IEEE TRANSACTIONS ON ROBOTICS, VOL. 22, NO.

1, FEBRUARY 2006 43

Kinematics for Multisection Continuum Robots

Bryan A. Jones, Member, IEEE and Ian D. Walker, Fellow, IEEE

classes of shapes and motions, restricting their potential appli-

AbstractWe introduce a new method for synthesizing kine- cability.
matic relationships for a general class of continuous backbone, or Hirose and colleagues designed the serpeniod curve, which
continuum, robots. The resulting kinematics enable real-time task closely matches the kinematics of a snakes body in 2-D as it
and shape control by relating workspace (Cartesian) coordinates
to actuator inputs, such as tendon lengths or pneumatic pres- crawls across the ground [15]. However, this work does not
sures, via robot shape coordinates. This novel approach, which address the 3-D kinematics of the many snake-like robots they
carefully considers physical manipulator constraints, avoids arti- designed, instead focusing on their planar dynamics. Chirik-
facts of simplifying assumptions associated with previous ap- jian and Burdick took the opposite approach, in which they
proaches, such as the need to fit the resulting solutions to the link the shape of a particular mathematical curve to a high
physical robot. It is applicable to a wide class of existing contin- degree of freedom robot [6], [7], [8]. Their modal decomposi-
uum robots and models extension, as well as bending, of individ-
ual sections. In addition, this approach produces correct results tion then reduced the number of degrees of freedom necessary
for orientation, in contrast to some previously published ap- to control this curve to which a robot was fitted, such as the
proaches. Results of real-time implementations on two types of 30 degree-of-freedom robot discussed in their work. However,
spatial multi-section continuum manipulators are reported. few continuum robots match the proposed curve, limiting its
applicability. Although recent work [11] extends this approach
Index TermsContinuum robot, kinematics, trunk, tentacle, by introducing a more convenient curve which eliminates the
biologically inspired robots
problem of mode switching, this method likewise relies on
approximating the shape of an actual robot. An alternate ap-
proach is to exactly compute the kinematics of a high degree
I. INTRODUCTION
of freedom manipulator composed of spherical joints and rigid

C ONTINUUM robots, in contrast to traditional rigid link

robots, feature a continuous backbone with no joints [24],
as illustrated in Fig. 1. In this sense, they are essentially inver-
links using standard D-H methods [32]; however, this ap-
proach cannot be applied to continuum robots, which contain
no prismatic or revolute joints. The authors of [14] returned to
tebrate, as opposed to the vertebrate-like structure of conven- Hiroses concept by choosing a curve an arc of a circle
tional robots. The wide range of abilities demonstrated by which closely matches the kinematics of continuum robots
invertebrate limbs such as elephant trunks [9], [14], [28], oc- due to the equal distribution of forces inherent in the design of
topus arms [19], [23], and squid tentacles [30] have motivated
a recent surge of research activity in continuum robots. Poten-
tial applications include navigation and operation inside com-
plex, congested environments, such as collapsed buildings in
search and rescue operations [4], [29], [32] or the human body
in medical applications [25], [26]. Continuum robots also en-
able novel forms of environmental interaction, via compliant
and/or whole arm manipulation [13].
Realization of continuum robots proves to be a very inter-
esting and challenging problem. A large proportion of efforts
in the area have focused on design and construction of contin-
uum robot hardware. To date, a number of interesting hard-
ware prototypes have been produced [1], [9], [14], [15], [19],
[22], [27], as well as several commercial products [3], [16],
[28]. These manipulators demonstrate the strong potential of
continuum robots. In comparison, efforts in modeling and
real-time algorithm development to extract the full physical
potential from these robots have lagged behind. For example,
almost all previous approaches either do not address trunk
orientation and either approximate positional kinematics or
solve them only for limited kinematic models, which constrain
the performance of continuum robot hardware to reduced Fig. 1. The Clemson Air-OCTOR multi-section continuum
robot.
44 IEEE TRANSACTIONS ON ROBOTICS, VOL. 22, NO. 1, FEBRUARY 2006

continuum robots [13], an assumption also made throughout joints, with the equivalent to rotations i and translations d i
this paper. However, the model in [14] assumes an inexten- being an arc of constant curvature modeled by trunk parame-
sible backbone, does not correctly determine the orientation of ters s, , and where s gives trunk length, determines cur-
the trunk tip due to a simplifying assumption in the model, vature, and determines the angle of curvature as shown in
and is limited to the particular physical design (a unique ac- Fig. 2. A critical question for continuum robot kinematics is
tuator arrangement) of hardware implemented by the authors how to modify conventional robot kinematics to fit this case.
of [14]. A fundamental concept exploited in this paper, first intro-
These approaches constrain the performance of contin- duced by [14], is to fit a conceptual conventional rigid-link
uum robot hardware to reduced classes of shapes and motions, manipulator to the continuous backbone of a continuum robot.
which in turn has restricted their potential applicability. In this Using simple geometric arguments, the work in [14] deter-
paper, we modify and extend some of the ideas in [14] as part mined the relationship [ d ] = f1 ( s, , ) between D-H ta-
T

of a new and quite general kinematic formulation for contin-

ble parameters i and di and the trunk parameters s, , and .
uum robots. This approach, unlike previous work, produces
Implicit substitution enables computation of trunk tip position
complete and correct results for both task-space position and
and orientation x as a function of s, , and as
orientation. The method fully models for the first time, to the
best of our knowledge, the exact kinematics of trunk sections ( )
A f1 ( s, , ) = A ( s, , ). This allows the use of modified D-
which can extend in length as well as bend in two dimensions H procedures to relate end effector task coordinates to robot
using closed-form equations. The approach is also modular, shape [14].
allowing its application to a wide range of physical realiza- Next, to map s, , and to specific continuum robot hard-
tions, including in particular the common three evenly-spaced ware, another transformation is necessary which varies based
tendons per section design. In addition, the resulting Jacobi- on the construction of the robot, demonstrating the modularity
ans, derived via a straightforward application of well-known of this approach. For the Clemson Elephant Trunk manipula-
D-H techniques and the chain rule, are relatively simple and tor, the work in [14] determined a mapping ( s, k , ) = f 2 a ( l )
computable in real time. This model also allows a straightfor- to compute s, , and given the length l of the four cables
ward analysis of singularities present in a continuum manipu-
which controlled each trunk section. That work, while innova-
lator. In the following sections, we discuss the modeling ap-
tive, was limited to the unique actuator arrangement in the
proach, describe the various kinematic transformations and
robot in [14], and assumed constant trunk length, which re-
Jacobian synthesis in detail, analyze uniqueness, singularities,
stricted the generality of the results. In this paper, we intro-
and accuracy, and report results from a real-time hardware
duce an analogous mapping initially reported in [18], [19] of
implementation on two different multi-section continuum ro-
bots.
( s, k , ) = f 2b ( l ) from the lengths l of each of the three cables
per section to trunk parameters s, , and illustrated in Fig. 2
for a class of designs typified by the Clemson Air-OCTOR
II. MODULAR APPROACH robot shown in Fig. 1. This particular actuator arrangement is
quite common across a wide variation in continuum robot
Deriving kinematics using the method introduced in this
hardware realizations [3], [16], [23], [27], [28]. Alternatively,
paper involves two steps. First, this work extends an earlier
method [14] by formulating the trunk kinematics problem as a ( )
a third mapping ( s, , ) = f 2 c p between the air pressure in
series of substitutions applied to a modified homogenous
transformation matrix computed using a Denavit-Hartenberg
(D-H) type approach. As in [14], this formulation also relies
on the assumption that the trunk bends with constant curva-
ture, due to the compliance of the trunk [13]. Second, velocity
kinematics are developed by first determining the Jacobian
using standard techniques based on the D-H table then by
chaining together the Jacobians produced by computing deri-
vates of each succeeding substitution. The following para-
graphs outline this approach, while the next section develops
the resulting equations in detail.
In conventional manipulator forward kinematics, a standard
homogenous transformation matrix A ( , d ) produced via a D-
H table is used to transform local shape coordinates ( , d ) into
task coordinates x, representing both end effector position and
orientation. This is based on a series of independent rotations Fig. 2. Manipulator variables s, and , where gives the
i and translations di with one independent variable per joint at rotation in the xy plane. Three cables of length l , running in
each of the i joints. However, continuum robots lack explicit straight lines between equidistant cable guides, are also
shown.
IEEE TRANSACTIONS ON ROBOTICS, VOL. 22, NO. 1, FEBRUARY 2006 45

each of the three McKibben actuators in a section of the Oct- III. VERTEBRATE-TO-INVERTEBRATE (SHAPE)
Arm manipulator [23] shown in Fig. 20 may be utilized. TRANSFORMATION: MODIFIED D-H KINEMATICS
Therefore, the work in this paper which builds on the above The following paragraphs discuss the derivation of a new
mapping from [18], [19] is more generally applicable then the D-H formulation which fits a conceptual virtual (but con-
results in [14]. venient for calculation, as we shall see) rigid-link robot to the
In summary, these forward kinematic transformations relate continuum backbone. Given this modified D-H formulation,
trunk tip position and orientation x to the cable length l or we then develop geometrical solutions which relate the D-H
pressure p of the robot according to the information flow model to the shape of a continuum trunk. The D-H table ex-
plicitly described in this paper is for a single section of a con-
f f
DH
x , d
1
s, ,
2
l / p, (1) tinuum arm. The model for the complete, multi-section arm is
formed by multiplying the single-section D-H transformation
or, more formally, matrices in the conventional manner.
This new model shown in Fig. 3 is based on a rigid-link

( ( ( ))).
x = f D H f1 f 2 l / p (2)
arm composed of two revolute joints with intersecting axes
followed by a prismatic joint and completed with two more
revolute joints, also with intersecting axes, whose joint vari-
The corresponding transformations are fully developed and ables are coupled to those of the first two revolute joints. This
analyzed in the next section. differs from the rigid-link system used in [14], where the ap-
In forward velocity kinematics, the goal in the context of proximation of a torsional term in the first link of the D-H
continuum robots is to relate trunk tip linear and angular ve- table produces incorrect angular information in the resulting
locity x to either cable velocities l or actuator pressures p A matrix as shown by Fig. 4. This torsional model, while a
through a Jacobian, so that x = J l or x = Jp. Following the good approximation (another interesting approximate Jaco-
approach of (1), this can be accomplished through a series of bian is introduced in [20]), does not reflect the torsion-free
Jacobians according to the information flow construction of continuum robot backbones, yielding an incor-
rect orientation term. The kinematic arrangement introduced
Jf Jf in this paper more precisely matches the capabilities of real
x
J DH
, d
1
s, ,
2
l / p. (3)
continuum arms by modeling trunk bending as a bending
about two axes just as the actual robotic mechanism bends,
That is, a straightforward application of the chain rule [21] to therefore producing physically correct results for orientation,
(2) yields compared to the approach in [14] which models bending as a
torsion followed by a single-axis bend.
dx
=
x (
( , d ) ( s, , ) d l / p )
or
(
dt ( , d ) ( s, , ) l / p ) dt (4)

( )
x =J D H J f1 J f2 l / p ,

where J = J D H J f1 J f2 . This formulation reveals an advantage

of this approach: different Jacobian modules such as J f2 a for
the Clemson Elephant Trunk, J f2b for Air-OCTOR, or J f2 c for 4
OctArm allow different hardware realizations across a range
of continuum hardware designs to be easily plugged in to the 5
derivation.
The use of standard techniques produces the Jacobian J D H
from an appropriate D-H formulation, while J f1 and J f2 are
directly computed as partial derivatives of f1 and f 2. Section V z
presents each of these Jacobians in more detail. Initially, in
Section III we describe the first transformations, relating x to d3
( , d ) and then relating ( , d ) to ( s, , ). Section V then pre- 2 y
sents transformations between ( s, , ) and actuator input vari-
ables.
1 x

Fig. 3. A continuum robot section, shown as a thick, semi-

transparent line, is modeled by coupled link/joint arrangement.
46 IEEE TRANSACTIONS ON ROBOTICS, VOL. 22, NO. 1, FEBRUARY 2006

(a) (b)
Fig. 4: Comparison of orientational terms. Illustration (a) shows the results due to rotation of the trunk from = [ 0, 2 ] using
formulas taken from [14]. Local coordinates axes at the trunk tip, which are drawn in ( x, y ) order using the colors (red, green),
indicate that the trunk torsionally rotates about the base frame z axis. However, the continuum robots discussed in this paper are
mechanically constructed to be torsion-free, resulting in a local coordinate frame which rotates as shown in illustration (b), seem-
ing to spin about the tip z axis. Videos at IEEE Xplore more clearly illustrate these effects.

The first two revolute joints which lie at the base of a trunk following the convention established in [14]. The fixed rota-
section point the local coordinate frame toward the sections tion at the end of the table then orients the tip coordinate
tip. Next, the prismatic joint translates the local frame to the frame to match the tip coordinate frame in [14].
tip of the trunk. The final two revolute joints then orient the This approach yields the D-H table given in TABLE I, illus-
local frame to point along the tangent of the trunk sections trated in Fig. 5, which results in the homogenous transforma-
tip, so that following trunk sections will be properly oriented. tion matrix A of
However, the first and last pairs of variables in this model are
coupled, due to the constraints of the conceptual rigid link c1 s2 s4 c5 + c1c2 c4 c5 s1 s5 c1 s2 s4 s5 c1c2 c4 s5 s1c5
robot fitting a constant curvature continuous backbone. There- s s s c + s c c c + c s s1 s2 s4 s5 s1c2 c4 s5 + c1c5
fore, the model contains only three independent variables, A= 1 2 4 5 1 2 4 5 1 5
rather than five. This demonstrates the symmetry in the rigid-
( c2 s4 + s2 c4 ) s5 ( c2 s4 + s2 c4 ) s5

link model, reflecting that in the trunk section. Although the 0 0
(5)
D-H approach above aligns trunk extension along the local x c2 s4 c1 + s2 c4 c1c2 d3
axis, geometrical transformations in (7) cause it to extend s1c2 s4 + s2 c4 s1c2 d3
along the y axis. Therefore, the addition of a fixed rotation ,
c2 c4 s2 s4 s2 d 3
about the x axis at the beginning of the D-H table causes the
trunk to extend along the z axis after transformations in (7), 0 1

TABLE I
where ci cos i and si sin i . We note again here that the
NEW D-H TABLE, WHERE ASTERISKS REPRESENT JOINT VARIABLES.
ADDITIONAL UNNUMBERED LINKS ORIENT TRUNK EXTENSION ALONG last two joint variables will be coupled to the first two, spe-
THE z AXIS. cifically that 4 = 2 and 5 = 1 + .
Link a d Thus (5) represents the kinematics for a virtual rigid-link
robot whose initial and final conditions align with the real
- 0 0 0
2 trunk section. The next stage is to relate the joint variables of
the virtual robot to the shape variables s, k , and for the ac-
1 0 0 1*
2 tual continuum section.

2 0 0 2* + Thus we now derive a transformation which relates the D-H
2 2 parameters and d to trunk parameters s, , and . For this,
*
3 0 d 3 0 consider Fig. 6. As shown in [14] by the translational terms in
2
the rows 1 and 2 of column 4 of the 2-D Jacobian given equa-
4 0 0 4* tion (19) of [14], the ( x, y ) coordinates of the trunk in the 2-D
2 2
5 0 0 0 5* cos s 1 sin s
plane A are labeled in Fig. 6 as , . Also
- 0

0 0
2 from [14], the distance from the base to the tip of the trunk is
IEEE TRANSACTIONS ON ROBOTICS, VOL. 22, NO. 1, FEBRUARY 2006 47

[1 d3 4 5 ] = f1 ( s, , ) =
T
2
2 s
d= sin , (6) s
2 cos
tan 1 2
as labeled in Fig. 6. Per the figure, note that the consideration s
sin s sin 2 cos

of simple triangles yields the identities tan 1 =
and 1 s
l sin sin sin
h 2
sin 2 = . Lengths h and l can also be found from tri- (7)
2 s 2 s
sin sin .
2 2
h
angles in the diagram as sin ( ) = and s
sin 1 sin sin

cos s 1
2

l s
cos ( ) = , where the choice of matches the cos
cos s 1 1 2
tan
+
s
convention established by [14] of defining the 2-D D-H table sin 2 cos

to bend along the y axis. Substituting and solving using the
trigonometric identities sin ( 2 ) = 2sin cos and Substituting (7) into (5) produces
1
sin A sin B = ( cos( A + B ) cos( A B ) ) gives
2 cos 2 ( cos s 1) + 1 sin cos ( cos s 1)

sin cos ( cos s 1) cos 2 (1 cos s ) + cos s
s A=
and 2 = sin 1 sin sin .
1
1 = tan 1 cos sin s sin sin s
s 2
tan 2 cos 0 0

Again from (6), the necessary translation d3 is the distance cos ( cos s 1)
cos sin s (8)
2 s
from the trunks base to its tip, so that d3 = sin . The final
2 sin ( cos s 1)
two rotations, 4 and 5, simply repeat the rotations determined sin sin s .

for 1 and 2 , but in reverse order, so that 4 = 2 and sin s
5 = 1 + . The addition of is necessary to correct for the cos s

choice made to bend the trunk along the y axis in [14]. Sum- 0 1
marizing, the f1 found for the second arrow of (1) is
Thus, by fitting a particular phantom rigid-link manipula-
tor to a constant curvature continuum robot section, we are
able to use a modified version of the classical D-H procedure
to relate end effector (task) coordinates to robot shape. For
2 s
Length sin
z 2
Plane A
y

z0 = x2 x3 = z4 = z5
U x0 = x1 = z2 U z2 = x4 = x5

2
sin s
h Length

1
l x
cos s 1
Length

z1 z3 Fig. 6. Determining a geometrical transformation between D-
Fig. 5. Coordinate axes for D-H table given in TABLE I, where H parameters and trunk parameters. The semi-transparent line
the circled U shows an axis extending upward from the page. shows the actual trunk section modeled by the geometry.
48 IEEE TRANSACTIONS ON ROBOTICS, VOL. 22, NO. 1, FEBRUARY 2006

fore not applicable to the more general design in this paper.

Likewise, this model extends the three-cable kinematics in
both the trimole sensor [29], Hiroses Elastor cable kinematics
[15], and various active endoscopes [5], [25], [26] by deter-
h3 mining trunk length s based on an assumption of constant cur-
vature.
h2
Given three known cable lengths l1, l2, and l3 in a section of
Plane H the design, the constant distance d from the center of the trunk
to the location of the cables, and n, the number of segments in
Cable 3 this trunk section, the following analysis allows computation
of the resulting trunk length s, curvature , and direction of
Cable 2 curvature as illustrated in Fig. 2. The three cables are as-
Cable 1
sumed to lie equally spaced around the trunk, separated by
angles of 120. In addition, the curvature in each section of the
trunk is assumed to be uniform, so that trunk sections can be
Fig. 7. Bent section of trunk. modeled as arcs of varying-radii circles. This is typically
achieved in hardware by distributing bending forces equally
conventional rigid link manipulators, this is directly achieved
along the trunk using pneumatic pressure or springs [13].
via the dependence of A on ( , d ). However, since shape has
This section discusses computation of , , and s in three
a different meaning for continuum robots, the extra transfor- steps. First, equations are presented which determine distances
mation f1 has been made so that A is now an explicit function r1, r2, and r3 related to the trunks curvature. Next, from these
of ( s, , ). The next section discusses the derivation of f 2b and distances the curvature at angle is derived. Finally, addi-
f 2 c, which allows the further transformation into an A as a tional equations compute the trunk length s from . Through-
function of either cable lengths or actuator pressures. out this analysis, the problem of solving geometrical problems
posed by the trunk is accomplished by analyzing a set of 2-D
slices of the problem which allow computation of the desired
IV. SHAPE TO INPUT TRANSFORMATIONS quantities.
Controlling the shape of the manipulator requires a kine- A. Computation of i
matic model relating shape in terms of extension and bending Consider one bent segment in a section of the trunk as illus-
to actuator inputs. This section outlines a new kinematic trated by Fig. 7. This figure shows three cables, spaced at 120
model, tailored to three-tendon, 120 configuration manipula- intervals around the trunk, connecting two segments in a sec-
tor hardware as illustrated by Fig. 2. The model is both new tion of the trunk. The figure includes a plane H which passes
and novel compared to previous models developed in [12], through the point at which cable 1 connects to the trunk and is
[14], which were tailored to four-tendon sections and there- perpendicular to the other two cables. This plane exists due to
Cable 1 attachment point the design of the trunk, which places the three cables at fixed
intervals about the circumference of the trunk, insuring that all
three cables parallel each other.
The curvature of the trunk can be calculated from hc, the
d 3 height of the point c in Fig. 8 above the plane H. Finding this
b= 30 30
2 2d value requires knowledge of h2 and h3, the heights of the at-
a=
3 tachment points of cables 2 and 3 above the plane as shown in
d Fig. 7. The lengths 2h2 and 2h3 by which cables 2 and 3 ex-

Point 2m Point 3m
Point c
Cable 2
attachment point d 3

2b = d 3
Point 2m
2d
l l 3
Cables 2 and 3 attachment points h2 = 2 1
2n h2m
Cable 1
Plane A attachment point
Fig. 8: Top-down view of trunk. Fig. 9. Height of points 1 and 2 above plane.
IEEE TRANSACTIONS ON ROBOTICS, VOL. 22, NO. 1, FEBRUARY 2006 49

l
Plane A outer l3 + l2 2l1
n hc =
6

d
1 d
s
l1
n
lc
1 n 2n
d
r1 lc

d n
l
inner
n
Plane B

Fig. 10. Geometry of bent trunk section.

ceed the length l1 of cable 1 are equally distributed above the Fig. 11. Determination of r1.
plane shown in Fig. 7 and below a second plane, not shown in
the figure, which contains the bottom attachment point of ca- section. Therefore, the curvature of each segment remains
ble 1 and is perpendicular to cables 2 and 3. Therefore, the constant along the section. The curvature of a section is de-
total length of cables 2 and 3 along the entire section of an n- 1
fined as = where r is the radius of a circle which fits the
segment trunk section can be expressed in terms of h2 and h3 as r
arc produced by the bending of a trunk section. Consider a
l2 l1 l
h2 = and (9) virtual cable of length c as shown in the figure, attached to
2n n
the center of a cross section of the trunk and traveling from
l3 l1 the top to the bottom of a segment of the trunk. At some as yet
h3 = . (10)
2n unknown angle the situation shown in Fig. 10 exists, where
1 l
Note that because the plane passes through the trunk at the an isosceles triangle with sides of length r = and c sub-
n
attachment point of cable 1, h1 = 0 by definition. tends degrees of an arc of the circle formed by the trunk.
Considering plane A containing the circle which forms the top
Examining Fig. 8 which shows a 2-D slice of the trunk of the trunk shown in Fig. 8, note the leg of this triangle which
2d lies in this plane. This leg can be viewed as the vector in
viewed looking down on its top segment reveals a = and
3 polar coordinates in plane A where the origin of the plane is at
d 3 point c. The magnitude of corresponding vectors 1, 2, and 3,
b= . As shown in Fig. 9, viewing a different 2-D slice
2 notated 1 3, which pass through the attachment points for
perpendicular to plane A in Fig. 8 which contains the dotted cables 1, 2, and 3 allow determination of via coordinate
l l transformations detailed in the next section.
line in Fig. 8 allows determination of h2 m = 2 1 . Likewise,
3n To compute 1, rotate a second plane B shown in Fig. 10
l3 l1 l l
h3m = . Height hc is equidistant from these two points, so which is perpendicular to plane A containing and c about
3n n n
its height is their average, producing lc
from angle to an absolute angle of 90, the location of
n
l3 + l2 2l1 l
hc = . (11) cable 1, so that the plane now contains l1 and c . Fig. 11 shows
6n n
the contents of plane B after rotation, which contains an isos-
Note that hc may be negative at some orientations of the trunk, l
where 2l1 > l2 + l3. Though unexpected, the following geomet- celes triangle with sides r1 and c . Similar to the reasoning in
n
rical uses of hc are correct in this case. lc
(9) and (10), the length of this virtual cable can be divided
n
Based on the assumptions, curvature of the trunk segment is l1
evenly distributed along each of the n segments in the trunk into , the length of cable 1 in one trunk segment, and hc, the
n
50 IEEE TRANSACTIONS ON ROBOTICS, VOL. 22, NO. 1, FEBRUARY 2006

l12 + l2 2 + l32 l1l2 l2 l3 l1l3

i1 = iy =2 and (15)
d ( l1 + l2 + l3 )
ix
3 l3 + l2 2l1
i2 = tan 1 . (16)
i3 l2 l3
3
C. Computation of s, the length of a trunk section
Each section of the trunk has n segments. In each segment,
Fig. 12. Unit vectors i1 3, x , y along which vectors r1 3, x , y lie. the cable runs in a straight line from one cable guide to the
next. Solving for the cable length in one section allows deter-
height this virtual cable extends above the plane anchored by mination of the cable length in all sections. Therefore, con-
l1, so that s
sider just one section of the trunk whose length is therefore .
n
lc l1
= + 2hc . (12)
n n Bending the trunk produces the geometry shown in Fig. 10.
From (12), lc = l1 + 2nhc . Substituting from (11) yields
Relating this to the smaller similar triangle in the figure yields
l l3 + l2 + l1
hc + 1
hc
= 2 n . Substituting in (11) gives r = d ( l1 + l2 + l3 ) . lc =
3
. (17)
1
d r1 l3 + l2 2l1
1 l3 + l2 2l1 Note that in a circle of radius r , the angle subtended by a
Noting that = , 1 = . Repeating the process
r d ( l1 + l2 + l3 ) portion of the circumference of length a obeys the relationship
with the assumption that h2 = 0 instead of h1 = 0 as before, 2 r 2
= . To compute for the outer circle shown in Fig.
a
l3 + l1 2l2 1 s
2 = . (13) 10, substitute r = , a = , and = into the above equation
d ( l1 + l2 + l3 ) n
and solve for to obtain
Similarly, by assuming h3 = 0,
s
= . (18)
l + l 2l3 n
3 = 2 1 . (14)
d ( l1 + l2 + l3 )
Examination of the triangle drawn inside the outer circle of
B. Computation of and , trunk segment curvature and lc
angle of curvature 2n
Fig. 10 shown in Fig. 14 reveals that sin = . Substitut-
To convert the i per-cable curvatures into , the angle at 2 1
which the trunk bends as illustrated in Fig. 2, and , the cur-
vature at angle , consider determining and based on a Cable 2, height is h2
coordinate transformation of values just derived for 1, 2, and Cable 3, height is h3
3. Note that the positions of cables 1, 2 and 3 on the circum-
ference of the trunk with respect to the x axis are 90, 210
Cable 1, height is h1
and 30 as shown in Fig. 12. Viewing these curvatures in
plane A as vector quantities by noting the angle relative to the Height h at angle .
x axis along which each curvature extends yields 1 = 190, Defined as h = 0
2 = 2 210 , and 3 = 3 30. Using the change of basis
Plane C
cos ( 210 ) sin ( 210 ) to plane H
matrix B = to convert to a standard
cos ( 30 ) sin ( 30 )
orthogonal basis so that x y = B 1 [ 2 3 ] yields
Trunk center, height is hc
x = 3 2 and y = 2 3 = 1. In polar form,
3 Line at which Plane A

1 y planes intersect
= tan and = x 2 + y 2 . Substituting (13) and (14)
x
then simplifying yields Fig. 13. Plane containing circle and reference plane.
IEEE TRANSACTIONS ON ROBOTICS, VOL. 22, NO. 1, FEBRUARY 2006 51

h1
Tangent
lc

2n h
d h1 p
1
90 a
2 d
hc
Fig. 14. Computing s.

l
2n
ing (18) and solving yields s = sin 1 c . Substituting

2n
from (15) and (17) yields s in terms of cable lengths as
Fig. 15. Projecting h1 along a line parallel to the tangent shown
nd ( l1 + l2 + l3 )
s= in Fig. 16.
l12 + l2 2 + l32 l1l2 l2 l3 l1l3
. (19) end-on as in Fig. 16, note that heights measured from plane C
l 2 +l 2 +l 2 l l l l l l
sin 1
1 2 3 1 2 2 3 1 3
to a point on A do not change when following any line on A
3nd parallel to this tangent. For example, as illustrated in Fig. 15,

projecting the height h1 along a line parallel to the tangent at h
Note that in the limiting case when the cable lengths are equal, gives h1 = h1 p , where h1 p is the height at a point along the line
the terms under the square root go to zero so that connecting hc and h . Knowledge of the slope of the line cre-
l +l +l ated by viewing plane A end-on and of a position on this line
lim s = 1 2 3 . Together, equations (15), (16), and (19)
2
l1 + 0 3 allows computation of the height of this position above plane
comprise the function ( s, k , ) = f 2b ( l ) for a cable-actuated C , which appears as horizontal line in the figure. Therefore,
robot. the height of an arbitrary point on the circle such as h1 = h1 p
Because pneumatic actuators bend continuously, the deriva- can be determined by projecting the point along a line parallel
tion for f 2 c as defined in (1) can be found by increasing the to the tangent of h onto the diagonal line in Fig. 16 then ap-
number of cable guides to infinity. Examining the limiting plying the slope-intercept form of that line to find the points
case when n of f 2b yields identical expressions for (15) height.
and (16), while (19) becomes In the forward kinematics derivation, the height h1 = 0 was
chosen to anchor the plane as shown in Fig. 7, Fig. 8, and Fig.
l1 + l2 + l3 9. For the inverse kinematics derivation, the plane will be an-
s= . (20) chored by h = 0 instead. Examining the isosceles triangle
3
1 l
with sides and c shown in Fig. 10 creates Fig. 17 which
Therefore, ( s, k , ) = f 2 c ( l ) is defined by (15), (16), and (20). n
hc
D. Inverse kinematics (manipulator curvature to cable reveals sin = . Substituting (18) and solving produces
2 d
lengths) s
Consider the manipulator section shown in Fig. 2. Given hc = d sin . Projecting these points onto the diagonal line
2n
the length s of one section of the manipulator, the curvature of Fig. 16 allows the line to be defined as
of that section, the angle at which the curvature lies , the dis- hc h
tance d from the center of the manipulator to the location of h1 p = h1 = ( d a ) + h . Also note from Fig. 15 that
d
the cables, and the number of segments in this manipulator a
section n, this section discusses the computation of the lengths cos = = sin . Substituting and simplifying shows
2 d
of the three cables l1, l2, and l3 due to the configuration of the s
manipulator section. This is the inverse kinematics problem. that h1 = d sin (1 sin ).
2n
As shown in Fig. 13, the top of one segment in a section of Similar to the reasoning in (9) and (10), the length of cable
the elephant trunk lines in plane A . Heights can be measured 1 l1 can be divided into l , the length of a virtual cable attached
from a second plane C which is parallel to plane H shown in
at angle on the circumference of the manipulator and run-
Fig. 7. These two planes intersect along a single line tangent
ning the length of the manipulator, and h1, the height cable 1
to height h as shown in the figure. Viewing these two planes
extends above the plane anchored by l , so that l1 = l + 2nh1.
52 IEEE TRANSACTIONS ON ROBOTICS, VOL. 22, NO. 1, FEBRUARY 2006

Plane A viewed end-on

Cable 2
Cable 3
Center
l
2n
a
h2 h d 1
3
hc Point at , d
h1 where h = 0 k
2
Cable 1
Plane C viewed end-on
Fig. 16. End-on view of planar intersection shown in Fig. 13. Fig. 18. Triangle from Fig. 10 used to determine length l .
To find l , note that the length of a cable l attached to the cir-
lim l2 = lim l3 = s. This completes the derivation for a cable-
cumference of the manipulator at the point of maximum cur- 0 0

vature forms the triangle shown in the inner circle of Fig. actuator robot such as Air-OCTOR of
f 2b ( s, , ) = [l1 l2 l3 ] as defined in (1) where l1, l2, and l3
1 T
l
are given in (21), (22), and (23) above.
10, redrawn in Fig. 18, yielding sin = 2n . Therefore,
2 1 d Because pneumatic actuators bend continuously, the deriva-
tion for f 2 c 1 as defined in (1) can be found by increasing the
substituting (18), the cable length can be computed as number of cable guides to infinity. Examining the limiting
1 s case when n of (21), (22), and (23) yields
l = 2n d sin . Substituting and simplifying,
2n

s 1
l1 = 2n sin d sin . (21) s (1 d sin )
2n 1
l

f 2 c 1 ( s, , ) = l2 = s 1 + d sin + . (24)
3
Likewise, rotating to 210 and 330 to find h2 and h3, the l3
heights of the attachment points of cables 2 and 3, then re- s 1 d cos +
solving produces
6

s 1 Now that all the relationships in (1) describing the forward

l2 = 2n sin + d sin + and (22)
2n 3 kinematics for a trunk have been determined, the next section
will discuss the derivation of the forward velocity kinematics.
s 1
l3 = 2n sin d cos + . (23)
2n 6
V. JACOBIAN DEVELOPMENT
Note that because lim sin x = x, in the limiting case As overviewed by (3) and (4), straightforward application of
x 0
s 1 the chain rule with necessary substitutions yields the Jacobian
lim l1 = 2n d sin = s . In the same way,
0

2 n
(
x=J l /p l /p ) (25)
d hc
where l and p are the time derivates of length or pressure, and
2
( )
the notation l / p implies the selection of either l or p . ( ) ( )
Each Jacobian in the chain of Jacobians must be computed.
Computing the Jacobian which relates x to and d produced
lc
by classical D-H techniques based on the D-H table in TABLE I
n
yields
1

Fig. 17. Triangle from Fig. 10 used to relate hc to .

IEEE TRANSACTIONS ON ROBOTICS, VOL. 22, NO. 1, FEBRUARY 2006 53

s1 c 2 d3 c1 s 2 d 4 c1 c 2 0 0 Finally, substituting f 2 for either cable lengths or pressures

0 c2 d3 s2 0 0
into (28) and multiplying by J f2 produces the final Jacobian
c c d s1 s 2 d3 s1 c 2 0 0 which satisfies (25). Due to its length, this result is omitted.
J DH = 1 2 3 (26)
0 s1 0 s1 c1 s 24 However, the procedure is quite straightforward, and leads to

1 0 0 0 c 24 computationally tractable solutions, as discussed in the fol-
lowing section. Notice that by including our more general
0 c1 0 c1 s1 s 24
formulation for f 2, the resulting approach is applicable to a

where cij cos ( i + j ) and sij sin ( i + j ) . Computing J f1

much wider class of continuum robots than the approach in
[14].
from (7) yields

J f1 = VI. DISCUSSION AND IMPLEMENTATION

sin sin s s cos cos
A. Uniqueness
g g g
There are a number of ways in which some of the results of
s s s
2 cos sin 2 s sin cos
2
sin cos
2
the derivation developed in this paper can be obtained. This
section reviews alternative representations, approaches, and
h h h derivations.
s s (27) First, the parameters chosen to represent an arc in 3-D are
s cos 2sin
2 2 s not unique. Rather than describe the arc in terms of its length
0 cos
2 2 s, curvature , and angle of curvature as shown in Fig. 2, a
s s s polar representation, the arc could be described using a Carte-
2 cos sin s sin cos sin cos
2 2 2 sian coordinates in terms of length s, curvature along the x
h h h axis x , and curvature along the y axis y . Or, instead of arc
sin sin s s cos cos
length s, the number of radians subtended by the arc could
g g g be used. A number of other approaches are possible [2]; the
choice of representation in this work was motivated by both a
where g = cos 2 cos s cos s cos 2 1 and desire to maintain a common scheme with previous work such
h = 2 cos + 2 + 2 cos s 2 cos cos s . Computing
2
J f2 as [14] and by the ease of modifying the angle of curvature
results in long terms and is omitted for brevity. to compensate for rotations relative to the home (typically
user) coordinate frame due to the way in which the robot is
Next, substituting (7) into (26) then multiplying by (27) mounted, where actuators are placed in each section of the
produces robot, and where the mobile platform base is placed relative to
the home frame.
sin ( cos s 1) Likewise, the backbone presented in section III as the D-H
table given by Table 1 and transformations in Fig. 6 and (7) is
not unique. An alternative approach, for example, is to append
0 a counter-torsional term of 5 = to the four link D-H table

cos ( cos s 1) developed by [14]. A second approach is to abandon the back-
J D H , f1 =
bone-fitting method, finding the position and orientation of

cos sin s
the trunk tip through other methods. For example, rotating the
1 cos s trunk tip position in plane A in Fig. 6 of
cos s 1 sin s
T

sin sin s 0 about the y axis by produces the

cos ( s sin s + cos s 1) (28) same positional results as the backbone approach. Using a
cos sin s
2
differential curve representation, such as the Serret-Frenet, can
s cos s sin s also produce identical positional results as shown by [14].
cos s
2
Likewise, an integral representation, such as those developed
sin ( s sin s + cos s 1) . by Chirikjian and Burdick [8] or in [13], can be used to obtain
sin sin s identical positional results. Following the approach of [8], this
2 integral representation can then be factored into a modal de-
s sin sin
composition whose modal functions consist of arc length s,
0 0
angle of curvature , and the product of arc length and curva-
s cos cos ture s = , where gives the angle subtended by the arc rep-
resenting the trunk. However, this model for the trunk is exact,
54 IEEE TRANSACTIONS ON ROBOTICS, VOL. 22, NO. 1, FEBRUARY 2006

instead of the approximate model introduced by [8], leading to each section of the robot [13]. Therefore, this approach does
greater ability to precisely specify trunk shape. Additionally, not suffer from significant errors due to a fitting process.
our directly physically motivated backbone approach pro- However, this model assumes the absence of external forces
duces enhanced insight into the mechanics of the trunk. For which can bend the trunk in a non-circular fashion, leading to
example, trunk sections in this paper are chosen not as arbi- errors. For example, when the trunk grasps a non-circular ob-
trary points as in [8] but at the interface between two groups ject such as a box, it no longer bends in a circular arc. Like-
of McKibben actuators, again carefully modeling the physical wise, when it supports a gravity load, either the mass of the
construction of the trunk. trunk itself or of a payload the trunk is carrying, the trunk sags
The orientation of the tip and the computation of its angular in a non-circular fashion [26]. Finally, this kinematic analysis
velocity requires additional work when the backbone ap- ignores all dynamics effects due to inertial, centrifugal, and
proach is not followed. One approach, detailed in [8], is to Coriolis forces. Unfortunately, dynamics in the 3-D case re-
include a rotated frame which allows specification of orienta- main in integral-differential form, rendering them unsuitable
tion independent of the frame produced by the curves for real-time control [13], [17], or can be computed only as an
mathematical representation. The angular velocity is then de- approximation in closed form [7]. However, as discussed
fined by S ( ) = RR T where R gives the rotation matrix of the above, these issues are less significant for continuum robots
frame and S ( ) describes factoring a skew-symmetric matrix designed for operation in unstructured environments than for
S to a three-parameter vector . Factoring the resulting into traditional rigid-link manipulators which operate primarily in
a Jacobian then yields results parallel to those obtained using structured environments.
the backbone approach. The backbone method was chosen in C. Singularities
our approach because it provides a good method for comput- Robots constructed of a combination of revolute joints typi-
ing the kinematics of wide range of continuum manipulators, cally exhibit singularities at their workspace boundaries, in
including those not considered in our work, simply by modify- addition to inherent singularities due to their mechanical de-
ing the appropriate D-H entries and geometrical transforma- sign. Because the continuum robots analyzed in this paper can
tions to match a new design. In comparison, finding a correct theoretically extend to any arbitrary arc length s, they do not
orientation frame is non-trivial and involves additional com- possess workspace boundary singularities. Actuator length
putations to develop angular velocity. limits, however, impose workspace constraints, but they do
B. Accuracy not enter into the kinematics formulation. The only singular
Two concerns affect the accuracy of a continuum robot, de- configuration occurs when the trunk is straight ( = 0). In this
fined as the distance between desired and achieved locations case, velocities normal to the angle of trunk curvature can-
in the workspace. First, error due to the compliance of the not be produced. However, the angle of curvature is essen-
manipulator, in contrast to the rigid construction of traditional tially meaningless when the trunk is not curved, so this con-
robot, is significant; second, the nature of the modeling proc- figuration presents no practical problems. An alternate repre-
ess must be examined for possible sources of error. sentation for the trunk, for example as two orthogonal curva-
The first category of accuracy errors as discussed in [33] tures x and y discussed in section VI-A, would eliminate
arises from the construction of the robot. Traditional robots, this singularity.
due to their extremely rigid design, can achieve very high ac- D. Simulation results
curacy and repeatability. For example, a trunk-like high de-
As a prelude to implementation on continuum robot hard-
gree of freedom manipulator [32] can be expected to achieve
ware such as the Clemson Air-Octor robot [18] or the Penn
very high accuracy. Continuum robots, due to their soft, com-
State OctArm robot [23], the Jacobians derived in the previous
pliant design, cannot. However, high accuracy is not the in-
section were symbolically computed using Maple 9.5. First, a
tended niche of continuum robots. They are generally poor as
D-H table for an n-section trunk is determined. Given that
high- accuracy positioning tools when compared to traditional
table, a second routine then computes J D H by applying stan-
rigid-link robots. Instead, their compliance can be used to re-
duce the accuracy needed to perform manipulation tasks in dard Jacobian techniques to the table. Next, a set of n-section
unstructured environments by allowing the inherent flexibility geometrical transformations representing f1 are computed and
of the manipulator to compensate for a lack of precise knowl- their Jacobian J f1 computed. The product J D H J J f is then
1

edge of objects to be grasped. computed by substituting f1 into J f1 then multiplying. Finally,

The second category of errors arises due to the modeling J f2 is formed for an n-section trunk, substitutions are made,
approach. Some approaches discussed in the introduction,
then the product J D H J f1 J f2 is formed. This symbolically com-
such as [8], model the robot as an arbitrary curve, then per-
form an error-minimizing fitting process to match the actual puted matrix is then transformed into C code using Maples
robot to the curve. In contrast, this paper chooses a curve an CodeGeneration package. The resulting code, which has un-
arc of a circle which the manipulator naturally assumes due dergone common subexpression elimination via the optimize
to its design, as forces due, for example, to pneumatic pres- option to CodeGeneration, is then combined with a CLapack
sure are equally distributed into a constant curvature along call to solve (25) as
IEEE TRANSACTIONS ON ROBOTICS, VOL. 22, NO. 1, FEBRUARY 2006 55

kinematics was transformed into a C program which outputs

l /p=J +
( l / p) x (29) computed motor torques as voltages through a Servo-To-Go
D/A board and feeds motor encoder positions back to the con-
trol loop. The C code was then compiled as real-time executa-
for l or p using a QR rotation with column pivoting. This
ble under the QNX Momentics real-time operating system.
code is then placed in a Simulink block, which becomes part The code runs efficiently in a real-time control loop operating
of a block diagram which reads x from a joystick, computes at 500 Hz, providing responsive control to user input through
the resulting l / p based on the Jacobian, then updates a Open- a joystick. The following paragraphs detail the approach to
Inventor-based simulation of the trunk and also computes ac- implementation for both the cable-driven Air-Octor and the
tuator torques or pneumatic pressures necessary to achieve the pneumatically-actuated OctArm continuum robots.
desired l / p . Fig. 19 illustrates the results of this simulation as The two-section Air-Octor robot [18] shown in Fig. 1 con-
the trunk is moved along the z axis. sists of a central hollow chamber, constructed of dryer hose,
The performance of the algorithm is excellent. On a 2.8 into which air pressure is blown, enabling variation of each
GHz Pentium 4 processor for a three section trunk, which has sections stiffness. Helically-wrapped metal coils within the
9 degrees of freedom, the Jacobian is computed, then l / p is hose allow the tube to extend along its length, but not to
change its diameter. Surrounding this chamber, a second
computed based on a given 3 degree of freedom linear veloc-
chamber of ferret tubing, again with helically-wrapped metal
ity x in 0.163 ms. Therefore, this algorithm can easily be in-
coils, provides attachment points for a series of 8 cable guides
cluded in a real-time control system, as detailed in the follow- for each section. Six motors, three for each of the two sec-
ing paragraphs. tions, are arranged around the top of the robot at 60 intervals.
E. Implementation Each motor also includes an encoder to measure the length of
Using the Mathworks Real Time Workshop, the Simulink cable paid out of a spool attached to the motor shaft through a
block diagram which contains code to solve inverse velocity gearing mechanism. Three cables terminate at the first section,
while the other three run through the first section, terminating
at the second section, allowing shape control of both sections.
A simple PID loop servos desired cable lengths to actual
lengths. Implementation of inverse cable kinematics embodied
in equations (21), (22), and (23) map from desired trunk
shape, specified by a joystick through several non-Jacobian
control modes useful for whole-arm manipulation [10], to
desired cable lengths. In addition, Jacobian equations (28) and
(29) allow end-point control of the two-section robot based on
joystick input.
The OctArm manipulator [23] shown in Fig. 20 consists of
three or four sections, depending on the design. Each section
is actuated by joining three McKibben actuators which extend
only in length between two end-plates, so that the trunk can
extend or bend in any direction. Air pressure is regulated us-
ing a set of valves which maintain a desired pressure, input as
a voltage. Shape feedback is still in development, so the arm
runs open-loop. The length of each section of the arm was
measured by inflating all three actuators in a section to a given
pressure then measuring length. A third-order polynomial fit
of this measured length to pressure relationship was then used
to transform desired actuator lengths to pressures. Equation
(24) allows control of trunk shape via the joystick for whole-
arm grasping algorithms [10] or, using equation (29), end-
point control using the Jacobian.
The arms prove quite capable in practice. Fig. 20 illustrates
OctArm, mounted on the Foster-Miller Talon mobile platform,
performing a simple grasping task. Extensive experiments
Fig. 19. OpenInventor-based trunk simulation showing the pictured in Fig. 21 were carried out to examine the abilities of
trunk, under Jacobian control, moving along a straight-line the manipulator. Use of the robot based on the algorithms dis-
workspace trajectory. Red and green arrows indicate the ( x, y ) cussed in this paper enabled biologically-inspired behavior,
axes local to each trunk section. such as a prey strike, whole-arm grasping over a wide range
of object sizes, tunnel observation using a tip-mounted cam-
56 IEEE TRANSACTIONS ON ROBOTICS, VOL. 22, NO. 1, FEBRUARY 2006

(a) (b)

(c) (d)
Fig. 20. OctArm mounted on the Talon mobile Fig. 21. Abilities of the robot, such as (a) hook-style grasping,
platform. Courtesy of Chris Rahn. (b) large object grasping and whole-arm manipulation, (c)
prey capture, and (d) hole exploration using a tip-mounted
era, and teleoperation using cameras mounted at the trunk camera view shown in the lower left corner of (d).
base.
computation in real time, enabling real-time kinematic control
Prior to the analysis and algorithm development reported in
at the velocity level. Implementations on two spatial contin-
this paper, the two hardware platforms were operated in a sec-
uum robots are summarized.
tion-by-section fashion, with the operator moving only one
section of the robot at any time. This proved to be a slow, la-
borious, and often unsuccessful approach. Implementation of
REFERENCES
the approach described in this paper transformed the operation
[1] T. Aoki, A. Ochiai, and S. Hirose, Study on slime robot, Proc. IEEE
of the hardware, enabling successful operation over a wide
Intl. Conf. on Robotics and Automation, New Orleans, April 2004, pp.
range of missions, as shown in Fig. 20 and Fig. 21. Addition- 28082813.
ally, the ability to directly match the underlying models to the [2] Bishop, R., There is more than one way to frame a curve, American
hardware constraints, particularly the constant curvature sec- Mathematical Monthly, vol. 82, no. 3, pp. 246-251, Mar. 1975.
[3] R. Buckingham, Snake arm robots, Industrial Robot: An Intl. Journal,
tions, via relatively simple models has allowed us to adapt vol. 29, no. 3, 2002, pp. 242245.
easily to other effects of strong practical importance such as [4] J. Casper and R. R. Murphy, Human-robot interactions during the ro-
actuator travel limits. This will be reported on in future pa- bot-assisted urban search and rescue response at the World Trade Cen-
ter, IEEE Trans. on Systems, Man, and Cybernetics, vol. 33, no. 3, June
pers. 2003, pp. 367385.
[5] G. Chen, M. T. Pham, T. Redarce, C. Prelle, and F. Lamarquem, Design
and modeling of a micro-robotic manipulator for colonoscopy, Proc. 5th
Intl. Workshop on Research and Education in Mechatronics, Annecy,
VII. CONCLUSION
France, June 2005, to appear.
In this paper, we have presented a new and complete ap- [6] G. S. Chirikjian, Design and analysis of some nonanthropomorphic,
biologically inspired robots: an overview, Journal of Robotic Systems,
proach to kinematic development for continuum robots. The
Vol. 18, No. 12 , pp 701713, 2001.
approach extends previous work by modeling extension, as [7] G. S. Chirikjian, Hyper-redundant manipulator dynamics: a continuum
well as bending, of sections of the arm, and by presenting a approximation, Advanced Robotics, vol. 9, no. 3, pp. 217243, 1995.
complete, correct solution for orientation as well as position. [8] G. S. Chirikjian and J. W. Burdick, A modal approach to hyper-
redundant manipulator kinematics, IEEE Trans. on Robotics and Auto-
The approach allows for simple adaptation to a wide class of mation, vol. 10, no. 3, June 1994, pp. 343354.
existing designs, through the inclusion of a shape to actua- [9] R. Cieslak and A. Morecki, Elephant trunk type elastic manipultor a
tor module. Novel examples of such modules for pneumati- tool for bulk and liquid type materials transportation, Robotica, vol. 17,
1999, pp. 1116.
cally actuated and tendon-based sections are described. The [10] M. Cscenstis, B. A. Jones, W. McMahan, V. Iyengar, and I. D. Walker,
resulting Jacobians are relatively simple and amenable to User interfaces for continuum robot arms, Proc. IEEE/RSJ Intl. Conf.
IEEE TRANSACTIONS ON ROBOTICS, VOL. 22, NO. 1, FEBRUARY 2006 57

on Intelligent Robots and Systems, Aug. 2005, Edmonton, Canada, sub- Bryan A. Jones (S00M00) received the B.S.E.E. and
mitted. M.S degrees in electrical engineering from Rice Univer-
[11] F. Fahimi, H. Ashrafiuon, and C. Nataraj, An improved inverse kine- sity in 1995 and 2002, and a Ph.D. in electrical engineer-
matic and velocity solution for spatial hyper-redundant robots, IEEE ing from Clemson University in 2005. From 1996 to
Trans. on Robotics and Automation, vol. 18, no. 1, Feb. 2002, pp. 103 2000, he worked as a hardware design engineer for Com-
107. paq, specializing in board layout for high-availability
[12] I. A. Gravagne and I. D. Walker. On the kinematics of remotely- RAID controllers. He is currently an assistant professor at
actuated continuum robots, Proc. IEEE Intl. Conf. on Robotics and Mississippi State University. His research interests in-
Automation, San Francisco, 2000, pp. 25442550. clude robotics, real-time control system implementation, rapid prototyping for
[13] I. A. Gravagne and I. D. Walker, Large deflection dynamics and control real-time systems, and modeling and analysis of mechatronic systems.
for planar continuum robots, IEEE/ASME Trans. on Mechatronics, vol.
8, no. 2, June 2003, pp 299307. Ian D. Walker (S84M85SM02F06) received the
[14] M. W. Hannan and I. D. Walker, Kinematics and the implementation of B.Sc. in mathematics from the University of Hull, U.K.,
an elephants trunk manipulator and other continuum style robots, and the M.S. and Ph.D. degrees in electrical engineering
Journal of Robotic Systems, vol. 20, no. 2, Feb. 2003, pp. 4563. from the University of Texas at Austin, in 1983, 1985,
[15] S. Hirose, Biologically inspired robots, Oxford University Press, 1993, and 1989, respectively.
pp. 147155. He is a Professor of Electrical and Computer Engi-
[16] G. Immega and K. Antonelli, The KSI tentacle manipulator, Proc. neering at Clemson University, in Clemson, SC USA. His
IEEE Conf. on Robotics and Automation, Nagoya, Japan, May 1995, vol. research interests are in robotics, particularly kinemati-
3, pp. 31493154. cally redundant robots, robot reliability and fault detection, and biologically
[17] M. Ivanescu, A variable length tentacle manipulator control system, inspired robots.
Proc. IEEE Intl. Conf. on Robotics and Automation, Barcelona, Spain,
April 2005, pp 32853290.
[18] B. A. Jones, W. McMahan, and I. D. Walker, Design and analysis of a
novel pneumatic manipulator, Proc. of 3rd IFAC Symposium on Mecha-
tronic Systems, Sydney, Australia, Sept. 2004, pp. 745750.
[19] W. McMahan, B. A. Jones, I. Walker, V. Chitrakaran, A. Seshadri, and
D. Dawson, Robotic manipulators inspired by cephalopod limbs, Proc.
CDEN Design Conf., Montreal, Canada, 2004, pp. 110.
[20] H. Mochiyama and H. Kobayashi, The shape Jacobian of a manipulator
with hyper degrees of freedom, Proc. IEEE Intl. Conf. on Robotics and
Automation, Detroit, 1999, pp. 28372842.
[21] R. M. Murray, Z. Li, and S. S. Sastry. A mathematical introduction to
robotic manipulation. CRC Press, Boca Raton, FL, 1994.
[22] H. Ohno and S. Hirose, Design of slim slime robot and its gait of loco-
motion, Proc. IEEE/RSJ Intl. Conf. on Intelligent Systems, Maui, Ha-
waii, Oct. 2001, vol. 2, pp. 707715.
[23] M. B. Pritts and C. D. Rahn, Design of an artificial muscle continuum
robot, Proc. IEEE Intl. Conf. on Robotics and Automation, New Or-
leans, Louisiana, April 2004, vol. 5, pp. 47424746.
[24] G. Robinson and J. B. C. Davies, Continuum robots a state of the art,
Proc. IEEE Intl. Conf. on Robotics and Automation, Detroit, Michigan,
May 1999, vol. 4, pp. 28492854.
[25] N. Simaan, Snake-like units using flexible backbones and actuation
redundancy for enhanced miniaturization, Proc. IEEE Intl. Conf. on
Robotics and Automation, Barcelona, Spain, April 2005, pp. 30233028.
[26] N. Simaan, R. Taylor, and P. Flint, A dexterous system for laryngeal
surgery, Proc. IEEE Intl. Conf. on Robotics and Automation, New Or-
leans, April 2004, pp. 351357.
[27] K. Suzumori, S. Iikura, H. Tanaka, Development of flexible microac-
tuator and its application to robotic mechanisms, Proc. IEEE Intl. Conf.
on Robotics and Automation, Sacramento, California, April 1991, vol. 2,
pp. 16221627.
[28] Temple Allen Industries. http://www.templeallen.com/.
[29] H. Tsukagoshi, A. Kitagawa, and M. Segawa, Active hose: an artificial
elephant's nose with maneuverability for rescue operation, Proc. IEEE
Intl. Conf. on Robotics and Automation, Seoul, Korea, May 2001, vol. 3,
pp. 24542459.
[30] J. F. Wilson, D. Li, Z. Chen, and R. T. George Jr., Flexible robot ma-
nipulators and grippers: relatives of elephant trunks and squid tentacles,
Robots and Biological Systems: Toward a New Bionics?, 1993, pp. 474
479.
[31] J. F. Wilson, J. M. Snyder, The elastica with end-load flip-over, Jour-
nal of Applied MechanicsTransactions of the ASME vol. 55 no. 4, pp.
845 848, Dec. 1988.
[32] A. Wolf, H. B. Brown, R. Casciola, A. Costa, M. Schwerin, E. Shamas,
and H. Choset, A mobile hyper redundant mechanism for search and
rescue tasks, Proc. IEEE/RSJ Intl. Conf. on Intelligent Robots and Sys-
tems, Taipei, Taiwan, May 2003, vol. 3, pp. 28892895.
[33] C. Wu, A kinematic CAD tool for the design and control of a robot
manipulator, Intl. Journal of Robotics Research, vol. 3, no. 1, pp. 58
67, 1984.