| With increasing demands for faster, more accurate and reliable robots, the field of robotics has faced the challenges of reducing the required on-line computational power, calibration time, and engineering cost when developing new robot controllers. The robot arm position control is a complex kinematic and dynamic problem and has received researchers' attention for quite some time. During the last several years, most research on robot control has resulted in effective but computationally expensive algorithms.;The focus of this dissertation is to systematically design an improved robot controller based on time-varying neural networks in place of designs using conventional and adaptive robot controllers. The utility of the approach was demonstrated by providing results of the gross motion control of the first two joints of a SCARA robot (AdeptOne) through simulation. The robustness of the controller to parameter variations (noise and mass disturbances) was also effectively demonstrated along with the ability to maintain tracking accuracy when faced with new trajectories. Unlike most other control schemes, learning is done iteratively, based on observations of input and output relationships of the system in motion rather than having to specify the explicit model of the system. The improved controller based on the temporal ability of the simple recurrent network structure has provided the time-varying attribute needed for dealing with real-time production tasks during implementation. |
