Neural networks have been studied to control robotic manipulators. Most researches aimed to internalize inverse dynamic models of controlled objects. It has been difficult, however, to obtain true teaching signals of neural networks for learning unknown controlled objects. In the case of robotic manipulators, approximate models of the controlled objects can be generally derived. We believe that the neural networks perform best when they are not required to learn too much. Thus, in this paper, we propose neural networks that do not learn inverse dynamic models but compensate nonlinearities of robotic manipulators with the computed torque method. Furthermore, we show a method to obtain true teaching signals of the neural network compensators.
ASJC Scopus subject areas
- Control and Systems Engineering
- Electrical and Electronic Engineering