WorksComputer Science
Learning the Dynamics of Deformable Linear Objects Through Coupled Oscillators for Bimanual Robotic Manipulation
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 -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 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, -CON achieves prediction accuracy comparable to neural ordinary differential equation (NODE) models and Lagrangian neural networks (LNNs), with and fewer parameters, respectively. Relative to a model-free PD baseline, the proposed controller reduces mean endpoint regulation error by - in simulation and achieves a mean per-target reduction of across both endpoints in real-world experiments, demonstrating the importance of explicitly accounting for the mechanical response of stiff and heavy objects.
Introduction
As robots are increasingly adopted across a wide range of activities, they are required to interact with a more diverse set of objects. Manipulating deformable linear objects (DLOs) such as cables, wires, or tree branches is key to applications like electrical installation, cable routing, and automated harvesting, yet it remains challenging [3, 7] (Zhu et al. 2022): DLOs are high-dimensional, underactuated during manipulation, and often stiff and heavy enough that their physical response cannot be neglected, particularly during rapid motions. A main body of existing work relies on simplified deformation models of the DLO, namely quasi-static assumptions (Bretl and McCarthy, [5]) and kinematic descriptions, usually through Jacobian-based deformation estimation (Yu et al., [28]) or visual servoing (Shetab-Bushehri et al., [23]; Cao et al., [6]), instead of its full dynamics. Such simplifications are hard to justify for stiff and heavy DLOs, whose physical response cannot be neglected, especially under rapid motions.
Recent approaches to DLO manipulation learn motion models from data using LSTMs (Yang et al., [27]), GNNs (Wang et al., [26]; Chu et al., [9]), or recurrent state-space models (RSSMs) (Missal et al., [18]) to predict shape changes from past states and control inputs. Such architectures can require millions of learned parameters (Missal et al., [18]). Using position or velocity commands as inputs assumes that the manipulator can impose the commanded endpoint motion regardless of DLO reaction forces, limiting applicability to lightweight, compliant DLOs. Moreover, deploying these learned models for control typically requires numerical optimization at runtime and yields no analytical closed-loop guarantees.
Among learning methods that preserve physical structure (Stölzle, [24]), Lagrangian neural networks (LNNs) (Cranmer et al., [10]) have been applied to modeling the dynamics of soft robots (Liu et al., [17]) and deformable objects (Zhong et al., 2021; Friedl et al., [14]; Pierallini et al., [20]). However, evaluating their forward dynamics requires runtime differentiation of the learned Lagrangian, including second-order derivatives, which incurs substantial computational cost (Cranmer et al., [10]). Moreover, Lagrangian structure alone does not guarantee stability of the learned dynamics (Müller, [19]) or the resulting closed loop (Liu et al., [17]). Coupled oscillator networks (CONs) combine Lagrangian structure with provably stable latent dynamics and explicit expressions for model-based control (Stölzle and Della Santina, [25]). However, their strictly convex latent potential does not accommodate the translational freedom of floating-base DLOs, while their separately learned observation and input mappings do not enforce consistency between physical and latent forces under the principle of virtual work.
To address these limitations, we propose -CON for learning force-driven DLO dynamics and enabling closed-form control of bimanual manipulation. We extend CONs to floating-base systems through a floating-base coupled oscillator network (FBCON), whose latent potential depends only on shape, accommodating base translation while retaining stable shape dynamics. A learned affine diffeomorphism connects observed marker positions to latent coordinates and determines the corresponding force transformation through the principle of virtual work. The resulting Lagrangian structure enables us to explicitly account for the object’s physical response and design a PD model-based controller using the learned potential forces, without requiring runtime optimization.
Our contributions are:
Closed-form, force-based bimanual regulation of DLOs through learned potential-force compensation and PD feedback, with closed-loop stability guarantees in observation space for compatible target equilibria of the learned model.
FBCON: an extension of CONs to floating-base systems with explicit kinetic and potential energies and open-loop stability guarantees for the shape dynamics.
-CON: joint learning of FBCON dynamics and an affine diffeomorphism that consistently transforms coordinates and forces, preserves Lagrangian structure, and transfers shape-stability guarantees to observation space.
We validate model learning and control using four simulated DLOs and a real bimanual robotic setup involving the manipulation of a heavy and stiff electrical cable, for which compensating for its physical dynamics is essential.
Problem Statement
Our goal is to learn the dynamics of a DLO from data to enable model-based regulation to a desired shape. We represent its shape by marker positions , , with actuation only at the two endpoints (Figure 6 in the appendix). Both endpoints are held at a constant orientation. We order the coordinates as , where stacks the actuated endpoints and stacks the unactuated interior markers. The Cartesian endpoint forces induce generalized forces , with .

Figure 6. A DLO represented by 3D marker positions and actuated by Cartesian forces at both endpoints.
Assumption 1 (Sufficient Observation of DLO Dynamics). Over the operating regime considered, the marker positions are sufficient generalized coordinates for the DLO, and is a complete dynamical state. The acceleration is uniquely determined by through the Lagrangian dynamics
where is the mass matrix, is the Coriolis matrix, a dissipation term, and the gradient of the potential energy collecting both elastic and gravitational contributions.
Physically, Assumption 1 means that the motion is dominated by deformation of the slender backbone, with negligible cross-sectional deformation, as in standard rod models [1]. It further assumes that independent rotational dynamics, such as twist not captured by the backbone shape, and deformation modes not resolved by the markers have negligible influence on the observed motion. Under this assumption, we seek to learn the unknown dynamics from sufficiently informative trajectory data .
Then, given a desired equilibrium shape supportable by endpoint forces, we seek a model-based feedback law such that
We assume that is a locally unique equilibrium of the unactuated coordinates when the endpoints are fixed at . For regulation near this equilibrium, we further assume that implies [21, 4].
Dynamical Model Learning with Phi-CON
-CON learns DLO dynamics by combining an invertible coordinate mapping with an FBCON. We first separate base position from relative shape, then apply a learned affine map to the shape coordinates. This representation allows the FBCON potential to depend only on shape, accommodating base translation while retaining stable shape dynamics. The mapping also determines the velocity and force transformations, preserving Lagrangian structure and transferring the stability of the relative shape to observation space. We train the mapping and latent dynamics jointly using rollout losses in both spaces and a steady-state holding force loss, which encourages accurate predictions of the holding forces used by the controller in § 4.
Translation Invariant Representation
We choose the first marker as the base, , and stack the remaining marker positions relative to it as , following the ordering of in § 2. Together, , where is a fixed invertible matrix (Figure 1). A common translation changes only , leaving the relative shape unchanged. We then standardize as using the training mean and a fixed diagonal scaling matrix with strictly positive diagonal entries, preserving invertibility and the separation of base and shape.
Figure 1. Coordinate transformations of -CON. In observation space (steps 1–3), fixed maps express the marker positions as a base and a shape relative to it, and standardize them. A learned affine map acting only on the shape leads to latent space (steps 4–5), where the potential is flat along the base, with equilibria at any base position.
Affine Diffeomorphism
To complete the mapping into latent space, we apply a learned affine transformation to the relative coordinates while retaining the standardized base:
Here, and are learned, giving and . We parameterize using lower and upper triangular factors with nonzero diagonal entries enforced during training. Together with the invertibility of and , this makes an affine diffeomorphism with an exact inverse.
Its Jacobian is independent of and determines the velocity and force transformations:
The inverse-transpose factor follows from the principle of virtual work. The mean endpoint force and the positive scale are computed from the training data and held fixed, with chosen so that the centered endpoint forces, expressed in the standardized coordinates , have unit root-mean-square value. The mean force estimates the static load, with , where is the total mass of the DLO and is the gravitational acceleration vector.
Latent Dynamics: Floating-Base Coupled Oscillator Network
We extend CON (Stölzle and Della Santina, [25]) to floating-base systems by making the latent potential depend only on relative shape:
where and are learned, constant, symmetric positive definite mass and damping matrices,2 and is the transformed input from § 3.2. We choose
with learned stiffness and bias . This gives translation-invariant dynamics with no potential restoring force along the base, while inertia and damping act on all coordinates.
Proposition 1 (Stability of Latent FBCON). For the latent dynamics in Eq. (5) with potential energy defined as in Eq. (6), let and . Then the equilibrium set is globally asymptotically stable, with arbitrary base position and unique shape satisfying .
The proof in Appendix § 14.§ 1 adapts the energy argument of Stölzle and Della Santina [25] to a potential that depends only on shape.
Observation Space Dynamics
Since is affine, . Substituting the coordinate and force transformations into Eq. (5) and premultiplying by gives
Here, and are constant positive definite matrices. The dynamics retain Lagrangian structure ([22 Prop. 2, App. A]), with kinetic energy and potential energy . Their sum is the total mechanical energy. The potential gradient represents the elastic and gravitational load that must be balanced at equilibrium.
Proposition 2 (Stability of Open-Loop -CON). Under the conditions of Proposition 1, the dynamics in Eq. (7) with (equivalently, ) have the globally asymptotically stable equilibrium set , where corresponds uniquely to under the affine shape map.
The relative shape converges uniquely, while the base converges to a position determined by the initial condition. Without endpoint forces (), the model cannot remain at rest because its estimated gravitational load is uncompensated. The proof in Appendix § 14.§ 2 transfers Proposition 1 through the invertible affine mapping.
Training Strategy
We jointly train the coordinate mapping and latent dynamics using . Figure 2 summarizes the training pipeline. Here, and measure rollout errors in latent and observation space, respectively, and is the steady-state holding force loss with weight . is the primary objective, while acts as a lightly weighted auxiliary term on the latent rollout, and $\ell_{\mathrm{ss}} supervises the static forces that rollouts barely constrain. Exact invertibility of eliminates the need for a reconstruction loss. The position and velocity terms in each rollout loss are weighted separately.
Figure 2. Training of Φ-CON. Φ maps observations and inputs from observation space to latent space, where the FBCON is rolled out for steps and mapped back by . The losses and compare the rollout with the data in each space, and fits the holding forces to at rest.
For each training trajectory window , we initialize and . We then predict recursively via for , where integrates Eq. (5) using the transformed input .
The latent rollout loss compares the encoded ground-truth states with the latent rollout:
The observation-space loss compares and with and , respectively, accounting for the effect of the inverse mapping on position and velocity prediction errors.
At equilibrium, Eq. (7) gives and hence , where denotes the potential forces at the actuated endpoints. We supervise these forces for regulation through
where contains training samples near steady state, selected using the marker velocity and acceleration thresholds in Appendix § 10.§ 3.
Closed-Loop Control with Phi-CON
Following prior approaches to model-based soft robot control ([4, 12, 24]), we combine proportional-derivative (PD) feedback with feedforward compensation of the learned potential forces:
The feedforward term compensates the learned elastic and gravitational loads at the target. The PD feedback term adds stiffness and damping in the actuated coordinates to mitigate the effects of modeling errors and external disturbances.
Substituting Eq. (9) into the learned dynamics gives
Since actuation is restricted to the endpoints, is an equilibrium of Eq. (10) if and only if , where collects the unactuated components of . This condition requires the interior coordinates to be in equilibrium when the endpoints are held at .
Assumption 2 The target satisfies .
Proposition 3 (Global Asymptotic Stability of Closed-Loop -CON). Under Assumption 2 and the matrix conditions of Proposition 1, consider Eq. (7) under the control law in Eq. (9), with gains and . Then is a globally asymptotically stable equilibrium of the closed-loop dynamics in Eq. (10).
The proof in Appendix § 14.§ 3 uses the learned model’s mechanical energy modified by the feedforward and proportional terms as a Lyapunov function. The global guarantee extends to the physical DLO if its dynamics match the learned model globally.
Verification in Simulation
Simulation Setup
We evaluate the method in MuJoCo using its discrete elastic rod plugin ([2, 16]), which excludes extension and shear. All four simulated DLOs have length m and observed markers. Relative to a baseline DLO, the heavy, low damping, and compliant variants change density, damping, and stiffness, respectively; simulator details and physical parameters are given in Appendix § 10.§ 2.
Evaluation Metrics and Baseline Methods
For simulation and real experiments, we split complete trajectories into training and validation sets before extracting rollout windows and evaluate on a separate test set (Appendix § 10.§ 1).
We evaluate dynamic prediction, potential-force prediction, and control performance. For dynamic prediction, we average the position root mean squared error (RMSE) between ground truth and rollout over test trajectories:
We compute velocity error analogously and report mean standard deviation over seeds. Following prior work on DLO dynamics ([14, 20]), we compare against neural ordinary differential equation (NODE) ([8]) and LNN ([10]). The NODE predicts from using a multilayer perceptron (MLP) without Lagrangian structure. The LNN learns and with separate networks and derives dynamics by automatic differentiation, providing for potential shaping. To assess the learned affine diffeomorphism, we also train an FBCON (§ 3.3) directly in observation space without a learned coordinate mapping.
For control, we report each endpoint’s final regulation error, , as mean standard deviation over targets, pooling both endpoints in simulation. In real experiments, we also report the per-target relative improvement . Figures also show the shape error , the mean marker distance to the target shape.
Dynamic Model Prediction of Simulated DLO
Within the 1 s training horizon, -CON achieves comparable accuracy to NODE and LNN and outperforms directly employing FBCON, supporting the benefit of the learned affine diffeomorphic mapping (Figure 7). Beyond this horizon, LNN diverges in the reported rollout comparison. LNN training also exhibits unstable runs and higher computational cost: on the same dataset and an NVIDIA RTX 3090, NODE, LNN, and -CON require 0.64, 9.35, and 1.06 seconds per epoch, respectively. Across all four simulated DLOs, -CON achieves the lowest mean position RMSE at 2 s (Table 1). Appendix § 11 illustrates all four models on a training trajectory.
| Position RMSE [mm] ↓ | NODE | LNN | -CON (ours) |
| Baseline | 14.6 0.6 | 17.6 1.3 | 14.0 0.6 |
| Heavy | 12.5 1.1 | 13.6 0.7 | 6.6 0.9 |
| Low damping | 15.4 1.6 | 24.8 1.6 | 14.8 1.0 |
| Compliant | 11.0 1.3 | 12.9 0.6 | 4.5 0.5 |
Table 1. Prediction error in simulation: Position RMSE [mm] for predictions 2 s into the future (mean std over seeds).
Figure 7. Prediction error in simulation: position (left) and velocity (right) RMSE over 2 s rollouts for NODE, LNN, FBCON, and -CON on the baseline dataset. The gray region marks the 1 s training horizon.
Control of the Simulated DLO
We evaluate a PD force controller ( N m, N s m) alone and with potential-force feedforward from an LNN or -CON (§ 4). At test-trajectory steady states, -CON achieves a median relative force-component error of 0.17%, compared with 0.13% for LNN and 20.47% for an MLP trained directly on forces (Appendix § 12). The models contain 20 031, 203 584, and 35 674 parameters, respectively.
Each experiment starts from a straight rest configuration and targets the final steady-state shape of a test trajectory. PD+FF computes feedforward from the full , whereas PD uses only the actuated coordinates . Both feedforward models reduce final endpoint errors relative to PD alone and achieve comparable errors across the four datasets (Table 2). Low gains keep the closed loop compliant ([11]). Both steady-state errors scale with the inverse of the endpoint stiffness, so the relative reduction barely depends on the gain: on the baseline DLO it is 99.4%, 99.1%, and 98.8% for , 50, and 200 N m, the gain of the real robots (Appendix § 13). Figure 3 shows an example regulation experiment.
| Endpoint error [mm] ↓ | PD | PD+FF (LNN) | PD+FF (-CON) |
| Baseline | 238.6 4.7 | 2.5 6.9 | 1.5 3.5 |
| Heavy | 459.4 14.9 | 2.3 5.5 | 2.7 5.9 |
| Low damping | 238.9 4.6 | 3.6 5.2 | 5.9 9.5 |
| Compliant | 240.5 10.1 | 0.7 1.4 | 0.8 0.6 |
| CPU time [] ↓ | 7.8 1.1 | 25.5 2.9 | 12.6 3.4 |
| GPU time [] ↓ | 47.2 8.9 | 114.6 4.1 | 72.1 8.1 |
Table 2. Regulation error and computational cost in simulation: Final endpoint regulation error [mm] across the four simulated DLO datasets (mean std over test targets and both endpoints), and computation time of one control step for a new target (mean std, Appendix § 12).
Figure 3. Shape regulation in simulation: regulation from a straight rest configuration to the target shape (dashed line). Left: shape snapshots (axes in meters), with trails showing the preceding 0.6 s of motion. Right: shape error over time.
Validation in a Laboratory Setting
Experimental Setup
Two Franka Emika Panda robots firmly grasp the endpoints of a stiff electrical cable of length m and radius mm. Force sensors measure endpoint forces , and motion capture tracks markers, including the attachment points defining . Identified material parameters are excluded from the modeling and control formulations; characterization, calibration, and preprocessing details are given in Appendix § 10.§ 5.
The robots follow the generated endpoint trajectories from Appendix § 10.§ 1 using a bimanual impedance controller ([13]) with stiffness 200 N m and damping ratio 0.77 (21.8 N s m).
Dynamic Model Prediction of Real DLO
Figure 4 compares NODE, LNN, FBCON, and -CON rollout errors on recorded test trajectories. NODE is more accurate within the 2 s training horizon, and -CON matches it beyond. LNN has the largest errors after approximately 1 s. As in simulation, the learned mapping reduces the errors of the FBCON.
Figure 4. Prediction error in experiments: position (left) and velocity (right) RMSE over 2.5 s rollouts on the test trajectories of the electrical cable for NODE, LNN, FBCON, and -CON. The gray region marks the 2 s training horizon.
Control of Real DLO
We evaluate regulation on 76 steady-state shapes extracted from the training and validation trajectories with the thresholds used for (Appendix § 10.§ 3), and on four unseen targets obtained by manually moving the robots and recording the cable at rest, which ensures that each target is a feasible equilibrium (§ 2). For each target, the feedforward in Eq. (9) is evaluated once from the recorded and added to the force command of the impedance controller of § 6.1, which provides the PD term. Across the steady-state targets, adding -CON feedforward to the PD impedance controller reduces final endpoint errors from to mm on the left and from to mm on the right. The corresponding per-target relative improvements are and , respectively.
For the four successive unseen targets in Figure 5, PD+FF produces faster transients, particularly for distant targets, and matches or reduces the PD-only steady-state shape error.
Figure 5. Shape regulation in experiments: regulation to four successive unseen target shapes (changes at dotted lines) with PD and PD+FF (-CON). Top: shape error. Middle: targets (green) and cable configurations at five instants (axes in meters). Bottom: PD+FF photographs, targets (red).
Conclusions
We presented -CON for learning force-driven DLO dynamics by linking an FBCON to observed marker positions through an affine diffeomorphism. Its Lagrangian structure enables PD control with explicit compensation of elastic and gravitational forces, without runtime optimization, and provides closed-loop stability guarantees for feasible equilibria of the learned model.
Across four simulated DLOs, learned feedforward compensation reduces mean endpoint regulation error by 97.5–99.7% relative to PD alone. On a real cable manipulated by two robots, the learned feedforward reduces the steady-state endpoint error of the robots’ impedance controllers by 27.9% per target on average, showing how a learned model can improve low-level manipulator control for DLOs whose elastic and gravitational loads otherwise cause substantial steady-state errors.
Within the simulation training horizon, -CON achieves prediction accuracy comparable to NODE and LNN, with 90.2% fewer parameters and 8.8× faster training epochs than LNN.
The approach assumes that marker positions and velocities sufficiently describe the DLO dynamics (Assumption 1) and currently addresses regulation to statically feasible shapes with a fixed grasp orientation, without obstacle interaction or self-contact. Future work will address material memory, including hysteresis and plastic deformation, through internal states and extend the controller to trajectory tracking.
Acknowledgments
This work was supported in part by the Dutch Research Foundation (NWO) through the VENI grant ROSES 20297, the European Union’s Horizon Europe Programs from Project EMERGE (Grant No. 101070918) and Project RIPLEY (Grant No. 101165078), the Singapore MIT Alliance on Research and Technology (SMART) under the Mens, Manus, et Machina (M3S) program, by the BARI EMERGE program under grant N00014-26-1-2304, and by the MIT-GIST collaboration.
We thank the company ROCSYS for sharing with us the electric cables used in the experimental validations of our method. The authors also thank Micah Prendergast for sharing the force-torque sensor needed for the validation experiments with the robots.
Declaration of AI Use
During the preparation of this work, the authors used GPT-6 Astra (OpenAI) and Claude Opus 5.5 (Anthropic) for language editing and proofreading. Additionally, the authors used Claude Opus 5.5 for code implementation, reproducibility tests, and media formatting, like figures, plots, and editing of the accompanying video. The authors reviewed the content generated by these tools and remain responsible for it.
Implementation Details
Dataset Generation
We generate endpoint trajectories over a horizon s. Each endpoint executes one linear or oscillatory motion primitive with randomly sampled class, direction, and amplitude or speed, followed by a null primitive (hold) of random duration, so that the DLO reaches a steady state. In simulation, the hold duration is sampled uniformly between s and 2 s. We enforce workspace limits and a maximum endpoint separation of , and use PyRoki [15] to discard trajectories that are kinematically infeasible for the robots in § 6. The endpoints track the remaining trajectories using a proportional force controller ( N m) in MuJoCo. Each simulation continues for 2 s after with the endpoints held, giving 12 s recordings at 1 kHz, which we subsample to 100 Hz for the dynamics models.
For simulation and real experiments, we randomly split complete trajectories into 80% training and 20% validation data before extracting rollout windows. A separate test set has the same size as the validation set. All four simulated DLOs use the same 137 endpoint trajectories for training and validation (109 and 28 per seed). Their test set contains 28 trajectories generated separately with different random seeds. The control experiments in § 5.4 target the final steady states of these 28 test trajectories, except on the heavy DLO, where two trajectories do not settle (26 targets). On the real cable, we use 166 recorded trajectories of about 10.5 s, after trimming the initial rest to 0.2 s before motion onset. We hold out 28 for testing and use the remaining 138 for training and validation (110 and 28 per seed).
Simulated DLO Parameters
The MuJoCo discrete elastic rod plugin represents each DLO as fixed-length segments connected by ball joints, excluding extension and shear. Elastic torques resist rotations from the rest configuration, with bending and twisting stiffness set by Young’s modulus , shear modulus , and the cross-sectional area moments determined by radius . Viscous joint damping has coefficient . All four simulated DLOs have length m, observed markers, and radius m. Table 3 lists the baseline parameters and the variants, which change density (heavy), damping (low damping), or stiffness (compliant); all other parameters remain at baseline values.
| DLO Dataset | [Pa] | [Pa] | [kg/m] | [N·m·s/rad] |
| Baseline | ||||
| Heavy | ||||
| Low damping | ||||
| Compliant |
Table 3. Physical parameters of the four simulated DLOs. Bold values differ from the baseline.
Training Pipeline
For each model we train five seeds, each setting the initialization and the train/validation split, and keep the checkpoint with the lowest validation loss.
For the steady-state holding force loss, contains training samples for which every marker satisfies m/s and m/s.
Hyperparameters
All models are trained with AdamW (constant learning rate, gradient-norm clipping at 1) on batches of 512 windows of 100 steps, integrated with explicit Euler at the data rate: ms (1 s windows) in simulation and ms (2 s windows) on the real cable. States and forces are standardized with training-set statistics (§ 3.1), and and are computed from the training data and kept fixed (§ 3.2). Hyperparameters are tuned on the baseline simulated DLO and reused unchanged for the other three DLOs, and tuned separately for the real cable. Table 4 lists the optimization and loss hyperparameters of all models, Table 5 the network architectures of the NODE and LNN, and Table 6 the latent-space settings of the FBCON and -CON. The NODE and LNN receive the base position as a zero input, which makes them translation invariant like the FBCON and -CON. For the FBCON, the learned map of is fixed to the identity (i.e., , ), so compares standardized coordinates.
Table 4: Optimization and loss hyperparameters of all models. : weight of the velocity term in (position term weighted 1); —: not part of the model.
Table 4. Optimization and loss hyperparameters of all models. : weight of the velocity term in (position term weighted 1); —: not part of the model.
Table 5: Hidden layers width of the NODE and of the LNN mass network and potential network (swish activations).
| NODE | LNN | LNN | |
| Simulation | |||
| Real cable |
Table 5. Hidden layers width of the NODE and of the LNN mass network and potential network (swish activations).
Table 6: Mapping and weights , of the position and velocity terms in .
| FBCON | -CON | |
| Mapping | identity | affine |
| Simulation | ||
| Real cable | ||
Table 6. Mapping and weights , of the position and velocity terms in .
The LNN is conservative, i.e., it has no dissipation term, as in Cranmer et al. [10]. We also trained a damped variant following Eq. (1) with a learned constant , parameterized and initialized like the damping of -CON, but its training diverged.
We tune each model with Optuna’s TPE sampler over 40–60 trials of 1000 epochs (500 for the LNN), pruning diverging and dominated trials early. The search includes the rollout loss weights for position and velocity and the steady-state holding force loss weight . Trials are scored by the validation rollout error in observation space, using fixed position and velocity weights so that trials with different loss weights remain comparable, and, for models with a steady-state holding force loss (LNN, FBCON, -CON), by the median validation endpoint-force error. We select a Pareto-optimal trial; the search ranges are listed in Table 7.
Table 7: Hyperparameter search ranges (log-uniform unless stated).
| Hyperparameter | Range | Models |
| Learning rate | / / | NODE, LNN / -CON / FBCON |
| Weight decay | / | NODE, LNN, FBCON / -CON |
| all | ||
| -CON, FBCON (optional) | ||
| LNN, FBCON, -CON | ||
| Hidden layers | NODE; LNN (each network) | |
| Width | / | NODE / LNN |
Table 7.
Experimental Implementation
The electrical cable has identified density kg m, Young’s modulus MPa, and elastic damping coefficient MPa s. These values characterize the material and are excluded from the modeling and control formulations. Two Bota Systems SenseOne sensors measure endpoint forces , with calibration before recording and drift correction using a per-trajectory offset. Five OptiTrack cameras track markers, including the attachment points defining . Each
trajectory starts from the same robot joint configuration to limit variations in force capability, since robot dynamics are not modeled. To reduce noise amplification, we downsample marker positions from 240 Hz to 50 Hz before computing velocities by central finite differences.
Dynamic Prediction in Simulation
Figure 8 shows a rollout of all four models along a training trajectory, complementing the baseline position and velocity prediction errors in Figure 7 and the discussion in § 5.3.
Figure 8. Motion prediction in simulation: predictions along a high-velocity trajectory from NODE, LNN, FBCON, and Φ-CON. Thin lines trace motion between snapshots. NODE accumulates error, LNN diverges beyond its 1 s training horizon, and FBCON lags behind the ground truth; Φ-CON follows the motion throughout.
Prediction of Potential Forces for Simulated DLO
We evaluate potential-force prediction, which provides the feedforward term in our controller, against an MLP trained directly on forces, , and an LNN [10]. At steady state, the MLP predicts potential forces as . The MLP has four hidden layers of width 64 with ReLU activations, and its inputs are expressed relative to the first marker with that marker’s absolute position removed, so that, like the other models, it is translation invariant. It is trained on the baseline training trajectories, sampled at 500 Hz, with a mean-squared force loss over all time steps, using AdamW (learning rate and weight decay , gradient-norm clipping at 1) on batches of 256 windows for 2000 epochs. We train three seeds and keep, for each, the checkpoint with the lowest validation loss. For each test trajectory, we extract the final steady state and endpoint force , and compute the relative error , for endpoint and axis . Table 8 reports the median over trajectories for all three models. -CON outperforms the MLP and matches the LNN with fewer parameters: 20 031, compared with 35 674 for the MLP and 203 584 for the LNN. We also report the computation time of one prediction for a single state, measured with JAX on an AMD Ryzen 9 5900HS CPU and an NVIDIA RTX 3050 Ti Laptop GPU as mean std over calls of the compiled function. For -CON, and depend only on the learned parameters, so we assemble them once after training instead of at every call, and reduces to a few matrix-vector products. For a single state, the GPU is slower than the CPU because kernel-launch overhead dominates the computation.
The control times in Table 2 are measured in the same way, for one PD or PD+FF step with a new target compiled as a single function.
| Model | Left endpoint X | Left endpoint Y | Left endpoint Z | Right endpoint X | Right endpoint Y | Right endpoint Z | All | CPU | GPU | Params |
| MLP | 39.77% | 12.90% | 16.86% | 40.78% | 6.55% | 24.89% | 20.47% | 15.6 2.4 | 66.5 7.6 | 35 674 |
| LNN | 0.12% | 0.23% | 0.09% | 0.11% | 0.27% | 0.08% | 0.13% | 25.4 3.9 | 105.2 5.0 | 203 584 |
| Φ-CON | 0.19% | 0.20% | 0.15% | 0.20% | 0.20% | 0.14% | 0.17% | 13.9 3.0 | 64.8 15.8 | 20 031 |
Table 8. Force prediction error and computational cost in simulation: Median relative endpoint potential-force prediction error, normalized by the endpoint force magnitude, computation time per prediction (mean std), and number of parameters.
Gain Sensitivity of Simulated Regulation
We repeat the baseline regulation experiment of § 5.4 with higher PD gains, scaling with to keep the damping ratio, while the model, targets, and settings are unchanged (Table 9). The PD error scales with , since PD alone never supplies the holding force. The PD+FF error also decreases with the gain, so the relative reduction stays between 98.8% and 99.4%, including at the 200 N m used on the real robots.
| [N/m] | [N s/m] | PD | PD+FF (-CON) | Reduction |
| 10 | 1.0 | 238.6 4.7 | 1.5 3.5 | 99.4% |
| 50 | 2.24 | 48.2 1.1 | 0.44 0.66 | 99.1% |
| 200 | 4.47 | 12.1 0.4 | 0.15 0.21 | 98.8% |
Table 9. Regulation error in simulation for varying control gains: Final endpoint regulation error [mm] on the baseline DLO for increasing PD gains, with to keep the damping ratio (mean std over the 28 test targets and both endpoints).
Proofs of Stability
Proof of Proposition 1: Stability of Latent FBCON
As in Stölzle and Della Santina [25], choose the mechanical energy
where . Since is constant, differentiating and substituting Eq. (5) with gives
The shape Hessian is
so is strongly convex in the shape, while it is independent of the base . Since the dynamics do not depend on either, evolves as an autonomous subsystem. Strong convexity and make positive definite about and radially unbounded in this reduced state, giving compact positively invariant sublevel sets and Lyapunov stability. Since , requires ; remaining there requires , hence . LaSalle’s invariance principle therefore gives global asymptotic stability of the reduced equilibrium and, equivalently, of , whose Euclidean distance is .
The base also converges to a finite position. Partitioning and into base and shape blocks, the base equation is
Integrating with respect to time gives
where is an integration constant determined by the initial state, encoding the balance between generalized base momentum and accumulated damping forces. As and , invertibility of yields
Proof of Proposition 2: Stability of Open-Loop -CON
The affine bijection preserves the separation of base and shape: is an invertible affine function of , and . Under , the constant input compensates the linear load term in . Expressing the Lyapunov function from Appendix § 14.§ 1 in observation coordinates gives
Its time derivative satisfies
The induced invertible linear map between reduced errors preserves both stability and convergence, proving global asymptotic stability of . Base-position convergence also transfers through the fixed affine standardization.
Proof of Proposition 3: Global Asymptotic Stability of Closed-Loop System
The proof uses the Lyapunov function
where is the potential of (§ 3.4). Let denote the position part of . Its gradient is
which vanishes at . Using the affine map in § 3.2,
The first term penalizes every displacement that changes the relative shape. Any other nonzero displacement translates all markers equally, moving the endpoints and thus being penalized by the second term. The constant positive-definite lower bound makes strongly convex, with its unique minimum . Together with , this makes positive definite about and radially unbounded. Feasibility gives , so the closed-loop dynamics are . Differentiating and using Eq. (7) and feasibility in the second line, then substituting the controller in the third, yields
Since , requires ; remaining there requires , hence by strong convexity. The compact positively invariant sublevel sets of establish Lyapunov stability and allow LaSalle’s invariance principle to give global convergence to .
References
- [1]Carlo Alessi, Camilla Agabiti, Daniele Caradonna, Cecilia Laschi, Federico Renda, and Egidio Falotico. Rod models in continuum and soft robot control: a review, 2024. URL https://arxiv.org/abs/2407.05886.
- [2]Miklós Bergou, Max Wardetzky, Stephen Robinson, Basile Audoly, and Eitan Grinspun. Discrete elastic rods. In ACM SIGGRAPH 2008 Papers, SIGGRAPH ’08, New York, NY, USA, 2008. Association for Computing Machinery. ISBN 9781450301121. doi: 10.1145/1399504.1360662. URL https://doi.org/10.1145/1399504.1360662.
- [3]Aude Billard and Danica Kragic. Trends and challenges in robot manipulation. Science, 364(6446): eaat8414, 2019. doi: 10.1126/science.aat8414. URL https://www.science.org/doi/abs/10.1126/science.aat8414.DOI
- [4]Pablo Borja, Azita Dabiri, and Cosimo Della Santina. Energy-based shape regulation of soft robots with unactuated dynamics dominated by elasticity. In 2022 IEEE 5th International Conference on Soft Robotics, RoboSoft 2022, pages 396–402. IEEE, April 2022. ISBN 978-1-6654-0828-8. doi: 10.1109/RoboSoft54090.2022.9762071. URL https://elib.dlr.de/197489/.DOI
- [5]Timothy Bretl and Zoe McCarthy. Quasi-static manipulation of a kirchhoff elastic rod based on a geometric analysis of equilibrium configurations. The International Journal of Robotics Research, 33(1):48–68, 2014. doi: 10.1177/0278364912473169. URL https://doi.org/10.1177/0278364912473169.
- [6]Bin Cao, Xizhe Zang, Xuehe Zhang, Zhuo Chen, Shouqiang Li, and Jie Zhao. Shape control of elastic deformable linear objects for robotic cable assembly. Advanced Intelligent Systems, 6, 04 2024. doi: 10.1002/aisy.202300835.DOI
- [7]Alessio Caporali, Ignacio Cuiral-Zueco, Gonzalo López-Nicolás, and Gianluca Palli. Robotic perception and manipulation of deformable linear objects: A survey. The International Journal of Robotics Research, page 02783649261432253, 2026.
- [8]Ricky T. Q. Chen, Yulia Rubanova, Jesse Bettencourt, and David Duvenaud. Neural ordinary differential equations, 2019. URL https://arxiv.org/abs/1806.07366.
- [9]Jian Chu, Wenkang Zhang, Bo Ouyang, Kunmiao Tian, Shuai Zhang, and Kai Zhai. A global-local graph attention network for deformable linear objects dynamic interaction with environment. IEEE Robotics and Automation Letters, 9(12):11106–11113, 2024. doi: 10.1109/LRA.2024.3486215.DOI
- [10]Miles Cranmer, Sam Greydanus, Stephan Hoyer, Peter Battaglia, David Spergel, and Shirley Ho. Lagrangian neural networks, 2020. URL https://arxiv.org/abs/2003.04630.
- [11]Cosimo Della Santina, Matteo Bianchi, Giorgio Grioli, Franco Angelini, Manuel Catalano, Manolo Garabini, and Antonio Bicchi. Controlling soft robots: Balancing feedback and feedforward elements. IEEE Robotics & Automation Magazine, 24(3):75–83, 2017. doi: 10.1109/MRA.2016.2636360.DOI
- [12]Cosimo Della Santina, Christian Duriez, and Daniela Rus. Model-based control of soft robots: A survey of the state of the art and open challenges. IEEE Control Systems Magazine, 43(3):30–65, 2023. doi: 10.1109/MCS.2023.3253419.DOI
- [13]Giovanni Franzese, Leandro de Souza Rosa, Tim Verburg, Luka Peternel, and Jens Kober. Interactive imitation learning of bimanual movement primitives. IEEE/ASME Transactions on Mechatronics, 29(5):4006–4018, 2024. doi: 10.1109/TMECH.2023.3295249.DOI
- [14]Katharina Friedl, Noémie Jaquier, Jens Lundell, Tamim Asfour, and Danica Kragic. A Riemannian framework for learning reduced-order Lagrangian dynamics. In International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=RoN6M3i7gJ.arxiv.org/abs/2410.18868
- [15]Chung Min Kim*, Brent Yi*, Hongsuk Choi, Yi Ma, Ken Goldberg, and Angjoo Kanazawa. Pyroki: A modular toolkit for robot kinematic optimization. In 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2025. URL https://arxiv.org/abs/2505.03728.
- [16]K. Korner, B. Audoly, and K. Bhattacharya. Simple deformation measures for discrete elastic rods and ribbons. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, 477(2256):20210561, 12 2021. ISSN 1364-5021. doi: 10.1098/rspa.2021.0561. URL https://doi.org/10.1098/rspa.2021.0561.
- [17]Jingyue Liu, Pablo Borja, and Cosimo Della Santina. Physics-informed neural networks to model and control robots: A theoretical and experimental investigation. Advanced Intelligent Systems, 6(5):2300385, 2024.arxiv.org/abs/2305.05375
- [18]Tim Missal, Lucas Domingues, Berk Guler, Simon Manschitz, Jan Peters, and Paula Dornhofer Paro Costa. Ropedreamer: A kinematic recurrent state space model for dynamics of flexible deformable linear objects, 2026. URL https://arxiv.org/abs/2604.28161.
- [19]Eike Hermann Müller. Exact conservation laws for neural network integrators of dynamical systems. Journal of Computational Physics, page 112234, 2023. doi: 10.1016/j.jcp.2023.112234.DOI
- [20]Michele Pierallini, Yaolei Shen, Yuri De Santis, Giorgia Pazzi, Manolo Garabini, Chiara Gabellieri, and Franco Angelini. Physics-informed neural networks for estimating the cable dynamics via cartesian measurements. IEEE Robotics and Automation Letters, 2026. doi: 10.1109/LRA.2026.3723260.DOI
- [21]Pietro Pustina, Cosimo Della Santina, and Alessandro De Luca. Feedback regulation of elastically decoupled underactuated soft robots. IEEE Robotics and Automation Letters, 7(2):4512–4519, 2022. doi: 10.1109/LRA.2022.3150829.DOI
- [22]Pietro Pustina, Cosimo Della Santina, Frédéric Boyer, Alessandro De Luca, and Federico Renda. Input decoupling of lagrangian systems via coordinate transformation: General characterization and its application to soft robotics. IEEE Transactions on Robotics, 40:2098–2110, 2024. ISSN 1941-0468. doi: 10.1109/tro.2024.3370089. URL http://dx.doi.org/10.1109/TRO.2024.3370089.
- [23]Mohammadreza Shetab-Bushehri, Miguel Aranda, Youcef Mezouar, and Erol Özgür. As-rigid-as-possible shape servoing. IEEE Robotics and Automation Letters, 7(2):3898–3905, 2022. doi: 10.1109/LRA.2022.3145960.
- [24]Maximilian Stölzle. Safe yet Precise Soft Robots: Incorporating Physics into Learned Models for Control. Dissertation (tu delft), Mechanical Engineering, Delft University of Technology, 09 2025.
- [25]Maximilian Stölzle and Cosimo Della Santina. Input-to-State Stable Coupled Oscillator Networks for Closed-form Model-based Control in Latent Space. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, editors, Advances in Neural Information Processing Systems, volume 37, pages 82010–82059. Curran Associates, Inc., 2024. URL https://proceedings.neurips.cc/paper_files/paper/2024/file/952ddaa9299d81cc307427edab034784-Paper-Conference.pdf.DOI
- [26]Changhao Wang, Yuyou Zhang, Xiang Zhang, Zheng Wu, Xinghao Zhu, Shiyu Jin, Te Tang, and Masayoshi Tomizuka. Offline-online learning of deformation model for cable manipulation with graph neural networks. IEEE Robotics and Automation Letters, 7(2):5544–5551, 2022. doi: 10.1109/LRA.2022.3158376.DOI
- [27]Yuxuan Yang, Johannes A. Stork, and Todor Stoyanov. Learning differentiable dynamics models for shape control of deformable linear objects. Robotics and Autonomous Systems, 158:1014258, 2022. ISSN 0921-8890. doi: https://doi.org/10.1016/j.robot.2022.104258. URL https://www.sciencedirect.com/science/article/pii/S0921889022001518.
- [28]Mingrui Yu, Kangchen Lv, Changhao Wang, Yongpeng Jiang, Masayoshi Tomizuka, and Xiang Li. Generalizable whole-body global manipulation of deformable linear objects by dual-arm robot in 3-d constrained environments. The International Journal of Robotics Research, 44(4):607–639, 2025.arxiv.org/abs/2310.09899






