Flat Outputs in Control Theory

• Flat outputs are defined as functions of states, inputs, and their derivatives that enable global parameterization of trajectories in nonlinear and hybrid systems.
• They are computed using methods like distribution flag techniques and Lie bracket checks to transform system dynamics into canonical triangular forms for effective control.
• They facilitate motion planning, trajectory generation, and real-time feedback design in applications such as robotics, aerospace, and hybrid systems.

A flat output in control theory is a set of functions—usually of the system’s state, inputs, and potentially their derivatives or forward-shifts—that allows global parametrization of all feasible system trajectories and inputs by means of the flat output and finitely many of its derivatives or shifts. Differential flatness, and its discrete-, delay-, and geometric analogues, underpins a comprehensive theoretical and practical framework for motion planning, feedback design, and trajectory generation in nonlinear and hybrid dynamical systems. Flat outputs provide a constructive solution to trajectory generation and feedback, but also introduce deep geometric, algebraic, and analytic structures whose characterization and computation remain active research topics.

1. Fundamental Definitions and Classes of Flat Outputs

Consider a general control system

$\dot{x} = f(x, u)$

with $x \in \mathbb{R}^n$, $u \in \mathbb{R}^m$. The system is differentially flat if there exists a vector of outputs $y$ (“flat output”) such that:

• The map $(x, u, \dot{u}, \dotsc, u^{(q)}) \mapsto y$ is invertible locally.
• All system variables (state and input) can be expressed as functions of the flat output and a finite number of its time derivatives:

$$x = F_x(y, \dot{y},\ldots, y^{(r-1)}), \quad u = F_u(y, \dot{y},\ldots, y^{(r)}).$$

The minimal regularity and functional dependence needed for flat outputs gives rise to a hierarchy (Gstöttner et al., 2022):

• (x, u, …, $u^{(q)}$)-flat outputs: $y$ depends on $(x, u, \dot{u},\dots,u^{(q)})$.
• (x, u)-flat outputs: $y$ depends only on $x$ and $u$ ($q=0$).
• x-flat outputs: $y$ depends only on $x$.

In discrete-time, the analog replaces time-derivatives by shift operators (Diwold et al., 2020, Schlacher, 2019):

• The flat output must allow representation of all system trajectories by a finite number of forward shifts of $y$.
• For differential-delay systems, π-flatness generalizes flatness to modules over Ore algebras, allowing dependence on delays and advances (Antritter et al., 2012).

2. Structural Results and Normal Forms

Many results establish canonical “triangular” or “block chain” normal forms for flat systems, showing that, under change of coordinates and feedback, all variables can be recursively solved from the flat output and its derivatives or shifts:

• For two-input continuous-time control-affine systems, any x-flat system can be put into a triangular normal form encompassing classical Brunovsky, chained, and extended chained forms, after finite prolongations if necessary (Hartl et al., 21 May 2025, Gstöttner et al., 2024). In these forms, the flat output appears as a subset of the new coordinates, and the remaining variables evolve recursively.
• For discrete-time systems, a finite block-triangular normal form exists for two-input systems, where each subsystem only involves higher-order shifts of the flat output, allowing explicit parameterization (Diwold et al., 2020, Schlacher, 2019).
• For multi-input ($m+1\ge3$) systems, fully constructive necessary and sufficient geometric conditions for static-feedback equivalence to a multi-chained triangular form yield both sufficient flatness conditions and a means of constructing all compatible flat outputs (Hartl et al., 30 Oct 2025).

These normal forms lead to systematic algorithms for both establishing flatness and explicitly computing flat outputs, generally relying on the stepwise construction of distribution flags, involutivity and rank checks, and successive transformations.

3. Flat Output Computation: Algorithms and Geometric Criteria

The construction of flat outputs can proceed by several approaches:

• Distributional flag methods and feedback linearization: For control-affine systems, the so-called distribution flags $D_i$ generated by the input vector fields and their iterated Lie brackets with the drift provide a sequence whose rank and involutivity properties are checked. Involutive closures and characteristics select annihilators yielding candidate flat outputs. Algorithms include stepwise prolongation, linearization, and bracket–rank checks (see distribution-based and invariant-flag algorithms) (Hartl et al., 21 May 2025, Gstöttner et al., 2024).
• Test vs. constructive approaches in discrete time: Existence can be checked via bracket inclusions, while full parameterization uses Pfaffian system integrability and the explicit solution of associated PDEs or ODEs (Schlacher, 2019).
• Algebraic module methods for delayed or hybrid systems: For linear differential-delay systems, flatness is characterized by hyper-regularity of polynomial matrices over Ore rings, checked via Smith–Jacobson reductions (Antritter et al., 2012).
• Geometric symmetry-based optimization: For mechanical systems with symmetry, geometric flat outputs can be formulated as sections of a principal bundle and computed numerically by optimizing over equivariant maps subject to geometric constraints induced by the system structure (Welde et al., 2023, Welde et al., 2022).

Many papers provide practical algorithms blending these ideas, starting from pre-tests (distribution ranks, curvature conditions), then, if appropriate, invoking explicit integration or optimization routines to construct the flat output parameterization.

4. Globality, Singularity, and Hierarchies

Flat outputs are generally only locally defined. Singularities arise when the Jacobian (or variational matrix) of the coordinate change degenerates:

• Apparent singularities: Flatness fails for one choice of output, but can be resolved by switching to another chart; an atlas of compatible charts enables piecewise flat-based planning.
• Intrinsic singularities: No local flat output exists at these points (e.g., system equilibria where linear controllability fails); these act as genuine barriers for flat-based planning (Kaminski et al., 2017, Kaminski et al., 2022). Singularities can be detected by hyper-regularity criteria for the variational matrix (Kaminski et al., 2017) or the vanishing of truncated Jacobian determinants in block-triangular or chained forms (Kaminski et al., 2022). For many systems, singular loci correspond to physically meaningful regimes (stall, zero velocity, loss of control authority).

A strict hierarchy obtains: not every flat system possesses a flat output of lower dependence (e.g., an $(x,u)$-flat output), as shown constructively via counterexamples (Gstöttner et al., 2022).

5. Classes of Systems and Applications

Flat outputs and the associated theory unify disparate system classes:

• Underactuated mechanical systems with symmetry: Global or almost-global geometric flat outputs exist when group variables can be chosen as outputs if the number of group symmetries matches the number of controls and orthogonality criteria are satisfied (Welde et al., 2022, Welde et al., 2023).
• Sampled-data, time-delay, and hybrid systems: Specializations of flatness characterize when such non-classical systems permit trajectory planning via finite-dimensional representations (Schlacher, 2019, Antritter et al., 2012).
• Partial differential equations and higher-order PDEs: The flatness property has been extended to infinite-dimensional boundary control systems such as the controlled Kawahara equation, where suitable flat outputs are analytic boundary traces and generating functions generate the state and controls via convergent series (Capistrano-Filho et al., 2024).

Typical control designs employ Bézier (polynomial) parametrizations of the flat output, reducing infinite-dimensional trajectory generation to finite algebraic constraint satisfaction in the output’s control points, with major advantages for real-time implementation and constraint handling (Bekcheva, 2020).

6. Impact, Limitations, and Open Directions

The flat output paradigm greatly simplifies motion planning, feedback design, and constraint satisfaction for a broad class of nonlinear, hybrid, and infinite-dimensional systems. By reducing trajectory generation to the specification of arbitrary (smooth or analytic) flat output functions, control design becomes computationally tractable and adaptively robust in feasible regions.

However, several limitations remain:

• Not all flat systems admit low-order or configuration-dependent flat outputs; the dependence on input (and its derivatives or shifts) is sometimes unavoidable (Gstöttner et al., 2022).
• Chart singularities can preclude global planning, requiring sophisticated chart-switching and careful avoidance of intrinsic barriers (Kaminski et al., 2017, Kaminski et al., 2022).
• Computational complexity, particularly for high-dimensional or mixed algebraic-differential systems, poses challenges for real-time or large-scale systems (Hartl et al., 30 Oct 2025).
• The search for minimal-order flat outputs, explicit normal forms in multi-input cases, and systematic extensions beyond principal-bundle symmetries (such as nonholonomic or higher-order input structures) are ongoing topics.

Despite these limits, flat output theory remains a central and evolving tool in advanced nonlinear control, motion planning, and system identification for robotics, aerospace, and beyond. Recent geometrically-informed and numerically-optimized algorithms promise further scalability and generality (Welde et al., 2023, Hartl et al., 30 Oct 2025).