arxiv: 2604.06337 · v1 · submitted 2026-04-07 · 📡 eess.SY · cs.SY

Recognition: 2 theorem links

· Lean Theorem

Improving INDI for Input Nonaffine Systems via Learning-Based Nonlinear Control Allocation

Adam Hallmark , Pan Zhao

Authors on Pith no claims yet

Pith reviewed 2026-05-10 18:57 UTC · model grok-4.3

classification 📡 eess.SY cs.SY

keywords incremental nonlinear dynamic inversioncontrol allocationinput nonaffine systemssupervised learningnonlinear programmingactuator model

0 comments

The pith

Standard INDI with incremental control allocation fails for input nonaffine systems because it depends on an invalid linear actuator approximation, which a supervised learning model can replace by approximating the required nonlinear solve.

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

The paper demonstrates that standard incremental nonlinear dynamic inversion combined with incremental control allocation cannot be applied directly to input nonaffine systems, as this requires an untenable linear approximation of the actuator model. Retaining the static allocation approach then demands solving a nonlinear programming problem online to find the correct control inputs. To make this feasible in real time, the authors introduce a supervised learning model trained offline to approximate the nonlinear programming solution. Numerical experiments on an example system confirm the breakdown of the standard method and show that the learning-based allocation maintains INDI performance while remaining computationally light. This addresses a practical barrier for INDI in systems with nonaffine actuator behavior.

Core claim

For input nonaffine systems, applying standard INDI with ICA relies on an untenable linear approximation of the actuator model. Avoiding this limitation while keeping the static control allocation paradigm requires solving a nonlinear programming problem online. A supervised learning-based approach approximates this solution in real time, and numerical experiments validate both the identified limitations of the standard method and the effectiveness of the proposed alternative without degrading the stability or performance of the underlying INDI controller.

What carries the argument

Supervised learning model trained offline to approximate the solution of the nonlinear programming problem for nonlinear control allocation.

If this is right

Standard INDI plus incremental control allocation produces incorrect commands on input nonaffine systems due to the linear actuator approximation.
Solving the full nonlinear programming problem restores correctness but creates an online computational burden.
The supervised learning approximation yields a computationally tractable implementation that preserves INDI stability and tracking properties.
Numerical experiments on an example problem confirm both the failure mode of the linear method and the viability of the learned allocation.

Where Pith is reading between the lines

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

The same learning approximation could be applied to other incremental control methods that currently rely on linear allocation steps.
If the training data densely covers the actuator operating range, the approach might extend to hardware-in-the-loop tests without retraining.
Replacing the NLP solve with a learned map could enable higher control rates on embedded processors that cannot run real-time optimization.

Load-bearing premise

A supervised learning model trained offline can accurately and stably approximate the solution of the required nonlinear programming problem for control allocation in real time without degrading the stability or performance properties of the underlying INDI controller.

What would settle it

Deploy the learned allocation on a closed-loop nonaffine actuator system and observe either closed-loop instability or tracking error that exceeds the performance of an exact online NLP solver run at the same rate.

Figures

Figures reproduced from arXiv: 2604.06337 by Adam Hallmark, Pan Zhao.

**Figure 2.** Figure 2: Control architecture for the planar bi-tilt tricopter [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Figure 3: State trajectories for the SingleStep reference trajectory under the LCA (top), NLP (middle), and NN (bottom) methods. infeasible allocation problems under the input constraints. However, the LCA method results in persistently large feedback linearization errors during the second doublet, even when the QP remains feasible under the linearized equality constraint (5). The LCA method also experiences spike… view at source ↗

**Figure 5.** Figure 5: Feedback linearization errors, defined as [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

read the original abstract

This paper first demonstrates that applying standard incremental nonlinear dynamic inversion (INDI) with incremental control allocation (ICA) to input nonaffine systems relies on an untenable linear approximation of the actuator model. It then shows that avoiding this issue, while retaining the static control allocation paradigm, generally requires solving a nonlinear programming (NLP) problem. To address the associated online computational challenges, the paper subsequently presents a supervised learning-based approach. Numerical experiments on an example problem validate the identified limitations of standard INDI + ICA for input nonaffine systems, while also demonstrating that the proposed learning-based method provides an effective and computationally tractable alternative.

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.

Desk Editor's Note private letter to a colleague

The paper flags a real linear-approx flaw in INDI+ICA for nonaffine actuators and offers a supervised-learning workaround that works on their example, but skips error bounds and stability checks.

read the letter

Hi, The paper's main contribution is spotting that incremental nonlinear dynamic inversion with incremental control allocation relies on a linear actuator model that doesn't hold for nonaffine systems. They explain why that leads to problems and note that a proper fix means solving a nonlinear optimization for allocation each time. To make it practical, they train a supervised learner to mimic the solution of that optimization. The work does well in making the limitation explicit and then showing through simulation on an example system that the standard method underperforms while their approach recovers good performance without the online computation cost. It's a direct response to a practical issue in applying INDI to more realistic actuator dynamics. Where it falls short is in the analysis of the learned approximator. The paper doesn't provide bounds on the approximation error or examine how those errors propagate through the INDI loop to affect stability or performance. The numerical validation is confined to a single problem, without comparisons to other allocation methods or tests on robustness to model mismatch or disturbances. Engineers dealing with control of systems like UAVs or robotic arms where actuators have nonlinear characteristics would find this relevant. It keeps the incremental nature of INDI intact while addressing the allocation bottleneck. The paper shows clear thinking on the control side and honest engagement with the limitations of current methods. It deserves peer review because the identified issue is important for the field and the proposed solution is straightforward to implement, even if additional theoretical support would make the claims stronger.

Referee Report

2 major / 2 minor

Summary. The manuscript first shows that standard INDI combined with ICA for input nonaffine systems rests on an untenable linear approximation of the actuator map. It then argues that retaining a static allocation structure requires solving a nonlinear program (NLP) at each step. To make this tractable, the authors introduce a supervised learning model trained offline to approximate the NLP solution and replace the online optimizer. Numerical experiments on a single example system are presented to illustrate both the failure of the linear approximation and the practical performance of the learned allocator.

Significance. If the learned approximator can be equipped with explicit error bounds that preserve the incremental inversion and closed-loop stability of INDI, the approach would offer a concrete route to extend INDI to nonaffine actuators while retaining real-time feasibility. The identification of the linear-approximation limitation itself is a useful diagnostic for the INDI literature.

major comments (2)

[Supervised learning approach] The section presenting the supervised learning approximator contains no approximation-error bounds, Lipschitz constants on the learned map, or Lyapunov-style argument that accounts for the residual between the learned allocation and the exact NLP solution. Because the central claim is that the substitution preserves the incremental inversion and stability properties of INDI, this omission is load-bearing.
[Numerical experiments] The numerical-experiments section reports validation on a single example but supplies neither quantitative metrics (e.g., tracking error norms, control effort, or stability margins), explicit baselines (exact NLP solver, other approximators), nor any closed-loop robustness checks under the learned residual. Without these, the claim that the method is 'effective' cannot be assessed.

minor comments (2)

[Abstract and Experiments] The abstract states that experiments 'validate' both the limitation and the new method but provides no quantitative summary; the same level of detail should appear in the main text's experiment description.
[Preliminaries and Learning Method] Notation for the actuator map and the learned allocation function is introduced without a clear table of symbols or explicit statement of the training loss and data-generation procedure.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which identify key areas where the manuscript can be strengthened. We address each major comment below, indicating the revisions we will make to the next version of the paper.

read point-by-point responses

Referee: [Supervised learning approach] The section presenting the supervised learning approximator contains no approximation-error bounds, Lipschitz constants on the learned map, or Lyapunov-style argument that accounts for the residual between the learned allocation and the exact NLP solution. Because the central claim is that the substitution preserves the incremental inversion and closed-loop stability properties of INDI, this omission is load-bearing.

Authors: We agree that the lack of explicit error bounds or a Lyapunov-style argument for the approximation residual is a substantive gap, as the central claim of the work depends on the learned allocator sufficiently preserving the incremental inversion properties. The manuscript emphasizes the conceptual demonstration that standard INDI+ICA relies on an untenable linearization for nonaffine actuators and that a learned nonlinear allocator offers a practical alternative. In the revision we will add a dedicated subsection discussing the empirical approximation error on the training and test data, including observed maximum residuals and an empirical check of local Lipschitz behavior of the learned map. We will also explicitly state that a full closed-loop stability proof that accounts for the residual is left for future work. This is a partial revision. revision: partial
Referee: [Numerical experiments] The numerical-experiments section reports validation on a single example but supplies neither quantitative metrics (e.g., tracking error norms, control effort, or stability margins), explicit baselines (exact NLP solver, other approximators), nor any closed-loop robustness checks under the learned residual. Without these, the claim that the method is 'effective' cannot be assessed.

Authors: We acknowledge that the numerical section is limited to a single illustrative example and does not yet contain the quantitative metrics, baselines, or robustness checks needed to fully substantiate the effectiveness claim. The experiments were designed to highlight the linear-approximation failure and to show that the learned allocator runs in real time while producing acceptable closed-loop behavior. In the revised manuscript we will augment the section with RMS tracking-error norms, integrated control effort, direct comparison against the exact NLP solver as baseline, and additional closed-loop simulations that inject the observed approximation residual as well as modest plant-parameter variations. These additions will be presented in new tables and figures. revision: yes

Circularity Check

0 steps flagged

No significant circularity in the derivation chain

full rationale

The paper first analyzes the standard INDI+ICA approach and shows via the actuator model that it relies on an untenable linear approximation for input nonaffine systems. It then states that retaining static allocation requires solving an NLP and introduces a supervised learning approximator to address online computation. Numerical validation on an example follows. No load-bearing step reduces by the paper's own equations or self-citation to its inputs: there is no self-definitional mapping, no fitted parameter renamed as a prediction, and no uniqueness theorem or ansatz imported from the authors' prior work. The claims rest on explicit modeling analysis and external numerical demonstration rather than construction from the method's own outputs.

Axiom & Free-Parameter Ledger

1 free parameters · 0 axioms · 0 invented entities

Only the abstract is available, so the ledger is necessarily incomplete. The learning-based method almost certainly introduces fitted parameters from supervised training; no other free parameters, axioms, or invented entities are identifiable from the given text.

free parameters (1)

supervised learning model parameters
The neural network or regressor weights are fitted to data generated by the nonlinear solver; these are required for the online approximation to work.

pith-pipeline@v0.9.0 · 5396 in / 1231 out tokens · 48459 ms · 2026-05-10T18:57:19.929921+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

applying standard incremental nonlinear dynamic inversion (INDI) with incremental control allocation (ICA) to input nonaffine systems relies on an untenable linear approximation of the actuator model
IndisputableMonolith/Foundation/BranchSelection.lean branch_selection unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

the proposed learning-based method provides an effective and computationally tractable alternative

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

18 extracted references · 18 canonical work pages

[1]

H. K. Khalil,Nonlinear Systems. Upper Saddle River, NJ, USA: Prentice Hall, 2002

work page 2002
[2]

J.-J. E. Slotine and W. Li,Applied Nonlinear Control. Englewood Cliffs, NJ, USA: Prentice Hall, 1991

work page 1991
[3]

Dynamic inversion: an evolving methodology for flight control design,

D. Enns, D. Bugajski, R. Hendrick, and G. Stein, “Dynamic inversion: an evolving methodology for flight control design,”International Journal of Control, vol. 59, no. 1, pp. 71–91, 1994

work page 1994
[4]

Cascaded incremental nonlinear dynamic inversion for mav disturbance rejection,

E. J. Smeur, G. C. de Croon, and Q. Chu, “Cascaded incremental nonlinear dynamic inversion for mav disturbance rejection,”Control Engineering Practice, vol. 73, pp. 79–90, 2018

work page 2018
[5]

Robust flight control using incremental nonlinear dynamic inversion and angular acceleration prediction,

S. Sieberling, Q. Chu, and J. Mulder, “Robust flight control using incremental nonlinear dynamic inversion and angular acceleration prediction,”Journal of Guidance, Control, and Dynamics, vol. 33, no. 6, pp. 1732–1742, 2010

work page 2010
[6]

Aircraft fault-tolerant trajectory control using incremental nonlinear dynamic inversion,

P. Lu, E.-J. Van Kampen, C. De Visser, and Q. Chu, “Aircraft fault-tolerant trajectory control using incremental nonlinear dynamic inversion,”Control Engineering Practice, vol. 57, pp. 126–141, 2016

work page 2016
[7]

Modeling and incremental nonlinear dynamic inversion control of a novel unmanned tiltrotor,

G. Di Francesco and M. Mattei, “Modeling and incremental nonlinear dynamic inversion control of a novel unmanned tiltrotor,”Journal of Aircraft, vol. 53, no. 1, pp. 73–86, 2016

work page 2016
[8]

From fundamentals to applications of incremental nonlinear dynamic inversion: A survey on INDI–part I,

A. Steinert, R. Stefan, S. Hafner, F. Holzapfel, and H. Haichao, “From fundamentals to applications of incremental nonlinear dynamic inversion: A survey on INDI–part I,”Chinese Journal of Aeronautics, p. 103553, 2025

work page 2025
[9]

Control allocation for over-actuated systems,

M. W. Oppenheimer, D. B. Doman, and M. A. Bolender, “Control allocation for over-actuated systems,” in2006 14th Mediterranean Conference on Control and Automation, pp. 1–6, 2006

work page 2006
[10]

Control allocation—a survey,

T. A. Johansen and T. I. Fossen, “Control allocation—a survey,” Automatica, vol. 49, no. 5, pp. 1087–1103, 2013

work page 2013
[11]

Minimum power control allocation for incremental control of over-actuated transition aircraft,

O. Pfeifle and W. Fichter, “Minimum power control allocation for incremental control of over-actuated transition aircraft,”Journal of Guidance, Control, and Dynamics, vol. 46, no. 2, pp. 286–300, 2023

work page 2023
[12]

Incremental nonlinear control allocation for a tailless aircraft with innovative control effectors,

I. Matamoros and C. C. de Visser, “Incremental nonlinear control allocation for a tailless aircraft with innovative control effectors,” in 2018 AIAA Guidance, Navigation, and Control Conference, p. 1116, 2018

work page 2018
[13]

Reconfigurable flight control using nonlin- ear dynamic inversion with a special accelerometer implementation,

B. Bacon and A. Ostroff, “Reconfigurable flight control using nonlin- ear dynamic inversion with a special accelerometer implementation,” inAIAA Guidance, Navigation, and Control Conference and Exhibit, p. 4565, 2000

work page 2000
[14]

Stability analysis for incremental nonlinear dynamic inversion control,

X. Wang, E.-J. Van Kampen, Q. Chu, and P. Lu, “Stability analysis for incremental nonlinear dynamic inversion control,”Journal of Guidance, Control, and Dynamics, vol. 42, no. 5, pp. 1116–1129, 2019

work page 2019
[15]

Nonlinear control allocation: A learning based approach,

H. Z. I. Khan, S. Mobeen, J. Rajput, and J. Riaz, “Nonlinear control allocation: A learning based approach,” in2024 IEEE 63rd Conference on Decision and Control (CDC), pp. 2833–2838, IEEE, 2024

work page 2024
[16]

Constrained nonlinear control allocation based on deep auto-encoder neural networks,

H. Huan, W. Wan, C. We, and Y . He, “Constrained nonlinear control allocation based on deep auto-encoder neural networks,” in2018 European Control Conference (ECC), pp. 1–8, IEEE, 2018

work page 2018
[17]

Constrained control allocation,

W. C. Durham, “Constrained control allocation,”Journal of Guidance, Control, and Dynamics, vol. 16, no. 4, pp. 717–725, 1993

work page 1993
[18]

Constrained control allocation - three-moment prob- lem,

W. C. Durham, “Constrained control allocation - three-moment prob- lem,”Journal of Guidance, Control, and Dynamics, vol. 17, no. 2, pp. 330–336, 1994

work page 1994