The behavior of a robotic manipulator is effected by the torques that drive the joints of the manipulator. Given a sequence of torque signals, the motion of the manipulator can be predicted based on an accurate mathematical model oi the manipulator. Controlling a manipulator represents the “inverse" situation whereby the desired motion of the manipulator is given and the sequence of torque signals which produces such motion is to be determined.

The desired torque signals could in principle be generated based on the model and the prescription of the desired motion. This method of deriving the torque signal is called the “computed torque” method. Because this control method is based on the mathematical model of the manipulator, any “mismatch" between the model and the real system degrades the performance of the manipulator. It is owing to such mismatch that uncertainty about the mode! arises. To improve the performance of the manipulator requires that the uncertainty be compensated.

A new approach to improving the performance of uncertain robotic systems using a neural network is presented in this dissertation, it is shown that this approach is applicable to ( 1) robot free motion, (2 ) robot compliant motion, and (3) multi-manipulator systems. In this approach, a neural network is used to “nullify” the uncertainty so that performance improvement can be achieved.

Using techniques from nonlinear system theory, closed-loop stability of each of the three types of robotic system (incorporated with a neural network) is analyzed. Results of the analyses confirm that the systems are stable in the sense that all signals in each system are bounded.

A new method for analyzing the performance of these systems is developed. Using this method, it is shown that the performance of all three types of system is improved as the learning process of the neural network is iterated. Numerical simulations are conducted. The results of the simulations confirm the conclusions of the theoretical analyses.

Two types of experiment, one involving robot free motion and the other involving robot compliant motion, are conducted. The results of the experiments clearly demonstrate the effectiveness of the proposed approach.