Identification of Deformable Linear Object Dynamics from Input-output Measurements in 3D Space

摘要

Controlling deformable linear objects requires reliable models that capture their complex and high-dimensional dynamics. This paper aims to obtain accurate state-space models of such objects from input-output measurements only, tailored for the use in a model predictive control setting. The proposed identification approach is initialized by a physics-inspired baseline model which approximates a deformable linear object by a serial chain of rigid bodies connected via elastic joints. The nonlinear residual dynamics of the baseline model are then modeled by an artificial neural network, where a key element is the so-called guided residual search that first infers the latent residual and state trajectories in the time domain, in this way facilitating supervised learning of the neural network. The efficacy of the modeling approach is demonstrated on experimental data of the 3D end-point motion of a flexible aluminum rod, excited by a 7-axis Franka Panda robot arm. Compared to a physics-inspired baseline model, the proposed method achieves a 14% prediction performance increase, and compared to a fully neural network-based approach with a similar number of parameters, it achieves a 58% improvement, while also reporting a training time that is more than 5 times faster.