Recent advances in machine learning have begun to embed oscillatory network principles within neural architectures, aiming to enhance computational efficiency and robustness in time-series regression. Building on these developments, we take a step toward applying such principles to the learning of physical dynamics. We introduce Neural Linear Oscillator Networks (nLON): a vision-based framework that extracts compact latent representations of complex motion directly from image sequences. A convolutional autoencoder encodes position and velocity into a low-dimensional manifold, whose temporal evolution is governed by coupled linear mechanical oscillators driven by a linear combination of the inputs. This strong structural prior not only promotes sample efficiency and interpretability but also guarantees that the learned model remains a mechanical, Wiener-type system. From this formulation, we derive closed-loop controllers that ensure stable regulation. We focus on soft robots-systems whose nonlinear, continuous, and high-dimensional dynamics make them both a challenging and ideal testbed for our approach. Using tentacle robots in high-fidelity simulations and real-world experiments, we validate that our framework delivers accurate long-horizon predictions and consistently surpasses state-of-the-art baselines, achieving superior structural fidelity and final-step accuracy. Finally, we leverage the learned dynamics for model-based control, demonstrating in simulation that the resulting scheme achieves robust and reliable tracking.

Reduced-order models are central to motion planning and control of quadruped robots, yet existing templates are often hand-crafted for a specific locomotion modality. This motivates the need for automatic methods that extract task-specific, interpretable low-dimensional dynamics directly from data. We propose a methodology that combines a linear autoencoder with symbolic regression to derive such models. The linear autoencoder provides a consistent latent embedding for configurations, velocities, accelerations, and inputs, enabling the sparse identification of nonlinear dynamics (SINDy) to operate in a compact, physics-aligned space. A multi-phase, hybrid-aware training scheme ensures coherent latent coordinates across contact transitions. We focus our validation on quadruped jumping—a representative, challenging, yet contained scenario in which a principled template model is especially valuable. The resulting symbolic dynamics outperform the state-of-the-art handcrafted actuated spring-loaded inverted pendulum (aSLIP) baseline in simulation and hardware across multiple robots and jumping modalities.

Obtaining dynamic models of continuum soft robots is central to the analysis and control of soft robots, and researchers have devoted much attention to the challenge of proposing both data-driven and first-principle solutions. Both avenues have, however, shown their limitations; the former lacks structure and performs poorly outside training data, while the latter requires significant simplifications and extensive expert knowledge to be used in practice. This paper introduces a streamlined method for learning low-dimensional, physicsbased models that are both accurate and easy to interpret. We start with an algorithm that uses image data (i.e., shape evolutions) to determine the minimal necessary segments for describing a soft robot's movement. Following this, we apply a dynamic regression and strain sparsification algorithm to identify relevant strains and define the model's dynamics. We validate our approach through simulations with various planar soft manipulators, comparing its performance against other learning strategies, showing that our models are both computationally efficient and 25x more accurate on out-of-training distribution inputs. Finally, we demonstrate that thanks to the capability of the method of generating physically compatible models, the learned models can be straightforwardly combined with model-based control policies.

This study explores a method for the dynamic modeling of soft robots, focusing on enhancing the deep learning-based Lagrangian modeling approach through the attention mechanism, which enriches the training process by allocating focused attention and analytical weighting to critical state features, thereby increasing the model's sensitivity to changes in the robot's state. We compared our method through simulation, demonstrating that the model is effective in long-term prediction and noise rejection.

This work concerns the application of physics-informed neural networks to the modeling and control of complex robotic systems. Achieving this goal requires extending physics-informed neural networks to handle nonconservative effects. These learned models are proposed to combine with model-based controllers originally developed with first-principle models in mind. By combining standard and new techniques, precise control performance can be achieved while proving theoretical stability bounds. These validations include real-world experiments of motion prediction with a soft robot and trajectory tracking with a Franka Emika Panda manipulator.