Learning an Accurate Neural Model of the Dynamics of a Typical Industrial Robot

To maintain their current signi cance neural networks must compete with other technical methods in realistic applications Control and modelling of nonlinear systems particularly of robots has been established as one of the main elds of application for neural networks Meanwhile little work has been done to introduce neural networks into control of real available machines and to evaluate their advantages over other control methods One should notice that in the eld of robotics there is a broad gap be tween the theoretical description of the systems dynamics e g the well known Newton Euler dynamics equation for rigid body motion and the real robot s behaviour Real e ects include static and coulomb friction with di cult depen dencies on joint positions and velocities nonlinear viscose friction nonlinear gear exibilities backlash and drive motor dynamics On the other hand accurate robot dynamics models are required e g to increase operation security by automatic fault detection Schneider et al or to improve tracking control accuracy by model based control Analytic model based robot control e g Freund has not yet found widespread application This is partly due to the fact that it is not very robust against unmodelled e ects particularly in the presence of controller delay Additional sensors for link positions would be required to deal with gear exibilities Furthermore it is very laborious to derive the complete dynamics equations and identify its parameters for a typical multi joint industrial robot Our goal is to contribute to the feasibility of model based control with a more accurate and more easily obtainable model which is generated by training a set of neural networks This model is capable of representing at least a part of the above mentioned non ideal force components Our previous work Jansen et al a b described a control scheme and a stability guarantee based on model mismatch estimates for the neural controller In the present paper we focus on experimental results of training a model of an industrial robot