Shape Tolerance for Robot Gripper Jaws 1

Tao Zhang, Lawrence Cheung and Ken Goldberg 2

ALPHA Lab, IEOR and EECS Dept., UC Berkeley

ABSTRACT for a given part: both align the part to the desired
orientation and achieve a form-closure grasp on the
In [19] we specified robot grippers that can orient part. The tolerance class is bounded by Jmax and Jmin.
and grasp parts with an arrangement of trapezoidal
jaw modules. Since jaw modules may be imprecisely
machined, we define a parametric tolerance class
such that part alignment is guaranteed for all jaw
geometry in the class. This tolerance class is derived Jmax
based on analysis of toppling, motion trajectory, and
form-closure. Given maximal jaw geometry from the
previous algorithm, we describe an O(n 3 ) algorithm
to compute the parametric tolerance class based on
maximal and minimal jaw specifications. We have
implemented the algorithm and report results from
physical experiments.
Jmin
1. INTRODUCTION

Although grippers are widely used for automated

manufacturing, assembly, and packing, the design of
gripper jaws is often ad-hoc and suboptimal. In Figure 1 Black trapezoids illustrate the maximum and
industry, 4 DOF robots, such as SCARA arms, and 1 minimum boundaries of the jaw tolerance class for
DOF parallel jaw grippers are common due to their
the part shown in gray.
low cost and high reliability. The combination of
these two devices is kinematically limited to orient- As illustrated in Figure 2, the gripper with trape-
ing parts in the horizontal plane. Zhang and Goldberg zoidal jaw modules rotates the part from its initial
[19] gave an algorithm to design jaws based on trape- resting orientation (a) to the desired final orientation
zoidal modules that will align parts in the vertical (b) for assembly. Our gripper design builds on recent
plane and grasp them in form closure. The algorithm results in toppling manipulation [8]. Zhang et al. [17]
finds jaws that achieve maximal contact at the final propose the toppling graph that can be used to iden-
grasp configuration to maximize resistance to applied tify the location of contacts permitting toppling.
forces. Zhang et al. [18] apply toppling to grasping and find
For many industrial applications, it may be pref- four frictionless point contacts that will align a given
erable to use jaws with smaller contact area, for ex- part in the vertical plane. [19] gives an O(n 5 ) algo-
ample to minimize gripper weight for high-velocity rithm to compute the maximal jaw design with linear
transfer. Furthermore, machined jaws may not pre- contacts that has the following properties: (1) It is
cisely comply with the specified maximal contact able to align the part from the initial orientation to the
geometry. In this paper we consider variations in jaw desired final orientation; (2) It has maximal (linear)
shape and define a tolerance class for jaws based on contact with the part at the desired orientation of the
maximal and minimal contact areas. part; and (3) It achieves a form-closure grasp on the
Let Jmax denote the jaw specification from [19] part at its desired orientation.
that achieves maximal linear contacts with the part at In this paper, we develop a tolerance class speci-
its desired final orientation. Let J denote an instance fied as a range of trapezoidal jaws. Each jaw module
of jaw geometry that is defined by an arrangement of is determined by the locations of two vertices that
trapezoidal jaw modules. We say J is admissible if it make contact with the part in its final grasp configu-
will rotate the part to the desired orientation and ration. The line segment between these two vertices
achieve form-closure. Let Jmin denote the admissible represents an accessible segment on an edge of the
jaw geometry with minimal contacts at the desired part at its desired orientation. The accessible segment
orientation of the part. Figure 1 shows Jmax and Jmin corresponds to an edge of the jaw module that is nei-
ther horizontal nor vertical.

1
This work was supported in part by the National Science Foundation under CDA-9726389 and Presidential Faculty
Fellow Award IRI-9553197. Research funding was also provided by Adept Technology, Ford Motors, and Califor-
nia State MICRO Grant 00-032.
2
For more information contact: [email protected].
 eralized configuration space and describes multi-step
error detection and recovery strategies. Joskowicz et
al. [6] present kinematic tolerance in term of configu-
ration space and develop a worst-case tolerance
analysis algorithm for 2-DOF planar pairs. Sack and
Joskowicz [13] extend the analysis to multi-pair pla-
nar mechanisms with statistical geometric variation.
They also model general planar part pairs using 3-
Figure 2 Gripper with trapezoidal jaw modules ro - dimensional configuration-space to capture both
tates the part in the gravitational plane to facilitate quantitative and qualitative kinematic variation [14].
assembly. Latombe et al. [7] considers assembly sequence
planning problem with toleranced parts. They give a
There are three types of vertices in the set of ac- polynomial time algorithm to decide if an assembly
cessible segments: one pushing vertex, one toppling sequence exists given the specified tolerances. Their
vertex, and other vertices named tips. As illustrated in tolerance model is similar to ours in that both
Figure 3, we define a single variational parameter λ approaches fix the relative orientation of edges.
along the edge at each tip. As is common in toler- Akella and Mason [1] develop a planner to generate
ance analysis, we assume perfect form: all jaw mo d- orienting plans for toleranced polygonal parts. Their
ules in the tolerance class have perfect linear edges. tolerance model is defined by circular uncertainty
We define the tolerance class by fixing the pushing zones around the nominal positions of the COM and
and toppling vertices and computing how far the tips the vertices. Chen et al. [4] propose parameteric
can be expanded or contracted along the accessible tolerance classes for sensorless part orienating and
segments. Note that we define a single common tol- fixturing. Each are defined by a uncertainty zone at
erance parameter λ for all tips. part vertices. They develop algorithms to compute the
boundaries of the tolerance class. Brost and Peters [3]
give an algorithm to design 3D modular fixtures for
λ toleranced parts that are specified by an uncertainty
polygon at each vertex. Bohringer et al. [2] show that
toleranced parts can be oriented using an elliptic
force field.
λ

3. PROBLEM DEFINITION

Let I denote the input to the maximal jaw design al-

gorithm in [19]: the n-sided convex projection of an
Figure 3 Variational parameters along an accessible extruded polygonal part, its COM, its initial and de-
segment of a jaw module. sired orientations, vertex clearance radius ε, µt and
µs : friction coefficients of gripper-part and surface-
We present an O(n 3 ) algorithm for testing if a
part, respectively.
jaw specification J is admissible. Given Jmax, we then
We first consider problem (1), testing if a given
present an O(n 3 ) algorithm to compute the tolerance J is admissible (will rotate the part and hold it in
class. form-closure). The input to problem (1) is <I, J>.
The output is binary: yes if J is admissible; no if not.
We then consider problem (2), finding the
2. RELATED WORK
lower boundary of the tolerance class. The input of
problem (2) is <I, Jmax>. The output is Jmin.
[9], [11], [12], and [15] survey the status quo and
mathematical approaches to tolerancing. Neumann
[10] describes a new standard, Y14.5M, which pro-
vides a mathematical basis for dimensioning and tol-
erancing.
A fundamental problem in geometric toleranc-
ing is classification: given a part, is it within toler-
ance? Yap and Chang [16] give an example using a
1-dimensional probe model.
Configuration space can provide a theoretical ba-
sis for tolerance analysis. Donald [5] studies part ma-
nipulation with geometric uncertainty. He considers
shape variations as an additional dimension in a gen- Figure 4 Notation.
 For toppling to be successful, there must exist a
As shown in Figure 4, the part sits on a work- horizontal line at height h that has the following
surface at an initial resting pose. We define the World characteristics:
frame, W, to be a Cartesian coordinate system origi-
nating at pivot point P with X-axis on the surface 1. h > Hi (θ), if Vi (θ) < h < Vi+1 (θ); #1
pointing right, Z-axis vertical to the surface pointing 2. h > Ji (θ), if Vi (θ) < h < Vi+1 (θ); #2
up. The pushing contact, A’, is a distance zA’ from the 3. h < max (Vi (θ)), where 0 < θ < θd . #3
surface; the toppling contact, A, is a distance zA from i

the surface. Starting from the pivot, we consider each where i is the index of visible edges.
edge of the part in counter-clockwise order, namely The first three criteria can be described as: the
e1 , e2 , …, en . The edge ei , with vertices vi at (xi , zi ) toppling contact A must be above the toppling func-
and v(i+1) at (x(i+1), z(i+1)), is in direction ψi from the X- tion, the jamming function, and the liftoff function.
axis. When the horizontal line crosses a vertex function,
Let θ denote the rotation angle of the part from there is a contact edge switch. Therefore, A must sat-
the +X direction; initially θ =0 and at the final orien- isfied criteria 1 and 2 for the new contact edge. The
tation θ =θd . We say an edge ek is visible if it can be third criterion requires that A must make contact with
seen from +X direction; invisible, otherwise. the part.
We assume the part can be treated as a rigid ex- For example, we want to rotate a sample part 35º
trusion of a polygon; both the part and the jaws are for assembly. Figure 5 illustrates the toppling graph
rigid; part geometry and location of the COM are of the part given zA’ = 0.5cm. Note that H2 and J2
known; part motion is sufficiently slow to apply equal 0. We can see that A at zA = h 1 is unable to top-
quasi-static analysis. ple the part to the desired orientation because the line
goes under H3 after rotating to θ1 ; A at zA = h 2 is ca-
pable to perform the task. Notice that A switches con-
4. TOLERANCE ANALYSIS tact edge from e2 to e1 at θ2 .
Given zA’, the toppling graph allows us to find
The tolerance analysis is a combination of toppling, the feasible range of zA such that the corresponding A
motion trajectory, and form-closure study. and A’ can rotate the part from the resting orientation
to the desired orientation.
4.1 Toppling graph
Our analysis involves the graphical construction 4.2 Trajectory analysis
of a set of shape functions that represent the mechan- To ensure no portion of the jaw blocks the part
ics of grasping. All of these functions are piecewise rotation, we define quasi-vertex functions to represent
sinusoidal and dependent on θ. They map from part the motion trajectory of vertices.
orientation to height: S1 →ℜℜ +, where S1 is the set of The part performs both rotation and linear trans-
planar orientations. The shape functions include ver- lation during toppling. We decompose the part mo-
tion into pure rotation and pure translation. The part
tex functions Vj (θ), toppling functions Hj (θ), and
first rotates about pivot point P to semi-position, and
jamming functions Jj (θ).
then translates to actual-position. Let (θ xj , θ zj ) and (θ
x’j , θ z’j ) denote the actual-position and the semi-
J3 Z position of vertex vj after the part is toppled by θ,
respectively. Let (d xj , d zj ) and (d x’j , d z’j ) denote the
V5 V4 actual-position and the semi-position of vertex vj after
H3 V3
80 h1 the part is toppled to its desired orientation, respec-
θ1
V2 tively. Let θxt and d xt denote the distance between the
H1 θ2 actual-position and the semi-position of any point
40 h2
after the part is toppled by θ and θd , respectively. To
obtain a quasi-vertex function, we define a frame of
0 30 60
J1 -
θ reference Fj at the desired orientation of the part
originating at vj . The Z-axis of Fj is the interior nor-
mal of edge e( j-1), and the X-axis is on edge e( j-1) obey-
Figure 5 Toppling Graph. ing the right-hand rule.
Given zA , the quasi-vertex function Qjk(θ, zA ) in-
The toppling graph, which consists of these dicates the location of vk in Fj as the part rotates,
shape functions, helps us to identify the range of the which can be shown to be:
contact permits toppling. Each function represents a
particular property of the part, and the graph de-
Qjk(θ) =  x Q jk (θ, z A ) 
scribes properties of the part during grasping.  Q (θ, z ) 
 z jk A 

=
 ( θ x' k −d x' j +θ xt − d xt ) cos(ψ j −1 + θd ) +( θ z 'k − d z ' j ) sin(ψj −1 + θd )  equivalent to solve a system of equations and can be
  done in O(n 3 ).
 
−( θ x' k −d x' j +θ xt − d xt ) sin(ψj −1 + θd ) +( θ z 'k − d z ' j ) cos(ψj −1 + θd )  Second, we need to test if J is able to rotate the
  part to the desired final orientation.
 
#4 Since the part is rolled by the pushing contact
and the toppling contact, we need to identify these
 d x' j   x j cosθ d − z j sinθ d  two points. These two contacts are only vertices that
where  0  =  0 , keep touch with the part during the toppling phase.
    This can be done easily in time O(n).
 d z' j   x j sinθ d + z j cosθ d 
    Known the height of the pushing contact zA’, we
 1   1  construct the corresponding toppling graph. If h = zA
satisfies inequality #1 ~ #3, the pair of A and A’ can
( z A − d z ' m )(d x ' m +1 − d x ' m ) rotate the part to the final orientation. Since the time
d xt = xA - - d x’m if d z’m <
to obtain a toppling graph is O(n), this step takes
d z ' m+ 1 − d z ' m
O(n).
zA < d z’m+1 , Finally, we need to consider if any portion of J
will block the part’s trajectory.
( z A −θ z ' l )(θ x' l + 1 −θ x ' l ) - x’ if z’ <
and θ xt = xA - θ l θ l
θ z ' l +1 − θ z ' l z x
zA < θ z’l+1 .

We represent the motion trajectory of the edges Fj Frame O

of the part based upon the quasi-vertex functions, and
then we derive the accessible segments of the jaws.
To guarantee no obstacle blocks the part rotation, the z
jaws should stay away from the motion trajectory of I
the edges.
The quasi-vertex function describes the motion
trajectory of the part. Note that the quasi-vertex func-
tion is the projection of configuration-space (x, z, θ)
onto the plane of (x, z); the shape function is the pro-
ject of configuration-space (x, z, θ) onto the plane of
(θ, z). Therefore, the shape function and the quasi- W Frame
vertex function are both the decomposition of the
configuration-space. The reason we apply C-space x
decomposition is that, in order to reduce complexity,
we only need to keep a portion of the configuration- Figure 6 Motion trajectory of the part.
space information that is necessary for certain analy-
sis. From equation #4, we have:

4.3 Problem (1): Checking if J is admissible

x Q jk 2 + z Q jk 2 = (θ x'k −d x' j +θ xt −d xt )2 + (θ z 'k −d z ' j )2 .
To solve problem (1) — checking if a given J is
admissible, we first test if J will achieve form-closure
on the part. Therefore, the intersection between Qjk(θ) and the X-
Starting from the one closest to P, we order the axis of Fj is at:
vertices of J in counter-clockwise (a 1 , b 1 ), (a 2 , b 2 ),
…, (a m , b m ). Let Vk denote the unit normal vector 2
xjk =  ( d x' j + d xt ) −( x'k + xt )  +  z' k − d z ' j 
2

pointing inward at (a k, b k), Vka (Vkb) denote the X (Z)-   θ 

~ ~ ~
θ θ
axis projection of this unit vector, and Tk denote the
torque of Vk respective to P. ~
V1a V2a V3a ... Vma  where θ satisfies:
Let M denote V V V3b ... Vmb  , let ω
 1b 2b  ~
 T1 T2 T3 ... Tm  z Qjk (θ, z A ) =
denote [ω1 ω 2 ω 3 ...ω m ]T . −( ~ x 'k − d x' j + ~ xt − d xt ) sin(ψj −1 + θd ) +( ~ z' k − d z ' j ) cos(ψj −1 + θd )
θ θ θ

It is well known that J generates a form-closure = 0.

grasp on the part if and only if ∃ ω >0, s.t. Mω
ω = 0.
Therefore, to check the form-closure grasps is Figure 6 illustrates the motion of the part. O is
the origin of Fj at vj , and I is the intersection between
the quasi-vertex function and the X-axis of frames Fj . λ must be nonnegative. We choose a small positive
Therefore, xjk is the length of OI . We project OI to number δ. Starting with λ = δ, we use the algorithm
the X-axis of W, and the length of the resulting seg- for Problem (1) to check if the corresponding J is
admissible. If so, we try λ = 2δ, and so on, until λ is
ment is ( d x' j + d xt ) −( ~ x ' k + ~ xt ) ; we project OI sufficiently large that J is not admissible. We then
θ θ
to the Z-axis of W, and the length of the resulting interpolate to a desired level of accuracy and the cor-
responding J is Jmin.
segment is ~ z 'k − d z ' j . If xjk of J is smaller than that
θ
of Jmax (where j and k correspond to all the visible
5. IMPLEMENTATION RESULTS
edges), no portion of the jaws will become an obsta-
cle in the trajectory. We compute xjk of J and that of
Jmax, and compare these two values for all j and k. We verify our shape tolerance algorithms by the fol-
lowing example. The part is initially at the stable
Thus the algorithm to solve Problem (1) runs in time
orientation defined by the vertices at (0,0), (51.2, 0),
O(n 3 ).
(64.1, 57.2), (37.5, 96.2), (-32.2, 44.6), and COM at
(21.9, 42.3). We need to rotate the part 20º to final
4.4 Problem (2): Computing the tolerance class
[Jmin, Jmax] defines a tolerance class: the un- orientation for assembly.
countable set of grippers with jaws having edges par-
allel to Jmin and Jmax and volume within these lower a1 46.41 b1 21.64
and upper boundaries. We first prove the convexity a2 33.94 b2 78.97
of the tolerance class: if J ∈ [Jmin, Jmax], J must be a3 33.94 b3 78.97
admissible. Then we describe an algorithm to find a4 5.93 b4 95.61
Jmin by searching the upper bond of λ. a5 -6.65 b5 103.08
a6 -39.76 b6 42.04
Lemma 1. If J ∈ [Jmin, Jmax], J must be admissible. a7 -48.04 b7 26.78
Proof: Note that both Jmin and Jmax are admissible
a8 -6.17 b8 4.55
by definition.
(1) Can J topple the part to desired orientation?
First, we consider the toppling conditions. Since the Table 1 Optimal jaw design: vertex location.
pushing contact and the toppling contact are the same
for all J in the class, the toppling condition of J is the We find the optimal gripper jaw design as shown
same as that of Jmax. Second, we consider the part’s in Figure 7. Table1 indicates the location of the jaw
motion trajectory conditions. The smaller J, the less module vertices. The toppling contact is at (a 2 , b 2 )
likely it will block the motion trajectory of the part. and the pushing contact is at (a 6 , b 6 ). We apply our
Since Jmax is the geometry shape that guarantees no algorithm to find the upper bond of λ equals 22 as
collision in the part’s motion trajectory and J is illustrated in Figure 8.
smaller than Jmax, J satisfies the motion trajectory
conditions. Therefore, J is able to topple the part to
desired orientation.
(2) Can J achieve a form-closure grasp on the part at
its desired orientation?
J is larger than Jmin. Since Jmin achieve a form-
closure grasp on the part at its desired orientation, J
must have the same property.
In summary, J can topple the part to desired orienta-
tion and achieve a form-closure grasp on the part at
its desired orientation. Therefore, J is admissible. g Figure 7 Jmax rotates the part and grasps it securely at
the desired orientation.
Lemma 2. λ must be nonnegative for J.
Proof: Assume that there exists a λ that is negative We conducted physical experiments to verify our
for certain jaw geometry J. Then, the total length of results on the example part. Two sets of jaws were
the contact edges of J are longer than that of Jmax machined from aluminum. The friction coefficients
because λ < 0. Since Jmax has maximal contacts with are µt = 0.0875 and µs = 0.0875. The first set is Jmax
the part at the desired orientation, some portion of J as shown in Figure 7. The second set is Jmin as shown
will become an obstacle in the part rotation trajec- in Figure 8. We installed these two sets of the jaws
onto an AdeptOne industrial robot. We tested each
tory. Therefore, λ can only be nonnegative. g
Jmax and Jmin 50 times to align the part and observed
zero failures. Figure 7 and 8 illustrate, in sequence,
Numerical Algorithm: We use binary search to find
the maximum variational parameter λ. By Lemma 2,
both jaw sets successfully rotate the part to the de- in Robotics: The Algorithmic Perspective,
sired orientation and grasp it securely. ed. P. Agarwal, L. Kavraki, and M. Mason.
(a2) A. K. Peters, 1999.
(a1)
[5] B. Donald. “Planning multi-step error detec-
tion and recovery strategies,” Int. J. Robot.
Res., vol. 9, no. 1, pp. 3-60, 1990.
[6] L. Joskowicz, E. Sacks, and V. Srinivasan.
“Kinematic tolerance analysis,” Computer-
( a3) Aided Design, Vol. 29, No. 2, pp. 147-157,
1996.
[7] J.-C. Latombe, R. Wilson, and F. Cazals.
“Assembly sequencing with toleranced
parts,” Computer-Aided Design, Vol. 29,
No. 2, pp. 159-174, 1997.
[8] K. Lynch. “Toppling manipulation,” in
Figure 8 Physical experiments for Jmin.
IEEE Int. Con. Robot. Automat., Detroit,
1999, pp. 2551-2557.
[9] J. Meadows. Geometric Dimensioning and
6. DISCUSSION AND FUTURE WORK
Tolerancing, Marcel Dekker, Inc., 1995.
[10] A. Neumann. “The new Y14.5M standard
It is very difficult to characterize the grasping pro-
on dimensioning and tolerancing,” Manufac-
prieties of the jaws with uncountable shape uncer-
turing Review, Vol. 7, No. 1, pp. 9-15, 1994.
tainty. Therefore, we intend to quickly check the
[11] A. Requicha. “Mathematical definition of
orientability of jaws during interactive design cycle.
tolerance specifications,” Manufacturing
Algorithms with low complexity, such as those de-
Review, Vol. 6, No. 4, pp. 269-274, 1993.
scribed in this paper, can provide rapid feedback to
[12] U. Roy, C. Liu, and T. Woo. “Review of
designers.
dimensioning and tolerancing: representa-
We propose a rigorous parametric tolerance class
tion and processing,” Computer-Aided De-
to address the shape uncertainty of the gripper jaws.
sign, Vol. 23, No. 7, pp. 466-483, 1991.
We study shape tolerance of the jaws in terms of top-
[13] E. Sacks and L. Joskowicz. “Parametric ki-
pling graph, part’s motion trajectory, and form-
nematic tolerance analysis of planar mecha-
closure. We present a fast checking algorithm, and
nisms,” Computer-Aided Design, Vol. 29,
use it to compute the tolerance class. We implement
No. 5, pp. 333-341, 1997.
the algorithms and illustrate with physical examples.
[14] E. Sacks and L. Joskowicz. “Parametric ki-
In the future, we will study sensitivity to changes
nematic tolerance analysis of general planar
in friction coefficient and consider alternative materi-
systems,” Computer-Aided Design, Vol. 30,
als for gripper jaws. We will also consider sensitivity
No. 9, pp. 707-714, 1998.
in jaw shape normal to the contacting surfaces, which
[15] H. Voelcker. “A current perspective on tol-
may justify use of deformable materials such as rub-
erancing and metrology,” Manufacturing
ber for the contacting surfaces. The idea is to design
Review, Vol. 6, No. 4, pp. 258-268, 1993.
jaws that are also robust to variations in part shape.
[16] C. Yap and E. Chang. “Issues in the metrol-
ogy of geometric tolerancing,” in J.-P. Lau-
mond and M. Overmas, editors, Algorithms
REFERENCES
for Robotic Motion and Manipulation, pp.
393-400, A. K. Peters, 1997.
[1] S. Akella and M. Mason. “Orienting toler-
[17] T. Zhang, G. Smith, R. Berretty, M. Over-
anced polygonal parts,” Int. J. Robot. Res.,
mars, and K. Goldberg. “The toppling
vol. 19, no. 12, pp. 1147-1170, 2000.
graph: designing pin sequences for part
[2] K. Bohringer, B. Donald, L. Kavraki, and F.
feeding,” in Proc. IEEE Int. Conf. Robot.
Lamiraux. “Part orientation with one or two
Automat., San Francisco, 2000, pp. 139-146.
stable equilibria using programmable force
[18] T. Zhang, G. Smith and K. Goldberg.
fields,” IEEE Trans. Robot. Automat., vol.
“Compensatory grasping with the parallel-
16, no. 2, pp. 157-170, 2000.
jaw gripper,” in 4 th Int. Workshop Algo-
[3] R. Brost and R. Peters. “Automatic design of
rithmic Foundations Robot., Hanover, NH,
3-D fixtures and assembly pallets,” Int. J.
2000.
Robot. Res., vol. 17, no. 12, pp. 1243-1281,
[19] T. Zhang and K. Goldberg. “Design of grip-
1998.
per jaws based on trapezoidal modules,” in
[4] J. Chen, K. Goldberg, M. Overmas, D. Hal-
Proc. IEEE Int. Conf. Robot. Automat.,
perin, K. Bohringer, and Y. Zhuang. “Shape
Seoul, Korea, 2001.
tolerance in feeding and fixturing,” in Ro-