Robotic Manipulation under Uncertainty and Limited Dexterity

Abstract: Robotic manipulators today are mostly constrained to perform fixed, repetitive tasks. Engineers design the robot’s workcell specifically tailoredto the task, minimizing all possible uncertainties such as the location of tools and parts that the robot manipulates. However, autonomous robots must be capable of manipulating novel objects with unknown physical properties such as their inertial parameters, friction and shape. In this thesis we address the problem of uncertainty connected to kinematic constraints and friction forces in several robotic manipulation tasks. We design adaptive controllers for opening one degree of freedom mechanisms, such as doors and drawers, under the presence of uncertainty in the kinematic parameters of the system. Furthermore, we formulate adaptive estimators for determining the location of the contact point between a tool grasped by the robot and the environment in manipulation tasks where the robot needs to exert forces with the tool on another object, as in the case of screwing or drilling. We also propose a learning framework based on Gaussian Process regression and dual arm manipulation to estimate the static friction properties of objects. The second problem we address in this thesis is related to the mechanical simplicity of most robotic grippers available in the market. Their lower cost and higher robustness compared to more mechanically advanced hands make them attractive for industrial and research robots. However, the simple mechanical design restrictsthem from performing in-hand manipulation, i.e. repositioning of objects in the robot’s hand, by using the fingers to push, slide and roll the object. Researchers have proposed thus to use extrinsic dexterity instead, i.e. to exploit resources and features of the environment, such as gravity or inertial forces,  that can help the robot to perform regrasps. Given that the robot must then interact with the environment, the problem of uncertainty becomes highly relevant. We propose controllers for performing pivoting, i.e. reorienting the grasped object in the robot’s hand, using gravity and controlling the friction exerted by the fingertips by varying the grasping force.