betaFeedback

WorksComputer Science

Constructive Infinitely Differentiable Guidance Vector Fields for Robotic Control

Abstract

Guidance vector fields (GVFs) are a popular framework for tasks such as robotic path-following and multi-robot motion coordination, but existing approaches may be non-differentiable or have large derivatives, limiting their use in controllers that require derivatives of the guidance signal, which is common when the GVF is used as an input in a cascaded controller. This paper proposes an infinitely differentiable constructive GVF. The resulting vector field depends on a smoothing parameter and is a generalization of [1], recovering the original GVF in the limit as this parameter approaches zero. It also admits a simple and efficient implementation that preserves smoothness under discretization and can reduce computation time by using substantially fewer curve samples while maintaining small path-following errors. We establish the main theoretical properties and demonstrate its practical advantages through numerical and real-world experiments.

Introduction

Guidance vector fields (GVFs) have been widely used for robot control, particularly for path following, owing to their conceptual simplicity and strong theoretical foundations [1, 2, 3, 4]. More recent developments include nonsingular constructions, multi-robot coordination, inverse-kinematics formulations, data-based representations, extensions to Lie groups, and safety-aware formulations [5, 6, 7, 8, 9, 10]. Given a static target curve in the configuration space, a GVF usually assigns to each configuration in its domain a desired velocity vector. Following this velocity field causes the system to approach and traverse the target curve. Among these approaches, [1] proposed a practical constructive GVF for generic curves with formal convergence guarantees. Its construction is based on the Euclidean distance and the closest-point map between the current configuration and the target curve.

In many robotic systems, however, the velocity generated by the GVF is not the physical control input. Instead, this desired velocity must be realized by a lower-level controller acting through accelerations, forces, or torques. A common solution is to employ a cascade architecture, for example through a backstepping design [11], in which the velocity generated by the GVF is supplied as a reference to a dynamics-level control layer. In a quadrotor, for instance, the propeller thrusts and the resulting torques can be controlled so that the vehicle tracks the desired linear velocity prescribed by the GVF.

Implementing such a cascade architecture often requires time derivatives of the GVF output [11]. In some designs,

derivatives of order higher than one are required; for example, differential-flatness-based quadrotor controllers may use the acceleration, jerk, and snap induced by the desired velocity field [12]. These derivatives appear in feedforward acceleration, force, or torque terms that are important both in formal stability analyses and in improving practical tracking performance [11, 12, 13].

The required derivatives, however, may fail to exist. The construction in [1], for example, relies on the Euclidean distance between the configuration and the curve and its derivative, which is related to the closest point between the current point and the curve. Although the Euclidean distance is continuous everywhere, it is generally non-differentiable at points having multiple closest points on the curve, a set called the medial axis. At the center of a circle, for instance, every point on the circle is a closest point. At such configurations, the GVF is discontinuous. See Figure 1 for an illustration of this phenomenon.

Figure 1. Illustration of the discontinuity problem in [1]. The target curve is the ellipse, and the red line segment represents its medial axis, where the distance function is nondifferentiable. The plots on the right show the two components of the GVF in [1] (orange) and of the proposed GVF (blue), evaluated along the magenta line. Note the discontinuity of the orange vector field at p2=0p_2=0, which lies on the medial axis.

These singular configurations form a set of measure zero and may not be reachable by trajectories initialized outside it. Nevertheless, they can still create practical difficulties. Near the medial axis, the vector field can change very rapidly. Consequently, a small perturbation in the robot configuration may produce a large change in the desired velocity and, therefore, in the feedforward control input generated from its derivatives.

A second source of non-smoothness, perhaps even more important in practice, arises from the numerical implementation of the closest-point computation. Except for very simple curves, the Euclidean distance and closest point cannot be obtained analytically. One possible implementation is to discretize the curve into a finite collection of samples and select the sample closest to the current configuration. This procedure introduces a hard nearest-neighbor decision. As the configuration crosses a boundary between the Voronoi regions of two samples, the selected sample may change abruptly, producing a non-smooth approximation of the GVF even in regions where the original GVF was smooth. Numerical differentiation of this approximate field can then generate very large values. Unlike the intrinsic medial axis singularity, this phenomenon may occur even arbitrarily close to the target curve, which is particularly concerning because a path-following robot is expected to operate close to the curve during most of its motion.

In this paper, we propose a construction which is a generalization of [1] when we consider static curves. The resulting GVF is parameterized by a positive scalar rr, which controls the trade-off between tracking precisely the original target curve and having smoother control inputs. For every r>0r > 0, the proposed vector field is infinitely differentiable at every point of the configuration space. This property is preserved when the defining integrals are replaced by a fixed numerical quadrature over sampled points of the curve, which is straightforward to implement. As r→+0r \to+0, the proposed vector field converges to the original vector field of [1] at configurations for which the closest-point evaluation operation is well defined.

The construction is based on what we call the rr-Differentiable Evaluation Operator (rr-DEO). The operator replaces evaluation of a function at a closest point on the curve with a normalized Gaussian-weighted average over the entire curve. It also admits a probabilistic interpretation. The normalized exponential weights used by the DEO are closely related to those arising from log-sum-exp and soft-min approximations of hard maximum and minimum operators [14]. Log-sum-exp regularization has been used in several control applications to replace non-smooth objects with differentiable approximations. In model predictive control, smooth maximum approximations have been employed to obtain differentiable soft penalties [15]. Smooth approximations of geometric distances have also been used together with control barrier functions to produce smoother distance-dependent safety constraints and control signals [16]. Similarly, soft-minimum and soft-maximum operators have been used to regularize minimum-based barrier functions, allowing them to be represented by a single continuously differentiable control barrier function [17].

In the more specific context of GVFs, [9] employs a smooth approximation of the robot-to-obstacle distance to augment the path-following vector field of [1] with obstacle-avoidance capabilities. This regularization is necessary because the Euclidean distance to an obstacle may be non-differentiable, whereas its derivative is required by the safety controller. Nevertheless, the path-following component remains based on the original closest-point construction for the target curve and therefore retains the regularity and numerical implementation issues discussed above. Furthermore, smoothing only a scalar distance function would not, by itself, regularize the selection of the closest point and of its corresponding tangent vector. In contrast, the DEO regularizes the complete closest-point-evaluation operation and can therefore be applied simultaneously to both the normal and tangential components of the GVF in [1].

Other approaches can also generate sufficiently smooth GVFs, although under different formulations. In [7], for example, a closed path represented by discrete data is approximated through a finite Fourier series. The path parameter is then introduced as an additional virtual state θ\theta, and a GVF is constructed in the resulting augmented space. The vector field can be smooth in this augmented state space, but the physical velocity command depends on both the physical configuration and the additional virtual variable. Hence, it is not an autonomous vector field defined solely on the original physical configuration space. The present method, by contrast, constructs an everywhere infinitely differentiable static vector field directly in the original configuration space, without introducing a path-parameter state.

A related probabilistic approach is proposed in [18], where a score-based diffusion model learns an attracting vector field from discrete path samples, while a second neural network learns a tangential field that induces motion along the path. This formulation accommodates unordered and branching path representations. Our normal component also admits a score interpretation through Gaussian smoothing. Here, however, we consider a prescribed regular curve and its tangent vectors, and compute both components directly through normalized Gaussian-weighted integrals, without training.

We establish the differentiability and limiting properties of the proposed operator and analyze the dynamics generated by the resulting GVF. In particular, for sufficiently small rr, we show that almost all trajectories converge to a tube around the target curve that shrinks with rr, and that as r→+0r \to+0, we recover the prescribed target curve. We also present a numerical implementation that preserves the smoothness of the construction and experimentally validate the method on a real robotic platform.

In summary, the contributions of this paper are:

  • Practical: our approach offers two main benefits. First, the GVF is infinitely differentiable, so cascaded controllers that require its derivatives are not subject to disturbances caused by points of nondifferentiability. Second, implementing the method in [1] typically requires sampling points on the target curve. To have a good performance (i.e., guaranteeing convergence and tracking and not having control inputs that are too noisy), a large number of sample points may be necessary, especially when computing derivatives, which can increase the computational cost. Our approach allows the GVF to be implemented using fewer sampled points at the cost of modifying the target curve, with this trade-off controlled by the parameter rr. Although the computational cost per sample of our approach is higher

than in [1], we show through extensive simulations that this trade-off can be very favorable: the computational time can drop sharply while keeping a very reasonable tracking error. Furthermore, as we will show, we obtain very competitive results even when, instead of implementing [1] using discrete points, we use polynomial segments obtained through regression.

  • Theoretical: for static curves, our GVF generalizes the formulation in [1], which is recovered in the limit r→0+r \to0^{+}. The stability analysis in [1] must account for discontinuities in the vector field, which requires a cumbersome mathematical analysis. We establish stability for every r>0r > 0 without this complication, since the resulting vector field is infinitely differentiable. In essence, our formulation smooths out the discontinuities in [1] through an adjustable smoothing parameter, enabling a more straightforward stability analysis. Finally, interestingly, we found that our proof follows more directly if instead of the classical Lyapunov arguments, the dual Lyapunov argument of [19] is applied.

Basic Definitions

We use R+\mathbb{R}^{+} for the set of nonnegative reals. We explicitly indicate vector dimensions. Thus, instead of writing v∈Rn×1v \in\mathbb{R}^{n \times1}, we make it clear that it is, for example, a column vector by writing v∈Rn×1v \in\mathbb{R}^{n \times1}. Given a square matrix MM, Tr⁡(M)\operatorname{Tr}(M) is its trace and det⁡(M)\det(M) its determinant. Given a function f:Ra×1↦Rb×1f : \mathbb{R}^{a \times1} \mapsto\mathbb{R}^{b \times1}, we use the Jacobian convention that ∂f∂p:Ra×1↦Rb×a\frac{\partial f}{\partial p} : \mathbb{R}^{a \times1} \mapsto\mathbb{R}^{b \times a}. We use M⊤M^{\top} to denote the transpose of a matrix. ∥⋅∥\| \cdot\| denotes the Frobenius norm of a matrix/vector. For a variable y∈Rny \in\mathbb{R}^{n}, a function F:Rn↦RmF : \mathbb{R}^{n} \mapsto\mathbb{R}^{m} and a differentiable one dimensional set Y⊆Rn\mathcal{Y} \subseteq\mathbb{R}^{n}, integrals like ∫YF(y)∥dy∥\int_{\mathcal{Y}} F(y)\lVert dy\rVert should be interpreted as taking any parametrization y:[0,L]↦Rny : [0,L] \mapsto\mathbb{R}^{n} of Y\mathcal{Y}, and evaluating ∫0LF(y(τ))∥y′(τ)∥dτ\int_{0}^{L} F(y(\tau))\lVert y'(\tau)\rVert d\tau. Any parametrization will provide the same result.

Methodology

The Original Vector Field

We begin with some important definitions. We consider a Euclidean configuration space, with p∈Rn×1p \in\mathbb{R}^{n \times1} denoting a point in this space represented as a column vector.

Definition 1. We introduce the following definitions:

  • A curve C⊂Rn×1\mathcal{C} \subset\mathbb{R}^{n \times1} is said to be regular if it is bounded, has finite length ℓC\ell_{\mathcal{C}}, is closed, one-dimensional, at least twice differentiable, has no self-intersections, and has finite maximum curvature κC\kappa_{\mathcal{C}}.

  • We define pC∗:Rn×1↦2Rn×1p_{\mathcal{C}}^{*} : \mathbb{R}^{n \times1} \mapsto2^{\mathbb{R}^{n \times1}} as the function that maps a point pp to the set of all points ξ∈C\xi\in\mathcal{C} that are closest to pp. Let UC⊂Rn×1\mathcal{U}_{\mathcal{C}} \subset\mathbb{R}^{n \times1} be the set of points pp for which pC∗(p)p_{\mathcal{C}}^{*}(p) is a singleton, that is, a set containing a single element (i.e., the complement of the medial axis).

  • Let TC:C↦Rn×1T_{\mathcal{C}} : \mathcal{C} \mapsto\mathbb{R}^{n \times1} be the function that maps each point on the regular curve C\mathcal{C} to its normalized tangent vector, according to a given predefined orientation.

  • Let DC:Rn×1↦R+D_{\mathcal{C}} : \mathbb{R}^{n \times1} \mapsto\mathbb{R}^{+} be the half-squared Euclidean distance from pp to C\mathcal{C}:

DC(p)≜min⁡ξ∈C∥ξ−p∥22.(1)D_{\mathcal{C}}(p) \triangleq\min_{\xi\in\mathcal{C}} \frac{\|\xi-p\|^{2}}{2}. \tag*{(1)}

The Original Vector Field

The GVF ΨC:UC↦Rn×1\Psi_{\mathcal{C}} : \mathcal{U}_{\mathcal{C}} \mapsto\mathbb{R}^{n \times1} proposed in [1] for tracking a regular curve C\mathcal{C}, in the special case of static curves and constant gains, can be constructed as follows. For positive scalars GG and HH, it is defined as a weighted sum of Normal and Tangent components, with the former responsible for convergence to the curve and the latter for traversing it. Precisely,

ΨC(p)≜G(pC∗(p)−p)+HTC(pC∗(p)).(2)\Psi_{\mathcal{C}}(p) \triangleq G\left(p_{\mathcal{C}}^{*}(p)-p\right)+H T_{\mathcal{C}}\left(p_{\mathcal{C}}^{*}(p)\right). \tag*{(2)}

It is shown in [1] that, if p˙(t)=ΨC(p(t))\dot{p}(t)=\Psi_{\mathcal{C}}(p(t)), then DC(p(t))→0D_{\mathcal{C}}(p(t)) \to0 as t→∞t \to\infty for almost all initial conditions p(0)p(0), and that the curve C\mathcal{C} is traversed according to the orientation chosen for TC(ξ)T_{\mathcal{C}}(\xi).

Although this GVF works well in many situations, it has an important practical limitation. Implementing this vector field in a control loop often requires derivatives of ΨC\Psi_{\mathcal{C}}. For instance, if the control input of the system is acceleration, p¨=u\ddot{p}=u, we can use

u(p,p˙)=Kv(ΨC(p)−p˙)+ddtΨC(p)(3)u(p,\dot{p})=K_{v}\left(\Psi_{\mathcal{C}}(p)-\dot{p}\right)+\frac{d}{dt}\Psi_{\mathcal{C}}(p) \tag*{(3)}

for Kv>0K_{v}>0. The GVF is undefined when p∉UCp\notin\mathcal{U}_{\mathcal{C}}. Although this set has measure zero and this issue may therefore be regarded as a mere technicality, the field ΨC\Psi_{\mathcal{C}} can vary rapidly near such points, causing ddtΨC(p)\frac{d}{dt}\Psi_{\mathcal{C}}(p) to become very large. Furthermore, the situation is aggravated by the fact that, when computing this vector field, it is seldom possible to determine the closest point pC∗(p)p_{\mathcal{C}}^{*}(p) analytically; one must instead resort to discretizations and numerical algorithms. These algorithms introduce numerical errors. In this case, when an approximation such as

ddtΨC(p)≈ΨC(p+p˙Δt)−ΨC(p−p˙Δt)2Δt(4)\frac{d}{dt}\Psi_{\mathcal{C}}(p) \approx\frac{\Psi_{\mathcal{C}}(p+\dot{p}\Delta t)-\Psi_{\mathcal{C}}(p-\dot{p}\Delta t)}{2\Delta t} \tag*{(4)}

is used, the numerical errors in each computation of ΨC(p)\Psi_{\mathcal{C}}(p) can cause disastrous results in the computation of the numerical derivative, because the numerical error may be of an order smaller than O(Δt)\mathcal{O}(\Delta t).

At the cost of not tracking the desired curve C\mathcal{C} perfectly, we propose a modification of ΨC(p)\Psi_{\mathcal{C}}(p) that guarantees that it is defined and infinitely differentiable everywhere, while also behaving well under numerical differentiation. This trade-off between tracking accuracy and smoothness will be governed by a parameter rr.

The Modified Vector Field

We begin by noting that the vector field in [1] can be constructed by defining a specific function ψC:C×Rn×1↦Rn×1\psi_{\mathcal{C}} : \mathcal{C} \times\mathbb{R}^{n \times1} \mapsto\mathbb{R}^{n \times1} as

ψC(ξ,p)≜G(ξ−p)+HTC(ξ)(5)\psi_{\mathcal{C}}(\xi,p) \triangleq G(\xi-p)+H T_{\mathcal{C}}(\xi) \tag*{(5)}

for positive constants GG, HH, and evaluating pC∗(p)p_{\mathcal{C}}^{*}(p) into the variable ξ\xi. The source of nondifferentiability is that pC∗(p)p_{\mathcal{C}}^{*}(p) is not uniquely defined at points p∉UCp \notin\mathcal{U}_C. Furthermore, it varies rapidly near such points. We proceed by replacing this closest-point-evaluation operation with one that is better behaved, that is, differentiable. For this purpose, we need some key definitions.

Definition 2. Consider a regular curve CC and let rr be a positive number. Define the function ρCr:C×Rn×1↦R+\rho_C^r : C \times\mathbb{R}^{n \times1} \mapsto\mathbb{R}^{+} as

ρCr(ξ,p)≜exp⁡(−∥ξ−p∥22r2)∫Cexp⁡(−∥χ−p∥22r2)∥dχ∥.(6)\rho_C^r(\xi,p) \triangleq\frac{\exp\left(-\frac{\lVert\xi-p\rVert^2}{2r^2}\right)}{\int_C \exp\left(-\frac{\lVert\chi-p\rVert^2}{2r^2}\right)\lVert\mathrm{d}\chi\rVert}. \tag*{(6)}

We then define the rr-Differentiable Evaluation Operator (DEO), which maps a function f:C×Rn×1↦Ra×bf : C \times\mathbb{R}^{n \times1} \mapsto\mathbb{R}^{a \times b} to a function fˉr:Rn×1↦Ra×b\bar{f}_r : \mathbb{R}^{n \times1} \mapsto\mathbb{R}^{a \times b} defined by fˉr(p)=ECr[f(ξ,p)](p)\bar{f}_r(p) = \mathbb{E}_C^r[f(\xi,p)](p), where

ECr[f(ξ,p)](p)≜∫Cf(ξ,p)ρCr(ξ,p)∥dξ∥.(7)\mathbb{E}_C^r[f(\xi,p)](p) \triangleq\int_C f(\xi,p)\rho_C^r(\xi,p)\lVert\mathrm{d}\xi\rVert. \tag*{(7)}

Sometimes, we omit the argument “(p)(p)” and simply write ECr[f(ξ,p)]\mathbb{E}_C^r[f(\xi,p)]. ■

The DEO is linear and approximately implements the operation f(ξ,p)↦f(pC∗(p),p)f(\xi,p) \mapsto f(p_C^*(p),p), that is, the closest-point-evaluation operation in ξ\xi, while preserving kk-times differentiability with respect to pp, provided that f(ξ,p)f(\xi,p) itself is kk-times differentiable with respect to pp. In particular, if ff is infinitely differentiable with respect to pp, then so is ECr[f(ξ,p)](p)\mathbb{E}_C^r[f(\xi,p)](p). Under some conditions on ff and, at least, for p∈UCp \in\mathcal{U}_C, the DEO reduces to the closest-point-evaluation operation as r→+0r \to+0; see Lemma 2 in the Appendix.

The choice of the symbol E\mathbb{E}, which resembles the expectation operator for a random variable, is not accidental. The quantity ECr[f(ξ,p)](p)\mathbb{E}_C^r[f(\xi,p)](p) can be interpreted as the expectation of the function f(Ξ,p)f(\Xi,p) with respect to a pp-parametrized random variable Ξ(p)∈C\Xi(p) \in C distributed according to the pp-parametrized distribution Ξ(p)∼ρCr(ξ,p)\Xi(p) \sim\rho_C^r(\xi,p). As r→+0r \to+0, ρCr(ξ,p)\rho_C^r(\xi,p) converges to a Dirac delta distribution concentrated at pC∗(p)p_C^*(p), namely, δ(ξ−pC∗(p))\delta(\xi-p_C^*(p)). Equivalently, the random variable becomes “deterministic,” with Ξ(p)=pC∗(p)\Xi(p)=p_C^*(p). This interpretation provides useful intuition about the operator.

Using this construction, we can define an infinitely differentiable version of (2), namely, the function ΨCr:Rn×1↦Rn×1\Psi_C^r : \mathbb{R}^{n \times1} \mapsto\mathbb{R}^{n \times1} given by

ΨCr(p)≜ECr[ψC(ξ,p)](p).(8)\Psi_C^r(p) \triangleq\mathbb{E}_C^r[\psi_C(\xi,p)](p). \tag*{(8)}

In Subsection II-C, we show how this quantity can be approximated with arbitrary accuracy. We now state the main result of this paper.

Proposition 1. We have the following properties:

  • For r>0r>0, ΨCr(p)\Psi_C^r(p) is infinitely differentiable with respect to pp everywhere.

  • For all p∈UCp \in\mathcal{U}_C,

lim⁡r→+0ΨCr(p)=ΨC(p).(9)\lim_{r\to+0}\Psi_C^r(p)=\Psi_C(p). \tag*{(9)}
  • There exists an rmax⁡>0r_{\max}>0 such that, for r≤rmax⁡r\le r_{\max} and for almost all initial conditions, that is, except for a set of measure zero, the dynamical system p˙=ΨCr(p)\dot{p}=\Psi_C^r(p), p∈Rn×1p\in\mathbb{R}^{n\times1} enters in finite time a neighborhood Cr⊇CC^r\supseteq C of CC and stays there forever, moving inside it, never stopping. Furthermore, lim⁡r→+0Cr=C\lim_{r\to+0}C^r=C.

Proof. See the Appendix. ■

Proposition 1 says that the vector field is infinitely differentiable, that it approaches the vector field in [1] in the limit r→+0r\to+0 and that it is guaranteed to drive the system, for almost all initial conditions, to a neighborhood CrC^r of CC, that can be shrunk to be arbitrarily close to CC.

This vector field provides a trade-off between perfectly tracking the desired curve CC and obtaining smoother, and in fact infinitely differentiable, control inputs. As rr increases, the tracking error relative to CC increases, whereas the derivatives of ΨCr(p)\Psi_C^r(p) with respect to pp decrease in magnitude.

Finally, it is important to make the following comparison with the result in [1]: our proposed approach reduces to [1] for static curves and constant gains. The case of time varying curves will be discussed in future work. Considering the gains, the full form of the vector field in [1] for static curves is:

ΨC(p)=G(p)(pC∗(p)−p)+H(p)TC(pC∗(p)).(10)\Psi_C(p)=G(p)(p_C^*(p)-p)+H(p)T_C(p_C^*(p)). \tag*{(10)}

In comparison with (2), note that G(p)G(p) and H(p)H(p) are allowed to be nonnegative functions of pp, whereas in our case GG and HH are constants. This flexibility, not present in our work, can be helpful to tune the controller. We believe our theoretical guarantee can be extended to also have functions and this will be investigated in a future work, nevertheless, note that although we cannot fully modify G(p)G(p) and H(p)H(p), we can trivially extend our result to have more flexibility than simply having a constant GG and HH. This is because if the whole vector field is multiplied by a kk-times positive differentiable function v:Rn↦R+v:\mathbb{R}^n\mapsto\mathbb{R}^+ the resulting vector field Ψ~Cr(p)=v(p)ΨCr(p)\widetilde{\Psi}_C^r(p)=v(p)\Psi_C^r(p) is still kk-times differentiable and keep the convergence properties. The latter fact comes because a multiplication by v(p)>0v(p)>0 just parametrizes the speed at which the robot will be moving, and thus do not modify the theoretical convergence guarantees.

As discussed in [1], in some applications, e.g., fixed wing aircrafts, it is desirable to have constant speed. We could achieve this by selecting v(p)=vc/∥ΨCr(p)∥v(p)=v_c/\lVert\Psi_C^r(p)\rVert. But then the vector field no longer can be guaranteed to be even continuous, because there can be points pp in which ΨCr(p)=0\Psi_C^r(p)=0. This is specially the case with closed curves, and is not a limitation of our work, but a topological constraint. There is no continuous nonvanishing autonomous (i.e., non time-varying) vector field in Rn\mathbb{R}^n whose globally asymptotically stable limit cycle is a closed curve [4].

Numerical Implementation

Analytical computation of the integral in ΨCr(p)\Psi_C^r(p) is usually difficult. However, it can be approximated as follows. Sample NN evenly spaced points ξi\xi_i on the curve, such that ∥ξi+1−ξi∥\lVert\xi_{i+1}-\xi_i\rVert is approximately constant. Then, for a function f(ξ,p)f(\xi,p),

ECr[f(ξ,p)](p)≈∑i=1NρCr(ξi,p)f(ξi,p)(11)\mathbb{E}_C^r[f(\xi,p)](p)\approx\sum_{i=1}^{N}\rho_C^r(\xi_i,p)f(\xi_i,p) \tag*{(11)}

where

ρ~Cr(i,p)≜exp⁡(−∥p−ξi∥22r2)∑j=1Nexp⁡(−∥p−ξj∥22r2)(12)\tilde{\rho}_{C}^{r}(i,p) \triangleq\frac{\exp\left(-\frac{\lVert p-\xi_i\rVert^2}{2r^2}\right)}{\sum_{j=1}^{N}\exp\left(-\frac{\lVert p-\xi_j\rVert^2}{2r^2}\right)} \tag*{(12)}

To avoid numerical instabilities that may occur when all the terms exp⁡(−∥p−ξi∥22r2)\exp\left(-\frac{\lVert p-\xi_i\rVert^2}{2r^2}\right) are too small, it is preferable to compute D~C(p)=min⁡j=1,…,N∥p−ξj∥2\widetilde{D}_{C}(p)=\min_{j=1,\ldots,N}\lVert p-\xi_j\rVert^2 and instead use the mathematically equivalent expression

ρ~Cr(i,p)≜exp⁡(D~C(p)−∥p−ξi∥22r2)∑j=1Nexp⁡(D~C(p)−∥p−ξj∥22r2)(13)\tilde{\rho}_{C}^{r}(i,p) \triangleq\frac{\exp\left(\frac{\widetilde{D}_{C}(p)-\lVert p-\xi_i\rVert^2}{2r^2}\right)}{\sum_{j=1}^{N}\exp\left(\frac{\widetilde{D}_{C}(p)-\lVert p-\xi_j\rVert^2}{2r^2}\right)} \tag*{(13)}

This approximation is analogous to computing the integrals using the rectangle rule. More accurate results could be obtained using more sophisticated integration rules. However, it is important to keep the points ξi\xi_i fixed. This ensures that the approximation Ψ~Cr(p)\widetilde{\Psi}_{C}^{r}(p) of ΨCr(p)\Psi_{C}^{r}(p) remains infinitely differentiable with respect to pp.

Experimental Results

The Model

Throughout this section, we will perform experiments and simulations to evaluate the claimed benefits of our proposed method, and also to help the user to select these parameters.

We use a custom-built quadrotor platform (see Figure 2). It uses a low-level controller to handle linear velocity commands, transforming them into angular velocity and total thrust inputs. We will present the model of the drone plus the low level controller in this subsection, as well as the parameters. The model and parameters used in simulation and real experiments are the same.

Drone used in the experiment and simulation

Figure 2. Drone used in the experiment and simulation

We consider a drone that is controlled through total thrust T∈RT \in\mathbb{R} and angular velocity vector ω∈R3×1\omega\in\mathbb{R}^{3\times1}, with the latter written in inertial frame. Let pp, vv, z∈R3×1z \in\mathbb{R}^{3\times1} be, respectively, the position, linear velocity and zz axis (aligned with the total thrust direction) of the drone measured the inertial frame. We also consider a virtual state vd∈R3×1v_d \in\mathbb{R}^{3\times1} needed to implement our controller. We assume the state evolving equations:

p˙=v,v˙=g+s(T)Mz,v˙d=ν,z˙=ω×z(14)\dot{p}=v,\quad\dot{v}=g+\frac{s(T)}{M}z,\quad\dot{v}_d=\nu,\quad\dot{z}=\omega\times z \tag*{(14)}

in which gg is the gravity vector, measured in the inertial frame, MM is the drone mass and ν\nu is a virtual control variable. The function s(T)s(T) implements a saturation of TT between Tmin⁡T_{\min} and Tmax⁡T_{\max}.

Following [20], given a velocity vector field Ψrr(p)\Psi_{r}^{r}(p) such that v=Ψrr(p)v=\Psi_{r}^{r}(p) is desired, we use the following control scheme, for positive constants KvK_v, KΔK_\Delta and KωK_\omega:

ad(p,v)≜Kv(ΨCr(p)−v)+ddtΨCr(p)a_d(p,v) \triangleq K_v\left(\Psi_{C}^{r}(p)-v\right)+\frac{d}{dt}\Psi_{C}^{r}(p)
ac(p,v,vd)≜ad−g+KΔ(vd−v),zd(p,v,vd)≜ac∥ac∥a_c(p,v,v_d) \triangleq a_d-g+K_\Delta(v_d-v),\quad z_d(p,v,v_d)\triangleq\frac{a_c}{\lVert a_c\rVert}
ν(p,v,vd)=ad,T(p,v,z,vd)=Mz⊤ac\nu(p,v,v_d)=a_d,\quad T(p,v,z,v_d)=Mz^\top a_c
ω(p,v,z,vd)=zd×(ddtzd−Kωz)(15)\omega(p,v,z,v_d)=z_d\times\left(\frac{d}{dt}z_d-K_\omega z\right) \tag*{(15)}

Note that this control scheme requires the second derivative of the field ΨCr(p)\Psi_{C}^{r}(p).

Even though the terms ddtΨCr\frac{d}{dt}\Psi_{C}^{r} and ddtzd\frac{d}{dt}z_d could be computed analytically, we estimated them numerically for the sake of simplicity and computational efficiency, also aiming to test the claim that our implementation is suitable for numerical differentiation. They are implemented as follows:

ddtΨCr(p)≈ΨCr(p+vΔt)−ΨCr(p−vΔt)2Δt\frac{d}{dt}\Psi_{C}^{r}(p)\approx\frac{\Psi_{C}^{r}(p+v\Delta t)-\Psi_{C}^{r}(p-v\Delta t)}{2\Delta t}
zd+=zd(p+vΔt,v+aΔt,vd+adΔt)z_d^{+}=z_d(p+v\Delta t,v+a\Delta t,v_d+a_d\Delta t)
zd−=zd(p−vΔt,v−aΔt,vd−adΔt)z_d^{-}=z_d(p-v\Delta t,v-a\Delta t,v_d-a_d\Delta t)
ddtzd(p,v,vd)≈zd+−zd−2Δt(16)\frac{d}{dt}z_d(p,v,v_d)\approx\frac{z_d^{+}-z_d^{-}}{2\Delta t} \tag*{(16)}

for a small Δt\Delta t. In this case, we use the drone acceleration given by a=v˙=g+(s(T)/M)za=\dot{v}=g+(s(T)/M)z.

The parameters used (real experiment and simulation) can be seen in Table I. The target curve CC, the value of rr and the number of sampled points NN will change during the simulations and experiments and thus are not reported in this table.

NameValueNameValueNameValue
GG2.0 s−12.0\,\mathrm{s}^{-1}HH1.0 m/s1.0\,\mathrm{m/s}KvK_v2.0 s−12.0\,\mathrm{s}^{-1}
KΔK_\Delta2.0 s−12.0\,\mathrm{s}^{-1}KωK_\omega2.0 s−12.0\,\mathrm{s}^{-1}MM2.09 kg2.09\,\mathrm{kg}
Δt\Delta t0.001 s0.001\,\mathrm{s}Tmin⁡T_{\min}0 N0\,\mathrm{N}Tmax⁡T_{\max}30 N30\,\mathrm{N}

Table I. Parameter values (simulation and experiment).

Choosing the Parameters of the GVF

The main design choice considered in this paper is the smoothing parameter rr, including whether to use r=0r=0, corresponding to the original approach in [1], or r>0r>0, corresponding to our proposed approach. Additional choices arise in the numerical implementation. For r>0r>0, we use the sampled-point approximation described in Subsection II-C and must select the number of samples NN. For r=0r=0, alternatives to searching through sampled points are available. Here, we also consider piecewise polynomial approximation: we use regression to approximate CC with NN polynomial segments of degree mm, matching derivatives through order m−1m-1 at the joins, and search for the closest point on each segment. For a unified notation, we use m=0m=0 to denote the sampled-point representation and m>0m>0 to denote polynomial segments. For m≥1m\geq1, finding the closest point on a polynomial segment requires finding the real roots of a polynomial of degree at most 2m−12m-1. Numerical root-finding algorithms can be used. For m=1m=1 and m=2m=2, the stationary-point equations are linear and cubic, respectively, and can be solved analytically, using Cardano’s formula in the cubic case.

These choices depend on several factors, including the target curve, the GVF parameters, and the low-level dynamics. We must balance steady-state tracking error, measured relative to the original curve, against computational time. In this subsection, we present a case study using the model in Subsection III-A, providing practical observations and simulation results to help guide these choices.

We first define the performance metrics. Let Φ(t,p0,C,Π)\Phi(t,p_0,C,\Pi) denote the position p(t)p(t) obtained by simulating the system in (14) and (15), with p(0)=p0p(0)=p_0, z(0)z(0) pointing upwards, and the remaining state variables initially zero. The controller parameters are those in Table I, the target curve is CC, and the implementation parameters are collected in Π=(r,N,m)\Pi=(r,N,m). This bundle contains the smoothing parameter r≥0r \ge0, the number NN of points or polynomial segments, and the polynomial degree mm, with m=0m=0 denoting points. For simplicity, we refer to both points and polynomial segments as “segments” when discussing their computational cost. We only consider m=0m=0 when r>0r>0, since the implementation studied here uses sampled points.1

We define the average steady-state tracking error as 2^2

E(Π)≜lim sup⁡T→∞1T∫0TE[2DC(Φ(t,P0,C,Π))] dt.(17)E(\Pi) \triangleq\limsup_{T\to\infty}\frac{1}{T}\int_0^T \mathbb{E}\left[\sqrt{2D_C\left(\Phi(t,P_0,C,\Pi)\right)}\right]\,dt. \tag*{(17)}

Here, the expectation is taken over the initial position P0P_0 and the target curve CC, conditioned on successful tracking for the given Π\Pi. Their distributions will be specified soon. A trajectory is classified as successful if its steady-state tracking error is below 20 cm and it traverses the curve rather than becoming stuck. Thus, E(Π)=E(r,N,m)E(\Pi)=E(r,N,m) measures the average steady-state error among successful trajectories. We report the failure rate separately as F(Π)∈[0,1]F(\Pi)\in[0,1]. Throughout this analysis, this distance is measured relative to the original target curve CC, evaluated numerically using a dense reference sampling, rather than relative to the polynomial approximation or the deformed set CrC^r caused by smoothing (see Proposition 1).

First, consider the ideal case without discretization:

Practical Observation 1: Even without discretization, small values of rr can produce large derivatives of the guidance signal, particularly near the medial axis. In a controller that requires these derivatives, actuator limits, finite bandwidth, and numerical approximations can then degrade tracking or cause failures. On the other hand, large values of rr can cause excessive smoothing and thus cause large errors as well. The choice of rr must balance these effects.

The difficulties associated with small rr depend on the low-level dynamics, including the number of GVF derivatives required and the system’s sensitivity to them. Consequently, a positive value of rr that performs well for one platform need not be suitable for another. For sufficiently small rr, Lemma 7 bounds the size of an invariant neighborhood of CC for the ideal first-order dynamics. However, this bound depends strongly on the curve geometry and does not directly predict the tracking error of the full closed-loop system.

Dimensional analysis of the ideal GVF provides another useful starting point. For geometrically similar curves, scaling all spatial coordinates by a factor α\alpha and scaling both rr and H/GH/G by the same factor preserves the relative geometry of the ideal first-order trajectories, with correspondingly scaled initial conditions. For example, if a particular choice produces an error equal to 5% of the curve length, doubling the curve size suggests doubling rr and H/GH/G to retain the same relative behavior. Thus:

Practical Observation 2: The value of rr should be commensurate with the “size” of the curve. Larger curves require larger values of rr.

With discretization, computational time must also be balanced against performance. Small values of NN, or low polynomial degrees when r=0r=0, reduce evaluation time but may produce poor tracking or failures, such as becoming stuck. Larger values can improve the approximation, at the cost of additional computation. For the smoothed field, the interaction between rr and the sample spacing is particularly important:

Practical Observation 3: The smoothing parameter rr and the number of samples NN should be chosen together. For approximately uniform arc-length sampling and being ℓC\ell_C the length of the curve, the sample spacing is h≈ℓC/Nh\approx\ell_C/N. Near the curve, a useful heuristic is to make this spacing small compared with rr, so that several nearby samples contribute appreciably to the Gaussian-weighted average. If the sampling is too sparse relative to rr, the weights can become concentrated on individual samples and change rapidly as the robot moves, even though the implemented field remains infinitely differentiable. Smaller values of rr therefore generally require denser sampling. The required sampling density depends on the curve geometry and the derivatives required by the controller.

For a fixed polynomial degree, all implementations considered here have evaluation complexity O(N)\mathcal{O}(N): the smoothed field sums contributions from all sampled points, while the original field searches through all points or polynomial segments.2 Their costs per segment, however, differ substantially. Quantifying these differences helps guide the choice of Π=(r,N,m)\Pi=(r,N,m).

Practical Observation 4: Let τ(r,m)\tau(r,m) denote the evaluation cost per segment in our optimized C++ implementations when using a smoothing factor rr and polynomials of degree mm (for r=0r=0). The measured cost ratios are approximately

τ(r,0)/τ(0,0)≈10\tau(r,0)/\tau(0,0) \approx10 for r>0r > 0, τ(0,1)/τ(0,0)≈20\tau(0,1)/\tau(0,0) \approx20, τ(0,2)/τ(0,0)≈200\tau(0,2)/\tau(0,0) \approx200, and τ(0,3)/τ(0,0)≈600\tau(0,3)/\tau(0,0) \approx600.

This means that, for example, implementing our proposed strategy with N=200N = 200 points is comparable in time to implementing [1] using N=10×200=2,000N = 10 \times200 = 2{,}000 sampled points or with polynomial segments of degree 2 with N=2,000/200=10N = 2{,}000/200 = 10 polynomial segments. These ratios are empirical estimates obtained from our implementation. They depend on the hardware and numerical algorithms used but they, nevertheless, provide a conservative bound according to our experiments.

We performed simulations using the model in Subsection III-A and the parameters in Table I. The initial position was drawn from a zero-mean normal distribution with covariance σ2I\sigma^2 I, where σ2=4 m2\sigma^2 = 4\ \mathrm{m}^2. The curve was selected with equal probability from five curves of roughly the same size, centered at the origin and contained in a ball of radius 1 m1\ \mathrm{m}. We used ten initial positions for each curve, giving 50 trajectories for each parameter bundle Π\Pi, simulated from t=0t = 0 to t=30 st = 30\ \mathrm{s}. Computations were performed on an 11th Gen Intel Core i7-11800H CPU at 2.30 GHz, with each GVF evaluation using a single core.

We then considered a computational-budget experiment: given a time budget tbt_b for one GVF evaluation on a single CPU core of our machine, which of the tested choices of rr, NN, and mm provides the best tracking performance? Note that the controller in (15) requires nine such evaluations. Figure 3 presents the results for budgets ranging from 0.1 μs0.1\ \mu\mathrm{s} to 200 μs200\ \mu\mathrm{s}. For each feasible choice of rr and mm, we use the largest affordable number of segments, estimated as N=⌊tb/τ(r,m)⌋N = \left\lfloor t_b/\tau(r,m) \right\rfloor (see “Practical Observation 4” for the meaning of τ\tau).

Figure 3. Computational-budget experiment using five target curves and ten initial positions per curve. Top: average steady-state tracking error among successful trajectories. Middle: failure rate. Bottom: best-performing tested smoothing parameter r>0r > 0. Each implementation uses the largest number of segments estimated to fit within the evaluation budget.

The top panel reports the average error EE, the middle panel reports the failure FF, and the bottom panel shows the best-performing tested value, named r∗r^*, of r>0r > 0 at each budget. Note that we only report averages of EE when the failure rate FF is less than 20% as otherwise the mean would be contaminated.

Figure 3 shows the expected behavior that, the larger is the time budget, the better is the results we can obtain, in the sense of having a smaller average error and failure rate. It also shows a substantial advantage for the smoothed approach in this case study. It achieves reasonable steady-state errors at small computational budgets, while the other implementations require substantially larger budgets to become viable. By that point, the smoothed approach has already reached E=0.0008 mE = 0.0008\ \mathrm{m}, an error likely to be below the tracking accuracy achievable on the physical platform described in Subsection III-A. Furthermore, no failures were observed for the selected smoothed configurations at any tested budget, whereas the other implementations exhibited nonzero failure rates. These results show that, for the tested curves and controller, smoothing can provide good tracking with a simple implementation and a small computational budget.

At large time budgets, the error of the smoothed field levels off at approximately E=0.0008 mE = 0.0008\ \mathrm{m}, with a best tested value of r=0.02 mr = 0.02\ \mathrm{m}. For time budgets tb>200 μst_b > 200\ \mu\mathrm{s}, the best error for r>0r > 0 becomes stable at this value, whereas the polynomial approach can potentially obtain smaller error values such as 0.0001 m0.0001\ \mathrm{m}. However, these results are unlikely to be useful in practice, since such an error is very unlikely to be obtained in a practical situation in the quadrotor platform considered here.

Smoothness of Numerical Derivatives

We have claimed and proved that the vector field ΨCr\Psi_C^r is smooth, and that, compared to the original ΨC\Psi_C, it behaves much better when computing derivatives. In this subsection, we will illustrate this.

To do so, we assumed a time budget that would allow us to compute [1] using 15,000 sampled points. Per “Practical Observation 4” in the previous subsection, this would allow us to use, still with r=0r = 0, 750 polynomial segments of degree 1, 75 polynomial segments of degree 2 or 25 polynomial segments of degree 3. Furthermore, it would allow us to have 1,500 points for r>0r > 0.

We have implemented the model in Subsection III-A using these five strategies, and computed the derivatives of the control input ΨCr\Psi_C^r numerically up to the 4th4^{\mathrm{th}} derivative. In the case of r>0r > 0, we took r=0.18r = 0.18 (as in the upcoming experiment) and, instead of 1,500 points, we took roughly one third, (30%) which 450 points, thus providing one third of the time budget of the other approaches.

Figure 4 shows the plots of the derivatives up to order 4 for each one of the approaches, computed with Δt=0.001 s\Delta t = 0.001\ \mathrm{s}. The smoothness of our approach is very clear: all high order derivative plots stay “stable”, whereas in the others the numerical derivatives start to degenerate due to numerical errors that come from the nondifferentiability. For example, when using r=0r = 0 and sampled points, the control input is barely differentiable even once, having already high frequency noise.

Figure 4. Comparison of the control inputs from the GVF and their numerical derivatives considering different implementations.

Experiment with Drone

We implemented the vector field on a custom-built quadrotor in an indoor motion-capture arena measuring approximately 13.8 m × 30.0 m. The overhead motion-capture system consisted of 54 Swift120A/AW cameras arranged in a regular rectangular layout. Each camera provides 12 MP resolution and operates at up to 320 fps, with a specified positioning accuracy of ±0.03 mm. The FZMotion system was operated using FZMotion’s Moca software, which was configured to publish pose data at 120 Hz. The pose data were forwarded through VRPN to the onboard ROS 2/PX4 control stack, with Micro XRCE-DDS providing the ROS 2-PX4 transport. The drone computes everything in a Raspberry Pi 4B with 2 GB RAM. The system model and controller parameters are as described in Subsection III-A and Table I, respectively.

Our target curve is shown in Figure 5. It is important to mention that this curve did not appear as one of the five curves used in the analysis in Subsection III-B. Furthermore, it is significantly more complex topology-wise and also larger, being contained in a radius of 3 m instead of 1 m. These are interesting, because we can then try to extrapolate from the analysis in Subsection III-B.

Target path and tracking trajectory in the flight experiment

Figure 5. Target path (black) and tracking trajectory (green) in the flight experiment

The experiments in Subsection III-B were performed on a 11th Gen Intel Core i7-11800H CPU at 2.30 GHz using a single core, whereas the drone has a Raspberry Pi 4B with 2 GB RAM. We estimate that the latter is roughly 8 times slower than the former. This implies that, if we have to adapt the xx axis of Figure 3 to the drone, also considering that each evaluation of the full controller requires 9 evaluations of the GVF, we arrive to the consideration that the time scale in Figure 3 should be multiplied by 100. If we want to operate at 1 KHz, this implies looking at tb=1t_b = 1ms/100 = 10μ10\mus in Figure 3. This suggests r=0.02r = 0.02m and that we would have an error 0.8cm. But our curve is roughly three times the size of the curves considered in Subsection III-B. “Practical Observation 2” implies that we should roughly triple r=0.02r = 0.02m then and expect roughly three times the average error. In the real experiment, considering real life behaviors not modeled by the model in Subsection III-A, we decided to have r=0.18r = 0.18m. The time budget of 1ms in the drone allow us to use roughly N=800N = 800 points. We decided to use N=300N = 300.

The tracking performance (distance to the curve) can be seen in Figure 6, and it is at most 0.09m and 0.03m on average. Figure 5 shows the agreement between the trajectory (green line) and the target curve (black).

Figure 6. Curve tracking error 2DC(p(t))\sqrt{2D_{\mathcal{C}}(p(t))} in the experiment, in relation to the original curve C\mathcal{C}.

Figure 7 shows the measured physical states. Figure 8 shows the control inputs in high level ΨC\Psi_C, along with its first and second derivatives (needed in the controller), and the lower level control inputs ω\omega and TT. Figure 9 shows auxiliary variables and the virtual state sds_d that helps understand the behavior. The high-level guidance signal ΨC\Psi_C and its first two derivatives vary smoothly at the scale shown. The angular velocity input ω\omega, however, exhibits more pronounced fast fluctuations. This behavior is consistent with the sensitivity of the low-level controller to fluctuations in the measured states, particularly the linear velocity vv. Indeed, the measured velocity enters both ada_d and aca_c, and its fluctuations propagate through zdz_d and z˙d\dot{z}_d into the angular velocity command. Spatial smoothing, as performed by our proposed GVF, does not filter these feedback signals. Its benefit is to provide a differentiable guidance field, avoiding discontinuities associated with the curve representation and allowing the required derivatives to be computed numerically. The experiment illustrates the use of this guidance field in a controller requiring second derivatives, under real measurement conditions.

Figure 7. Physical measured states in the experiment.

Figure 8. Control inputs in the experiment.

Auxiliary variables and virtual state $v_d$

Figure 9. Auxiliary variables and virtual state vdv_d

Conclusion

In this paper, we proposed a generalization of the GVF in [1] that allows a smoothing parameter rr. For r>0r > 0 the GVF is infinitely differentiable, which is beneficial for control loops that require derivatives. Perhaps even more important in practice, this smoothing allows the implementation of the GVF with much less sampled points, and thus much faster, at a modest expense of tracking error.

For future works, we would like to:

  • Extend to the case in which G/HG/H can be varying with pp. Currently, G/HG/H should be constant, as opposed to [1].

  • Extend to the time varying case, as in [1].

  • For static curves and r=0r = 0, one could use structures as KD-Trees and BVH (Bounding Volume Hierarchy) to speed up the calculation from O(N)O(N) to O(log⁡(N))O(\log(N)). A speed up like this is probably possible for our method as well, at the expense of some numerical precision, since points ξ\xi at the curve far from pp contributes very little to the computation of ΨCr(p)\Psi_{\mathcal{C}}^r(p). This will be studied in a future work.

Appendix

To prove Proposition 1 we will need a series of lemmas.

Lemma 1. Let C\mathcal{C} be a regular curve and f:C×Rn×1↦Ra×bf : \mathcal{C} \times\mathbb{R}^{n \times1} \mapsto\mathbb{R}^{a \times b}. Furthermore, let

MCf(p)≜max⁡ξ∈Cmax⁡χ∈PC∗(p)∥f(ξ,p)−f(χ,p)∥.(18)M_{\mathcal{C}}^f(p) \triangleq\max_{\xi\in\mathcal{C}} \max_{\chi\in\mathcal{P}_{\mathcal{C}}^*(p)} \left\|f(\xi,p)-f(\chi,p)\right\|. \tag*{(18)}
gC(ξ,p)≜∥ξ−p∥22−DC(p).(19)g_{\mathcal{C}}(\xi,p) \triangleq\frac{\left\|\xi-p\right\|^2}{2}-D_{\mathcal{C}}(p). \tag*{(19)}
GCr(p)≜{ξ∈C | ∣gC(ξ,p)∣≤3r2ln⁡(ℓC2r)}.(20)G_{\mathcal{C}}^r(p) \triangleq\left\{\xi\in\mathcal{C}\ \middle|\ \left|g_{\mathcal{C}}(\xi,p)\right| \le3r^2\ln\left(\frac{\ell_{\mathcal{C}}}{2r}\right)\right\}. \tag*{(20)}
BC(p)≜max⁡ξ∈C(1+κC∥p−ξ∥).(21)B_{\mathcal{C}}(p) \triangleq\max_{\xi\in\mathcal{C}}\left(1+\kappa_{\mathcal{C}}\left\|p-\xi\right\|\right). \tag*{(21)}
mCf(p,r)≜max⁡ξ∈GCr(p)max⁡χ∈pC∗(p)∥f(ξ,p)−f(χ,p)∥(22)m_C^f(p,r) \triangleq\max_{\xi\in G_C^r(p)}\max_{\chi\in p_C^*(p)}\|f(\xi,p)-f(\chi,p)\| \tag*{(22)}

where recall that ℓC\ell_C is the length of the curve CC. Then, for any 0<r<ℓC/20<r<\ell_C/2, any p∈Rn×1p\in\mathbb{R}^{n\times1} and any χ∈pC∗(p)\chi\in p_C^*(p):

∥ECr[f(ξ,p)](p)−f(χ,p)∥≤mCf(p,r)+r2MCf(p)8BC(p)KℓC2,(23)\left\|\mathbb{E}_C^r[f(\xi,p)](p)-f(\chi,p)\right\|\leq m_C^f(p,r)+r^2M_C^f(p)\frac{8\sqrt{B_C(p)}}{K\ell_C^2}, \tag*{(23)}

where K=∫−11e−u2/2duK=\int_{-1}^{1}e^{-u^2/2}du.

Proof. Preliminary definitions: Fix any pp and any χ∈pC∗(p)\chi\in p_C^*(p). We start by noting that the factor exp⁡(−DC(p)/r2)\exp(-D_C(p)/r^2) cancels from the numerator and denominator of the DEO. Thus, define

L=∫Ce−gC(ξ,p)/r2∥dξ∥(24)L=\int_C e^{-g_C(\xi,p)/r^2}\|d\xi\| \tag*{(24)}

and Δ=ECr[f(ξ,p)](p)−f(χ,p)\Delta=\mathbb{E}_C^r[f(\xi,p)](p)-f(\chi,p). We have that

Δ=1L∫Ce−gC(ξ,p)/r2(f(ξ,p)−f(χ,p))∥dξ∥.(25)\Delta=\frac{1}{L}\int_C e^{-g_C(\xi,p)/r^2}(f(\xi,p)-f(\chi,p))\|d\xi\|. \tag*{(25)}

We bound this quantity by lower bounding the denominator and upper bounding the numerator’s norm.

Lower bounding the denominator: Let ξ(s)\xi(s) be an arc-length parametrization such that ξ(0)=χ\xi(0)=\chi. Define h(s)=gC(ξ(s),p)h(s)=g_C(\xi(s),p). We have that h(0)=0h(0)=0 and, since h(s)h(s) attains its minimum at s=0s=0, dhds(0)=0\frac{dh}{ds}(0)=0. Moreover

d2ds2h(s)=(1+(ξ(s)−p)⊤ξ′′(s))≤BC(p)(26)\frac{d^2}{ds^2}h(s)=\left(1+(\xi(s)-p)^\top\xi''(s)\right)\leq B_C(p) \tag*{(26)}

where we used the fact that, since ss is the arc-length, from the Frenet-Serret formulae we have ∥ξ′′(s)∥≤κC\|\xi''(s)\|\leq\kappa_C for all ss.

Note that if a twice differentiable function k(s)k(s) satisfies k′′(s)≤Bk''(s)\leq B everywhere then double integration yields k(s)≤k(0)+k′(0)s+Bs2/2k(s)\leq k(0)+k'(0)s+Bs^2/2. Therefore, we have that:

gC(ξ(s),p)=h(s)≤BC(p)2s2.(27)g_C(\xi(s),p)=h(s)\leq\frac{B_C(p)}{2}s^2. \tag*{(27)}

Assume 0<r<ℓC/20<r<\ell_C/2. Thus, since BC(p)≥1B_C(p)\geq1, the interval ∣s∣≤r/BC(p)|s|\leq r/\sqrt{B_C(p)} is contained in an arc-length interval of length ℓC\ell_C. Hence,

L≥∫−r/BC(p)r/BC(p)e−BC(p)s2/(2r2)ds=rBC(p)∫−11e−u2/2du=KrBC(p).(28)L\geq\int_{-r/\sqrt{B_C(p)}}^{r/\sqrt{B_C(p)}}e^{-B_C(p)s^2/(2r^2)}ds =\frac{r}{\sqrt{B_C(p)}}\int_{-1}^{1}e^{-u^2/2}du =\frac{Kr}{\sqrt{B_C(p)}}. \tag*{(28)}

Upper bounding the numerator’s norm: Since 0<r<ℓC/20<r<\ell_C/2, we have that 3r2ln⁡(ℓC/(2r))>03r^2\ln(\ell_C/(2r))>0, and thus the set GCr(p)G_C^r(p) is nonempty because it contains pC∗(p)p_C^*(p). From (25) and the triangle inequality, we upper bound the norm of the numerator of Δ\Delta by U1+U2U_1+U_2, in which:

U1=∫GCr(p)e−gC(ξ,p)/r2∥f(ξ,p)−f(χ,p)∥∥dξ∥.U_1=\int_{G_C^r(p)}e^{-g_C(\xi,p)/r^2}\|f(\xi,p)-f(\chi,p)\|\|d\xi\|.
U2=∫C−GCr(p)e−gC(ξ,p)/r2∥f(ξ,p)−f(χ,p)∥∥dξ∥.(29)U_2=\int_{C-G_C^r(p)}e^{-g_C(\xi,p)/r^2}\|f(\xi,p)-f(\chi,p)\|\|d\xi\|. \tag*{(29)}

By the definition of mCf(p,r)m_C^f(p,r), U1≤mCf(p,r)LU_1\leq m_C^f(p,r)L. For U2U_2, we note that, on C−GCr(p)C-G_C^r(p),

e−gC(ξ,p)/r2≤e−(3r2ln⁡(ℓC/(2r)))/r2=(2rℓC)3.(30)e^{-g_C(\xi,p)/r^2}\leq e^{-(3r^2\ln(\ell_C/(2r)))/r^2}=\left(\frac{2r}{\ell_C}\right)^3. \tag*{(30)}

Therefore, U2≤ℓC⋅MCf(p)(8r3/ℓC3)U_2\leq\ell_C\cdot M_C^f(p)(8r^3/\ell_C^3).

The quantity ∥Δ∥\|\Delta\| is thus upper bounded by U1/L+U2/LU_1/L+U_2/L. The first term is bounded by mCf(p,r)m_C^f(p,r), while the second term satisfies

8MCf(p)r3LℓC2≤8MCf(p)r3BC(p)KrℓC2=r2MCf(p)8BC(p)KℓC2.(31)\frac{8M_C^f(p)r^3}{L\ell_C^2}\leq\frac{8M_C^f(p)r^3\sqrt{B_C(p)}}{Kr\ell_C^2}=r^2M_C^f(p)\frac{8\sqrt{B_C(p)}}{K\ell_C^2}. \tag*{(31)}

And the result in (23) is proved. ■

This lemma establishes bounds, but we now show that for some points pp and assuming continuity of ff, this bound can be made arbitrarily small by making rr small.

Lemma 2. Let CC be a regular curve and f:C×Rn×1↦Ra×bf:C\times\mathbb{R}^{n\times1}\mapsto\mathbb{R}^{a\times b} a continuous function in both arguments. Define the subset of Rn×1\mathbb{R}^{n\times1} for all points pp in which multiple values of pC∗(p)p_C^*(p) does not create a multi-valued ff4:

UCf≜{p∈Rn×1∣f(ξ1,p)=f(ξ2,p) ∀ξ1,ξ2∈pC∗(p)}.(32)\mathcal{U}_C^f\triangleq\{p\in\mathbb{R}^{n\times1}\mid f(\xi_1,p)=f(\xi_2,p)\ \forall\xi_1,\xi_2\in p_C^*(p)\}. \tag*{(32)}

Given a point p∈UCfp\in\mathcal{U}_C^f, denote the common evaluation of ff at pp by f∗(p)f^*(p). Then, when r→+0r\to+0, ECr[f(ξ,p)](p)→f∗(p)\mathbb{E}_C^r[f(\xi,p)](p)\to f^*(p).

Proof. We invoke Lemma 1 and analyze the behavior of (23). Note that for a fixed p∈UCfp\in\mathcal{U}_C^f, lim⁡r→+0mCf(p,r)=0\lim_{r\to+0}m_C^f(p,r)=0, because gC(ξ,p)≥0g_C(\xi,p)\geq0 and gC(ξ,p)=0g_C(\xi,p)=0 implies ξ∈pC∗(p)\xi\in p_C^*(p), and, therefore, for any such ξ\xi, f(ξ,p)=f∗(p)f(\xi,p)=f^*(p). This and the continuity of ff guarantee that lim⁡r→+0mCf(p,r)=0\lim_{r\to+0}m_C^f(p,r)=0. Clearly, the other term also vanishes when r→+0r\to+0. ■

As we get closer and closer to a point p∉UCfp\notin\mathcal{U}_C^f, the convergence becomes slower, because mCfm_C^f falls to 00 slower.

Lemma 3. Let CC be a regular curve and f:C×Rn×1↦Rm×1f:C\times\mathbb{R}^{n\times1}\mapsto\mathbb{R}^{m\times1} be a function that is differentiable in the second variable. Define Cov⁡Cr[f(ξ,p),g(ξ,p)](p)\operatorname{Cov}_C^r[f(\xi,p),g(\xi,p)](p) as the covariance

matrix between f(ξ,p)f(\xi,p) and another function g:C×Rn×1↦Rm×1g:C\times\mathbb{R}^{n\times1}\mapsto\mathbb{R}^{m\times1}:

Cov⁡Cr[f(ξ,p),g(ξ,p)]≔ECr[f(ξ,p)g(ξ,p)T]−ECr[f(ξ,p)]ECr[g(ξ,p)]T(33)\operatorname{Cov}^{r}_{C}[f(\xi,p),g(\xi,p)] \coloneqq\mathbb{E}^{r}_{C}[f(\xi,p)g(\xi,p)^{\mathsf{T}}]-\mathbb{E}^{r}_{C}[f(\xi,p)]\mathbb{E}^{r}_{C}[g(\xi,p)]^{\mathsf{T}} \tag*{(33)}

Then:

∂∂pECr[f(ξ,p)]=ECr[∂f∂p(ξ,p)]+Cov⁡Cr[f(ξ,p),ξ−p]r2(34)\frac{\partial}{\partial p}\mathbb{E}^{r}_{C}[f(\xi,p)]=\mathbb{E}^{r}_{C}\left[\frac{\partial f}{\partial p}(\xi,p)\right]+\frac{\operatorname{Cov}^{r}_{C}[f(\xi,p),\xi-p]}{r^{2}} \tag*{(34)}

Proof. By inspection. ■

Lemma 4. Let CC be a regular curve. Define the rr-DEO-variance of a function f:C×Rn×1↦Rm×1f:C\times\mathbb{R}^{n\times1}\mapsto\mathbb{R}^{m\times1}.

Var⁡Cr[f(ξ,p)]≔ECr[∥f(ξ,p)∥2]−∥ECr[f(ξ,p)]∥2.(35)\operatorname{Var}^{r}_{C}[f(\xi,p)]\coloneqq\mathbb{E}^{r}_{C}[\lVert f(\xi,p)\rVert^{2}]-\lVert\mathbb{E}^{r}_{C}[f(\xi,p)]\rVert^{2}. \tag*{(35)}

Consider ΨCr(p)\Psi^{r}_{C}(p) as in (8), with p∈Rn×1p\in\mathbb{R}^{n\times1}. Then:

Tr⁡(∂ΨCr∂p(p))=G(Var⁡Cr[ξ−p]r2−n)−Hr2ECr[TC(ξ)]TECr[ξ−p].(36)\operatorname{Tr}\left(\frac{\partial\Psi^{r}_{C}}{\partial p}(p)\right)=G\left(\frac{\operatorname{Var}^{r}_{C}[\xi-p]}{r^{2}}-n\right)-\frac{H}{r^{2}}\mathbb{E}^{r}_{C}[T_{C}(\xi)]^{\mathsf{T}}\mathbb{E}^{r}_{C}[\xi-p]. \tag*{(36)}

Proof. Take f(ξ,p)=ψC(ξ,p)=G(ξ−p)+HTC(ξ)f(\xi,p)=\psi_{C}(\xi,p)=G(\xi-p)+HT_{C}(\xi). Thus, using Lemma 3, the Jacobian of ΨCr(p)=ECr[ψC(ξ,p)](p)\Psi^{r}_{C}(p)=\mathbb{E}^{r}_{C}[\psi_{C}(\xi,p)](p) is:

G(Cov⁡Cr[ξ−p,ξ−p]r2−In×n)−Hr2ECr[TC(ξ)]ECr[ξ−p]T+Hr−2ECr[TC(ξ)(ξ−p)T].(37)G\left(\frac{\operatorname{Cov}^{r}_{C}[\xi-p,\xi-p]}{r^{2}}-I_{n\times n}\right)-\frac{H}{r^{2}}\mathbb{E}^{r}_{C}[T_{C}(\xi)]\mathbb{E}^{r}_{C}[\xi-p]^{\mathsf{T}}+Hr^{-2}\mathbb{E}^{r}_{C}[T_{C}(\xi)(\xi-p)^{\mathsf{T}}]. \tag*{(37)}

Now, take the trace of this matrix, use the linearity of ECr\mathbb{E}^{r}_{C} and the fact that Tr⁡(abT)=bTa\operatorname{Tr}(ab^{\mathsf{T}})=b^{\mathsf{T}}a for two column vectors a,ba,b, so we can conclude that the trace of the Jacobian is:

G(Var⁡Cr[ξ−p]r2−n)−Hr2ECr[TC(ξ)]TECr[ξ−p]+Hr−2ECr[(ξ−p)TTC(ξ)].(38)G\left(\frac{\operatorname{Var}^{r}_{C}[\xi-p]}{r^{2}}-n\right)-\frac{H}{r^{2}}\mathbb{E}^{r}_{C}[T_{C}(\xi)]^{\mathsf{T}}\mathbb{E}^{r}_{C}[\xi-p]+Hr^{-2}\mathbb{E}^{r}_{C}[(\xi-p)^{\mathsf{T}}T_{C}(\xi)]. \tag*{(38)}

The last term can be shown to be zero. This is because:

∫Ce−∥p−ξ∥22r2(ξ−p)TTC(ξ)∥dξ∥=−r2∫C∂∂ξ(e−∥p−ξ∥22r2)dξ=0(39)\int_{C}e^{-\frac{\lVert p-\xi\rVert^{2}}{2r^{2}}}(\xi-p)^{\mathsf{T}}T_{C}(\xi)\lVert d\xi\rVert = -r^{2}\int_{C}\frac{\partial}{\partial\xi}\left(e^{-\frac{\lVert p-\xi\rVert^{2}}{2r^{2}}}\right)d\xi=0 \tag*{(39)}

since the curve CC is closed. This concludes the proof. ■

Lemma 5. Let CC be a regular curve. Define:

ZCr(p)≔∫Ce−∥ξ−p∥22r2∥dξ∥.(40)Z^{r}_{C}(p)\coloneqq\int_{C}e^{-\frac{\lVert\xi-p\rVert^{2}}{2r^{2}}}\lVert d\xi\rVert. \tag*{(40)}

Also, define the set:

ΩCr≔{p∈Rn×1∣DC(p)≥nr2}.(41)\Omega^{r}_{C}\coloneqq\{p\in\mathbb{R}^{n\times1}\mid D_{C}(p)\geq nr^{2}\}. \tag*{(41)}

Let Ψ~Cr(p)≔ZCr(p)ΨCr(p)\widetilde{\Psi}^{r}_{C}(p)\coloneqq Z^{r}_{C}(p)\Psi^{r}_{C}(p), with ΨCr(p)\Psi^{r}_{C}(p) as in (8). Then, for all p∈ΩCrp\in\Omega^{r}_{C}:

Tr⁡(∂Ψ~Cr∂p(p))≥GnZCr(p).(42)\operatorname{Tr}\left(\frac{\partial\widetilde{\Psi}^{r}_{C}}{\partial p}(p)\right)\geq GnZ^{r}_{C}(p). \tag*{(42)}

Proof. Using standard calculus rules and the properties of the trace operator, we have that:

Tr⁡(∂Ψ~Cr∂p(p))=ZCr(p)Tr⁡(∂ΨCr∂p(p))+∂ZCr∂p(p)ΨCr(p).(43)\operatorname{Tr}\left(\frac{\partial\widetilde{\Psi}^{r}_{C}}{\partial p}(p)\right)=Z^{r}_{C}(p)\operatorname{Tr}\left(\frac{\partial\Psi^{r}_{C}}{\partial p}(p)\right)+\frac{\partial Z^{r}_{C}}{\partial p}(p)\Psi^{r}_{C}(p). \tag*{(43)}

By noting that ∂ZCr∂p(p)=ZCr(p)r2ECr[ξ−p]T\frac{\partial Z^{r}_{C}}{\partial p}(p)=\frac{Z^{r}_{C}(p)}{r^{2}}\mathbb{E}^{r}_{C}[\xi-p]^{\mathsf{T}}, and using the result in Lemma 4, we obtain that:

Tr⁡(∂Ψ~Cr∂p(p))=GZCr(p)(ECr[∥ξ−p∥2]r2−n).(44)\operatorname{Tr}\left(\frac{\partial\widetilde{\Psi}^{r}_{C}}{\partial p}(p)\right)=GZ^{r}_{C}(p)\left(\frac{\mathbb{E}^{r}_{C}[\lVert\xi-p\rVert^{2}]}{r^{2}}-n\right). \tag*{(44)}

Note that, since ∥ξ−p∥2≥2DC(p)\lVert\xi-p\rVert^{2}\geq2D_{C}(p) for ξ∈C\xi\in C and ECr\mathbb{E}^{r}_{C} is a monotone operator, ECr[∥ξ−p∥2]≥2DC(p)\mathbb{E}^{r}_{C}[\lVert\xi-p\rVert^{2}]\geq2D_{C}(p). Thus, for p∈ΩCrp\in\Omega^{r}_{C}, it is clear that the desired result holds. ■

Lemma 6. Let CC be a regular curve and consider the dynamical system p˙=ΨCr(p)\dot{p}=\Psi^{r}_{C}(p) with p∈Rn×1p\in\mathbb{R}^{n\times1} and ΨCr\Psi^{r}_{C} as in (8). Then, the set of initial conditions p0p_{0} that stays in the set ΩCr\Omega^{r}_{C} (see (41)) has measure zero in Rn×1\mathbb{R}^{n\times1}.

Proof. The key point is to define a weighted measure in Rn×1\mathbb{R}^{n\times1} instead of the traditional one. Let Vol⁡Cr:2Rn×1↦R+\operatorname{Vol}^{r}_{C}:2^{\mathbb{R}^{n\times1}}\mapsto\mathbb{R}^{+} be defined as:

Vol⁡Cr(A)=∫AZCr(p) dp(45)\operatorname{Vol}^{r}_{C}(A)=\int_{A}Z^{r}_{C}(p)\,dp \tag*{(45)}

with ZCr(p)Z^{r}_{C}(p) as in (40). It is finite for all sets, indeed:

Vol⁡Cr(Rn×1)≤(2π)n/2rnℓC(46)\operatorname{Vol}^{r}_{C}(\mathbb{R}^{n\times1})\leq(2\pi)^{n/2}r^{n}\ell_{C} \tag*{(46)}

in which the nn-dimensional Gaussian integral was employed.

Let Φ(t,p0)\Phi(t,p_{0}) be the flow of the differential equation p˙=ΨCr(p)\dot{p}=\Psi^{r}_{C}(p). Now consider the set A={p0∈Rn×1∣DC(Φ(t,p0))≥nr2 ∀t≥0}A=\{p_{0}\in\mathbb{R}^{n\times1}\mid D_{C}(\Phi(t,p_{0}))\geq nr^{2}\ \forall t\geq0\}. We will show that this set has zero measure in Rn×1\mathbb{R}^{n\times1}.

Using Ψ~Cr(p)=ZCr(p)ΨCr(p)\widetilde{\Psi}^{r}_{C}(p)=Z^{r}_{C}(p)\Psi^{r}_{C}(p) We use the weighted Liouville Formula [19]:

ZCr(Φ(t,p0))det⁡(∂Φ∂p0(t,p0))=ZCr(p0)exp⁡(∫0t1ZCr(Φ(s,p0))Tr⁡(∂Ψ~Cr∂p(Φ(s,p0)))ds)(47)Z^{r}_{C}(\Phi(t,p_{0}))\det\left(\frac{\partial\Phi}{\partial p_{0}}(t,p_{0})\right)=Z^{r}_{C}(p_{0})\exp\left(\int_{0}^{t}\frac{1}{Z^{r}_{C}(\Phi(s,p_{0}))}\operatorname{Tr}\left(\frac{\partial\widetilde{\Psi}^{r}_{C}}{\partial p}(\Phi(s,p_{0}))\right)ds\right) \tag*{(47)}

Now take a p0∈A⊆ΩCrp_{0}\in A\subseteq\Omega^{r}_{C} (see (41)). There, we can use Lemma 5 to conclude that:

ZCr(Φ(t,p0))det⁡(∂Φ∂p0(t,p0))≥ZCr(p0)eGnt.(48)Z^{r}_{C}(\Phi(t,p_{0}))\det\left(\frac{\partial\Phi}{\partial p_{0}}(t,p_{0})\right)\geq Z^{r}_{C}(p_{0})e^{Gnt}. \tag*{(48)}

Now, we use the change of variable formula and (48) to conclude that:

Vol⁡Cr(Φ(t,A))=∫AZCr(Φ(t,p0)) dΦ(t,p0)\operatorname{Vol}^{r}_{C}(\Phi(t,A))=\int_{A}Z^{r}_{C}(\Phi(t,p_{0}))\,d\Phi(t,p_{0})
=∫AZCr(Φ(t,p0))det⁡(∂Φ∂p0(t,p0))dp0≥eGntVol⁡Cr(A).(49)=\int_{A}Z^{r}_{C}(\Phi(t,p_{0}))\det\left(\frac{\partial\Phi}{\partial p_{0}}(t,p_{0})\right)dp_{0} \geq e^{Gnt}\operatorname{Vol}^{r}_{C}(A). \tag*{(49)}

eq. (48)

Thus, for all t≥0t \ge0, Vol⁡Cr(Φ(t,A))≥eGntVol⁡Cr(A)\operatorname{Vol}_{C}^{r}(\Phi(t,A)) \ge e^{Gnt}\operatorname{Vol}_{C}^{r}(A). Now, we note that from triangle inequality ∥ΨCr(p)∥≤G∥p∥+Gmax⁡ξ∈C∥ξ∥+H\|\Psi_{C}^{r}(p)\| \le G\|p\| + G\max_{\xi\in C}\|\xi\| + H. Thus ∥ΨCr(p)∥\|\Psi_{C}^{r}(p)\| has a linearly-bounded growth and thus the vector field is forward complete, i.e., Φ(t,p0)\Phi(t,p_{0}) is finite for all finite t,p0t,p_{0}. Thus, if we make t→∞t \to\infty, and use the fact that Gn>0G n > 0 and that Vol⁡Cr(Φ(t,A))\operatorname{Vol}_{C}^{r}(\Phi(t,A)) is bounded, we find that Vol⁡Cr(A)=0\operatorname{Vol}_{C}^{r}(A) = 0. Thus, it has zero measure in Rn×1\mathbb{R}^{n \times1}, since ZCr(p)>0Z_{C}^{r}(p) > 0.

Hence, any set of points that DC(p)≥nr2D_{C}(p) \ge nr^{2} forever has zero measure in Rn×1\mathbb{R}^{n \times1}. And the proof is concluded. ■

Lemma 7. Let CC be a regular curve and consider the dynamical system p˙=ΨCr(p)\dot{p} = \Psi_{C}^{r}(p) with p∈Rn×1p \in\mathbb{R}^{n \times1} and ΨCr(p)\Psi_{C}^{r}(p) as in (8). Then, for rr small enough, there is a compact set CrC^{r} that is positively invariant. Furthermore, Rn×1−ΩCr⊆Cr\mathbb{R}^{n \times1} - \Omega_{C}^{r} \subseteq C^{r}, Cr⊆UCC^{r} \subseteq U_{C} and lim⁡r→+0Cr=C\lim_{r \to+0} C^{r} = C.

Proof. Let δC=inf⁡p∉UCDC(p)\delta_{C} = \inf_{p \notin U_{C}} D_{C}(p). By the tubular-neighborhood theorem [21] for compact embedded twice differentiable curves (which is the case of regular curves), there exists a neighborhood of CC in which every point has a unique closest point on CC. Consequently, δC>0\delta_{C} > 0.

Now, let C^={p∈Rn×1∣DC(p)≤δC2}⊆UC\widehat{C} = \{p \in\mathbb{R}^{n \times1} \mid D_{C}(p) \le\frac{\delta_{C}}{2}\} \subseteq U_{C}. Consider a p∈C^p \in\widehat{C}, and thus p∈UCp \in U_{C}. In UCU_{C}, we can make rr small as possible (Lemma 2) so f(ξ,p)→f(pC∗(p),p)f(\xi,p) \to f(p_{C}^{*}(p),p) for any continuous function ff. Furthermore, since the set C^\widehat{C} is compact, we can make this bound independent of pp.

Let eN(p)=Eξ[ξ−p]−(pC∗(p)−p)e_{N}(p) = \mathbb{E}_{\xi}[\xi-p] - (p_{C}^{*}(p)-p) and eT(p)=Eξ[TC(ξ)]−TC(pC∗(p))e_{T}(p) = \mathbb{E}_{\xi}[T_{C}(\xi)] - T_{C}(p_{C}^{*}(p)). Thus, we can have functions ϵN(r),ϵT(r):R+↦R+\epsilon_{N}(r),\epsilon_{T}(r):\mathbb{R}^{+} \mapsto\mathbb{R}^{+} that decrease to 00 as r→+0r \to+0. Therefore, for all p∈C^p \in\widehat{C}:

∥eN(p)∥≤ϵN(r),∥eT(p)∥≤ϵT(r).(50)\|e_{N}(p)\| \le\epsilon_{N}(r), \quad\|e_{T}(p)\| \le\epsilon_{T}(r). \tag*{(50)}

Let ϵ(r)=ϵN(r)+HGϵT(r)\epsilon(r) = \epsilon_{N}(r) + \frac{H}{G}\epsilon_{T}(r) and c(r)=max⁡(nr2,12ϵ(r)2)c(r) = \max(nr^{2},\frac{1}{2}\epsilon(r)^{2}). Pick rr small enough so c(r)≤12δCc(r) \le\frac{1}{2}\delta_{C} and define

Cr≜{p∈Rn×1∣DC(p)≤c(r)}(51)C^{r} \triangleq\{p \in\mathbb{R}^{n \times1} \mid D_{C}(p) \le c(r)\} \tag*{(51)}

and thus Cr⊆UCC^{r} \subseteq U_{C}. Consider now that p∈Crp \in C^{r}. Thus, using the fact that ΨCr(p)=G(pC∗(p)−p+eN)+H(TC(pC∗(p))+eT)\Psi_{C}^{r}(p) = G(p_{C}^{*}(p)-p+e_{N}) + H(T_{C}(p_{C}^{*}(p))+e_{T}), the fact that (p−pC∗(p))TTC(pC∗(p))=0(p-p_{C}^{*}(p))^{T}T_{C}(p_{C}^{*}(p)) = 0 and Schwartz’s inequality:

ddtDC(p)≤−2GDC(p)+G2DC(p)ϵ(r)(52)\frac{d}{dt}D_{C}(p) \le-2GD_{C}(p) + G\sqrt{2D_{C}(p)}\epsilon(r) \tag*{(52)}

and one can check ddtDC(p)≤0\frac{d}{dt}D_{C}(p) \le0 at the boundary of CrC^{r}. This implies that CrC^{r} is positively invariant. This set is contained in UCU_{C} and is compact by construction. Furthermore, since c(r)≥nr2c(r) \ge nr^{2}, we have that Cr⊇Rn×1−ΩCrC^{r} \supseteq\mathbb{R}^{n \times1} - \Omega_{C}^{r}. In addition, by the definition of CrC^{r}, clearly lim⁡r→+0Cr=C\lim_{r \to+0} C^{r} = C. ■

Lemma 8. Let CC be a regular curve and consider the dynamical system p˙=ΨCr(p)\dot{p} = \Psi_{C}^{r}(p) with p∈Rn×1p \in\mathbb{R}^{n \times1} and ΨCr(p)\Psi_{C}^{r}(p) as in (8). Let A⊆UCA \subseteq U_{C} be a compact subset. For rr small enough, no point of AA can be such that ΨCr(p)=0\Psi_{C}^{r}(p) = 0.

Proof. The proof follows by contradiction. Suppose that there exists a compact A⊆UCA \subseteq U_{C} and that there exists p∈Ap \in A such that ΨCr(p)=0\Psi_{C}^{r}(p) = 0. Applying Lemma 2 with f=ψCf = \psi_{C}, since AA is compact, we can have rr small enough so, for all p∈Ap \in A, we have

∥ΨCr(p)−ΨC(p)⏟→0∥<H(53)\|\underbrace{\Psi_{C}^{r}(p)-\Psi_{C}(p)}_{\to0}\| < H \tag*{(53)}

in which H>0H > 0. Thus, ∥ΨC(p)∥<H\|\Psi_{C}(p)\| < H. But this is impossible, because ∥ΨC(p)∥=2G2DC(p)+H2≥H\|\Psi_{C}(p)\| = \sqrt{2G^{2}D_{C}(p)+H^{2}} \ge H, which contradicts (53). ■

Finally, we can prove Proposition 1:

First statement: comes trivially from the definition of ΨCr(p)\Psi_{C}^{r}(p).

Second statement: comes from Lemma 2, by using the function f(ξ,p)=ψC(ξ,p)f(\xi,p) = \psi_{C}(\xi,p).

Third statement: from Lemma 6, for almost all initial conditions, we cannot stay inside ΩCr\Omega_{C}^{r} forever. Thus we must enter, in finite time, Rn×1−ΩCr\mathbb{R}^{n \times1}-\Omega_{C}^{r}. Using Lemma 7, if rr is small enough, this implies that we entered CrC^{r}, which is positively invariant, and thus we never leave this set. Furthermore, this set CrC^{r} is such that lim⁡r→+0Cr=C\lim_{r \to+0}C^{r}=C. Finally, since Cr⊆UCC^{r} \subseteq U_{C} and it is compact, Lemma 8 says that the vector field never vanishes there. Thus, the point stay moving, with a vector field which is an approximation of the tangent field because, for rr small inside CrC^{r}, Eξ[ξ−p](p)≈0\mathbb{E}_{\xi}[\xi-p](p) \approx0 and Eξr[TC(ξ)](p)≈TC(pC∗(p))\mathbb{E}_{\xi}^{r}[T_{C}(\xi)](p) \approx T_{C}(p_{C}^{*}(p)).

References

  1. [1]A. M. C. Rezende, V. M. Gonçalves, and L. C. A. Pimenta, “Constructive time-varying vector fields for robot navigation,” IEEE Transactions on Robotics, vol. 38, no. 2, pp. 852–867, 2022.DOI
  2. [2]V. M. Gonçalves, L. C. A. Pimenta, C. A. Maia, B. C. O. Dutra, and G. A. S. Pereira, “Vector fields for robot navigation along time-varying curves in n-dimensions,” IEEE Transactions on Robotics, vol. 26, no. 4, pp. 647–659, 2010.DOI
  3. [3]W. Yao, H. G. de Marina, B. Lin, and M. Cao, “Singularity-free guiding vector field for robot navigation,” IEEE Transactions on Robotics, vol. 37, no. 4, pp. 1206–1221, 2021.arxiv.org/abs/2012.01826
  4. [4]W. Yao, B. Lin, B. D. O. Anderson, and M. Cao, “Topological analysis of vector-field guided path following on manifolds,” IEEE Transactions on Automatic Control, vol. 68, no. 3, pp. 1353–1368, 2023.
  5. [5]Z. Chen and Z. Zuo, “Non-singular cooperative guiding vector field under a homotopy equivalence transformation,” Automatica, vol. 171, p. 111962, 2025.arxiv.org/abs/2308.06934
  6. [6]Y. Zhou, J. Bautista, W. Yao, and H. G. de Marina, “Inverse kinematics on guiding vector fields for robot path following,” in 2025 IEEE International Conference on Robotics and Automation (ICRA), pp. 8635–8641, 2025.
  7. [7]Z. Chen, Z. Zuo, G. Wang, P. Li, and Y. Liu, “Discrete data-based frequency-domain closed planar path following method,” IEEE/ASME Transactions on Mechatronics, vol. 30, no. 6, pp. 5330–5341, 2025.DOI
  8. [8]F. Bartel, V. M. Gonçalves, and L. C. A. Pimenta, “Constructive vector fields for path following in fully-actuated systems on matrix lie groups,” arXiv preprint arXiv:2602.21450, 2026.
  9. [9]A. H. D. Nunes, V. M. Gonçalves, and L. C. A. Pimenta, “Safe vector field for robot navigation in n-dimensions,” IEEE Robotics and Automation Letters, 2026.DOI
  10. [10]W. Yao, H. G. de Marina, Z. Sun, and M. Cao, “Guiding vector fields for the distributed motion coordination of mobile robots,” IEEE Transactions on Robotics, vol. 39, no. 2, pp. 1119–1135, 2022.arxiv.org/abs/2209.09478
  11. [11]A. H. D. Nunes, G. V. Raffo, and L. C. A. Pimenta, “Integrated vector field and backstepping control for quadcopters,” in 2023 IEEE International Conference on Robotics and Automation (ICRA), pp. 1256–1262, 2023.DOI
  12. [12]D. Zhou and M. Schwager, “Vector field following for quadrotors using differential flatness,” in 2014 IEEE International Conference on Robotics and Automation (ICRA), (Hong Kong, China), pp. 6567–6572, IEEE, May 2014.DOI
  13. [13]Y. Gao, C. Bai, and Q. Quan, “Distributed and differentiable vector field control within a curved virtual tube for a robotic swarm under field-of-view constraints,” IEEE Transactions on Automatic Control, vol. 70, no. 8, pp. 5190–5205, 2025.DOI
  14. [14]P. Blanchard, D. J. Higham, and N. J. Higham, “Accurately computing the log-sum-exp and softmax functions,” IMA Journal of Numerical Analysis, vol. 41, no. 4, pp. 2311–2330, 2021.DOI
  15. [15]A. Richards, “Fast model predictive control with soft constraints,” European Journal of Control, vol. 25, pp. 51–59, 2015.
  16. [16]V. M. Gonçalves, A. Tzes, F. Khorrami, and P. Fraisse, “Safe smooth distances for second order kinematic robot control,” IEEE Transactions on Robotics, 2024. (in production, DOI: 10.1109/TRO.2024.3400924).DOI
  17. [17]P. Rabiee and J. B. Hoagg, “Soft-minimum and soft-maximum barrier functions for safety with actuation constraints,” Automatica, vol. 171, p. 111921, 2025.DOI
  18. [18]Z. Chen, S. Guo, and S. Zhao, “Guiding vector field generation via score-based diffusion model,” in 2026 IEEE International Conference on Robotics and Automation (ICRA), 2026.arxiv.org/abs/2604.24487
  19. [19]A. Rantzer, “A dual to Lyapunov’s stability theorem,” Systems & Control Letters, vol. 42, no. 3, pp. 161–168, 2001.DOI
  20. [20]V. M. Gonçalves, D. Chazelas, A. Tzes, and F. Khorrami, “Safe multi-agent drone control using control barrier functions and acceleration fields,” Robotics and Autonomous Systems, vol. 172, p. 104601, 2024.DOI
  21. [21]R. L. Foote, “Regularity of the distance function,” Proceedings of the American Mathematical Society, vol. 92, no. 1, pp. 153–155, 1984.DOI

Paper details

vlnc:W_dt34w7sz
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents