Page 21 - Human Inspired Dexterity in Robotic Manipulation

Digital Hand: Interface Between the Robot Hand and Human Hand 17

medical images such as MRI or CT. When joint angles change, the surface
mesh needs to deform accordingly. Each data point in the surface mesh
(sometimes called a node) is described as a function of related joint angles
to deform the surface mesh. For each node, joints that contribute to the
movement of this node and skinning parameters are defined. Anatomical
landmarks on the skin surface that are later used to define hand dimensions
and virtual markers are also assigned on the nodes.
This surface-mesh model with anatomical landmarks and joint centers
that is obtained by the previously mentioned methods is used as a
template-mesh model for homologous modeling, in which all models con-
sist of the same number of nodes with the same mesh topology. Such
template-mesh models are also called the “generic model.”

2.3 DIGITAL-HAND MODEL WITH SIZE VARIATIONS
2.3.1 Individual Digital-Hand Model

The digital-hand model of a specific person can be generated by the methods
described in the previous section, links and joints from the anatomical image,
medical images, or MoCap data, and the surface mesh from medical images.
Another method to create an individual model is to deform the generic
model by using hand dimensions. This method is more practical than the
first one because acquiring and processing the medical images is usually
difficult and time consuming.
To generate an individual-hand model by deforming the generic model,
hand dimensions, link lengths, and joint centers of the target subject are nec-
essary. Hand dimensions are obtained from traditional manual methods by
using a caliper. Some dimensions can be measured from a two-dimensional
palmar image scanned by a flatbed scanner [12]. As the generic model is
deformed by hand dimensions, the posture of the generic model must be
close to that for a manual hand anthropometry, fully extended position.
The surface mesh of the target person is generated by scaling the generic
model according to the scaling factor at each joint, which are calculated by a
nonlinear optimization method that minimizes the sum of the differences of
hand dimensions between the generic model and target person.
A “deformation transfer” is also used to deform the generic model to the
target model [13]. In the case of a deformation transfer, target dimensions are
reproduced more accurately than the scaling-based method. To reduce the
computational cost while keeping the difference of hand dimensions
between the generic model and the target subject small enough, the