Path-following Control of a Quadrotor using Quasi-Static Transverse Feedback Linearization

Abstract: We propose a quasi-static transverse feedback linearization (QSTFL) controller for a quadrotor to follow a prescribed geometric path, rather than a time-parameterized trajectory. In contrast to existing dynamic-feedback approaches, the controller does not introduce additional controller states. The thrust input is computed algebraically from the current state, eliminating the need for thrust-derivative measurements and numerical integration. The proposed design renders the path-following manifold invariant, ensuring that trajectories initialized on the path remain on it for all future time, while simultaneously regulating tangential velocity and yaw. We establish a diffeomorphic coordinate transformation and prove local exponential stability of the path-following manifold. In addition, closed-form expressions are derived for the thrust and torque inputs. Compared with dynamic-feedback constructions, the controller requires inversion of only a $3\times 3$ decoupling matrix rather than a $4\times 4$ one, leading to a simpler control law and reduced computational complexity. Numerical simulations demonstrate the effectiveness of the proposed method. Code and animations are publicly available at \footnotesize{\texttt{\href{https://gitlab.com/a5akhtar/quasistatic-tfl-uav/}{https://gitlab.com/a5akhtar/quasistatic-tfl-uav/}}}.

• The paper develops a quasi-static transverse feedback linearization controller that jointly regulates path errors, tangential speed, and yaw without introducing thrust dynamics.
• The method reduces torque decoupling from a 4×4 to a 3×3 matrix and creates four independently stabilized Brunovský error blocks containing 11 transverse states.
• Simulations confirm path invariance and local convergence on circular and twisted spatial paths, while stability remains limited by Euler-angle singularities, positive-thrust requirements, and local operating conditions.

Overview and motivation

This paper addresses the path-following problem (PFP) for a quadrotor UAV: driving the vehicle's center of mass onto a prescribed geometric curve in $\mathbb{R}^3$ — a set of points with no time parameterization — while guaranteeing that motions initialized on the path remain on it for all time. This contrasts with trajectory tracking, where a time-parameterized reference continues to evolve even if the vehicle slows or stops, generating spurious tracking error. The authors' central contribution is a quasi-static transverse feedback linearization (QSTFL) controller that achieves exponential stability of the path-following manifold without the dynamic extension used in prior work (2606.23218).

The design philosophy is monolithic rather than cascaded: translational and rotational dynamics are treated jointly, avoiding the well-known caveat that stability of individual cascade loops does not imply closed-loop stability of the interconnection. The approach is coordinate-based, using a 2-1-3 Euler parameterization chosen specifically because it yields a thrust direction $b_3$ free of yaw dependence and contains no singularity at hover. This choice trades away the global validity of coordinate-free $SO(3)$ formulations in exchange for analytical tractability.

Problem formulation

The desired path $\mathcal{C}$ is modeled as the regular intersection of two surfaces, $$\mathcal{C} = \{p \in U_\mathcal{C} : h_1(p) = h_2(p) = 0\}$$, with transversality condition $\nabla h_1(p) \times \nabla h_2(p) \neq 0$, making it a smooth embedded one-dimensional submanifold. A unit tangent field $\rho(p)$ is then well defined on a neighborhood of the curve. Two assumptions underpin the design: A1 requires strictly positive thrust ($u_t \neq 0$), consistent with safe flight; A2 requires $\langle \nabla h_2, R_3 \rangle \neq 0$, i.e., the thrust axis must not lie tangent to the surface $h_2 = 0$, which would render the input ineffective along that constraint direction.

The control objective is encoded through a four-component output

$b_3$0

where $b_3$1 and $b_3$2 are position-dependent desired tangential speed and yaw. Zeroing this output simultaneously achieves convergence to $b_3$3, controlled invariance of on-path motion, tangential velocity regulation, and yaw regulation — the three goals G1–G3 of the PFP.

Quasi-static feedback construction

The key structural observation is that the standard decoupling matrix obtained by differentiating the outputs until inputs appear is singular with constant rank 2: two affine relations between inputs and output derivatives always hold away from the model singularity $b_3$4. Rather than resorting to dynamic extension — as in prior TFL designs that introduce thrust and its derivative as controller states — the authors exploit quasi-static feedback, where inputs are expressed algebraically in terms of auxiliary inputs and their time derivatives evaluated at the current state.

Two auxiliary inputs are assigned first, $b_3$5 and $b_3$6. Because these two relations are independent even at hover (the authors select $b_3$7 so that $b_3$8 there), the thrust can be solved directly:

$b_3$9

with $SO(3)$0 the Hessian of $SO(3)$1. This is the pivotal step: precomputing $SO(3)$2 algebraically eliminates both the need for thrust-derivative measurements/double integration and one column of the decoupling matrix. Differentiating further to obtain $SO(3)$3 and $SO(3)$4 yields auxiliary inputs $SO(3)$5 and $SO(3)$6, producing a reduced $SO(3)$7 torque decoupling matrix $SO(3)$8 whose determinant is

$SO(3)$9

which is nonzero near the curve under Assumptions A1–A2 and the embedded-curve regularity. Compared with the $\mathcal{C}$0 decoupling matrices of dynamic-feedback constructions, this is a concrete simplification in both control-law structure and computational cost. The resulting error dynamics split into four decoupled linear time-invariant Brunovský blocks of dimensions 4, 2, 3, and 2 (11 states total), stabilized by linear gains $\mathcal{C}$1 with arbitrary eigenvalue placement.

Stability analysis

Since the linearized error dynamics occupy 11 of the 12 state dimensions, a one-dimensional zero dynamics remains, evolving on the zero-dynamics manifold $\mathcal{C}$2 — precisely the set of motions along the path. The analysis proceeds in three steps:

1. Compactness: The authors prove $\mathcal{C}$3 is diffeomorphic to $\mathcal{C}$4. On $\mathcal{C}$5, the velocity is forced to satisfy $\mathcal{C}$6, which uniquely determines roll, pitch, yaw, and angular velocity as functions of position $\mathcal{C}$7; hence the projection map restricted to $\mathcal{C}$8 is injective.
2. Diffeomorphism: Using the generalized inverse function theorem (which extends the classical result from points to submanifolds), the transformation $\mathcal{C}$9 — with $$\mathcal{C} = \{p \in U_\mathcal{C} : h_1(p) = h_2(p) = 0\}$$0 the arc-length coordinate on $$\mathcal{C} = \{p \in U_\mathcal{C} : h_1(p) = h_2(p) = 0\}$$1 obtained via closest-point projection composed with a diffeomorphism $$\mathcal{C} = \{p \in U_\mathcal{C} : h_1(p) = h_2(p) = 0\}$$2 — is shown to be a diffeomorphism on a tubular neighborhood of $$\mathcal{C} = \{p \in U_\mathcal{C} : h_1(p) = h_2(p) = 0\}$$3. The Jacobian determinant is explicitly computed and shown nonzero under the standing assumptions.
3. Exponential stability: In $$\mathcal{C} = \{p \in U_\mathcal{C} : h_1(p) = h_2(p) = 0\}$$4 coordinates, the $$\mathcal{C} = \{p \in U_\mathcal{C} : h_1(p) = h_2(p) = 0\}$$5-subsystem is exponentially stable by gain design while the $$\mathcal{C} = \{p \in U_\mathcal{C} : h_1(p) = h_2(p) = 0\}$$6-dynamics evolves on the compact set $$\mathcal{C} = \{p \in U_\mathcal{C} : h_1(p) = h_2(p) = 0\}$$7 and is therefore bounded. Diffeomorphic transport of this property yields local exponential stability of $$\mathcal{C} = \{p \in U_\mathcal{C} : h_1(p) = h_2(p) = 0\}$$8.

This establishes all three PFP goals: exponential convergence to the path, controlled invariance of on-path motion, and regulation of tangential velocity and yaw. The stability guarantee is local — restricted to the tubular neighborhood where the closest-point projection is smooth — and the positive-thrust condition in the algebraic thrust law is only guaranteed near $$\mathcal{C} = \{p \in U_\mathcal{C} : h_1(p) = h_2(p) = 0\}$$9; for large initial errors, the authors note that gains must be designed, or $\nabla h_1(p) \times \nabla h_2(p) \neq 0$0 saturated, to avoid $\nabla h_1(p) \times \nabla h_2(p) \neq 0$1, at the cost of temporarily forfeiting the exponential stability certificate.

Simulation results

Simulations are run on the non-parameterized geometric model (rotation matrix dynamics) rather than the Euler-parameterized design model, providing a check against parameterization mismatch, with $\nabla h_1(p) \times \nabla h_2(p) \neq 0$2 and $\nabla h_1(p) \times \nabla h_2(p) \neq 0$3.

Example 1 (invariance): On a horizontal circle at constant traversal speed with heading toward the center, the UAV is initialized on the path but at rest with heading pointing outward. The controller corrects the heading while the vehicle never leaves the path, empirically confirming controlled invariance.

Example 2 (convergence): On a twisted spatial curve defined by $\nabla h_1(p) \times \nabla h_2(p) \neq 0$4 and $\nabla h_1(p) \times \nabla h_2(p) \neq 0$5, with $\nabla h_1(p) \times \nabla h_2(p) \neq 0$6 and $\nabla h_1(p) \times \nabla h_2(p) \neq 0$7, the UAV starts off-path at $\nabla h_1(p) \times \nabla h_2(p) \neq 0$8 with nonzero attitude error and body rates. Closed-loop eigenvalues are placed between $\nabla h_1(p) \times \nabla h_2(p) \neq 0$9 and $\rho(p)$0 across the four blocks, and the outputs converge to zero, demonstrating path convergence from arbitrary local initial conditions.

Code and animations are publicly available, supporting reproducibility.

Limitations and open questions

Several restrictions bound the applicability of the results. The Euler parameterization fails at $\rho(p)$1, so aggressive maneuvers approaching vertical attitudes are outside the valid region — a fundamental cost of the coordinate-based formulation relative to $\rho(p)$2 methods. Assumption A2 excludes configurations where the thrust axis is tangent to $\rho(p)$3, and the guarantee of $\rho(p)$4 holds only locally near $\rho(p)$5. The stability result is local exponential, confined to the tubular neighborhood of the path, and no robustness analysis (e.g., to aerodynamic disturbances, motor saturation, or parameter uncertainty) is provided. Finally, validation is purely numerical; experimental demonstration on hardware remains open, as does extension of the QSTFL framework to time-varying or non-compact paths, for which the compactness argument via diffeomorphism to $\rho(p)$6 does not directly apply.

Conclusion

The paper formulates quadrotor path following as transverse feedback linearization and solves it with quasi-static feedback, yielding closed-form thrust and torque laws without dynamic extension. The main theoretical results are a diffeomorphic coordinate transformation on a tubular neighborhood of the path, compactness of the zero-dynamics manifold, and local exponential stability of the motion-invariant path-following set. The reduction of the decoupling matrix inversion from $\rho(p)$7 to $\rho(p)$8, together with the elimination of thrust-derivative estimation, constitutes a practical simplification over existing dynamic-feedback designs, substantiated by simulations on the full geometric model.