Learning Systems Behavior For The Automatic Correction And Optimization Of Off-line Robot Programs

Robot program generation by means of graphical 08lane simulation tools has become a widely accepted alternative (or additional technique) to on-line teach-in methods. Unfortunately, the generated programs have to be adjusted (re-taught) manually to the production cell’s environment before they finally can be used in the factory plant. The positioning errors resulting from the execution of the generated programs are due to the inaccurate modeling of the robot (geometry, kinematics and esp. the dynamics), the parts and the tools in the graphical simulation package. This paper presents an approach for correcting and optimizing inaccurate robot programs automatically. Sensors measure the di$erence between the robot’s intended and actual path. The errors are used to learn the robot’s system model, without explicitly modeling factors such as damping, backlash, friction etc. After the robot model has been learned, the positioning errors can be compensated for by adjusting the coordinates of the MOVE STRAIGHT commands of the robot programs. The paper describes approaches where lanear and non-linear neural networks are used for the forward and inverse robot system models. Finally simulation results are presented and discussed.