← Paper Digest

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 Santina

November 2026

TL;DR Φ-CON learns the force-driven dynamics of a stiff cable with a coupled oscillator network that is stable by construction, then uses its learned potential forces as feedforward in a PD controller. On a real two-arm setup this cuts steady-state endpoint error by 27.9% per target on average compared with PD alone.

What they do

Stiff, heavy cables push back on the robots holding them, so quasi-static or kinematic models with position-command inputs fall short. Φ-CON learns the dynamics of a deformable linear object (DLO) from marker positions and endpoint forces. Its structure is stable by construction and gives explicit kinetic and potential energies. The learned potential then drives a closed-form bimanual controller that needs no runtime optimization.

How it works

  • Floating-base CON (FBCON). Extends coupled oscillator networks so the latent potential depends only on shape. The cable can translate freely while its shape dynamics stay globally asymptotically stable.
  • Affine diffeomorphism Φ. Markers are split into a base and a relative shape, standardized, then passed through a learned invertible affine map. Its Jacobian also transforms velocities and, by virtual work, forces, so the Lagrangian structure carries over to observation space.
  • Training. Rollout losses in latent and observation space, plus a loss that fits the learned holding forces to measured forces at rest.
  • Control. PD feedback plus feedforward of the learned potential forces at the target. The closed loop is provably stable for feasible equilibria of the learned model.

Results

  • Simulation (four 1 m MuJoCo rods): lowest 2 s position RMSE on all four, e.g. 4.5 mm vs 11.0 mm (NODE) and 12.9 mm (LNN) on the compliant rod. It uses 20,031 parameters vs 710,234 (NODE) and 203,584 (LNN).
  • Simulated regulation: feedforward reduces endpoint error by 97.5-99.7% across the four rods compared with PD alone (238.6 mm to 1.5 mm on the baseline rod), comparable to LNN feedforward.
  • Real robots (two Franka Pandas, 0.76 m electrical cable, 12 mocap markers): on 76 steady-state targets taken from training and validation trajectories, endpoint error falls from 63.7 to 48.4 mm (left) and 48.3 to 32.3 mm (right), a 27.9% mean reduction per target. On four unseen targets, transients are faster and shape error is matched or reduced.
  • On real rollouts, NODE is more accurate within the 2 s training horizon; Φ-CON matches it beyond.

Limitations

Assumes marker positions and velocities capture the dynamics. Regulation only to statically feasible shapes with a fixed grasp orientation, with no obstacle contact or self-contact. Trajectory tracking and material memory such as hysteresis are left to future work.

Videos

Full video (YouTube)

Figures

From Figure 1: Coordinate transformations of Φ-CON. Fixed maps split the marker positions into a base and a shape rel...
From Figure 1: Coordinate transformations of Φ-CON. Fixed maps split the marker positions into a base and a shape relative to it and standardize them. A learned affine map acting only on the shape leads to latent space, where the potential is flat along the base.
From Figure 2: Training of Φ-CON. Φ maps observations and inputs to latent space, where the FBCON is rolled out for H...
From Figure 2: Training of Φ-CON. Φ maps observations and inputs to latent space, where the FBCON is rolled out for H steps and mapped back by the inverse of Φ. Losses compare the rollout with the data in each space and fit the learned holding forces to the measured forces at rest.
From Figure 3: Shape regulation in simulation, from a straight rest configuration to the target shape (dashed line). ...
From Figure 3: Shape regulation in simulation, from a straight rest configuration to the target shape (dashed line). Left: shape snapshots, with trails showing the preceding 0.6 s of motion. Right: shape error over time for PD only and PD with LNN or Φ-CON feedforward.
From Figure 5: Shape regulation in experiments. Two robots regulate the real cable to successive unseen target shapes...
From Figure 5: Shape regulation in experiments. Two robots regulate the real cable to successive unseen target shapes with PD plus Φ-CON feedforward; photographs at five instants, with targets in red.
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.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.