Chapter 2. Prehension

90 THE PHASES OF PREHENSION
able to generate an optimal hand posture for a given object. The input
to the network consisted of two dimensional visual images of different
sized and shaped objects. Cylinders, prisms, and spheres were used,
varying in size from 3 cm to 7 cm. The other input to the network
during the learning phase was prehensile hand shapes, using data
collected from a DataGlove (VPL, Inc., California). The DataGlove
had sixteen sensors recording thirteen flexor/extensor joint angles and
three abductiodadduction angles. Two types of hand postures were
used: palm opposition and pad opposition. During the learning phase,
objects were grasped repeatedly using a trial and error approach. The
network learned the relationships between the objects and postures. As
seen in Figure 4.10, the third layer of neurons is examined in order to
see the internal representation. The level of neuronal activity increased
with object size. The activation patterns for the same objects were
similar. In terms of oppositions, the neuronal activation patterns were
different for pad vs palm opposition. Choosing the correct hand
posture to use based on object properties is an ill-posed problem, since
there are many possible solutions. Therefore, during the optimization
phase, Uno et al. used a criterion function to make the selection. The
function C(y)=I;y2i, where i is the ith output of the sixteen DataGlove
sensors, is mimimized when the hand is flexed as much as possible.
Using relaxation techniques during the optimization phase, the
network minimized this function and computed an optimal hand
posture for a given object.
A problem with these models is the limited number of inputs and
outputs. The versatile performance of human prehension, in contrast,
can be viewed as emerging from a large multi-dimensional constraint
space. In Chapter 7, sources of these constraints are identified and
grouped together into ten different categories. The list brings in the
notion of higher level goals working together with harder constraints.
4.5.2 Choosing virtual finger mappings
An important component to the grasp strategy is the number of real
fingers that will be used in a virtual finger. Newel1 et al. (1989)
studied the number of fingers used in opposition to the thumb as
adults and children grasped cubic objects ranging in width from .8 to
24.2 cm. For the largest objects, two hands were used; more children
than adults used two hands. In general, the number of fingers used in
opposition to the thumb increased with object size. As expected, one
finger was used with the thumb at very small object sizes, and all four
fingers were used in opposition to the thumb for the largest cubes.