide both feedforward control inputs and reference state profiles, which are tracked using a Model Predictive Controller (MPC). The MPC eliminates the need for controller switching or gain scheduling across flight envelopes, enabling a single universal controller for hover, transition, and cruise. A nonlinear Dynamic Inversion (DI) controller is also designed for comparison. Two numerical schemes for MPC are implemented and evaluated. Simulation results across all flight modes demonstrate that MPC achieves superior robustness to parameter uncertainties compared to DI. A computational cost analysis further highlights the trade-off between execution time and performance for the different MPC solvers.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. The paper develops a unified optimal control framework for a Quadrotor Biplane tailsitter UAV. It formulates a trajectory optimization problem using nonlinear programming and direct collocation that incorporates nonlinear dynamics, actuator limits, and angle-of-attack constraints to produce safe, time-efficient transition trajectories. These trajectories, generated for a range of initial cruise velocities in the cruise-to-hover case, are learned by a feedforward multilayer neural network to enable real-time reference generation. The learned references are tracked by a single Model Predictive Controller (MPC) that operates without mode switching or gain scheduling across hover, transition, and cruise. The approach is compared to a nonlinear Dynamic Inversion (DI) controller, with simulations claiming superior MPC robustness to parameter uncertainties and an analysis of computational trade-offs for different MPC solvers.

Significance. If the quantitative claims hold, the work would demonstrate a practical path to a single universal controller for tailsitter UAVs, reducing the engineering burden of mode-specific designs while respecting actuator and aerodynamic constraints during rapid transitions. The combination of direct-collocation optimization, neural-network learning of optimal references, and MPC tracking is a coherent strategy for real-time feasibility. However, the absence of supporting numerical evidence in the manuscript limits its assessed significance at present.

major comments (3)
  1. [Abstract] Abstract: the statement that 'simulation results across all flight modes demonstrate that MPC achieves superior robustness to parameter uncertainties compared to DI' supplies no quantitative metrics, error bars, uncertainty magnitudes, or details on how constraints were enforced or violated, preventing verification of support for the central claim of a universal controller.
  2. [Abstract] Abstract: the claim that the feedforward neural network 'generalizes across operating conditions and enables real-time generation of constraint-satisfying transition trajectories' is not accompanied by reported generalization errors, performance on unseen initial velocities, or explicit checks that angle-of-attack and actuator limits remain satisfied, which is load-bearing for the assertion that a single MPC formulation works universally without implicit retuning.
  3. [Abstract] The central claim that the same MPC (prediction model, horizon, costs, and constraints) tracks references in hover, transition, and cruise without any mode-dependent retuning rests on the fidelity of the nonlinear aerodynamics model during rapid variation; no sensitivity analysis or validation against higher-fidelity aero data is described to confirm this assumption holds.
minor comments (1)
  1. [Abstract] The abstract mentions 'Two numerical schemes for MPC are implemented and evaluated' and a 'computational cost analysis' but does not name the schemes or report the specific execution times versus performance trade-offs.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback, which identifies opportunities to strengthen the quantitative support and clarity of our claims. We address each major comment point by point below, indicating the revisions we will make to the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the statement that 'simulation results across all flight modes demonstrate that MPC achieves superior robustness to parameter uncertainties compared to DI' supplies no quantitative metrics, error bars, uncertainty magnitudes, or details on how constraints were enforced or violated, preventing verification of support for the central claim of a universal controller.

    Authors: We agree that the abstract would benefit from explicit quantitative metrics to support the robustness claim. The full manuscript (Section 5) presents simulation results comparing MPC and DI under parameter uncertainties, including specific tracking error statistics, robustness margins, and constraint satisfaction rates across flight modes. We will revise the abstract to incorporate key quantitative findings from these simulations, such as average error reductions and violation frequencies, ensuring the central claim is verifiable. revision: yes

  2. Referee: [Abstract] Abstract: the claim that the feedforward neural network 'generalizes across operating conditions and enables real-time generation of constraint-satisfying transition trajectories' is not accompanied by reported generalization errors, performance on unseen initial velocities, or explicit checks that angle-of-attack and actuator limits remain satisfied, which is load-bearing for the assertion that a single MPC formulation works universally without implicit retuning.

    Authors: The manuscript describes training the neural network on optimal trajectories for a range of initial cruise velocities and evaluating it on unseen conditions, with results confirming low generalization errors and constraint satisfaction. However, these details are not summarized in the abstract. We will update the abstract to report the generalization error metrics (e.g., test-set MSE), performance on held-out velocities, and explicit verification that angle-of-attack and actuator limits are respected in the generated trajectories. revision: yes

  3. Referee: [Abstract] The central claim that the same MPC (prediction model, horizon, costs, and constraints) tracks references in hover, transition, and cruise without any mode-dependent retuning rests on the fidelity of the nonlinear aerodynamics model during rapid variation; no sensitivity analysis or validation against higher-fidelity aero data is described to confirm this assumption holds.

    Authors: The MPC formulation employs identical prediction model, horizon, costs, and constraints across all regimes, with simulations in the manuscript demonstrating successful tracking without retuning. We acknowledge that a dedicated sensitivity analysis against higher-fidelity aerodynamic data is not currently included. We will add a discussion section addressing the model's assumptions and, where feasible, include sensitivity results by varying key aerodynamic parameters; otherwise, we will explicitly note this as a limitation of the present validation. revision: partial

Circularity Check

0 steps flagged

No circularity; derivation chain is self-contained

full rationale

The paper outlines a standard optimal control pipeline: offline nonlinear programming with direct collocation to generate trajectories from nonlinear dynamics, actuator limits, and AoA constraints; training a feedforward NN on those trajectories for real-time generalization; and using the resulting references in an MPC formulation that is designed to operate without mode switching. No step reduces a claimed prediction or result to its own inputs by construction (e.g., no fitted parameter is relabeled as a prediction, no self-definitional loop in the dynamics or controller, and no load-bearing self-citation or imported uniqueness theorem). The universal-controller claim rests on the MPC being formulated once with the same prediction model, horizon, costs, and constraints across regimes, supported by simulation comparisons to DI. This is a conventional separation of offline optimization, learning-based approximation, and online tracking, with no evident reduction to tautology.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The framework rests on standard assumptions from nonlinear control theory and optimization, with fitted elements introduced during trajectory generation and neural network training.

free parameters (1)
  • Optimization cost weights and neural network hyperparameters
    Weights in the nonlinear programming objective and parameters for training the multilayer neural network on optimal trajectories are chosen or fitted to data.
axioms (1)
  • domain assumption The nonlinear vehicle dynamics model sufficiently captures angle-of-attack dependent aerodynamics during transition
    Invoked to ensure that trajectories produced by the optimizer are physically realizable and constraint-satisfying.

pith-pipeline@v0.9.0 · 5543 in / 1374 out tokens · 68816 ms · 2026-05-09T17:42:39.277863+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

22 extracted references

  1. [1]

    Stone and G Clarke

    R H. Stone and G Clarke. Optimization of transition manoeuvres for a tail-sitter unmanned air vehicle (uav). 04 2019

    2019

  2. [2]

    Transition between level flight and hovering of a tail-sitter vertical takeoff and landing aerial robot.Advanced Robotics, 24:763–781, 04 2010

    Koichi Kita, Atsushi Konno, and Masaru Uchiyama. Transition between level flight and hovering of a tail-sitter vertical takeoff and landing aerial robot.Advanced Robotics, 24:763–781, 04 2010

    2010

  3. [3]

    Attitude control of a novel tailsitter: Swiveling biplane-quadrotor.Journal of Guidance, Control, and Dynamics, 43(3):599–607, 2020

    Nidhish Raj, Ravi Banavar, Abhishek, and Mangal Kothari. Attitude control of a novel tailsitter: Swiveling biplane-quadrotor.Journal of Guidance, Control, and Dynamics, 43(3):599–607, 2020

    2020

  4. [4]

    Model based transition control for biplane tailsitter

    Shubhanshu Gupta, Sreekar Doranala, Mangal Kothari, and Abhishek. Model based transition control for biplane tailsitter. InProceedings of the AIAA SciTech Forum, 2025

    2025

  5. [5]

    H inf ty robust control of a quadrotor biplane tailsitter uav

    Tanay Kumar, Mangal Kothari, and Raktim Bhattacharya. H inf ty robust control of a quadrotor biplane tailsitter uav. InProceedings of the AIAA SciTech Forum, 2024

    2024

  6. [6]

    Optimal transition trajectory of a quadrotor biplane tailsitter

    Shubhanshu Gupta, Mangal Kothari, and Abhishek. Optimal transition trajectory of a quadrotor biplane tailsitter. InProceedings of the IF AC World Congress, 2023

    2023

  7. [7]

    Iterative learning based feedforward control for transition of a biplane-quadrotor tailsitter uas

    Nidhish Raj, Abhishek, and Mangal Kothari. Iterative learning based feedforward control for transition of a biplane-quadrotor tailsitter uas. InProceedings of the IEEE International Conference on Robotics and Automation, 2020

    2020

  8. [8]

    Piecewise polynomial modeling for control and analysis of aircraft dynamics beyond stall.Journal of Guidance, Control, and Dynamics, 12 2018

    Torbjørn Cunis, Laurent Burlion, and Jean-Philippe Condomines. Piecewise polynomial modeling for control and analysis of aircraft dynamics beyond stall.Journal of Guidance, Control, and Dynamics, 12 2018

    2018

  9. [9]

    Mathematical modelling of aircraft un- steady aerodynamics at high incidence with account of wing-tail interaction

    Alexander Khrabrov, Yu Vinogradov, and Nikolay Abramov. Mathematical modelling of aircraft un- steady aerodynamics at high incidence with account of wing-tail interaction. 08 2004

    2004

  10. [10]

    Mcclamroch and Ilya Kolmanovsky

    N.H. Mcclamroch and Ilya Kolmanovsky. Hybrid switched mode control approach for v/stol flight control problems. volume 3, pages 2648 – 2653 vol.3, 01 1997

    1997

  11. [11]

    Switching control approach for stable transition state process on hybrid vertical take-off and landing uav

    Ghozali Hadi, Harish Putra, Puspita Triana Dewi, Aris Budiyarto, and Agus Budiyono. Switching control approach for stable transition state process on hybrid vertical take-off and landing uav. 07 2017

    2017

  12. [12]

    Switching control approach for stable transition state process on hybrid vertical take-off and landing uav

    Ghozali Hadi, Harish Putra, Puspita Triana Dewi, Aris Budiyarto, and Agus Budiyono. Switching control approach for stable transition state process on hybrid vertical take-off and landing uav. 07 2017. 15

    2017

  13. [13]

    K. Kita, A. Konno, and M. Uchiyama. Transition between level flight and hovering of a tail-sitter vertical takeoff and landing aerial robot.Advanced Robotics, 24(5-6):763–781, 2010

    2010

  14. [14]

    Hernandez-Garcia and Hugo Rodriguez

    Rogelio G. Hernandez-Garcia and Hugo Rodriguez. Transition flight control of a cyclic tiltrotor uav based on the gain-scheduling strategy. pages 951–956, 07 2015

    2015

  15. [15]

    Gain scheduling pid control of the quad-rotor helicopter

    Jing Qiao, Zhixiang Liu, and Youmin Zhang. Gain scheduling pid control of the quad-rotor helicopter. pages 1594–1601, 10 2017

    2017

  16. [16]

    Survey of advances in guidance, navigation, and control of unmanned rotorcraft systems

    Farid Kendoul. Survey of advances in guidance, navigation, and control of unmanned rotorcraft systems. Journal of Field Robotics, 29:315 – 378, 03 2012

    2012

  17. [17]

    Nonlinear model predictive tracking control for rotorcraft- based unmanned aerial vehicles

    H Jin Kim, David Shim, and Shankar Sastry. Nonlinear model predictive tracking control for rotorcraft- based unmanned aerial vehicles. volume 5, pages 3576 – 3581 vol.5, 02 2002

    2002

  18. [18]

    Nonlinear model predictive control technique for unmanned air vehicles.Journal of Guidance, Control, and Dynamics, 29:1179–1188, 09 2006

    Nathan Slegers, Jason Kyle, and Mark Costello. Nonlinear model predictive control technique for unmanned air vehicles.Journal of Guidance, Control, and Dynamics, 29:1179–1188, 09 2006

    2006

  19. [19]

    Hierarchical and hybrid model predictive control of quadcopter air vehicles

    Alberto Bemporad, Carlo Pascucci, and C Rocchi. Hierarchical and hybrid model predictive control of quadcopter air vehicles. volume 3, 09 2009

    2009

  20. [20]

    Linear tracking for a fixed-wing uav using nonlinear model predictive control.Control Systems Technology, IEEE Transactions on, 17:1202 – 1210, 10 2009

    Yeonsik Kang and J Karl Hedrick. Linear tracking for a fixed-wing uav using nonlinear model predictive control.Control Systems Technology, IEEE Transactions on, 17:1202 – 1210, 10 2009

    2009

  21. [21]

    Biplane-quadrotor tail- sitter uav: Flight dynamics and control.Journal of Guidance, Control, and Dynamics, 41(5):1049–1067, 2018

    Swati Swarnkar, Hardik Parwana, Mangal Kothari, and Abhishek Abhishek. Biplane-quadrotor tail- sitter uav: Flight dynamics and control.Journal of Guidance, Control, and Dynamics, 41(5):1049–1067, 2018

    2018

  22. [22]

    ∂Fax ∂˙z(t) ∂Faz ∂˙z(t) # . Derivative w.r.t.θ(t) ∂H ∂θ(t) = 2(θ(t)−θ(t f))QX(5,5) + u1(t) m (λ2(t) cosθ(t)−λ 4(t) sinθ(t)) + 1 m λ2(t)λ 4(t) cosθ(t) sinθ(t) −sinθ(t) cosθ(t)

    Martin Fodslette Moller. A scaled conjugate gradient algorithm for fast supervised learning.Neural Networks, 6:525–533, 1993. A Derivatives for the Indirect MPC Solver This appendix provides the full expressions required for the indirect optimal control method used in the MPC solver. These include the costate dynamics, the partial derivatives of the Hamil...

    1993