Learning and generalizing control-based grasping and manipulation skills

One of the main challenges in the field of robotics is to build machines that can function intelligently in unstructured environments. Because of this, the field has witnessed a trend away from the sense-plan-act paradigm where the robot makes an attempt to model everything before planning and acting. Nevertheless, few approaches to robotic grasping and manipulation have been proposed that do not require detailed geometric models of the manipulation environment. One exception is the control-based approach where closed-loop controllers reactively generate grasping and manipulation behavior. This thesis develops and extends the control-based approach to grasping and manipulation and proposes a new framework for learning control-based skills based on generalized solutions. This thesis extends control-based approaches to grasping and manipulation in several ways. First, several new controllers relevant to reaching and grasping are proposed, including a grasp controller that slides contacts over the surface of an object toward good grasp configurations by using haptic feedback. The number of different grasps that can be generated using grasp controllers is expanded through the use of virtual contacts. In addition, a new approach to statically-stable dexterous manipulation is proposed whereby the robot navigates through a space of statically stable grasp configurations by executing closed-loop controllers. In a series of experiments, grasp controllers are shown to be a practical approach to synthesizing different grasps from a variety of different starting configurations. This thesis also proposes a new approach to learning control-based behaviors by applying a generalized solution in new situations. Instead of searching the entire space of all controller sequences and combinations, only variations of a generalized solution, encoded by an action schema, are considered. A new algorithm, known as schema structured learning, is proposed that learns how to apply the generalized solution in different problem contexts through a process of trial and error. This approach is applied to the grasp synthesis problem, enabling a robot to learn grasp skills with relatively little training experience. The algorithm learns to select an appropriate reach-grasp strategy based on coarse visual context. In an experiment where a dexterous humanoid robot grasps a range of grocery items it had no prior experience with, the learned grasp skills are shown to generalize well to new objects and object configurations