Embedding the Physics in Black-box Inverse Dynamics Identification: a Comparison Between Gaussian Processes and Neural Networks

• Published in: IFAC-PapersOnLine 2023-01, Vol.56 (2), p.1584-1590
• Main Authors: Giacomuzzo, Giulio, Turcato, Niccolo', Libera, Alberto Dalla, Carli, Ruggero
• Format: Article
• Language: English
• Subjects: Learning for control; Machine learning; Nonlinear system identification
• ISSN: 2405-8963; 2405-8963
• DOI: 10.1016/j.ifacol.2023.10.1858

In recent years, black-box estimators for robot inverse dynamics have drawn the attention of the robotics community. This paper compares two recent black-box approaches that try to improve generalization and data efficiency by embedding the physical laws governing the system dynamics in two different ways. The so-called Deep Lagrangian Networks (DeLaNs) impose the structure of the Lagrangian equations but do not constrain the basis functions used to model the dynamics. Instead, the Gaussian process model based on the recently introduced Geometrically Inspired Polynomial (GIP) kernel constrains the basis functions of the regression problem to a physically inspired finite-dimensional space but does not force structural properties to be guaranteed. We carried out extensive experiments both on simulated and real manipulators with increasing degrees of freedom (DOF). Our results show that: (i) the accuracy of the DeLaNs model deteriorates much more rapidly than the one of the GIP kernel model with the DOF increase. (ii) the GIP kernel model better estimates the different components of the dynamics, namely, the inertial, Coriolis, and gravitational torques, despite not directly imposing structural properties.