Grasp planning under uncertainty

Abstract

Advanced robots such as mobile manipulators offer nowadays great opportunities for realistic manipulators. Physical interaction with the environment is an essential capability for service robots when acting in unstructured environments such as homes. Thus, manipulation and grasping under uncertainty has become a critical research area within robotics research. This thesis explores techniques for a robot to plan grasps in presence of uncertainty in knowledge about objects such as their pose and shape. First, the question how much information about the graspable object the robot can perceive from a single tactile exploration attempt is considered. Next, a tactile-based probabilistic approach for grasping which aims to maximize the probability of a successful grasp is presented. The approach is further extended to include information gathering actions based on maximal entropy reduction. The combined framework unifies ideas behind planning for maximally stable grasps, the possibilities of sensor-based grasping and exploration. Another line of research is focused on grasping familiar object belonging to a specific category. Moreover, the task is also included in the planning process as in many applications the resulting grasp should be not only stable but task compatible. The vision-based framework takes the idea of maximizing grasp stability in the novel context to cover shape uncertainty. Finally, the RGB-D vision-based probabilistic approach is extended to include tactile sensor feedback in the control loop to incrementally improve estimates about object shape and pose and then generate more stable task compatible grasps. The results of the studies demonstrate the benefits of applying probabilistic models and using different sensor measurements in grasp planning and prove that this is a promising direction of study and research. Development of such approaches, first of all, contributes to the rapidly developing area of household applications and service robotics.