Learning the Dynamics of Deformable Linear Objects Through Coupled Oscillators for Bimanual Robotic Manipulation

Iván López-Broceño, Maximilian Stölzle, Daniel Feliu-Talegon, Daniela Rus, Cosimo Della SantinaarXiv 2026 · arXiv

TL;DR

The paper introduces $\Phi$-CON, a physics-informed neural network that learns the complex, force-driven dynamics of stiff and heavy cables. By combining a stable latent oscillator network with a learned coordinate mapping, it enables high-precision bimanual robotic manipulation with 90% fewer parameters than standard models and explicit closed-loop stability guarantees.

What Problem Does This Paper Address?

Manipulating Deformable Linear Objects (DLOs)—like industrial cables, wires, or tree branches—is a fundamental challenge in robotics. Most existing approaches rely on simplified "quasi-static" models that assume the object is always at rest, or purely kinematic descriptions that ignore the object's mass and stiffness.

While these simplifications work for lightweight, floppy strings moved slowly, they fail for the stiff and heavy objects common in industrial settings. For these objects, the physical response (inertia, elasticity, and gravity) cannot be ignored, especially during rapid motions. Existing deep learning models like Lagrangian Neural Networks (LNNs) can capture these dynamics but often lack stability guarantees, are computationally heavy, or struggle with the "floating-base" nature of DLOs (where the whole object can translate in space).

Methods

The authors propose a framework called $\Phi$-CON, which stands for a learned affine diffeomorphism ($\Phi$) coupled with a Floating-Base Coupled Oscillator Network (FBCON).

  1. FBCON (The Latent Engine): This extends previous "Coupled Oscillator Networks" to systems that can move freely in space. It separates the object's base position from its relative shape. Crucially, the shape dynamics are designed to be stable by construction.
  2. $\Phi$ (The Coordinate Bridge): A learned affine mapping connects real-world marker positions to the latent space of the FBCON. This mapping is designed to be consistent with the principle of virtual work, ensuring that forces and velocities are transformed correctly while preserving the underlying physical (Lagrangian) structure.
  3. Model-Based Control: Because $\Phi$-CON provides explicit expressions for kinetic and potential energy, the authors can design a Proportional-Derivative (PD) controller that "knows" the cable's physics. It uses the learned model to provide feedforward compensation for the cable's internal elastic and gravitational forces.
Coordinate transformations of $\Phi$-CON. It maps observations through a standardized floating-base representation into a latent space where shape potential is easily defined.

The training process (shown in Figure 2) involves rolling out the dynamics in latent space and comparing the results to ground-truth data in both latent and observation spaces, alongside a specific loss term for "holding forces" at steady state.

The training pipeline for $\Phi$-CON, showing how observations are mapped to latent space, integrated over time, and mapped back for loss calculation.

Datasets

The researchers evaluated their method on both simulated and real-world data:

  • Simulation: Four distinct DLO variants created in the MuJoCo physics engine: Baseline, Heavy (increased density), Low Damping, and Compliant (lower stiffness).
  • Real-World: A stiff electrical cable manipulated by two Franka Emika Panda robotic arms. The setup used motion capture (12 markers) and force-torque sensors at the robot end-effectors.

Key Insights and Ideas

  • Physics-Informed Sparsity: By baking physical laws (like the Lagrangian structure) directly into the network architecture, $\Phi$-CON achieves state-of-the-art accuracy with a fraction of the parameters. It requires 97.2% fewer parameters than Neural ODEs and 90.2% fewer than LNNs.
  • Stability by Design: Most neural networks are "black boxes" that could theoretically command a robot to do something unstable. $\Phi$-CON’s latent dynamics are provably stable, and the authors provide a mathematical proof that this stability transfers to the real-world observation space.
  • The Importance of Force Compensation: The results highlight that for stiff objects, simply knowing where the cable is isn't enough. You must account for the forces it exerts. The feedforward term in their controller is what allows the robots to overcome the cable's resistance and reach precise target shapes.

Key Results

  • Prediction Accuracy: In simulation, $\Phi$-CON achieved the lowest mean position error for 2-second future predictions across all four DLO types, significantly outperforming LNNs which often diverged over longer horizons.
Table 1: Position RMSE comparison across different models and DLO types.
  • Control Performance: Compared to a standard model-free PD controller, the $\Phi$-CON-based controller reduced endpoint regulation error by a staggering 97.5–99.7% in simulation.
  • Real-World Success: On the actual robotic setup, the method achieved a 27.9% mean reduction in endpoint error compared to the robots' built-in impedance controllers, proving its utility in noisy, real-world environments.
Table 2: Regulation error and computational cost. Note that $\Phi$-CON is significantly faster than LNN on both CPU and GPU.

Limitations and Future Work

  • State Representation: The model assumes that marker positions and velocities are enough to describe the state. It does not currently account for "material memory" effects like hysteresis or plastic deformation (where a cable stays bent after being moved).
  • Environment Interaction: The current work focuses on free-space manipulation. It does not yet handle collisions with obstacles or the cable touching itself (self-contact).
  • Grasp Constraints: The experiments were conducted with a fixed grasp orientation. Future iterations could extend this to allow the robots to re-orient their "hands" during manipulation.

Summary updated

Abstract

Many learning-based approaches for deformable linear objects (DLOs) rely on simplified deformation models that neglect dynamic effects and the mechanical response of the objects, limiting their applicability to lightweight, compliant objects and slow motions. Inspired by recent advances in learning physically structured dynamics with coupled oscillator networks (CONs), we propose $Φ$-CON for learning force-driven DLO dynamics. The model combines a floating-base coupled oscillator network (FBCON), whose latent potential energy depends only on the DLO's shape, with a learned affine diffeomorphism $Φ$ between observed positions along the DLO and latent coordinates. Its shape dynamics are stable by construction, and its Lagrangian structure yields explicit expressions for kinetic and potential energy. We exploit this structure to design a model-based proportional-derivative (PD) controller that compensates for the learned potential forces, thereby establishing closed-loop stability guarantees and ensuring convergence. We evaluate the approach on four simulated DLOs and a real bimanual robot setup. In simulation, $Φ$-CON achieves prediction accuracy comparable to neural ordinary differential equation (NODE) models and Lagrangian neural networks (LNNs), with $97.2\%$ and $90.2\%$ fewer parameters, respectively. Relative to a model-free PD baseline, the proposed controller reduces mean endpoint regulation error by $97.5$-$99.7\%$ in simulation and achieves a mean per-target reduction of $27.9\%$ across both endpoints in real-world experiments, demonstrating the importance of explicitly accounting for the mechanical response of stiff and heavy objects.
BibTeX
@misc{lopezbroceno2026learning,
  author = {Iván López-Broceño and Maximilian Stölzle and Daniel Feliu-Talegon and Daniela Rus and Cosimo Della Santina},
  title = {{Learning the Dynamics of Deformable Linear Objects Through Coupled Oscillators for Bimanual Robotic Manipulation}},
  year = {2026},
  eprint = {2611.23563},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.23563},
}
Figure