Researchers led by Iván López-Broceño, with authors including Daniela Rus and Cosimo Della Santina, have built Φ-CON, a model that learns how deformable linear objects such as cables and rods move when forces act on them, according to a paper on arXiv. Many learned models of such objects ignore dynamics and the objects' mechanical response, so they work only for light, soft objects moved slowly.

Φ-CON combines a network of coupled oscillators, whose stored energy depends only on the object's shape, with a learned mapping between measured points along the object and hidden coordinates. Its shape dynamics are stable by design, and the team used its structure to build a controller that compensates for the learned forces and is guaranteed to converge.

On four simulated objects it predicted as accurately as neural ordinary differential equations and Lagrangian neural networks while using 97.2% and 90.2% fewer parameters. Against a model-free controller, it cut the mean error at the object's ends by 97.5 to 99.7% in simulation and by 27.9% on average on a real two-armed robot.

Source: arXiv Robotics (cs.RO)

Background

Iván López-Broceño

Robotics researcher

Robotics researcher.