A study of neural network approaches for modeling industrial robot manipulators
Faculty of Computing, Health and Science
School of Engineering and Mathematics / Centre for Communications Engineering Research
In order to effectively apply industrial robots as part of computer controlled systems, it is essential to model robot dynamics. Using neural networks we can formulate the equations governing the dynamics of the robot as a system identification problem. Various features such as degrees of freedom, control system to be adopted, work volume, load-carrying capacity and accuracy and repeatability needs to be considered as part of the process to model robot dynamics. Since recurrent neural networks can be used to approximate any non-linear function, it is possible to use such neural networks to emulate the functions describing the robot dynamics. In this paper we explore the neural networks that emulate the robot dynamic functions. We will analyse the training methods using data obtained by exciting the robot with input signals and measuring the joint positions, velocities, accelerations and input torques. During the training process, joint accelerations in the resulting data sets are used as outputs whereas joint positions, velocities and torque signals are treated as inputs. This paper explores the features and neural network techniques that are suitable.for modeling industrial robot manipulators.