Two arms handling a cable get a learned force model: compensating elastic and gravity loads lowers real-robot endpoint error per target
Synopsis
When two robot arms regulate a deformable linear object such as a cable, prior work mostly relied on quasi-static or kinematic models that cannot capture the mechanical response of stiff, heavy objects in fast motion. The new approach jointly learns a floating-base coupled oscillator network and an invertible affine coordinate map, so the latent potential depends only on shape and elastic and gravitational loads get explicit expressions. Feeding those potential forces forward into a proportional-derivative controller reduced mean endpoint regulation error relative to PD alone in simulation, and per-target endpoint error also fell on a stiff real cable.
Interpretation
Shape regulation of deformable linear objects such as cables and wires can now draw its feedforward force from a learned mechanical model instead of runtime numerical optimization. Earlier learned approaches took position or velocity commands as input, implicitly assuming the manipulator can impose endpoint motion regardless of the object's reaction forces, which limited them to lightweight, compliant objects and slow motions. Across four simulated deformable linear objects, the controller with this feedforward reduced mean endpoint regulation error relative to a model-free proportional-derivative baseline; on a real stiff cable, per-target endpoint error also fell.
The model writes object motion in Lagrangian form, so kinetic and potential energy have explicit expressions and the controller can directly compensate the learned elastic and gravitational loads. A floating-base coupled oscillator network makes the latent potential depend only on relative shape, accommodating overall translation; a learned affine diffeomorphism maps observed marker coordinates to latent coordinates and transforms forces consistently through the principle of virtual work. Training combines rollout errors in latent and observation space with a steady-state holding-force loss; the real experiments use two Franka Emika Panda robots grasping a stiff cable, with force sensors measuring endpoint forces and motion capture tracking markers.
Within the training horizon the learned model matches neural ordinary differential equation models and Lagrangian neural networks in prediction accuracy, with fewer parameters and per-epoch training time between the two. Lagrangian neural networks require runtime second-order automatic differentiation of the learned Lagrangian, which is costly, and Lagrangian structure alone does not guarantee closed-loop stability. Within the training horizon the model is comparable in accuracy to both comparators; across the four simulated objects it attains the lowest mean position root-mean-square error; on the same dataset, per-epoch time is 0.64 seconds for NODE, 9.35 seconds for LNN and 1.06 seconds for this model.
Perspective
The demonstrated scope is shape regulation of four simulated deformable linear objects and of one stiff cable on a real bimanual setup, with statically feasible target shapes, fixed grasp orientation, and no obstacle interaction or self-contact.
The method assumes marker positions and velocities sufficiently describe the object's dynamics over the operating regime, neglecting cross-sectional deformation and twist modes not resolved by the markers; the closed-loop stability guarantee concerns feasible equilibria of the learned model and extends to the physical object only if its dynamics match the learned model globally.
