The research goal is to design a new class of learning controllers for non-linear mechanical systems. A learning controller, similar to the human motor control system, learns the desired feedforward control input through practicing the given task over several cycles. From one cycle to the next, the controller improves its performance based on errors from previous cycles and once converged, under ideal conditions, it executes the task without any error. This research is divided into five tasks. Prior to this study, the Principal Investigator has developed several classes of learning controllers, including the Fourier Based and Finite Element Based repetitive controllers. The first task of this project will further refine these two approaches, and will include proof of stability and experimental verifications. The second task, neural network based learning control, attempts to develop a learning controller which can use the information learned from one task to perform a different task. The third task extends the second to include force control capability. The modification needed for this task is development of a third layer of neural nets to recognize the output pattern of force sensors and learn about characteristics of contact surfaces, such as stiffness and local geometry. The fourth task focuses on the experimental verification of the learning controllers developed earlier. It requires assembling an articulated two degree of freedom manipulator as well as hardware and software implementation of neural network architecture. The fifth task will extend the activity of the first two tasks to include mechanical system flexibility. This work will contribute to the field of dynamic controls and learning systems, and will have immediate application to the fields of robotics and automation.
