A Criteria-Based Approach to Grasp Synthesis
R. D. Hester, M. Cetin, C. Kapoor and D. Tesar
Robotics Research Group
The University of Texas at Austin
Mail Code R9925, Austin, TX 78712-1100
Recently, NASA-JSC1 designed a robotic hand
to be used as the end-effector of the Robonaut, a dual
arm robot to be used in space station extra-vehicular
activities. The Robonaut Hand [13] is a 12-degrees of
freedom (dof), five-fingered, anthropomorphic
grasping device based on the 95th percentile man's
hand.
The Robonaut Hand design includes a
forearm, wrist with pitch and yaw capability, palm,
palm dof, two primary fingers, two secondary fingers
and a thumb finger (See Figure 1). Robotics Research
Group (RRG) at UT Austin was asked to study the
grasp synthesis problem for the NASA-JSC hand.
This paper presents the approach developed and
applied for the grasp synthesis problem of the
NASA-JSC hand. In Section 2, the common
definitions used in grasp synthesis are introduced and
the previous studies in grasp planning are surveyed
with an emphasis on analytical approaches. In
Section 3, the approach to the grasp synthesis
problem is explained. Although the approach
presented was applied only to the NASA-JSC hand, it
is general and should apply to any arbitrary dof
robotic hand. In Section 4, a case study is introduced
and the results are discussed. Section 5 contains the
summary and conclusions for future work.
Abstract
This study introduces a criteria-based
methodology for grasp synthesis. The method allows
using multiple performance criteria both at the finger
and the hand levels, which are used to generate a
preliminary grasp and an optimum grasp,
respectively. The approach offers reduced complexity
by decomposing grasp synthesis into manageable
phases. A case study using a physical hand
demonstrates the effectiveness of the method.
1.0 Introduction
Grasp synthesis can be defined as constructing a
set of contact locations and a corresponding hand
configuration for a given set of task attributes
associated with a target object. Grasp synthesis, also
known as grasp planning, has enjoyed the attention of
researchers. The approaches to grasp planning can be
classified as empirical and analytical depending on
the procedure used to construct a target grasp. The
viability of these approaches has been tested in a
number of cases with the physical hands designed
and popularized in grasp research [11],[14].
2.0 Preliminaries
The study of grasps can be categorized into two
broad groups. Empirical approaches mimic human
grasping by using heuristics to select one of a set of
standard hand shapes that best conforms to task
requirements and the target object geometry
[4][8][12]. Analytical approaches, on the other hand,
determine the contact locations on the object and the
hand configuration that satisfy task requirements
through kinematic and dynamic formulations. The
method discussed in this paper is an analytical
approach and the remainder of this section reviews
analytical approaches to grasping.
2.1 Definitions
This section introduces various definitions used
in the rest of the paper. Most of these definitions can
Figure 1 NASA-JSC Robonaut Hand
1
National Aeronautics and Space Administration –
Johnson Space Center
1
be found in [6]. The following assumptions are made
for the grasp synthesis approach
1. The grasp is limited to a precision (fingertip)
grasp
2. The object and finger links are rigid bodies.
3. Accurate models of the object and the hand exist.
4. Relative positioning of the wrist frame with
respect to the object frame is available and fixed.
5. No rolling or slipping occurs at the contact
points.
Task objectives must be defined to guide the
grasp synthesis process. The expected loads on the
object can be characterized by a task wrench space
{
2
2
ν 2 ν y ν 2 ω2 ωy ω2
B = (ν x ,L,ω z ) ∈ R 6 , x2 + 2 + z2 + 2x + 2 + 2z ≤ β
δ1 δ 2 δ 3 δ 4 δ 5 δ 6
0 U1
δ 1
M ≡ U∆U T ,U ∈ R 6 , δ ≥ 0
B = [U1 , L,U 6 ]
O
i
i
δ 6 U 6
0
(1)
where Ji is the Jacobian of the ith finger. The grip
map G describes the resultant force on the object due
to forces at the contacts
f c1
Fo = G M , G = [G1 L G n ]
fc
n
(3)
where G + is the transpose of the generalized inverse
of the grip map.
The contacts between the fingertips and object
are assumed point contacts with friction. The friction
cone FC represents the forces at a contact.
T
{
}
f12 + f 22 ≤ µf 3 , f 3 ≥ 0
(4)
Many grasp analysis methods utilize a linearized
form of the friction cone
µ
µ
LFC ci = f ∈ R 3 : f 12 ≤
f 3 , f 22 ≤
f 3 , f 3 ≥ 0
2
2
(8)
2.2 Previous Work
The computational intensity of grasp synthesis
stems from two factors. Analysis of any particular
grasp is complex and the solution space of feasible
grasps is typically very large.
Shimoga [11] surveyed current grasp synthesis
algorithms available in literature and categorized the
popular grasp properties into five basic classes: force
closure, equilibrium, dexterity, stability and dynamic
response. The methods that were reviewed focus on
the analysis of an arbitrary grasp rather than the grasp
synthesis.
Force closure, the ability to resist arbitrary loads,
is a necessary but not a sufficient condition for an
admissible grasp. Ferrari [2] and Pollard [10] use
convexity theory to verify force closure and calculate
a “grasp quality” factor, GQ. This “grasp quality”
measure furnishes a means for comparing force
closure grasps. GQ is a function of the contact
locations on the object. The hand configuration need
not be known.
In addition to force closure, the grasp must
satisfy the equilibrium requirements. External loads
must be balanced without exceeding friction limits at
the contacts or torque limits at the finger joints.
Finding the contact forces that satisfy all the
constraints is complex since the contact force
solution is the sum of the forces that balance the
external load and the internal grasping forces for
which there is a multiplicity of solutions. Kerr [5],
where Gi is the grip map of the i-th contact. The
grasp matrix Gh defines the velocity of the object
with respect to joint velocity and is a function of the
grip map and hand Jacobian
FC ci = f ∈ R 3 :
(7)
where δi is the relative importance ratio of motion in
the direction Ui in the object frame. The twist
ellipsoid is centered on the object frame origin.
(2)
T
Vo = G hθ&, G h = G + J h
(6)
The task wrench space W defines the hyperplanes
orthogonal to the wrench space axes that bound the
set of task loads. Hsu, Li and Sastry [3] represent the
task motion requirements with a task twist ellipsoid:
A grasp is fully defined by the hand Jacobian,
grip map and contact model. The hand Jacobian Jh
describes the contact motion with respect to joint
velocity
0
J 1
x& C = J h (θ )θ&, J h (θ ) =
O
J n
0
}
W = ( f1 ,L,τ 3 ) ∈ R 6 , f i ≤ f i max , τ i ≤ τ i max , i = 1,2,3
(5)
The linearized friction cone LFC defines the largest
pyramid enclosed by the friction cone and
independently constrains the friction terms.
2
Shimoga [11] formulate the solution for the contact
forces as an optimization problem in which the
internal grasping forces are minimized.
These
methods assume a single load condition and
repetitive analysis is necessary for sets of external
loads.
Dexterity criteria provide a means to evaluate
different grasps for a particular task. Shimoga [11]
presents a list of dexterity measures that involve the
grasp matrix Gh. Most of the reviewed criteria had
been developed for redundant manipulators, but their
extensions to multi-fingered hands were discussed.
Stability, the ability to reject load disturbances,
is not essential to grasp synthesis and therefore not
included in this study. In fact, Nguyen [9] determined
that any 3D force closure grasp could be made stable.
Dynamic response, a subset of stability analysis, is
also not considered.
Hsu, Li and Sastry [3][7] present a method for
grasp synthesis based on specifying the grasp map G
and hand Jacobian Jh so that their performance
measure is maximized. The task ellipsoids, grasp
map and hand Jacobian are used to develop
structured quality measures in the twist space and
wrench space.
The grasp analysis methods mentioned so far
evaluate the whole grasp by examining the grasp
matrix Gh or the grip map G, and the hand Jacobian
Jh. This complicates grasp synthesis since it expands
the solution space of feasible grasps by the product of
the feasible finger solutions. We attempt to avoid
this complexity by deconstructing the grasp analysis
to the finger level and reducing the solution space to
the sum of the feasible finger solutions.
(internal) grasping forces. Therefore, force criteria at
the finger level can evaluate the ability to generate
forces at the contact, but provide no information on
the overall effect on the object.
A hand level
criterion that is an extension of a finger level force
criterion must be combined with a force criterion of
the grip map G to provide complete information
about the grasp.
In this study, we propose a two-phase grasp
synthesis methodology that incorporates the
advantages of grasp deconstruction (Figure 2). The
preliminary grasp phase constructs an initial grasp by
selecting the optimum finger configurations based on
finger level criteria. This initial grasp serves as the
starting point for the second phase. The Grasp
optimization phase then searches the grasp solution
space for a locally optimum configuration. The
objective function for this phase is composed of hand
level criteria and GQ, which is a force criterion for
the grip map.
Environment
Constraints
Wrist and
Object
Position
Hand
Constraints
DH Parameters
Fwd Kinematics
Hand
Optimization
Contact point,
Contact normal
Object
Model
Finger
Level
Criteria
Initial Hand
Configuration
Preliminary Grasp
Grasp Optimization
Hand
Level
Contact Model
Grasp
Optimization
3.0 Criteria-based Grasp Planning
Deconstructing grasp synthesis to the finger level
provides two computational advantages:
• Motion and force criteria use the finger Jacobian
Ji and contact map Gi.
• Grasp synthesis is reduced to finding the best
configuration of each finger.
The former advantage reduces computational
intensity by using smaller matrices.
More
importantly, the latter advantage reduces the size of
the search space from the product of the number of
finger solutions to their sum.
The rigid body assumptions allow the grasp to be
fully deconstructed in the velocity domain. The
desired object motion is mapped to motion at a
contact through Gi. The finger motion criteria can
then be evaluated individually using the contact
motion requirements.
The grasp cannot be deconstructed in the force
domain since the forces at a contact are a
combination of the external loads and the inter-finger
Task
Wrench/
Twist Space
Task
Requirements
Contact Points,
Hand Configuration
Grasp
Equilibrium
Contact Points
Hand Config.
Joint Torques
Figure 2 Two-Phase Grasp Synthesis
In the rest of the section, we discuss the finger
level criteria incorporated in this study, their
extension to hand level criteria, and the grasp quality
measure GQ. The implementation of these criteria in
the preliminary grasp and grasp optimization phases
is also presented.
3.1 Velocity Transmission
3
The velocity transmission ratio is a measure of
how effectively the joint velocities of a manipulator
are conveyed to end-effector velocity. This ratio has
been defined in robotics literature [1][15] as follows:
(
( )
α = eiT JJ T
−1
ei
)
−1 / 2
point contacts with friction, are bounded by the
friction cone FC. These directions are approximated
using the vectors, Ufc,j (j=1,…,4), aligned with the
four edges of the linearized friction pyramid LFC.
The force transmission ratio for the direction Ufc,j is
expressed as follows:
(9)
(
(
(
Α j = U Tj Gi J i J iT
)
−1
GiTU j
)
−1 / 2
( )
B j = U Tfc , j JJ T U fc , j
where ei is the direction of interest for velocity
transmission in the manipulator base frame and J is
the Jacobian matrix of the manipulator.
In the case of grasp synthesis, the directions of
interest are defined by the columns Uj (j=1…6) of the
twist ellipsoid matrix U in (8). The object velocities
in these directions can be transformed to velocities at
a contact by the transpose of the contact map Gi. The
velocity transmission ratio for the direction Uj is
expressed as follows:
)
−1 / 2
(14)
The force transmission ratio for a finger is
constructed from the sum of Bj2.
6
η ftr ,i = ∑ B 2j
(15)
j =1
This criterion is equivalent to the velocity term of the
Task Compatibility Index proposed by Chiu [1]. The
velocity transmission ratio for the hand is simply the
weighted sum of the finger criteria.
(10)
n
η ftr = ∑ wiη ftr ,i
where Ji is the finger Jacobian matrix.
The velocity transmission ratio for a finger is
constructed as the sum of Aj2 weighted by the relative
importance ratios δi of the twist ellipsoid directions.
(16)
i =1
The weights wi are assigned the relative importance
of force transmission through the individual fingers.
6
ηvtr ,i = ∑ δ j Α 2j
(11)
3.3 Grasp Quality
An intuitive definition of the grasp quality factor
is that it is a measure of how hard the fingers must
squeeze the grasped object in order to resist external
loads and maintain a grasp. The grasp of an object
can be characterized as a linear combination of the
primitive wrenches imposed by a unit force at each of
the n contacts in a grasp. For point contacts with
friction, four unit forces aligned with the edges of the
linearized friction pyramid are defined.
j =1
This criterion is equivalent to the velocity term of the
Task Compatibility Index proposed by Chiu [1].
The velocity transmission ratio for the hand is simply
the weighted sum of the finger criteria.
n
ηvtr = ∑ wiηvtr ,i
(12)
i =1
The weights wI are assigned the relative importance
of velocity transmission through the individual
fingers.
4n
GS = w | w = ∑ α i wi , α i ≥ 0,
i =1
( ( ))
β = eiT JJ T ei
i =1
i
≤ 1, f ci = 1
(17)
The convex hull of the grasp wrench space GS forms
the bound of external loads that can be resisted by
unit forces at the contacts. The grasp quality factor
GQ is the reciprocal of the amount by which the
grasp wrench space GS is scaled so that it encloses
the task wrench space W.
3.2 Force Transmission
Similar to the velocity transmission criterion of
the previous section, the force transmission ratio a
measure of how effectively joint torques are
transformed into forces at the end-effector. This
criterion has been defined in robotics literature
[1][15] as follows:
−1 / 2
4n
∑α
(13)
where ei is the direction of interest in the manipulator
base frame and J is the manipulator Jacobian matrix.
The directions of interest for a finger, in the case of
k m = min( k ) | W ∈ kGS
(18)
GQ = k m−1
(19)
Interested readers should refer to Pollard [10] and
Ferrari [2] who extensively discuss the calculation
and the derivation of GQ.
4
τ
JTR = ∑ i
τ i max
i =1
n
3.4 Preliminary Grasp
The preliminary grasp phase aims to construct a
preliminary grasp by
• determining the accessible surface areas of the
object for each finger,
• selecting the best configuration for each finger
based on finger-related criteria, and
The surface of the given object is partitioned by
laying a grid on the surface. The grid dimensions are
chosen by considering the size of the search space in
accordance with the actual fingertip contact area.
The finger configuration sets are created by
determining the accessible points on the grid. The
finger configurations are then ranked according to the
following selection criterion:
η f ,i = γη ftr ,i + (1 − γ )η vtr ,i
(22)
where τimax is the torque limit of the ith joint. JTR
represents the joint torques needed to grasp a given
object, and thus is related to the power requirements
for a grasp synthesis. Lower values of JTR are
desired for which the torque requirements on the
joints are decreased. For this case study, the task
wrench space consisted of unit forces in the object
frame.
Second grasp property, grasp matrix velocity
transmission, is defined for motion related tasks
along the major axis of the twist ellipsoid B. It is
given by
(
(
T
vtr (Gh ) = U major
GhGhT
(20)
)
−1
U major
)
−1 / 2
(22)
where Umajor is the direction of greatest importance of
the motion ellipsoid B. For this case study, U was
aligned with the object reference frame and
(δ1,…,δ6)=(1,10,1,1,1,1).
After laying a grid on the cylinder object, all the
collision free, force closure grasps (3240
configurations) were identified. The scatter plots in
Figures 4 and 5 show the variation of the grasp
properties JTR and vtr(Gh) with respect to ηftr and ηvtr
for all these configurations. Also displayed in these
plots are the logarithmic trend lines for the scattered
data.
Figure 3 shows that Joint Torque Ratio has a
decreasing trend with respect to the increasing value
of (ηftr. x GQ)
where γ is the parameter that determines if the task is
motion-oriented (γ<0.5) or force-oriented (γ>0.5).
The finger configurations that maximize the selection
criterion form the initial grasp.
3.5 Grasp Optimization
The resulting hand configuration of the
preliminary phase is chosen as a starting
configuration for the final search routine. A directsearch is conducted on the grid surface by
considering the hand level criteria. The search is
controlled by the following objective function:
η g = γη ftr GQ + (1 − γ )η vtr
2
(21)
1
The task selection parameter γ is the same as for the
finger selection criterion. The search ends when ηg
reaches a local maximum. The calculation intensity
of the search phase is significantly reduced since the
hand level criteria ηftr and ηvtr are algebraic sums of
the previously calculated finger criteria. The convex
hull analysis to determine GQ is the only complex
computation necessary for each grasp evaluation.
0.9
0.8
0.7
JTR
0.6
0.5
0.4
0.3
0.2
0.1
0
0
4.0 Evaluation of the Proposed Method
In order to justify using the performance criteria
of the previous section to get desirable grasp
properties, the grasp synthesis problem for a cylinder
object is considered. The results in this section show
the variation of two grasp properties with respect to
varying force and velocity transmission values.
First grasp property, Joint Torque Ratio (JTR), is
defined for force related tasks. Joint Torque Ratio for
an n-dof hand is given by
0.5
1
1.5
2
2.5
3
ηftr x GQ
Figure 3 Torque Ratio vs. Force Transmission
Figure 4 shows that vtr(Gh) has an increasing
trend with respect to the increasing value of ηvtr.
These results indicate that by optimizing the
hand level criteria ηftr and ηvtr, the desired grasp
properties for the velocity and force related tasks are
enhanced. The final criteria values, and therefore the
5
[2]
grasp properties, also depend on the search routine
used in the final phase.
[3]
0.57
0.56
0.55
vtr(Gh)
0.54
[4]
0.53
0.52
0.51
[5]
0.50
0.49
0.48
40
45
50
[6]
55
η vtr
[7]
Figure 4 Grasp Matrix Velocity Transmission
5.0 Summary and Conclusions
A synthesis approach has been introduced for
grasp planning problem of robotic hands. The
approach tries to decompose the synthesis problem
into two interconnected problems in finger and hand
levels, preliminary and final grasp, respectively. At
each level, a grasp configuration is selected using
different performance criteria. In the case of
preliminary grasp, the selected configuration is fed
into the final grasp phase, which constitutes a starting
point for the grasp optimization. The finger and hand
level criteria are connected, so the two-phase
approach is justified.
The methodology has also been implemented as
an algorithm to be used in the grasp planning of the
NASA-Robonaut hand. Although the computational
requirements/properties of the algorithm is not in the
scope of this study, near real-time performance has
been observed and the final local optimum of the
algorithm has been satisfactory (under the trend lines
presented in Section 4.0).
Future work should focus on developing better
search techniques than the direct-search employed in
this study so that the efficiency of the algorithm can
be enhanced.
Acknowledgements
This research was conducted under the NASA
Grant no. NAG9-809. We thank Dr. H. Aldridge of
NASA-JSC for his comments and discussions during
this project. We also acknowledge the efforts of M.
Pryor and C. Cocca of RRG at UT.
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
References
[1]
S. Chiu, “Task Compatibility of Manipulator
Postures”, Int. J. of Robotics Research, Vol. 7,
No. 5, 1988, pp. 13-21.
6
C. Ferrari and J. Canny, “Planning Optimal
Grasps”, Proc. 1992 Int. Conf. on Robot. and
Auto., pp. 2290-2295.
P. Hsu, Z. Li and S. Sastry, “ On Grasping and
Coordinated Manipulation by a Multifingered
Robot Hand”, Proc. 1988 Int. Conf. on Robot.
and Auto., pp. 384-389.
T. Iberall, et.al. “ Knowledge-Based Prehension:
Capturing Human Dexterity”, Proc. 1988 Int.
Conf. on Robot. and Auto., pp. 82-87.
J. Kerr and B. Roth, “ Analysis of Multifingered
Hands”, Int. J. of Robotics Research, Vol. 4,
1986, pp.3-17.
R. M. Murray, Z. Li and S. S. Sastry, “A
Mathematical
Introduction
to
Robotic
Manipulation”, CRC Press, Boca Raton, (1994).
Z. Li and S. Sastry, “Task Oriented Optimal
Grasping by Multifingered Robot Hands”, Proc.
1987 Int. Conf. on Robot. and Auto., pp. 389393.
H. Liu, T. Iberall and G. Bekey, “The Multidimensional Quality of Task Requirements for
Dextrous Robot Hand Control”, Proc. 1989 Int.
Conf. on Robot. and Auto., pp. 452-457.
V. Nguyen, “ Constructing Stable Grasps in
3D”, Proc. 1987 Int. Conf. on Robot. and Auto.,
pp. 234-239.
N. S. Pollard, “Parallel Methods for
Synthesizing Whole-Hand Grasps from
Generalized Prototypes”, MIT AI Lab. Tech.
Rep. AI-TR 1464, (1994).
K. B. Shimoga, “ Robot Grasp Synthesis
Algorithms: A Survey”, Proc. 1996 Int. Conf.
on Robot. and Auto., pp. 230-266.
R. Tomovic, G.Bekey and W. Karplus, “A
Strategy for Grasp Synthesis with Multifingered
Hands”, Proc. 1987 Int. Conf. on Robot. and
Auto., pp. 83-89.
C. Lovchik and M. Difler, “The Robonaut
Hand: A Dextrous Robotic Hand for Space”,
submitted to the 1999 Int. Conf. on Robot. and
Auto..
G. Hirzinger, “Status Report 1997”, Institute of
Robotics and System Dynamics, DLR German
Aerospace Res. Est., (1997).
M.J. Van Doren and D. Tesar, "Criteria
Development To Support Decision Making
Software for Modular, Reconfigurable Robotic
Manipulators", M.S.E. Thesis, University of
Texas at Austin.