Force control of robot manipulators with neural networks compensation: a comparative study

This article presents a comparative study between the hybrid impedance control approach and the impedance force control approach with neural networks compensation of manipulator modeling errors. The proposed algorithm consists of an outer hybrid impedance control loop that generates the reference (target) acceleration to an inner inverse dynamics control loop. In order to improve the controller robustness, a compensation action of the manipulator modeling errors is introduced, acting on the target acceleration. This compensation action is based on a neural network model achieved by minimizing the modeling errors along the manipulator trajectory. The neural network algorithm uses an error training signal to model errors, that is minimized along the trajectory. The performance of the hybrid impedance control system with neural network compensation, is compared with the impedance force controller with neural network compensation, and is illustrated by computer simulations with a two degree-of-freedom PUMA 560 robot, whose end-effector is forced to move along a frictionless surface located perpendicular to the horizontal plane. The results obtained reveal even better performance when the desired force profile is not constant along the trajectory. This situation is very important in fine motion tasks where the desired force must have a varying profile along the trajectory, namely at the beginning and the end of the task. The results have also shown accurate force tracking and position control, even when significant uncertainties in the robot dynamic model are assumed