Learning Flatness-Preserving Residuals for Pure-Feedback Systems

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 $5\times$ lower than the nominal flat model, while being $20\times$ faster over a structure-agnostic alternative.

💡 Research Summary

The paper addresses a fundamental gap in the application of differential flatness to real‑world nonlinear systems: while flatness enables simple planning and linear‑feedback design, the nominal models on which flatness is established often omit important dynamics such as aerodynamic drag, actuator saturation, or environmental disturbances. Conventional residual‑learning approaches augment the nominal dynamics with a learned correction term, but they typically use unrestricted function approximators. This freedom can destroy the flatness property, rendering flatness‑based planners and controllers inapplicable or requiring costly redesign.

To solve this, the authors focus on a class of differentially flat systems that can be expressed in a pure‑feedback form. In this representation, the state vector is partitioned into a cascade of sub‑states (x_1,\dots,x_r) where each sub‑state’s dynamics depend only on the previous sub‑states and the next one (or the control input at the last stage). Under mild regularity conditions (smoothness and nonsingular Jacobians), such systems are known to be locally flat with the flat output simply (y = x_1).

The key insight is that if the residual dynamics (\Delta(x)) are constrained to a lower‑triangular structure—i.e., (\Delta_i) depends only on ((x_1,\dots,x_i))—the augmented dynamics (\bar f + \Delta) retain the pure‑feedback cascade and the same Jacobian nonsingularity. The authors prove (Theorem 2) that this structural restriction guarantees that the augmented system remains differentially flat with the same flat output (y = x_1). Consequently, the flatness‑preserving residual does not alter the semantic meaning of the flat output, which is crucial for downstream tasks such as trajectory generation.

Beyond the existence proof, the paper provides a constructive method (Theorem 3) to obtain the flatness diffeomorphism of the augmented system directly from the known diffeomorphism of the nominal model. By recursively correcting the known implicit functions (h_k) (which map flat output derivatives to the missing sub‑states) with the learned residual terms, the authors derive explicit formulas for the new transformation (\hat\Phi_k). This eliminates the need for ad‑hoc derivations of new flatness maps whenever a residual is learned.

For learning, the authors propose parameterizing each (\Delta_i) with smooth function approximators (e.g., neural networks, splines) that are explicitly built to respect the lower‑triangular dependency. Because the structure is hard‑wired, standard supervised learning pipelines (least‑squares, stochastic gradient descent) can be employed without additional constraints or regularization to enforce flatness.

The methodology is validated on a planar (2‑D) quadrotor. The nominal model captures rigid‑body dynamics; the residual captures unmodeled aerodynamic effects such as rotor drag and down‑wash. By dynamically extending the thrust input twice, the system is cast into a pure‑feedback form with flat output equal to the planar position. After learning the lower‑triangular residual from simulated trajectories, the authors apply a flatness‑based planner and a simple feedback law. The results show a five‑fold reduction in tracking error compared to using the nominal model alone. Moreover, when the same learned residual is incorporated into a structure‑agnostic NMPC scheme, the computational load is roughly twenty times higher, highlighting the efficiency gains of preserving flatness.

In summary, the paper makes three major contributions: (1) identification of a broad class of flatness‑preserving residuals for pure‑feedback systems; (2) a constructive algorithm to update flatness diffeomorphisms after residual learning; and (3) empirical evidence that the approach yields both superior control performance and dramatically lower computational cost. This work opens a practical pathway for integrating data‑driven model corrections into flatness‑based control pipelines, with potential impact across aerial robotics, automotive systems, and any domain where differential flatness is exploited but model fidelity remains a challenge.