💻 computer science

SoFFT: Spatial Fourier Transform for Modeling Continuum Soft Robots

This paper proposes SoFFT, a novel modeling framework that applies the spatial Fourier transform to Cosserat Rod Theory to represent continuum soft robot backbones as signals, thereby unifying existing strategies, enabling data-driven deformation capture, and reducing degrees of freedom while maintaining accuracy.

Original authors: Daniele Caradonna, Diego Bianchi, Franco Angelini, Egidio Falotico

Published 2026-05-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Daniele Caradonna, Diego Bianchi, Franco Angelini, Egidio Falotico

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

The Big Picture: Taming the "Infinite" Robot

Imagine a soft robot made of silicone or rubber. Unlike a traditional robot arm made of rigid metal joints, this robot is like a piece of spaghetti or a human arm: it can bend, twist, and stretch in infinite ways.

The problem? If a robot has infinite ways to move, it is incredibly hard to write a computer program to control it. It's like trying to write a recipe for a soup where you have to list every single water molecule's movement.

For years, scientists have tried to solve this by chopping the robot into small, rigid "segments" (like cutting a snake into pieces) and pretending each piece is a simple curve. This works, but it's a bit of a hack.

The Paper's New Idea:
Instead of chopping the robot, the authors suggest treating the robot's shape as a musical signal. Just as a complex sound wave can be broken down into simple notes (frequencies), the robot's complex bending shape can be broken down into simple "spatial waves."

They call this method SoFFT (Spatial Fourier Transform).

The Core Concept: The Robot as a Song

Think of the robot's backbone (its spine) as a long string.

The Old Way: You look at the string and say, "This part is bent like a circle, this part is straight, this part is bent again." You describe it piece by piece.
The SoFFT Way: You look at the string and ask, "What musical notes make up this shape?"

In this analogy:

Low Notes (Low Frequencies): These are the big, slow curves of the robot (like a gentle "C" shape).
High Notes (High Frequencies): These are the tiny, wiggly details (like a sharp kink or a small ripple).

The authors realized that most soft robots don't actually need "high notes" to look right. They mostly move in big, smooth curves. By using a mathematical tool called the Fourier Transform, they can listen to the robot's shape, identify which "notes" (waves) are loud and important, and ignore the quiet, tiny ones.

The "Data-Driven" Magic: Listening to the Robot

One of the coolest parts of this paper is how they figure out which notes to keep. They don't just guess; they let the robot tell them.

The "Motor Babbling" Phase: They make the robot wiggle around randomly (like a baby babbling) using standard signals.
The "Listening" Phase: They use sensors to record how the robot moves.
The "Spectrum" Analysis: They run the data through a Fast Fourier Transform (FFT). This is like looking at a music equalizer on a stereo. It shows them exactly which "spatial frequencies" (shapes) the robot actually uses.

The Result: They can see that 95% of the robot's movement is made up of just three or four simple waves. This allows them to build a model that is much simpler (fewer calculations) but just as accurate as the complex, infinite models.

Why This Matters (According to the Paper)

The paper validates this idea in two ways:

Computer Simulations: They tested it on a virtual robot (called "H-Support") and found that their method could predict the robot's shape accurately while using fewer "degrees of freedom" (simpler math).
Real-World Experiments: They built a real robot with pneumatic (air) muscles and tendon cables. They hit it with a stick (an external force) and watched it bounce. The SoFFT method successfully captured how the robot reacted to the hit, identifying new "wiggles" (high-frequency waves) that appeared only when the robot was hit.

The "Aha!" Moments from the Research

Unifying the Old Methods: The authors show that the old "chop-it-into-segments" methods are actually just a clumsy way of trying to do what the Fourier Transform does naturally. Their method explains why those old methods work and tells you exactly how many segments you need.
The "Aliasing" Warning: If you try to measure a very wiggly robot with sensors that are too far apart, you get a "glitch" (aliasing), just like when a spinning wheel looks like it's moving backward in a movie. The paper uses math to tell you exactly how close your sensors need to be to avoid this.
Noise vs. Signal: When they tested the real robot, they found that friction and the way the robot was built created "noise" (unwanted high-frequency wiggles). The Fourier method helped them separate the "real" movement from the "noise," allowing them to build a cleaner model.

Summary

In short, this paper proposes a new way to model soft robots. Instead of treating them as a collection of rigid blocks, it treats them as a sound wave. By using the Fourier Transform, the authors can:

Simplify the math needed to control the robot.
Automatically figure out the best way to model the robot based on real data.
Understand exactly how the robot reacts to pushes, pulls, and its own internal motors.

It turns the "infinite" complexity of a soft robot into a manageable, finite set of "notes" that a computer can easily understand and control.

Technical Summary: SoFFT: Spatial Fourier Transform for Modeling Continuum Soft Robots

Problem Statement
Continuum Soft Robots (CSRs), constructed from flexible materials, possess theoretically infinite Degrees of Freedom (DoFs), offering high adaptability in unstructured environments. However, this infinite dimensionality complicates controller design and simulation. While Cosserat Rod Theory (CRT) provides an effective framework for modeling CSRs as time-varying curves (backbones) via Partial Differential Equations (PDEs), practical implementation requires spatial discretization to reduce DoFs. Existing methods, such as Piecewise Constant Curvature (PCC), Piecewise Constant Strain (PCS), and Geometrically Variable Strain (GVS), rely on heuristic choices regarding the number of segments or the selection of functional bases (e.g., polynomials, trigonometric functions). These approaches often lack a unified theoretical justification for selecting the optimal basis or determining the minimum number of segments required to balance accuracy and computational efficiency. Furthermore, data-driven reduction methods like Proper Orthogonal Decomposition (POD) often couple all strain modes, potentially obscuring mode-specific dynamics and scaling issues.

Methodology
The authors propose "SoFFT," a framework that treats the CSR backbone as a spatial signal and applies the Fourier Transform (FT) to analyze its deformation. The core methodology involves:

Spatial Signal Formulation: The strain field $\xi(s, t)$ is treated as a signal defined over space $s \in [0, L]$ . The Continuous Spatial Fourier Transform (SFT) is defined to map the strain field into the spatial frequency domain (wavenumber $k$ ).
Frequency Domain Analysis: The paper reinterprets the CRT statics and dynamics equations within the spatial frequency domain. This reveals that the stiffness matrix acts as a convolutional filter on the strain field. The authors demonstrate that existing discretization techniques (e.g., PCS, Piecewise Affine Strain) can be mathematically interpreted as specific signal reconstructors (e.g., Zero-Order Hold, First-Order Hold) in the frequency domain.
Data-Driven Spectrum Extraction: A novel methodology is proposed to extract the optimal modeling basis directly from experimental data without prior knowledge of physical parameters. This involves:
- Sensorization: Measuring strain samples along the robot.
- Motor Babbling: Applying standard excitation signals (step, chirp, white noise) to actuators to explore the workspace.
- Spectral Analysis: Computing the Space-Time Fourier Transform (STFT) to identify dominant spatial harmonics ( $k_{max}$ ) and dynamic characteristics (resonance/anti-resonance).
- Basis Selection: Using the Basis Pursuit Denoising (BPD) algorithm to select a sparse set of optimal basis functions from a predefined dictionary (polynomials, trigonometric functions) that best reconstructs the strain field while minimizing DoFs.
Truncation Criteria: An energy-based truncation index is defined to determine the minimum number of segments or basis functions required, ensuring that the retained energy meets a specific accuracy threshold.

Key Contributions

Unified Framework: The paper unifies existing CSR modeling strategies (PCC, PCS, GVS) under a single CRT framework by interpreting them as spatial signal reconstruction problems.
Theoretical Justification: It provides a theoretical foundation for heuristic methods, such as determining the minimum number of segments required for a given robot length and strain profile, based on the Nyquist-Shannon theorem applied to spatial signals.
Mode-Specific Analysis: Unlike POD, which derives a coupled basis for all modes, the proposed FT approach treats strain modes (bending, twisting, shearing, stretching) independently, avoiding scaling issues and allowing for mode-specific truncation.
Data-Driven Modeling: The introduction of a data-driven workflow that uses STFT and BPD to identify optimal functional bases and reduce model order based on experimental data, independent of the robot's physical parameters.
Validation: The approach is validated through numerical simulations (using the SoRoSim toolbox) and real-world experiments on an H-Support prototype equipped with a motion capture system.

Results

Numerical Validation: Simulations on cylindrical and conical H-Support robots demonstrated that the STFT effectively captures resonance and anti-resonance peaks. The analysis revealed that the choice of functional basis significantly influences the distribution of deformation energy across spatial harmonics. For instance, polynomial bases distribute energy inversely proportional to the harmonic order, while Fourier bases concentrate energy in specific harmonics.
Experimental Validation: Experiments on a physical prototype confirmed the method's ability to reconstruct the backbone shape with high accuracy. Using BPD with a mixed dictionary of polynomial and trigonometric bases, the model achieved a Mean Squared Error (MSE) comparable to the full-order model while reducing the DoFs (e.g., reducing from 30 to 24 DoFs with a 1% truncation threshold).
Robustness to Perturbations: The method successfully identified new spatial harmonics (in the 6–8 $m^{-1}$ range) induced by external impulsive forces, demonstrating its capability to model contact interactions and non-uniform deformations.
Error Analysis: Position and orientation errors were found to increase along the robot's length due to error propagation in forward kinematics, but remained within acceptable limits (e.g., < 5% of robot length for position error with 5% truncation).

Significance and Claims
The paper claims that the SoFFT framework offers a compact and efficient representation of CSR deformation by leveraging the spectral properties of the strain field. It asserts that this approach:

Unifies disparate modeling strategies under a common mathematical language.
Provides a rigorous, data-driven alternative to heuristic model selection, balancing accuracy and computational efficiency.
Facilitates the identification of optimal basis functions that capture the specific dynamic and static behaviors of the robot, including the effects of actuator dynamics and external forces.
Does not replace the CRT PDEs but rather exploits spectral information to optimize the basis within the GVS framework.

The authors conclude that while the current implementation is primarily offline, the computational efficiency of the Fast Fourier Transform (FFT) suggests potential for future real-time applications, such as model-based or learning-based control and real-time adaptation to changing environmental conditions.