Compliance and learning: control skills for a robot operating in an uncertain world

I address problems of uncertainty in robotic systems, in the context of the control of the dynamic behavior of the manipulator, and from an implementation orientation. My thesis is that adaptability rather than explicit modelling and attempted prediction of all phenomena is a viable solution. To show the credibility of this thesis in the robotic domain, two problems involving uncertainty have been attacked. The first is the problem of not knowing exactly the position or motion of an object that lies in the robot work space and that is to manipulated. The second source of perplexity is the phenomenon of unmodelled external torques at the robot arm joints that affect manipulator behavior. These are seen to be representative of problems at the control level in that they involve both environmental and internal states. In the study of positional uncertainties of objects, two tasks were chosen, both involving the application of forces by an actual robot arm. These tasks are surface tracing and crank turning. The necessary adaptability is achieved by implementing compliant behavior. The compliant architecture formed for these tasks is based on a robot position controller. Uncertainties occurring at the joint torque level of the arm itself that affect the performance of such a position controller inspired the second part of the research. This part entails experiments with, in simulation, combinations of linear control algorithms with stochastic learning automata. The goal of the learning controller is to counteract torques occurring at the joints that the linear compensator by itself does not expect. I preface the description of these experiments with an examination of existing schemes for learning and adaptive control.