Dynamics Modeling of Robot Manipulators Based on Deep Lagrangian Network and Torque Separation Technique

Abstract

In recent years, learning robot manipulator dynamics with deep networks has been extensively studied, as it avoids deriving the analytical expression of the robot dynamics equations. In particular, deep Lagrangian networks that incorporate the prior knowledge of Lagrangian mechanics into the deep networks have shown prominent advantages in robot dynamics modeling. Inspired by this, this paper proposes a robot dynamics modeling approach based on a deep Lagrangian network and a torque separation technique. First, friction modeling is incorporated into deep Lagrangian networks to more comprehensively learn the dynamics of the robot manipulator. Meanwhile, the inputs of deep Lagrangian networks are reconstructed using the prior knowledge of robot joint types, which can entail valuable information for the model learning process that may result in better approximation and generalization. Moreover, a torque separation technique is introduced to extract the inertial–Coriolis–centrifugal torques and gravity–friction torques from the input torques, which allows us to accurately estimate each part of the robot dynamics and helps us understand the in-depth dynamic characteristics of robot manipulators. The simulation results on two-degrees-of-freedom (2-DOF) and 5-DOF robot manipulators demonstrate the feasibility and effectiveness of the proposed dynamics modeling method.