Stable and semantic robotic grasping using tactile feedback

Abstract

This thesis covers two topics of robotic grasping: stable grasping and semantic grasping. The first part of the thesis is dedicated to the stable grasping problem, where we focus on a grasping pipeline that robustly executes a planned-to-be stable grasp under uncertainty. To this end, we first present a learning method which estimates the stability of a grasp based on tactile feedback and hand kinematic data. We then show our hand adjustment algorithm which works with the grasp stability estimator and synthesizes hand adjustments to optimize a grasp towards a stable one. With these two methods, we obtain a grasping pipeline with a closed-loop grasp adjustment process which increases the grasping performance under uncertainty. The second part of the thesis considers how robotic grasping should be accomplished to facilitate a manipulation task that follows the grasp. Certain task-related constraints should be satisfied by the grasp in use, which we refer to as semantic constraints. We first develop an example-based method to encode semantic constraints and to plan stable grasps according to the encoded semantic constraints. We then design a task description framework to abstract an object manipulation task. Within this framework, we also present a method which could automatically construct this manipulation task abstraction from a human demonstration.