Page 189 - Human Inspired Dexterity in Robotic Manipulation

188 Human Inspired Dexterity in Robotic Manipulation

by using the right or left hand and then regrasping it by using the right or left
hand, and (3) once placing an object by using both hands and then regrasping
it by using both hands, etc. Moreover, dual-arm manipulator motions
induce a special topology in the manipulation space. Indeed, the manipula-
tion space is structured into four foliated manifolds: the manifold in which
the robot moves alone, the manifolds in which the left (respectively, right)
hand moves the object, and the manifold in which both hands move the
object. A regrasping strategy should be selected according to the context
of the task. So to seamlessly generate a regrasping motion, we have to con-
struct a manipulation graph that accounts for the special topology of the
manipulation space of dual-arm manipulation.
In this section, we describe the dexterous manipulation planner for dual-
arm manipulators [1]. Here, the manipulation graph is originally proposed in
[2] for a single manipulator. Grasp and placement configurations are, respec-
tively, computed by the grasp planner proposed in [3, 4] and the object
placement planner proposed in [5].

10.2 DEFINITIONS
In this section, we introduce the definitions required to describe a manip-
ulation problem.

10.2.1 Notation

Fig. 10.1 shows the manipulation space used to plan a manipulation task. Let
us consider a robot having n-DOF right arm and n-DOF left arm. Let CS r ,
CS l , and CS o be the configuration space of the right arm, the left arm, and
the object, respectively. The manipulation space is defined by CS ¼ CS r
CS l CS o . Let CS free be the collision-free subset of CS. Let CP be the
domain in CS where the object is stably placed on the environment. Also,
let CG r and CG l be the domain in CS where the object is stably grasped by
the right and the left hands, respectively. CP, CG r , and CG l are subdimen-
sional manifolds in CS.

10.2.2 Grasp Planner

A stable grasp is realized by using the grasp planner proposed in [3, 4]. Before
executing the manipulation planner, the grasp planner generates multiple
candidates of stable grasping postures for a given object. More concretely
speaking, the output of the grasp planner is a set of position/orientation