Physics-Informed Neural Networks for Robotic Dual-Arm Manipulation of Deformable Linear Objects

The manipulation of deformable linear objects (DLOs) presents several challenges due to the high dimensionality of their configuration space and their complex dynamics. In particular, model-based dual-arm robotic manipulation requires to accurately predict a DLO configuration in response to the applied actions. This thesis proposes a method to model DLOs in 3D space, considering both their shape and orientation. In the presented work, Physics-Informed Neural Networks (PINNs) are introduced as an alternative to standard Neural Networks (NNs). The PINN training framework consists of boundary losses for the ends and a set of physics losses based on the Cosserat Rod theory. Additionally, supervised data from a numerical solver is added to the training. In this experiment, both a PINN and a classic NN have been trained with decreasing amounts of labeled data. All the models are tested on data from standard simulation, simulation with out-of-distribution actions, and real cables. On standard simulation data PINNs achieved significantly better results, especially with low amounts of labeled data. Better accuracies are reached on out-of-distribution data as well, showing lowest absolute errors and a generally lower performance degradation. On the real data PINNs got better results only when trained on smaller datasets, while data driven NNs reached better performances along with the increase of supervised data.