betaFeedback

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 ΦΦ-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 ΦΦ 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, ΦΦ-CON achieves prediction accuracy comparable to neural ordinary differential equation (NODE) models and Lagrangian neural networks (LNNs), with 97.2%97.2\% and 90.2%90.2\% fewer parameters, respectively. Relative to a model-free PD baseline, the proposed controller reduces mean endpoint regulation error by 97.597.5-99.7%99.7\% in simulation and achieves a mean per-target reduction of 27.9%27.9\% 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 Φ\Phi-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 Φ\Phi 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:

  1. 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.

  1. FBCON: an extension of CONs to floating-base systems with explicit kinetic and potential energies and open-loop stability guarantees for the shape dynamics.

  1. Φ\Phi-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 N>2N > 2 marker positions xi∈R3x_i \in\mathbb{R}^3, i=1,…,Ni = 1,\ldots,N, 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 x=[xa⊤,xu⊤]⊤∈R3Nx = [x_a^\top,x_u^\top]^\top\in\mathbb{R}^{3N}, where xa=[x1⊤,xN⊤]⊤∈R6x_a = [x_1^\top,x_N^\top]^\top\in\mathbb{R}^6 stacks the actuated endpoints and xu=[x2⊤,…,xN−1⊤]⊤∈R3N−6x_u = [x_2^\top,\ldots,x_{N-1}^\top]^\top\in\mathbb{R}^{3N-6} stacks the unactuated interior markers. The Cartesian endpoint forces τ∈R6\tau\in\mathbb{R}^6 induce generalized forces τx=Bτ\tau_x = B\tau, with B=[I6 06×(3N−6)]⊤∈R3N×6B = [I_6\ 0_{6\times(3N-6)}]^\top\in\mathbb{R}^{3N\times6}.

A DLO represented by $N$ 3D marker positions and actuated by Cartesian forces at both endpoints

Figure 6. A DLO represented by NN 3D marker positions {x1,…,xN}\{x_1,\ldots,x_N\} and actuated by Cartesian forces at both endpoints.

Assumption 1 (Sufficient Observation of DLO Dynamics). Over the operating regime considered, the NN marker positions xx are sufficient generalized coordinates for the DLO, and (x,x˙)(x,\dot{x}) is a complete dynamical state. The acceleration x¨\ddot{x} is uniquely determined by (x,x˙,τ)(x,\dot{x},\tau) through the Lagrangian dynamics

M(x)x¨+C(x,x˙)x˙+D(x)x˙+∂xU(x)=Bτ,(1)M(x)\ddot{x} + C(x,\dot{x})\dot{x} + D(x)\dot{x} + \partial_x U(x) = B\tau, \tag*{(1)}

where M(x)M(x) is the mass matrix, C(x,x˙)C(x,\dot{x}) is the Coriolis matrix, D(x)D(x) a dissipation term, and ∂xU\partial_x U the gradient of the potential energy U(x)U(x) 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 {x(tk),x˙(tk),τ(tk)}\{x(t_k),\dot{x}(t_k),\tau(t_k)\}.

Then, given a desired equilibrium shape xd∈R3Nx^d \in\mathbb{R}^{3N} supportable by endpoint forces, we seek a model-based feedback law τ(x,x˙)\tau(x,\dot{x}) such that

lim⁡t→∞∥x(t)−xd∥=0,lim⁡t→∞∥x˙(t)∥=0.(2)\lim_{t\to\infty}\lVert x(t)-x^d\rVert= 0,\qquad\lim_{t\to\infty}\lVert\dot{x}(t)\rVert= 0. \tag*{(2)}

We assume that xdx^d is a locally unique equilibrium of the unactuated coordinates when the endpoints are fixed at xadx_a^d. For regulation near this equilibrium, we further assume that (xa,x˙a)→(xad,0)(x_a,\dot{x}_a)\to(x_a^d,0) implies (xu,x˙u)→(xud,0)(x_u,\dot{x}_u)\to(x_u^d,0) [21, 4].

Dynamical Model Learning with Phi-CON

Φ\Phi-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, qb=x1∈R3q_{\mathrm{b}} = x_{1} \in\mathbb{R}^{3}, and stack the remaining marker positions relative to it as qr=[(xN−x1)T,(x2−x1)T,…,(xN−1−x1)T]T∈R3(N−1)q_{\mathrm{r}} = [(x_{N} - x_{1})^{\mathsf{T}}, (x_{2} - x_{1})^{\mathsf{T}}, \ldots, (x_{N-1} - x_{1})^{\mathsf{T}}]^{\mathsf{T}} \in\mathbb{R}^{3(N-1)}, following the ordering of xx in § 2. Together, q=[qb,qrT]T=Txq = [q_{\mathrm{b}}, q_{\mathrm{r}}^{\mathsf{T}}]^{\mathsf{T}} = Tx, where TT is a fixed invertible matrix (Figure 1). A common translation changes only qbq_{\mathrm{b}}, leaving the relative shape qrq_{\mathrm{r}} unchanged. We then standardize qq as q^=Sq(q−qˉ)\hat{q} = S_{q}(q - \bar{q}) using the training mean qˉ\bar{q} and a fixed diagonal scaling matrix SqS_{q} with strictly positive diagonal entries, preserving invertibility and the separation of base and shape.

Figure 1. Coordinate transformations of Φ\Phi-CON. In observation space (steps 1–3), fixed maps express the marker positions xx as a base qbq_{\mathrm{b}} and a shape qrq_{\mathrm{r}} 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 zw,r∗z_{\mathrm{w,r}}^{*} 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:

zw=Φ(x)=Aq^+c,A=diag⁡(I3,Ar),c=[0T,crT]T.(3)z_{\mathrm{w}} = \Phi(x) = A\hat{q} + c,\qquad A = \operatorname{diag}(I_{3}, A_{\mathrm{r}}),\qquad c = [0^{\mathsf{T}}, c_{\mathrm{r}}^{\mathsf{T}}]^{\mathsf{T}}. \tag*{(3)}

Here, Ar∈R(3N−3)×(3N−3)A_{\mathrm{r}} \in\mathbb{R}^{(3N-3)\times(3N-3)} and cr∈R3N−3c_{\mathrm{r}} \in\mathbb{R}^{3N-3} are learned, giving zw,b=q^bz_{\mathrm{w,b}} = \hat{q}_{\mathrm{b}} and zw,r=Arq^r+crz_{\mathrm{w,r}} = A_{\mathrm{r}}\hat{q}_{\mathrm{r}} + c_{\mathrm{r}}. We parameterize Ar=LrUrA_{\mathrm{r}} = L_{\mathrm{r}}U_{\mathrm{r}} using lower and upper triangular factors with nonzero diagonal entries enforced during training. Together with the invertibility of TT and SqS_{q}, this makes Φ\Phi an affine diffeomorphism with an exact inverse.

Its Jacobian JΦ=ASqTJ_{\Phi} = AS_{q}T is independent of xx and determines the velocity and force transformations:

z˙w=JΦx˙,τw=αJΦ−TB(τ−τˉ).(4)\dot{z}_{\mathrm{w}} = J_{\Phi}\dot{x},\qquad\tau_{\mathrm{w}} = \alpha J_{\Phi}^{-\mathsf{T}}B(\tau- \bar{\tau}). \tag*{(4)}

The inverse-transpose factor follows from the principle of virtual work. The mean endpoint force τˉ\bar{\tau} and the positive scale α\alpha are computed from the training data and held fixed, with α\alpha chosen so that the centered endpoint forces, expressed in the standardized coordinates q^\hat{q}, have unit root-mean-square value. The mean force estimates the static load, with τˉL+τˉR≈−mDLOg0\bar{\tau}_{\mathrm{L}} + \bar{\tau}_{\mathrm{R}} \approx-m_{\mathrm{DLO}}g_{0}, where mDLOm_{\mathrm{DLO}} is the total mass of the DLO and g0g_{0} 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:

Mwz¨w+Dwz˙w+∇Vw(zw)=τw,(5)M_{\mathrm{w}}\ddot{z}_{\mathrm{w}} + D_{\mathrm{w}}\dot{z}_{\mathrm{w}} + \nabla V_{\mathrm{w}}(z_{\mathrm{w}}) = \tau_{\mathrm{w}}, \tag*{(5)}

where MwM_{\mathrm{w}} and DwD_{\mathrm{w}} are learned, constant, symmetric positive definite mass and damping matrices,2 and τw\tau_{\mathrm{w}} is the transformed input from § 3.2. We choose

Uw(zw)=12zw,rTKrzw,r+1Tlog⁡cosh⁡(zw,r+br),∇Uw(zw)=[0Krzw,r+tanh⁡(zw,r+br)],(6)U_{\mathrm{w}}(z_{\mathrm{w}}) = \frac{1}{2}z_{\mathrm{w},r}^{\mathsf{T}}K_{r}z_{\mathrm{w},r} + \mathbf{1}^{\mathsf{T}}\log\cosh(z_{\mathrm{w},r}+b_{r}), \qquad\nabla U_{\mathrm{w}}(z_{\mathrm{w}}) = \begin{bmatrix} 0 \\ K_{r}z_{\mathrm{w},r}+\tanh(z_{\mathrm{w},r}+b_{r}) \end{bmatrix}, \tag*{(6)}

with learned stiffness Kr≻0K_{r} \succ0 and bias brb_{r}. 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 Uw(zw)U_{\mathrm{w}}(z_{\mathrm{w}}) defined as in Eq. (6), let τw=0\tau_{\mathrm{w}} = 0 and Mw,Dw,Kr≻0M_{\mathrm{w}}, D_{\mathrm{w}}, K_{r} \succ0. Then the equilibrium set Ew={(zw,z˙w):zw,r=zw,r⋆,z˙w=0}\mathcal{E}_{\mathrm{w}} = \{(z_{\mathrm{w}}, \dot{z}_{\mathrm{w}}) : z_{\mathrm{w},r} = z_{\mathrm{w},r}^{\star}, \dot{z}_{\mathrm{w}} = 0\} is globally asymptotically stable, with arbitrary base position and unique shape zw,r⋆z_{\mathrm{w},r}^{\star} satisfying Krzw,r⋆+tanh⁡(zw,r⋆+br)=0K_{r}z_{\mathrm{w},r}^{\star} + \tanh(z_{\mathrm{w},r}^{\star} + b_{r}) = 0.

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 Φ\Phi is affine, z¨w=JΦx¨\ddot{z}_{\mathrm{w}} = J_{\Phi}\ddot{x}. Substituting the coordinate and force transformations into Eq. (5) and premultiplying by α−1JΦT\alpha^{-1}J_{\Phi}^{\mathsf{T}} gives

Mxx¨+Dxx˙+g(x)=Bτ,g(x):=Bτˉ+α−1JΦT∇Uw(Φ(x)).(7)M_{x}\ddot{x}+D_{x}\dot{x}+g(x)=B\tau,\qquad g(x):=B\bar{\tau}+\alpha^{-1}J_{\Phi}^{\mathsf{T}}\nabla U_{\mathrm{w}}(\Phi(x)). \tag*{(7)}

Here, Mx=α−1JΦTMwJΦM_{x}=\alpha^{-1}J_{\Phi}^{\mathsf{T}}M_{\mathrm{w}}J_{\Phi} and Dx=α−1JΦTDwJΦD_{x}=\alpha^{-1}J_{\Phi}^{\mathsf{T}}D_{\mathrm{w}}J_{\Phi} are constant positive definite matrices. The dynamics retain Lagrangian structure ([22 Prop. 2, App. A]), with kinetic energy Tx(x˙)=12x˙TMxx˙T_{x}(\dot{x})=\frac{1}{2}\dot{x}^{\mathsf{T}}M_{x}\dot{x} and potential energy Ux(x)=α−1Uw(Φ(x))+τˉTBTxU_{x}(x)=\alpha^{-1}U_{\mathrm{w}}(\Phi(x))+\bar{\tau}^{\mathsf{T}}B^{\mathsf{T}}x. Their sum is the total mechanical energy. The potential gradient g(x)=∇Ux(x)g(x)=\nabla U_{x}(x) represents the elastic and gravitational load that must be balanced at equilibrium.

Proposition 2 (Stability of Open-Loop Φ\Phi-CON). Under the conditions of Proposition 1, the dynamics in Eq. (7) with τ=τˉ\tau=\bar{\tau} (equivalently, τw=0\tau_{\mathrm{w}}=0) have the globally asymptotically stable equilibrium set Ex={(x,x˙):qr(x)=qr⋆,x˙=0}\mathcal{E}_{x}=\{(x,\dot{x}):q_{r}(x)=q_{r}^{\star},\dot{x}=0\}, where qr⋆q_{r}^{\star} corresponds uniquely to zw,r⋆z_{\mathrm{w},r}^{\star} 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 (τ=0\tau=0), 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 ℓ=ℓz+ℓx+λssℓss\ell=\ell_{z}+\ell_{x}+\lambda_{\mathrm{ss}}\ell_{\mathrm{ss}}. Figure 2 summarizes the training pipeline. Here, ℓz\ell_{z} and ℓx\ell_{x} measure rollout errors in latent and observation space, respectively, and ℓss\ell_{\mathrm{ss}} is the steady-state holding force loss with weight λss>0\lambda_{\mathrm{ss}}>0. ℓx\ell_{x} is the primary objective, while ℓz\ell_{z} 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 Φ\Phi 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 HH steps and mapped back by Φ−1\Phi^{-1}. The losses ℓz\ell_z and ℓx\ell_x compare the rollout with the data in each space, and ℓss\ell_{ss} fits the holding forces ga(x)g_a(x) to τ\tau at rest.

For each training trajectory window ξ={(xt,x˙t,τt)}t=0H∼D\xi= \{(x_t,\dot{x}_t,\tau_t)\}_{t=0}^{H} \sim\mathcal{D}, we initialize z^w,0=Φ(x0)\hat{z}_{w,0} = \Phi(x_0) and z˙^w,0=JΦx˙0\hat{\dot{z}}_{w,0} = J_{\Phi}\dot{x}_0. We then predict recursively via (z^w,t+1,z˙^w,t+1)=fθ(z^w,t,z˙^w,t,τw,t)(\hat{z}_{w,t+1},\hat{\dot{z}}_{w,t+1}) = f_{\theta}(\hat{z}_{w,t},\hat{\dot{z}}_{w,t},\tau_{w,t}) for t=0,…,H−1t = 0,\ldots,H-1, where fθf_{\theta} integrates Eq. (5) using the transformed input τw,t\tau_{w,t}.

The latent rollout loss ℓz\ell_z compares the encoded ground-truth states with the latent rollout:

ℓz=Eξ∼D[1H+1∑t=0H(∥Φ(xt)−z^w,t∥22+∥JΦx˙t−z˙^w,t∥22)].(8)\ell_z = \mathbb{E}_{\xi\sim\mathcal{D}}\left[\frac{1}{H+1}\sum_{t=0}^{H}\left(\left\|\Phi(x_t)-\hat{z}_{w,t}\right\|_2^2+\left\|J_{\Phi}\dot{x}_t-\hat{\dot{z}}_{w,t}\right\|_2^2\right)\right]. \tag*{(8)}

The observation-space loss ℓx\ell_x compares xtx_t and x˙t\dot{x}_t with Φ−1(z^w,t)\Phi^{-1}(\hat{z}_{w,t}) and JΦ−1z˙^w,tJ_{\Phi}^{-1}\hat{\dot{z}}_{w,t}, respectively, accounting for the effect of the inverse mapping on position and velocity prediction errors.

At equilibrium, Eq. (7) gives g(x)=Bτg(x) = B\tau and hence ga(x)=τg_a(x) = \tau, where ga(x):=BTg(x)g_a(x) := B^{\mathsf{T}}g(x) denotes the potential forces at the actuated endpoints. We supervise these forces for regulation through

ℓss=1∣Dss∣∑(x,τ)∈Dss∥ga(x)−τ∥22,\ell_{ss} = \frac{1}{\lvert\mathcal{D}_{ss}\rvert}\sum_{(x,\tau)\in\mathcal{D}_{ss}}\left\|g_a(x)-\tau\right\|_2^2,

where Dss\mathcal{D}_{ss} 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:

τ=ga(xd)+Kp(xad−xa)−Kdx˙a.(9)\tau= g_a(x^d) + K_p(x_a^d-x_a)-K_d\dot{x}_a. \tag*{(9)}

The feedforward term ga(xd)g_a(x^d) 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

Mxx¨+Dxx˙+g(x)=B(ga(xd)+Kp(xad−xa)−Kdx˙a).(10)M_x\ddot{x}+D_x\dot{x}+g(x)=B\left(g_a(x^d)+K_p(x_a^d-x_a)-K_d\dot{x}_a\right). \tag*{(10)}

Since actuation is restricted to the endpoints, (xd,0)(x^d,0) is an equilibrium of Eq. (10) if and only if gu(xd)=0g_u(x^d)=0, where gug_u collects the unactuated components of gg. This condition requires the interior coordinates to be in equilibrium when the endpoints are held at xadx_a^d.

Assumption 2 The target satisfies gu(xd)=0g_u(x^d)=0.

Proposition 3 (Global Asymptotic Stability of Closed-Loop Φ\Phi-CON). Under Assumption 2 and the matrix conditions of Proposition 1, consider Eq. (7) under the control law in Eq. (9), with gains Kp≻0K_p \succ0 and Kd⪰0K_d \succeq0. Then (xd,0)(x^d,0) 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 L=1L = 1 m and N=30N = 30 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 xtx_t and rollout x^t\hat{x}_t over test trajectories:

RMSE⁡x(t)=1∣D∣∑ξ∈D13N∥x^t−xt∥22.\operatorname{RMSE}_{x}(t)=\frac{1}{|\mathcal{D}|}\sum_{\xi\in\mathcal{D}}\sqrt{\frac{1}{3N}\left\|\hat{x}_t-x_t\right\|_2^2}.

We compute velocity error RMSE⁡x˙(t)\operatorname{RMSE}_{\dot{x}}(t) analogously and report mean ±\pm 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 x¨\ddot{x} from (x,x˙,τx)(x,\dot{x},\tau_x) using a multilayer perceptron (MLP) without Lagrangian structure. The LNN learns M(x)M(x) and U(x)U(x) with separate networks and derives dynamics by automatic differentiation, providing ∂xU\partial_x U 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, ej=∥xa,j(tf)−xa,jd∥2e_j=\left\|x_{a,j}(t_f)-x^d_{a,j}\right\|_2, as mean ±\pm standard deviation over targets, pooling both endpoints in simulation. In real experiments, we also report the per-target relative improvement (ejPD−ejPD+FF)/ejPD(e_j^{\mathrm{PD}}-e_j^{\mathrm{PD+FF}})/e_j^{\mathrm{PD}}. Figures also show the shape error 1N∑i=1N∥xi−xid∥2\frac{1}{N}\sum_{i=1}^{N}\left\|x_i-x_i^d\right\|_2, the mean marker distance to the target shape.

Dynamic Model Prediction of Simulated DLO

Within the 1 s training horizon, Φ\Phi-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 Φ\Phi-CON require 0.64, 9.35, and 1.06 seconds per epoch, respectively. Across all four simulated DLOs, Φ\Phi-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] ↓NODELNNΦ\Phi-CON (ours)
Baseline14.6 ±\pm 0.617.6 ±\pm 1.314.0 ±\pm 0.6
Heavy12.5 ±\pm 1.113.6 ±\pm 0.76.6 ±\pm 0.9
Low damping15.4 ±\pm 1.624.8 ±\pm 1.614.8 ±\pm 1.0
Compliant11.0 ±\pm 1.312.9 ±\pm 0.64.5 ±\pm 0.5

Table 1. Prediction error in simulation: Position RMSE [mm] for predictions 2 s into the future (mean ±\pm std over seeds).

Figure 7. Prediction error in simulation: position (left) and velocity (right) RMSE over 2 s rollouts for NODE, LNN, FBCON, and Φ\Phi-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 (Kp=10K_{\mathrm{p}} = 10 N m−1^{-1}, Kd=1K_{\mathrm{d}} = 1 N s m−1^{-1}) alone and with potential-force feedforward from an LNN or Φ\Phi-CON (§ 4). At test-trajectory steady states, Φ\Phi-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 xdx^{d} of a test trajectory. PD+FF computes feedforward from the full xdx^{d}, whereas PD uses only the actuated coordinates xax_{a}. 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 Kp=10K_{\mathrm{p}} = 10, 50, and 200 N m−1^{-1}, the gain of the real robots (Appendix § 13). Figure 3 shows an example regulation experiment.

Endpoint error [mm] ↓PDPD+FF (LNN)PD+FF (Φ\Phi-CON)
Baseline238.6 ±\pm 4.72.5 ±\pm 6.91.5 ±\pm 3.5
Heavy459.4 ±\pm 14.92.3 ±\pm 5.52.7 ±\pm 5.9
Low damping238.9 ±\pm 4.63.6 ±\pm 5.25.9 ±\pm 9.5
Compliant240.5 ±\pm 10.10.7 ±\pm 1.40.8 ±\pm 0.6
CPU time [μs\mu\mathrm{s}] ↓7.8 ±\pm 1.125.5 ±\pm 2.912.6 ±\pm 3.4
GPU time [μs\mu\mathrm{s}] ↓47.2 ±\pm 8.9114.6 ±\pm 4.172.1 ±\pm 8.1

Table 2. Regulation error and computational cost in simulation: Final endpoint regulation error [mm] across the four simulated DLO datasets (mean ±\pm std over test targets and both endpoints), and computation time of one control step for a new target (mean ±\pm std, Appendix § 12).

Figure 3. Shape regulation in simulation: regulation from a straight rest configuration to the target shape xdx^d (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 L=0.76L = 0.76 m and radius r=9r = 9 mm. Force sensors measure endpoint forces τ\tau, and motion capture tracks N=12N = 12 markers, including the attachment points defining xax_{a}. 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−1^{-1} and damping ratio 0.77 (21.8 N s m−1^{-1}).

Dynamic Model Prediction of Real DLO

Figure 4 compares NODE, LNN, FBCON, and Φ\Phi-CON rollout errors on recorded test trajectories. NODE is more accurate within the 2 s training horizon, and Φ\Phi-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 Φ\Phi-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 Dss\mathcal{D}_{\mathrm{ss}} (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 ga(xd)g_a(x^d) in Eq. (9) is evaluated once from the recorded xdx^d and added to the force command of the impedance controller of § 6.1, which provides the PD term. Across the steady-state targets, adding Φ\Phi-CON feedforward to the PD impedance controller reduces final endpoint errors from 63.7±8.063.7 \pm8.0 to 48.4±5.648.4 \pm5.6 mm on the left and from 48.3±8.548.3 \pm8.5 to 32.3±5.232.3 \pm5.2 mm on the right. The corresponding per-target relative improvements are 23.5%±8.8%23.5\% \pm8.8\% and 32.2%±10.5%32.2\% \pm10.5\%, 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 xdx^d (changes at dotted lines) with PD and PD+FF (Φ\Phi-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 Φ\Phi-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, Φ\Phi-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 T=10T = 10 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 Tnullmin⁡=0.5T_{\mathrm{null}}^{\min} = 0.5 s and 2 s. We enforce workspace limits and a maximum endpoint separation of LL, 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 (Kp=50K_p = 50 N m−1^{-1}) in MuJoCo. Each simulation continues for 2 s after TT 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 EE, shear modulus GG, and the cross-sectional area moments determined by radius rr. Viscous joint damping has coefficient dd. All four simulated DLOs have length L=1L = 1 m, N=30N = 30 observed markers, and radius r=0.005r = 0.005 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 DatasetEE [Pa]GG [Pa]ρ\rho [kg/m3^3]dd [N·m·s/rad]
Baseline6×1076 \times10^71×1071 \times10^73×1033 \times10^31.5×10−21.5 \times10^{-2}
Heavy6×1076 \times10^71×1071 \times10^77.8×1037.8 \times10^31.5×10−21.5 \times10^{-2}
Low damping6×1076 \times10^71×1071 \times10^73×1033 \times10^33×10−33 \times10^{-3}
Compliant3×1063 \times10^61×1061 \times10^63×1033 \times10^31.5×10−21.5 \times10^{-2}

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, DssD_{\mathrm{ss}} contains training samples (xt,τt)(x_t, \tau_t) for which every marker satisfies ∥x˙i,t∥<0.01\lVert\dot{x}_{i,t}\rVert< 0.01 m/s and ∥x¨i,t∥<0.1\lVert\ddot{x}_{i,t}\rVert< 0.1 m/s2^2.

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: Δt=10\Delta t = 10 ms (1 s windows) in simulation and Δt=20\Delta t = 20 ms (2 s windows) on the real cable. States and forces are standardized with training-set statistics (§ 3.1), and τˉ\bar{\tau} and α\alpha 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 Φ\Phi-CON. The NODE and LNN receive the base position as a zero input, which makes them translation invariant like the FBCON and Φ\Phi-CON. For the FBCON, the learned map of Φ\Phi is fixed to the identity (i.e., A=IA = I, c=0c = 0), so ℓz\ell_z compares standardized coordinates.

Table 4: Optimization and loss hyperparameters of all models. wxvelw_{x}^{\mathrm{vel}}: weight of the velocity term in ℓx\ell_x (position term weighted 1); —: not part of the model.

NODELNNFBCONΦ\Phi-CON
Simulation (all four DLOs)
Learning rate6.5×10−46.5 \times10^{-4}1.0×10−41.0 \times10^{-4} 32.9×10−32.9 \times10^{-3}2.2×10−32.2 \times10^{-3}
Weight decay003.9×10−43.9 \times10^{-4}3.8×10−23.8 \times10^{-2}
wxvelw_{x}^{\mathrm{vel}}1.2729.013.131.3
λss\lambda_{\mathrm{ss}}—0.360.650.50
Epochs5000200050005000
Parameters710 234203 58412 37520 031
Real electrical cable
Learning rate1.1×10−31.1 \times10^{-3}2.0×10−42.0 \times10^{-4} 31.1×10−31.1 \times10^{-3}4.3×10−34.3 \times10^{-3}
Weight decay005.5×10−45.5 \times10^{-4}2.4×10−42.4 \times10^{-4}
wxvelw_{x}^{\mathrm{vel}}0.5326.96.091.04
λss\lambda_{\mathrm{ss}}—1.8×10−31.8 \times10^{-3}0.184.2×10−44.2 \times10^{-4}
Epochs2000200050005000
Parameters336 932200 60320343156

Table 4. Optimization and loss hyperparameters of all models. wxvelw_{x}^{\mathrm{vel}}: weight of the velocity term in ℓx\ell_x (position term weighted 1); —: not part of the model.

Table 5: Hidden layers ×\times width of the NODE and of the LNN mass network M\mathcal{M} and potential network U\mathcal{U} (swish activations).

NODELNN M\mathcal{M}LNN U\mathcal{U}
Simulation3×5123 \times5125×325 \times324×1284 \times128
Real cable2×5122 \times5126×1286 \times1287×647 \times64

Table 5. Hidden layers ×\times width of the NODE and of the LNN mass network M\mathcal{M} and potential network U\mathcal{U} (swish activations).

Table 6: Mapping Φ\Phi and weights wzposw_z^{\mathrm{pos}}, wzvelw_z^{\mathrm{vel}} of the position and velocity terms in ℓz\ell_z.

FBCONΦ\Phi-CON
Mapping Φ\Phiidentityaffine
Simulation
wzposw_z^{\mathrm{pos}}2.2×10−32.2 \times10^{-3}5.6×10−55.6 \times10^{-5}
wzvelw_z^{\mathrm{vel}}1.6×10−51.6 \times10^{-5}5.8×10−45.8 \times10^{-4}
Real cable
wzposw_z^{\mathrm{pos}}1.9×10−31.9 \times10^{-3}1.3×10−51.3 \times10^{-5}
wzvelw_z^{\mathrm{vel}}3.2×10−43.2 \times10^{-4}1.5×10−31.5 \times10^{-3}

Table 6. Mapping Φ\Phi and weights wzposw_z^{\mathrm{pos}}, wzvelw_z^{\mathrm{vel}} of the position and velocity terms in ℓz\ell_z.

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 D≻0D \succ0, parameterized and initialized like the damping of Φ\Phi-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 λss\lambda_{\mathrm{ss}}. 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, Φ\Phi-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).

HyperparameterRangeModels
Learning rate[10−6,10−2][10^{-6}, 10^{-2}] / [10−5,10−2][10^{-5}, 10^{-2}] / [10−5,3×10−3][10^{-5}, 3 \times10^{-3}]NODE, LNN / Φ\Phi-CON / FBCON
Weight decay{0}∪[10−4,10−2]\{0\} \cup[10^{-4}, 10^{-2}] / [10−4,10−1][10^{-4}, 10^{-1}]NODE, LNN, FBCON / Φ\Phi-CON
wzvelw_z^{\mathrm{vel}}[10−2,102][10^{-2}, 10^{2}]all
wzpos,wzvelw_z^{\mathrm{pos}}, w_z^{\mathrm{vel}}[10−5,102][10^{-5}, 10^{2}]Φ\Phi-CON, FBCON (optional)
λss\lambda_{\mathrm{ss}}[10−4,1][10^{-4}, 1]LNN, FBCON, Φ\Phi-CON
Hidden layers{2,…,8}\{2, \ldots, 8\}NODE; LNN (each network)
Width{64,128,256,512}\{64, 128, 256, 512\} / {32,64,128}\{32, 64, 128\}NODE / LNN

Table 7.

Experimental Implementation

The electrical cable has identified density ρ≈3.5×103\rho\approx3.5 \times10^3 kg m−3^{-3}, Young’s modulus E≈22E \approx22 MPa, and elastic damping coefficient ν≈2\nu\approx2 MPa s. These values characterize the material and are excluded from the modeling and control formulations. Two Bota Systems SenseOne sensors measure endpoint forces τ\tau, with calibration before recording and drift correction using a per-trajectory offset. Five OptiTrack cameras track N=12N = 12 markers, including the attachment points defining xax_a. 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, τ^x=MLP⁡(x,x˙,x¨)\hat{\tau}_x = \operatorname{MLP}(x, \dot{x}, \ddot{x}), and an LNN [10]. At steady state, the MLP predicts potential forces as τ^xss=MLP⁡(xss,0,0)\hat{\tau}_{x_{\mathrm{ss}}} = \operatorname{MLP}(x_{\mathrm{ss}}, 0, 0). 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 10−410^{-4}, 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 xssix_{\mathrm{ss}}^{i} and endpoint force τssi∈R6\tau_{\mathrm{ss}}^{i} \in\mathbb{R}^{6}, and compute the relative error ej,ki=∥g^a,j,k(xssi)−τss,j,ki∥/∥τss,ji∥2e_{j,k}^{i} = \left\lVert\hat{g}_{a,j,k}(x_{\mathrm{ss}}^{i}) - \tau_{\mathrm{ss},j,k}^{i} \right\rVert/ \left\lVert\tau_{\mathrm{ss},j}^{i} \right\rVert_{2}, for endpoint j∈{left,right}j \in\{\mathrm{left}, \mathrm{right}\} and axis k∈{X,Y,Z}k \in\{X, Y, Z\}. Table 8 reports the median over trajectories for all three models. Φ\Phi-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 ±\pm std over 10410^{4} calls of the compiled function. For Φ\Phi-CON, KrK_r and Ar=LrUrA_r = L_r U_r depend only on the learned parameters, so we assemble them once after training instead of at every call, and g(x)g(x) 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 xdx^d compiled as a single function.

ModelLeft endpoint XLeft endpoint YLeft endpoint ZRight endpoint XRight endpoint YRight endpoint ZAllCPUGPUParams
MLP39.77%12.90%16.86%40.78%6.55%24.89%20.47%15.6 ±\pm 2.466.5 ±\pm 7.635 674
LNN0.12%0.23%0.09%0.11%0.27%0.08%0.13%25.4 ±\pm 3.9105.2 ±\pm 5.0203 584
Φ-CON0.19%0.20%0.15%0.20%0.20%0.14%0.17%13.9 ±\pm 3.064.8 ±\pm 15.820 031

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 ±\pm std), and number of parameters.

Gain Sensitivity of Simulated Regulation

We repeat the baseline regulation experiment of § 5.4 with higher PD gains, scaling KdK_d with Kp\sqrt{K_p} to keep the damping ratio, while the model, targets, and settings are unchanged (Table 9). The PD error scales with 1/Kp1/K_p, 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−1^{-1} used on the real robots.

KpK_p [N/m]KdK_d [N s/m]PDPD+FF (Φ\Phi-CON)Reduction
101.0238.6 ±\pm 4.71.5 ±\pm 3.599.4%
502.2448.2 ±\pm 1.10.44 ±\pm 0.6699.1%
2004.4712.1 ±\pm 0.40.15 ±\pm 0.2198.8%

Table 9. Regulation error in simulation for varying control gains: Final endpoint regulation error [mm] on the baseline DLO for increasing PD gains, with Kd∝KpK_d \propto\sqrt{K_p} to keep the damping ratio (mean ±\pm 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

V=12z˙w⊤Mwz˙w+Uw(zw)−Uw(zw∗),(11)V = \frac{1}{2}\dot{z}_w^{\top}M_w\dot{z}_w + U_w(z_w)-U_w(z_w^\ast), \tag*{(11)}

where (zw⋆,0)∈Ew(z_w^\star, 0) \in\mathcal{E}_w. Since MwM_w is constant, differentiating VV and substituting Eq. (5) with τw=0\tau_w = 0 gives

V˙=z˙wTMwz¨w+∇UwTz˙w=z˙wT(Mwz¨w+∇Uw)=−z˙wTDwz˙w≤0.(12)\dot{V} = \dot{z}_w^{\mathsf{T}} M_w \ddot{z}_w + \nabla U_w^{\mathsf{T}} \dot{z}_w = \dot{z}_w^{\mathsf{T}} \left(M_w \ddot{z}_w + \nabla U_w\right) = -\dot{z}_w^{\mathsf{T}} D_w \dot{z}_w \le0. \tag*{(12)}

The shape Hessian is

∇zw,r2Uw=Kr+diag⁡(sech⁡2(zw,r+br))⪰Kr≻0,(13)\nabla^2_{z_w,r} U_w = K_r + \operatorname{diag}\left(\operatorname{sech}^2(z_{w,r} + b_r)\right) \succeq K_r \succ0, \tag*{(13)}

so UwU_w is strongly convex in the shape, while it is independent of the base zw,bz_{w,b}. Since the dynamics do not depend on zw,bz_{w,b} either, (zw,r,z˙w)(z_{w,r}, \dot{z}_w) evolves as an autonomous subsystem. Strong convexity and Mw≻0M_w \succ0 make VV positive definite about (zw,r⋆,0)(z_{w,r}^\star, 0) and radially unbounded in this reduced state, giving compact positively invariant sublevel sets and Lyapunov stability. Since Dw≻0D_w \succ0, V˙=0\dot{V} = 0 requires z˙w=0\dot{z}_w = 0; remaining there requires ∇Uw=0\nabla U_w = 0, hence zw,r=zw,r⋆z_{w,r} = z_{w,r}^\star. LaSalle’s invariance principle therefore gives global asymptotic stability of the reduced equilibrium and, equivalently, of Ew\mathcal{E}_w, whose Euclidean distance is (∥zw,r−zw,r⋆∥2+∥z˙w∥2)1/2(\lVert z_{w,r} - z_{w,r}^\star\rVert^2 + \lVert\dot{z}_w\rVert^2)^{1/2}.

The base also converges to a finite position. Partitioning MwM_w and DwD_w into base and shape blocks, the base equation is

Mw,bbz¨w,b+Mw,brz¨w,r+Dw,bbz˙w,b+Dw,brz˙w,r=0.(14)M_{w,bb}\ddot{z}_{w,b} + M_{w,br}\ddot{z}_{w,r} + D_{w,bb}\dot{z}_{w,b} + D_{w,br}\dot{z}_{w,r} = 0. \tag*{(14)}

Integrating with respect to time gives

Mw,bbz˙w,b+Mw,brz˙w,r+Dw,bbzw,b+Dw,brzw,r=cb,(15)M_{w,bb}\dot{z}_{w,b} + M_{w,br}\dot{z}_{w,r} + D_{w,bb}z_{w,b} + D_{w,br}z_{w,r} = c_b, \tag*{(15)}

where cbc_b is an integration constant determined by the initial state, encoding the balance between generalized base momentum and accumulated damping forces. As z˙w→0\dot{z}_w \to0 and zw,r→zw,r⋆z_{w,r} \to z_{w,r}^\star, invertibility of Dw,bb≻0D_{w,bb} \succ0 yields

zw,b→Dw,bb−1(cb−Dw,brzw,r⋆).(16)z_{w,b} \to D_{w,bb}^{-1}\left(c_b - D_{w,br}z_{w,r}^\star\right). \tag*{(16)}

Proof of Proposition 2: Stability of Open-Loop Φ\Phi-CON

The affine bijection Φ\Phi preserves the separation of base and shape: zw,rz_{w,r} is an invertible affine function of qrq_r, and x˙=JΦ−1z˙w\dot{x} = J_{\Phi}^{-1}\dot{z}_w. Under τ=τˉ\tau= \bar{\tau}, the constant input compensates the linear load term τˉTBTx\bar{\tau}^{\mathsf{T}}B^{\mathsf{T}}x in UxU_x. Expressing the Lyapunov function from Appendix § 14.§ 1 in observation coordinates gives

Vx(x,x˙)=α−1V(Φ(x),JΦx˙).(17)V_x(x,\dot{x}) = \alpha^{-1}V\left(\Phi(x), J_{\Phi}\dot{x}\right). \tag*{(17)}

Its time derivative satisfies

V˙x=−α−1x˙⊤JΦ⊤DwJΦx˙=−x˙⊤Dxx˙≤0.(18)\dot{V}_x=-\alpha^{-1}\dot{x}^{\top}J_{\Phi}^{\top}D_wJ_{\Phi}\dot{x}=-\dot{x}^{\top}D_x\dot{x}\le0. \tag*{(18)}

The induced invertible linear map between reduced errors preserves both stability and convergence, proving global asymptotic stability of Ex\mathcal{E}_x. 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

V=12x˙⊤Mxx˙+Ux(x)−Ux(xd)−g(xd)⊤(x−xd)+12(xa−xad)⊤Kp(xa−xad),(19)V=\frac{1}{2}\dot{x}^{\top}M_x\dot{x}+U_x(x)-U_x(x^d)-g(x^d)^{\top}(x-x^d)+\frac{1}{2}(x_a-x_a^d)^{\top}K_p(x_a-x_a^d), \tag*{(19)}

where UxU_x is the potential of gg (§ 3.4). Let W(x)W(x) denote the position part of VV. Its gradient is

∇W(x)=g(x)−g(xd)+BKp(xa−xad),(20)\nabla W(x)=g(x)-g(x^d)+BK_p(x_a-x_a^d), \tag*{(20)}

which vanishes at x=xdx=x^d. Using the affine map in § 3.2,

∇2W(x)⪰α−1T⊤Sq⊤diag⁡(03×3,Ar⊤KrAr)SqT+BKpB⊤≻0.(21)\nabla^2W(x)\succeq\alpha^{-1}T^{\top}S_q^{\top}\operatorname{diag}(0_{3\times3},A_r^{\top}K_rA_r)S_qT+BK_pB^{\top}\succ0. \tag*{(21)}

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 WW strongly convex, with its unique minimum W(xd)=0W(x^d)=0. Together with Mx≻0M_x\succ0, this makes VV positive definite about (xd,0)(x^d,0) and radially unbounded. Feasibility gives g(xd)=Bga(xad)g(x^d)=Bg_a(x_a^d), so the closed-loop dynamics are Mxx¨+(Dx+BKdB⊤)x˙+∇W(x)=0M_x\ddot{x}+(D_x+BK_dB^{\top})\dot{x}+\nabla W(x)=0. Differentiating VV and using Eq. (7) and feasibility in the second line, then substituting the controller in the third, yields

V˙=x˙⊤(Mxx¨+g(x)−g(xd))+x˙a⊤Kp(xa−xad)=−x˙⊤Dxx˙+x˙a⊤[τ−ga(xad)+Kp(xa−xad)]=−x˙⊤Dxx˙−x˙a⊤Kdx˙a≤0.(22)\begin{aligned} \dot{V}&=\dot{x}^{\top}(M_x\ddot{x}+g(x)-g(x^d))+\dot{x}_a^{\top}K_p(x_a-x_a^d)\\ &=-\dot{x}^{\top}D_x\dot{x}+\dot{x}_a^{\top}\left[\tau-g_a(x_a^d)+K_p(x_a-x_a^d)\right]\\ &=-\dot{x}^{\top}D_x\dot{x}-\dot{x}_a^{\top}K_d\dot{x}_a\le0. \tag*{(22)} \end{aligned}

Since Dx≻0D_x\succ0, V˙=0\dot{V}=0 requires x˙=0\dot{x}=0; remaining there requires ∇W(x)=0\nabla W(x)=0, hence x=xdx=x^d by strong convexity. The compact positively invariant sublevel sets of VV establish Lyapunov stability and allow LaSalle’s invariance principle to give global convergence to (xd,0)(x^d,0).

References

  1. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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

Paper details

vlnc:W_s6h863yc
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 6

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

GPT-6 Astra (bare), Claude Opus 5.5 (bare)

View statement
Looking for related papers…
Contents