Constructive Infinitely Differentiable Guidance Vector Fields for Robotic Control

Vinicius Mariano Goncalves, Peilin Wu, Adriano Rezende, Weijia YaoarXiv 2026 · arXiv

Guidance Vector Fields (GVFs) are a cornerstone of robotic path-following. They provide a simple, elegant way to tell a robot where to go: at every point in space, the field points toward the path and along it. However, a hidden "sharpness" in the math behind these fields has long troubled engineers working with high-performance robots like quadrotors.

In their paper, "Constructive Infinitely Differentiable Guidance Vector Fields for Robotic Control," Gonçalves et al. propose a clever mathematical "smoothing" technique that turns these jagged fields into perfectly smooth, infinitely differentiable signals.

TL;DR

The authors propose a new "constructive" Guidance Vector Field that replaces non-differentiable "closest-point" lookups with a smooth, Gaussian-weighted average over the target path. This creates an infinitely differentiable ($C^\infty$) guidance signal that is more robust to numerical errors and enables stable, high-order control (like acceleration and jerk) for complex robotic systems.

What Problem Does This Paper Address?

Most GVFs rely on finding the closest point on a target curve to the robot's current position. While this sounds simple, it creates a major problem at the medial axis—the set of points equidistant from two or more parts of the curve (e.g., the exact center of a circle).

At these points, the "closest point" can jump abruptly from one side of the curve to the other. This creates a discontinuity in the vector field. For a simple ground robot, this might just cause a small jitter. But for a quadrotor using a "cascaded" controller, these jumps are disastrous. Because the controller needs derivatives of the guidance signal (to calculate required acceleration, jerk, or snap), a tiny jump in velocity becomes an infinite spike in acceleration.

Figure 1: The discontinuity problem. The orange line shows the abrupt jump in the traditional GVF at the medial axis ($p_2=0$), while the blue line shows the proposed smooth version.

Furthermore, when curves are discretized into samples for computer implementation, these "hard" jumps happen every time the robot moves from the "influence" of one sample to the next, creating high-frequency noise in the control signal.

Methods

The authors introduce the $r$-Differentiable Evaluation Operator ($r$-DEO). Instead of a "hard" winner-take-all search for the single closest point, the $r$-DEO calculates a normalized Gaussian-weighted average over the entire curve.

Key features of the approach:

  • Smoothing Parameter ($r$): A user-defined scalar that controls the trade-off. A smaller $r$ tracks the curve more precisely (approaching the original GVF), while a larger $r$ provides a smoother field that is easier to differentiate.
  • Infinite Differentiability: The resulting vector field is $C^\infty$ everywhere, even at the medial axis.
  • Numerical Robustness: Because it uses an integral-like approach, it is much less sensitive to the number of points used to represent the curve.

Datasets

The researchers evaluated their method using:

  1. Synthetic Curves: Five different 2D target curves (ellipses, cloverleafs, etc.) used to test computational efficiency and tracking error across thousands of simulated trials.
  2. Real-world 3D Path: A complex, knot-like 3D trajectory used for flight experiments with a custom-built quadrotor in a motion-capture arena.

Key Insights and Ideas

  • Integration as a Filter: By treating the closest-point problem as an integration task rather than a search task, the authors effectively "filter" out the geometric singularities of the path.
  • Efficiency through Sparsity: Surprisingly, the smooth field works better with fewer samples than the traditional method. Because the Gaussian weights "blend" nearby samples, the robot doesn't experience the "jumps" typical of nearest-neighbor lookups.
  • Stability Guarantees: The paper provides formal proofs that robots using this field will converge to a "tube" around the target curve, and that this tube shrinks to the actual curve as the smoothing parameter $r$ goes to zero.

Key Results

  • Clean Derivatives: In tests computing up to the 4th derivative (snap), the proposed method remained stable and "clean," while traditional methods (even those using high-degree polynomials) produced unusable, noisy signals.
Fig. 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 nondifferen- tiable. 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 = 0, which lies on the medial axis.
  • Computational Budget: The smoothed GVF achieved lower tracking errors at low computational budgets. It could run effectively with significantly fewer curve segments than traditional approaches.
  • Real-world Flight: A quadrotor successfully tracked a complex 3D path with an average error of just 3 cm, demonstrating that the theoretical smoothness translates directly to stable physical flight.
Figure 2: The custom quadrotor used to validate the method in real-world flight tests.

Limitations and Future Work

  • Constant Gains: Currently, the ratio between the "push toward the curve" and "move along the curve" must be constant. The authors want to extend this to allow these gains to vary based on the robot's position.
  • Static Curves: The current math assumes the path isn't moving. Future work will address time-varying curves (e.g., following a moving target).
  • Scaling: While the method is efficient, calculating the weighted average over very long, complex curves could become slow ($O(N)$ complexity). The authors suggest exploring spatial data structures like KD-trees to speed this up to $O(\log N)$.

Summary updated

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.
BibTeX
@misc{goncalves2026constructive,
  author = {Vinicius Mariano Goncalves and Peilin Wu and Adriano Rezende and Weijia Yao},
  title = {{Constructive Infinitely Differentiable Guidance Vector Fields for Robotic Control}},
  year = {2026},
  eprint = {2611.24424},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.24424},
}
Figure