Ph.D. Thesis Defense by Phillip J. Heil

(Dr. Nader Sadegh, advisor)

"Non-Repetitive Learning Force Controller for a Robotic Manipulator"

Abstract

As robotic manipulators become widespread in manufacturing, force regulation will play an important role in the overall robot controller. Accurate and repeatable force control can reduce part and robot breakage and cut the programming requirement by sensing and reacting to the environment. Learning force algorithms offer an opportunity to model robots and configure controllers without having to go through precise parameter measurement or extensive programming. Learning controllers are already being used for both position and force control, but they have their drawbacks. Often the workpiece has to be carefully placed or precisely modeled. In almost all cases the task the robot has to perform needs to be repetitive and tied to cycle time in some way. The learning controller discussed in this dissertation takes a different approach to learning the characteristics of its environment. The structure of the controller is a proportional-derivative force controller with a feed forward term that anticipates upcoming control actions. Learning takes place as the end effector comes into contact and moves along the surface monitoring and correcting the force error. The learning steps the controller makes are based on the position of the end effector instead of using time as an index, like most other learning force controllers. The result is that once the controller has learned the unknown surface, it can be directed to perform non-repetitive tasks provided the velocity trajectory of the end effector remains the same. A model for a 3-degree of freedom system is presented as well as the synthesis of the proposed controller. A simulation of this system, tracking a sinusoidal surface, is presented with a plain PD controller and the learning controller. The PD controller could not maintain a desired contact force with the surface while the learning controller could track adequately within 8-10 cycles along the sinusoidal surface. To prove the actual usefulness of the learning controller, it was implemented on a gantry-style 2-dof robot with a force sensor end effector. Results showed that the controller could learn an unknown surface, reducing error, despite sensor noise and unmodelled contact forces.