Soft manipulators, renowned for their compliance and adaptability, hold great promise in their ability to engage safely and effectively with intricate environments and delicate objects. Nonetheless, controlling these soft systems presents distinctive hurdles owing to their nonlinear behavior and complicated dynamics. Learning‐based controllers for continuum soft manipulators offer a viable alternative to model‐based approaches that may struggle to account for uncertainties and variability in soft materials, limiting their effectiveness in real‐world scenarios. Learning‐based controllers can be trained through experience, exploiting various forward models that differ in physical assumptions, accuracy, and computational cost. In this article, the key features of popular forward models, including geometrical, pseudo‐rigid, continuum mechanical, or learned, are first summarized. Then, a unique characterization of learning‐based policies, emphasizing the impact of forward models on the control problem and how the state of the art evolves, is offered. This leads to the presented perspectives outlining current challenges and future research trends for machine‐learning applications within soft robotics.

Remotely operated vehicles (ROVs) can exploit contact with the substrate to enhance their station keeping capabilities. A negatively buoyant underwater legged robot can perform passive station keeping, relying on the frictional force to counteract disturbances acting on the robot. Unlike conventional propeller-based ROVs, this approach has similar, slightly higher efficiency while reducing disturbances to the substrate. Detailed analysis on the passive station keeping performance of an underwater legged robot was performed using Seabed Interaction Legged Vehicle for Exploration and Research 2 (SILVER2) as a reference platform, investigating the effect of leg configuration, net weight, and the nature of the substrate on station keeping performance. A numerical model was developed to study the effect of both geometrical and physical parameters on the station keeping performance, which accurately predicted the station keeping behavior of the robot during field tests. Finally, we defined a metric called station keeping efficiency for the evaluation of station keeping performance; the underwater legged robots showed higher station keeping efficiency (28%) than commercial propeller-based ROVs (11%), showing they could present an alternative for tasks such as environmental monitoring.

Catastrophic forgetting is a phenomenon in which a neural network, upon learning a new task, struggles to maintain its performance on previously learned tasks. It is a common challenge in the realm of continual learning (CL) through neural networks. The mammalian brain addresses catastrophic forgetting by consolidating memories in different parts of the brain, involving the hippocampus and the neocortex. Taking inspiration from this brain strategy, we present a CL framework that combines a plastic model simulating the fast learning capabilities of the hippocampus and a stable model representing the slow consolidation nature of the neocortex. To supplement this, we introduce a variational autoencoder (VAE)-based pseudo memory for rehearsal purposes. In addition by applying lateral inhibition masks on the gradients of the convolutional layer, we aim at damping the activity of adjacent neurons and introduce a sleep phase to reorganize the learned representations. Empirical evaluation demonstrates the positive impact of such additions on the performance of our proposed framework; we evaluate the proposed model on several class-incremental and domain-incremental datasets and compare it with the standard benchmark algorithms, showing significant improvements. With the aim to showcase practical applicability, we implement the algorithm in a physical environment for object classification using a soft pneumatic gripper. The algorithm learns new classes incrementally in real time and also exhibits significant backward knowledge transfer (KT).

Soft robot perception integrates information from distributed, multi-modal sensors, broadening their application to active interaction. Our work introduces recurrent learning models for tactile-based object recognition, demonstrating comparable performance in virtual and real-world scenarios. The work focuses on soft grippers, which facilitate adaptation to objects of varying shapes and sizes thanks to passive finger compliance. Our model successfully identifies over sixteen heterogeneous objects. Findings underscore the significance of sensory multi-modality over single. We highlight how spatial distribution and sensory signal dynamics influence overall estimation accuracy, and what the minimal grasp set is to achieve certain recognition.

The elephant trunk is a highly dexterous muscular hydrostat whose continuous, distributed deformations pose significant challenges for mathematical modeling. We introduce linear “stereotypical” laws that map desired trunk configurations, parameterized by curvature and length, directly to the internal muscle-analogue forces required in our rod-based dynamic model. The trunk is represented as a simplified multi-segment structure of point masses linked through longitudinal and radial muscle analogues and connective tissue, all modeled using rods. Using these laws, the model predicts biological reaching trajectories with tip-position errors below 8% while maintaining hydrostatic volume across trials. The resulting force-shape mappings reveal consistent, repeatable internal force patterns underlying trunk postures, providing a compact representation of actuation strategies that generate specific planar shapes. By reducing high-dimensional continuum dynamics to simple linear relationships, this framework preliminarily enables the inference of muscle-force distributions from shape configurations, laying the groundwork for deeper exploration of the elephant trunk motion strategies and their translation into advanced robotic systems control.