Real-time control of a robot using neural networks

The real-time computation of the robot kinematics is very important. The basic transformations are the direct kinematic transformation (DKT) and the inverse kinematic transformation (IKT). The DKT can be computed in a straightforward way using the Denavit-Hartenberg notation. No such general method yet exists for the IKT, although this transformation is of major interest for control purposes. In this paper a neural network is presented that maps the IKT independent of the type of robot. After training, the network achieves very good accuracy and may easily be implemented in real-time. The performance of the algorithm is tested an the RTX robot, a SCARA-type robot with six degrees of freedom. This robot is controlled by a distributed control system. A host computer realizes the continuous path control and a network of 5 slave-transputers is used to compute the local controls and to drive the DC servomotors