Relatively robust grasping

Abstract

This thesis presents an approach for grasping objects robustly under significant positional uncertainty. In the field of robot manipulation there has been a great deal of work on how to grasp objects stably, and in the field of robot motion planning there has been a great deal of work on how to find collision-free paths to those grasp positions. However, most of this work assumes exact knowledge of the shapes and positions of both the object and the robot; little work has been done on how to grasp objects robustly in the presence of position uncertainty. To reason explicitly about uncertainty while grasping, we model the problem as a partially observable Markov decision process (POMDP). We derive a closed-loop strategy that maintains a belief state (a probability distribution over world states), and select actions with a receding horizon using forward search through the belief space. Our actions are world-relative trajectories (WRT): fixed trajectories expressed relative to the most-likely state of the world. We localize the object, ensure its reachability, and robustly grasp it at a specified position by using information-gathering, reorientation, and goal-seeking WRT actions. This framework is used to grasp objects (including a power drill and a Brita pitcher) despite significant uncertainty, using a 7-DOF Barrett Arm and attached 4-DOF Barrett Hand equipped with force and contact sensors. Our approach is generalizable to almost any sensor type, as well as wide ranges of sensor error and pose uncertainty.