💻 computer science

Hölder Signed Distance: A Differentiable, Signed, Parallelizable Metric for Robotics

This paper introduces the Hölder signed distance, a novel closed-form, differentiable, and GPU-parallelizable metric for convex polyhedra that overcomes the non-differentiability and computational limitations of existing Euclidean SDFs to enable efficient real-time robotic motion planning and control.

Original authors: Felipe Bartelt, Ali Umut Kaypak, Anthony Tzes, Farshad Khorrami, Luciano C. A. Pimenta, Vinicius M. Gonçalves

Published 2026-08-11

📖 8 min read🧠 Deep dive

CC BY 4.0

Original authors: Felipe Bartelt, Ali Umut Kaypak, Anthony Tzes, Farshad Khorrami, Luciano C. A. Pimenta, Vinicius M. Gonçalves

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot how to move through a crowded room without bumping into anything. To do this safely, the robot's brain needs to constantly ask a very specific question: "How close am I to that chair?" In the world of robotics, this isn't just about knowing the distance; it's about knowing the direction of that distance and how fast it's changing. If the robot is getting too close, it needs to know exactly how to turn away, instantly and smoothly. This is where a mathematical tool called a "Signed Distance Function" comes in. Think of it as a magical ruler that doesn't just tell you the distance to an object, but also whispers, "You are safe," or "You are touching," or even "You are already inside!" The problem is that the traditional way of calculating this ruler has a glitch: when the robot is in a tricky spot, like exactly between two walls, the ruler suddenly gets confused and stops giving smooth instructions. It's like a GPS that suddenly jumps from "Turn left" to "Turn right" without a smooth transition, causing the robot to shake, jitter, or crash.

This paper introduces a brand-new, super-smooth version of that magical ruler, designed specifically for robots made of flat, blocky shapes (like boxes or pyramids). The authors, a team of researchers, realized that the old ruler breaks because it relies on a "minimum" operator—a mathematical way of picking the smallest number—which acts like a sharp corner in a graph. When the robot moves past that corner, the math gets stuck. To fix this, they invented a new kind of "soft" minimum and maximum, which they call the Hölder Minimum and Hölder Maximum. Imagine these as a pair of magical gloves that can gently round off any sharp corners in the math, making the ruler's instructions flow like water instead of snapping like a twig. They tested this new ruler on a real robot arm, and it worked beautifully, letting the robot glide through tight spaces without the annoying shaking that happens with the old method.

The Problem: The Robot's Shaky Ruler

Robots need to know exactly how far they are from obstacles to avoid crashing. In the past, engineers used a standard method called the Euclidean Signed Distance Function (SDF). Think of this as a ruler that measures the shortest line between a robot and an obstacle. If the number is positive, the robot is safe. If it's zero, they are touching. If it's negative, the robot is already inside the obstacle (a crash!).

However, this standard ruler has a major flaw: it isn't smooth everywhere. Imagine a robot moving between two square boxes. As it slides past the corner of one box, the "closest point" on the box suddenly jumps from one edge to another. In math terms, this creates a sharp "kink" or a corner in the data. When a robot's control system tries to use this kinked data to decide how to move, it gets confused. It might suddenly jerk the robot's arm left, then right, then left again, causing a violent shaking known as "chattering." This is dangerous and makes precise tasks, like inserting a peg into a hole, nearly impossible.

The Solution: Softening the Sharp Edges

The authors of this paper decided to fix the ruler by replacing the sharp corners with smooth curves. They started by looking at the math behind the ruler, which relies heavily on finding the "minimum" (the closest point) and the "maximum" (the furthest separating line). The problem is that the standard "min" and "max" functions are like sharp cliffs; if you walk up to them, you hit a wall.

To solve this, the team invented two new mathematical tools: the Hölder Minimum and the Hölder Maximum.

The Hölder Minimum is a "soft" version of the minimum. Instead of picking the single smallest number and stopping, it gently blends the numbers together.
The Hölder Maximum does the same for the largest numbers.

These new tools are "differentiable," which is a fancy math way of saying they are perfectly smooth. You can slide your finger along their graph without ever hitting a bump. The authors also added a special "shaping function," which acts like a final polish, ensuring that even when the distance is exactly zero (touching), the math remains smooth and doesn't break.

How It Works: The Magic of Polyhedra

The paper focuses on robots and obstacles that are shaped like convex polyhedra—think of them as shapes made of flat faces, like a cube, a pyramid, or a dodecahedron. The new ruler, which the authors call the Hölder Differentiable Signed Distance Function (HD-SDF), works in three clever steps:

Checking the Directions: Instead of checking every single possible direction to find the closest point (which would take forever), the HD-SDF checks a specific, finite list of directions based on the faces and edges of the shapes.
Smoothing the Math: It uses the new Hölder Minimum and Maximum to find the closest and furthest points along those directions, but it does so without creating sharp corners.
Polishing the Result: It applies the shaping function one last time to ensure the final distance number is smooth no matter what.

The result is a distance calculator that is signed (it knows if you are inside or outside), differentiable (it never jitters), and closed-form (it gives the answer directly without needing to guess and check over and over again).

The Proof: Speed and Smoothness

The researchers didn't just draw pretty pictures; they put their new ruler to the test.

Speed Test:
They compared their HD-SDF against the old standard Euclidean SDF and another smooth method called GDF.

On a standard computer processor (CPU), the HD-SDF was slower than the old methods, taking about 5 milliseconds for complex shapes. This is because it checks many directions at once.
However, because their method is built to run many calculations at the same time, they tested it on a Graphics Processing Unit (GPU). When running on a GPU, the speed improved dramatically. For simple shapes, it was up to 13 times faster than the CPU version, and for complex shapes, it became competitive with the fastest existing methods. This means that with the right hardware, this smooth ruler can be just as fast as the old, shaky ones.

The "False Alarm" Test:
Because the HD-SDF is designed to be smooth, it is slightly "conservative." This means it might say "You are touching" (distance ≤ 0) even when there is still a tiny gap (distance > 0). The authors wanted to know how often this happens.

They ran 500,000 random tests with different shapes.
They found that if the gap between objects is larger than 10 centimeters, the new ruler almost never makes a mistake.
Even if the gap is very small, down to 100 micrometers (which is thinner than a human hair), the "false alarm" rate is only about 0.7%.
The authors note that this is a very small price to pay for a ruler that never shakes, and the error can be made even smaller by checking more directions.

The Robot Experiment:
The ultimate test was a real robot arm, a Franka Emika Panda, trying to squeeze through a narrow corridor made of boxes.

The Old Way: When using the standard Euclidean distance, the robot's control system got confused by the sharp corners. The robot's joints started shaking violently, with the speed of the last joint jumping back and forth rapidly. It looked like the robot was having a seizure.
The New Way: When the researchers switched to the HD-SDF, the shaking stopped completely. The robot moved smoothly and steadily through the corridor, reaching its target without a single jitter. The control signals were continuous and calm, proving that the smooth math translates directly to smooth motion.

Why This Matters

This paper shows that we don't have to choose between a ruler that is fast and one that is smooth. By using the Hölder Minimum and Maximum, the authors created a tool that is mathematically perfect for modern robotics. It allows robots to perform delicate tasks, like assembly or insertion, where even a tiny vibration can cause a failure. The method is flexible (you can choose how smooth you want it to be), it works on standard computer chips but shines on powerful GPUs, and it keeps robots safe and steady. While the paper focuses on blocky, convex shapes, the success of this approach suggests a bright future for robots that can move with the grace of a dancer, even in the tightest of spaces.

Technical Summary: Hölder Signed Distance

1. Problem Statement

In robotic motion planning and control, particularly for safety-critical applications like Control Barrier Functions (CBFs), it is essential to compute distances between sets (e.g., robot links and obstacles) that are not only accurate but also differentiable. The standard Euclidean Signed Distance Function (SDF) is not differentiable everywhere; non-differentiability occurs when the "witness points" (the pair of points on two objects attaining the minimum distance) are not unique, such as when faces are parallel or edges align.

Existing alternatives attempt to solve this but suffer from significant trade-offs:

Iterative methods: Some differentiable metrics rely on iterative algorithms (e.g., alternating projections), which introduce convergence issues and computational overhead.
Unsigned metrics: Many differentiable alternatives are unsigned (non-negative only), failing to distinguish between separation and penetration depth, which is crucial for CBFs and planning in collision scenarios.
Approximations: Other approaches use discretized fields or surrogate potentials, which can compromise geometric fidelity near boundaries.

The paper identifies a gap for a differentiable, signed, non-iterative, and parallelizable distance metric between convex polyhedra.

2. Methodology

The proposed solution, the Hölder Differentiable Signed Distance Function (HD-SDF), constructs a smooth distance by modifying the classical min-max formulation of the Euclidean SDF. The methodology proceeds in three main stages:

A. Hölder Minimum and Maximum Operators

The core innovation is the introduction of Hölder Minimum ( $HMin_\gamma$ ) and Hölder Maximum ( $HMax_\gamma$ ) operators.

These operators generalize the traditional $\min$ and $\max$ functions.
They are defined using Hölder norms and are differentiable everywhere except on a zero-measure set (specifically, when the output is zero).
They preserve key algebraic properties of the standard operators, including associativity, commutativity, and sign preservation.
A shaping function ( $\Phi_{\gamma, \epsilon}$ ) is applied to the output of these operators. This function is designed to be flat near zero (eliminating the remaining non-differentiability) while asymptotically approaching the identity function for large values, thereby preserving the original distance shape.

B. Discretization of the Separating Axis Theorem (SAT)

To compute the distance between two convex polyhedra $A$ and $B$ , the authors adapt the Separating Axis Theorem.

Instead of maximizing over the continuous unit sphere, the method maximizes over a finite set of candidate normals $\tilde{N}(A, B)$ .
This set includes face normals and edge direction vectors of both polyhedra.
The authors replace the cross-product normalization (which causes discontinuities when edges become parallel) with the raw edge direction vectors to ensure continuity.
The inner minimization over the polyhedra is restricted to their vertex sets $V(A)$ and $V(B)$ , leveraging the Fundamental Theorem of Linear Programming.

C. Construction of HD-SDF

The final metric $D_{\gamma, \epsilon}(A, B)$ is constructed by:

Replacing the standard $\min$ and $\max$ in the SAT formulation with $HMin_\gamma$ and $HMax_\gamma$ .
Applying the shaping function $\Phi_{\gamma, \epsilon}$ to both the inner (vertex) and outer (normal) operations.
The parameter $\gamma$ controls the order of differentiability, and $\epsilon$ controls the trade-off between smoothness and fidelity.

3. Key Contributions

The paper claims the following specific contributions:

Differentiable Operators: The definition of $HMin_\gamma$ and $HMax_\gamma$ , which are differentiable almost everywhere and can be made fully differentiable via a shaping function, unlike traditional min/max.
Closed-Form Signed Metric: The construction of the HD-SDF, a signed distance function between convex polyhedra that is $\gamma$ -times differentiable everywhere. Unlike prior works (e.g., [7], [8]), this metric is signed (providing penetration depth) and non-iterative (computed in closed form).
Parallelizability: The formulation relies on independent dot products and associative reductions, making it naturally amenable to GPU parallelization, unlike sequential iterative projection methods.
Conservative Safety Guarantee: The metric is proven to be conservative: a positive HD-SDF guarantees strict separation. While it may occasionally yield a non-positive value when objects are actually separated (false positive), this occurs rarely and only near contact.

4. Experimental Results

The authors validate the HD-SDF through runtime comparisons and a robotic manipulation experiment.

Computational Performance

Complexity: The evaluation complexity is $O(|V(A)|^2|V(B)| + |V(A)||V(B)|^2)$ , which is cubic in the number of vertices.
CPU vs. GPU: On a CPU, the method is slower than projection-based exact SDFs (e.g., FCL) and unsigned differentiable metrics (GDF), with a worst-case slowdown of roughly 500x compared to FCL. However, it remains practical.
Acceleration: With GPU acceleration, the HD-SDF achieves significant speedups (up to 13x over CPU for simple shapes) and approaches the speed of exact SDFs for complex cases due to massive parallelization.

Conservativeness

In tests with 500,000 random configurations of Platonic solids, the false-positive rate (indicating collision when none exists) was found to be negligible.
For a separation tolerance of 10 cm, the false-positive rate was below 0.2%. Even at a strict tolerance of 100 $\mu$ m, the rate remained around 0.7%.

Robotic Experiment

Setup: A Franka Emika Panda robot was tasked with navigating a corridor of boxes using a CBF-based controller.
Comparison: The experiment compared the HD-SDF against the standard Euclidean distance.
Outcome: The Euclidean distance caused "chattering" in control inputs due to non-differentiable jumps in witness points when faces were parallel. In contrast, the HD-SDF produced smooth, continuous control signals, allowing the robot to complete the task without vibration. The experiment also demonstrated the metric's utility in the negative (penetration) region, where it remained smooth, facilitating tasks like insertion or assembly.

5. Significance and Claims

The paper positions the HD-SDF as a practical tool for robotics that bridges the gap between geometric accuracy and mathematical smoothness required for optimization-based control.

Differentiability: It solves the non-differentiability problem of the Euclidean SDF without resorting to iterative solvers or sacrificing sign information.
Control Safety: By providing a smooth, signed gradient, it enables the robust application of Control Barrier Functions (CBFs), preventing the large or noisy control inputs that arise from non-differentiable distance functions.
Scalability: Its closed-form nature and compatibility with GPU parallelization make it suitable for real-time applications involving multiple object pairs.
Limitations: The authors acknowledge the metric is "conservative" (can signal false collisions) and currently applies only to convex polyhedra, noting that extension to non-convex shapes is a direction for future work.