Learning Flatness-Preserving Residuals for Pure-Feedback Systems
Fengjun Yang, Jake Welde, Nikolai Matni · 2025
Read Learning Flatness-Preserving Residuals for Pure-Feedback Systems by Fengjun Yang, Jake Welde, Nikolai Matni (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-04324.txt.
Introduction
Some machines are flat: from a few chosen outputs, every part of their motion can be planned and steered, which makes control design far easier. Fengjun Yang, Jake Welde, and Nikolai Matni ask what happens when engineers bolt a learned correction onto such a machine's model. A careless correction can quietly destroy the flatness that made control easy.
They show that corrections with a staircase structure, each layer depending only on earlier ones, keep flatness intact for the pure-feedback machines they study. They also give a recipe to rebuild the key change of variables for the corrected model from the original one. On a simulated quadrotor with unmodeled air effects, the learned flat model tracked paths five times better than the nominal one.
Catalog record
| CREATOR | Fengjun Yang, Jake Welde, Nikolai Matni |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 43 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-04324
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Learning Flatness-Preserving Residuals for Pure-Feedback Systems Fengjun Yang, Jake Welde, Nikolai Matni Abstract We study residual dynamics learning for differentially flat systems, where a nominal model is augmented with a learned correction term from data. A key challenge is that generic residual parameterizations may destroy flatness, limiting the applicability of flatness-based planning and control methods. To address this, we propose a framework for learning flatness-preserving residual dynamics in systems whose nominal model admits a pure-feedback form. We show that residuals with a lower-triangular structure preserve both the flatness of the system and the original flat outputs. Moreover, we provide a constructive procedure to recover the flatness diffeomorphism of the augmented system from that of the nominal model. Building on these insights, we introduce a parameterization of flatness-preserving residuals using smooth function approximators, making them learnable from trajectory data with conventional algorithms. Our approach is validated in simulation on a 2D quadrotor subject to unmodeled aerodynamic effects. We demonstrate that the resulting learned flat model achieves a tracking error 5 × 5\times lower than the nominal flat model, while being 20 × 20\times faster over a structure-agnostic alternative. 1 Introduction Differential flatness enables the use of simple and efficient linear techniques for trajectory planning and feedback design in nonlinear systems ( Fliess et al., 1995 ) . For systems such as quadrotors, the existence of a flatness diffeomorphism allows the system to be transformed into a chain of integrators, greatly simplifying control ( Mellinger and Kumar, 2011 ; Greeff and Schoellig, 2018 ) . However, flatness is typically established for idealized or nominal models ( Welde et al., 2023 ) , which often fail to capture real-world disturbances and unmodeled dynamics that arise in practice. In certain examples, known perturbations or unmodeled components can be incorporated into an augmented flat model ( Faessler et al., 2017 ) , but there is currently no general methodology for doing so, or for assessing whether these perturbations will preserve flatness. Motivated by this shortcoming, we propose augmenting a differentially flat nominal model by learning to approximate the unmodeled (i.e., residual ) dynamics from trajectory data. However, existing residual learning algorithms ( Shi et al., 2019 ; Carron et al., 2019 ; Torrente et al., 2021 ; Chee et al., 2022 ) typically use generic functional forms for the residual, which can lead to augmented dynamics that no longer enjoy the flatness property. Furthermore, even if flatness is preserved, it is often unclear how to construct the flatness diffeomorphism for the resulting system, rendering flatness-based algorithms inapplicable. To overcome this, we restrict our attention to a class of differentially flat nominal dynamics that admit a pure-feedback form and derive a sufficient condition under which the residual preserves flatness. We show that when the residual is lower-triangular—a property that preserves the pure-feedback structure of the nominal dynamics—the flatness diffeomorphisms of the augmented system can be constructed using those of the nominal model. We validate our approach on planar quadrotor dynamics, achieving significantly lower tracking error than without learning while requiring an order of magnitude less computation time than a structure-agnostic approach that incorporated the learned residual via nonlinear model predictive control (NMPC). To the best of our knowledge, this is the first work to address learning flatness-preserving residual dynamics for differentially flat systems. 1.1 Related Work 1.1.1 Learning Residual Dynamics Leveraging learned residual dynamics in the control loop has been shown to improve tracking performance on various robotic systems, including manipulators ( Carron et al., 2019 ) and quadrotors ( Shi et al., 2019 ; Torrente et al., 2021 ; Chee et al., 2022 ) . Prior work incorporated residual dynamics by designing system-specific controllers or applying general techniques such as NMPC. In Shi et al. (2019) , the authors designed a fix-point iteration-based algorithm tailored to quadrotors to compute the control input. In Carron et al. (2019) ; Torrente et al. (2021) ; Chee et al. (2022) , the authors compute the control input with NMPC, which is drastically more computationally expensive than flatness-based control methods for quadrotors ( Sun et al., 2022 ) . To address these challenges, we restrict our nominal dynamics to a class of differentially flat pure-feedback systems and seek to learn flatness-preserving residuals, thereby retaining the computational benefits of flatness-based approaches while eliminating the need for case-by-case controller design. 1.1.2 Learning to Linearize Nonlinear Systems Learning-based techniques have been applied to differentially flat systems to derive linearizing coordinate transformations from data. Representative examples include seeking to identify flatness diffeomorphisms from input-output data ( Sferrazza et al., 2016 ; Ma et al., 2022 ) , and directly learning a feedback-linearizing control law ( Umlauft et al., 2017 ; Westenbroek et al., 2019 ; Greeff and Schoellig, 2020 ) . For pure-feedback systems, adaptive control methods have also been proposed to deal with parametric uncertainty in the dynamics, including model reference adaptive control ( Nam and Araposthathis, 1988 ) and neural control methods that leverage backstepping ( Ge and Wang, 2002 ; Ge and Wang, 2004 ) . In a related vein, model-free control techniques exist for flat systems to estimate and reject more general disturbances online ( Fliess and Join, 2013 ; Join et al., 2024 ) . In contrast to these approaches, which are complementary to our work, we focus on learning residual dynamics from data while ensuring that the resulting augmented model preserves flatness. 1.2 Contributions Our contributions are three-fold: • Flatness-preserving residuals: We identify a broad class of residual dynamics that preserve the differential flatness of pure-feedback form systems (Theorem 2 ) and show how to systematically construct the flatness diffeomorphisms for the augmented model from those of the nominal model (Theorem 3 ). • Learning algorithm: We develop a parameterization of the residual dynamics that enforces this structure, enabling flatness-preserving residual models to be learned directly from trajectory data. • Empirical validation: We demonstrate in simulation that our method significantly reduces tracking error relative to the nominal controller while being an order-of-magnitude more computationally efficient than an NMPC-based structure-agnostic approach. 2 Preliminaries and Problem Formulation Consider the nonlinear system \dot{\mathbf{x}}=f(\mathbf{x},\mathbf{u}) (1) with state \mathbf{x}\in\mathbb{R}^{n} and control input \mathbf{u}\in\mathbb{R}^{m} . We recall the definition of a differentially flat system, as originally introduced by Fliess et al. (1995) , but follow the formulation given in Hagenmeyer and Delaleau (2003) . Definition 1 (See Hagenmeyer and Delaleau (2003) ) . The nonlinear system ( 1 ) is (differentially) flat if a flat output \mathbf{y}=\Lambda(\mathbf{x},\mathbf{u},\dot{\mathbf{u}},\ddot{\mathbf{u}},\cdots,\mathbf{u}^{(\alpha)}) and a finite number of its time derivatives can be used to reconstruct the state \mathbf{x} and control input \mathbf{u} : \displaystyle\mathbf{x} \displaystyle=\Phi(\mathbf{y},\dot{\mathbf{y}},\ddot{\mathbf{y}},\dots,\mathbf{y}^{(\beta)}), \displaystyle\mathbf{u} \displaystyle=\Psi(\mathbf{y},\dot{\mathbf{y}},\ddot{\mathbf{y}},\dots,\mathbf{y}^{(\beta)}). Here, \Lambda , \Phi , and \Psi are smooth functions, with (\Phi,\Psi) referred to as the flatness diffeomorphism . The quantities \alpha and \beta are finite positive integers, and \mathbf{z}^{(c)} denotes the c -th time derivative of \mathbf{z} . We say that this system is locally flat around a point (\mathbf{x}^{*},\mathbf{u}^{*}) if the equalities above hold in a neighborhood \mathcal{N} of (\mathbf{x}^{*},\mathbf{u}^{*}) . As outlined in the introduction, differential flatness enables efficient planning and control by allowing for the nonlinear system ( 1 ) to be rewritten as an equivalent linear one (see, e.g. , Fliess et al. (1995) ; Murray