Control Synthesis for a Class of Light and Agile

Robotic Tensegrity Structures

J.B. Aldrich, R.E. Skelton K. Kreutz-Delgado
Dept. of Mechanical and Aerospace Engineering Dept. of Electrical and Computer Engineering
University of California, La Jolla, CA 920930411 University of California, La Jolla, CA 920930407
Abstract
For a new class of tendon-driven robotic systems that is gen-
eralized to include tensegrity structures, this paper focuses
on a method to determine the tendon force inputs from a set
of admissible, non-saturating inputs, that will move the rigid-
body system from point A to point B along a prescribed path
in minimum time. The approach utilizes the existence condi-
tions and solution of a linear algebra problem that describes
how the set of admissible tendon forces is mapped onto the
set of path-dependent torques. Since this mapping is not one-
to-one, free parameters in the control law always exist. An
innity-norm minimization with respect to these free param-
eters is responsible for saturation avoidance. Characterizing
and optimizing these free parameters is the new contribu-
tion. Feedback is introduced to attenuate disturbances aris-
ing from the tensegrity paradigm. Examples illustrate meth-
ods and validate tensegritys superior saturation avoidance
capability.
1 Introduction
Robotic automation of repetitive assembly processes contin-
ues to gain more acceptance as an effective means to re-
duce labor costs and increase productivity in many man-
ufacturing industries. This is especially true in pick-and-
place applications where the objective is to move from one
position to another quickly and accurately. As a result of
driving the trajectories faster and faster, the inertial dynam-
ics of typical robots become too large to be ignored, and
therefore must be compensated by a feedback controller.
Assuming that a sufciently accurate plant model is avail-
able, the standard approach to this control problem is to im-
plement a version of feedback linearization known as the
computed-torque controller, [79,12]. Theoretically, this ap-
proach allows arbitrarily large conguration changes within
the robots workspace to occur quickly. As a practical mat-
ter, however, reconguration in near-zero time is not possible
since the inertial forces to be carried by the actuators would
have excessive magnitude causing actuator saturation. One
way to circumvent this problem is to make the robot lose
weight. For instance, a lighter design is possible by plac-
ing the heavy actuators at the base of the manipulator where
a pulley-tendon system transmits torque remotely [8], Fig.
(1a,b). Unfortunately, tendon compliance makes it difcult
Figure 1: Proposed tensegrity robot evolution
to transmit torque with sufcient bandwidth [5]. One ap-
proach that circumvents this problem is to design a mech-
anism that reduces tendon usage [13]. Alternatively, band-
width can also be recovered by reducing system mass [4].
For instance, the tensegrity systems in Fig. (1c-h) can be
designed with exceptionally low system mass and superior
saturation avoidance capability since large bending moments
normally present in the links get absorbed in the tendon net-
work [10]. The work herein suggests that tensegrity con-
cepts will revolutionize the manner in which tendon-driven
systems are designed, controlled and utilized. We believe
this will become especially true in environments where agile
maneuvering and delicate object handling require a soft
touch.
In the sections that follow, we address the following ques-
tions: Given a set of admissible tendon forces, how should
the control law be designed? For a given robot, which ten-
don network sustains more torque? What is its minimum-
time trajectory along a prescribed path? How can feedback
be used to keep it on track? Lastly, tensegrity model building
methods and control simulations are illustrated by example.
0-7803-7896-2/03/$17.00 2003 IEEE 5245
Proceedings of the American Control Conference
Denver, Colorado June 4-6, 2003
2 Admissible tendon forces
A necessary condition for a tendon-driven robotic system is
that all tendons be taut and unbroken. When this condition
holds, we say that the tendon actuation system is in a state of
unsaturation as follows.
Denition 1 An m-tendon actuation system is said to be un-
saturated if t A where
A :=t R
m
: 0 <t
i
< f
yi
,
and t
i
is the tension in the i
th
tendon, and f
yi
is the yield force
(max allowable tension) for the i
th
tendon. A system that is
not unsaturated, is saturated. Set A is the set of admissible
tendon forces.
3 Control problem statement
Given a desired reference trajectory, q
d
, for a tendon-driven
rigid-body system modelled as
M(q) q+V(q, q) q+g(q) = G(q)t
(q R
n
, t A R
m
, M is square), we seek to answer the
question: Does there exist admissible tendon forces t to yield
the following closed loop system?
z +K
v
z +K
p
z +K
i
v = 0
z = qq
d
= v
Yes, if and only if there exists t A to solve
Gt = M( q
d
K
v
z K
p
z K
i
) +V q+g = . (1)
There exists a t R
m
solving (1) iff G
L
= 0 where G
L
G = 0
and G
L
G
T
L
~0. If t R
m
exists to solve (1), then all solutions
are given by
t = G
+
+G
R
(2)
where GG
R
= 0 and G
T
R
G
R
~0.
It is important to recognize two facts. First, the dimension of
the nullspace of G is (G) =mrank(G). Hence, contains
(G) free parameters that have neither been characterized
nor optimized by the robotics community [1, 6, 8]. Second,
even if a solution t R
m
exists there is no guarantee that
there exists a choice for such that t A . An approach that
resolves these facts is the main contribution of this paper.
4 How should free parameters be optimized?
In order to understand how to optimize the free parameters,
, a more suitable saturation denition is helpful. Toward
this end, we introduce
D = diag

2/ f
y1
2/ f
y2
2/ f
ym

e = [ 1 1]
T
, t = [ t
1
t
m
]
T
to establish a series of equivalent statements as follows.
Theorem 1 The following statements are equivalent:
(i) The m-tendon actuation system is unsaturated.
(ii) There exists a < 1 such that |Dt e|

.
(iii) t +d A if |d|

< (1)min
i
f
yi
/2.
Corollary 1 If the tendons are uniform, i.e. f
yi
= f
y
for i =
1: m, then D
1
e = ( f
y
/2)e =: t
d
and the following statements
are equivalent:
(i) The m-tendon actuation system is unsaturated.
(ii) := f
y
/2|t t
d
|

> 0
(iii) t +d A if and only if |d|

<
Proof: Thm.: t A holds iff (2/ f
yi
)t
i
+ 1 <
1 and (2/ f
yi
)t
i
1 < 1 for all i = 1, . . . , m, which
shows the m-tendon actuation system is unsaturated iff
max[(2/ f
y1
)t
1
1[, . . . , [(2/ f
ym
)t
m
1[ 1. Cor.: (ii)
|t t
d
|

< f
y
/2 |Dt e|

< 1 with f
yi
= f
y
for i =
1, . . . , m. If (ii) holds, the perturbation force of least
magnitude which causes either the onset of slackness or
yield for the i
th
tendon has magnitude: f
y
/2 [t
i
f
y
/2[.
The saturating perturbation of least magnitude among all
tendons has magnitude: min
i
f
y
/2[t
i
f
y
/2[ = f
y
/2
max
i
[t
i
f
y
/2[ = f
y
/2|t ( f
y
/2)e|

= .
In order to keep a tendon actuation system in a state of un-
saturation that is robust to perturbations, d, we look to part
(iii) of the corollary to motivate the following robust control
objective.
Robust control objective: Maximize saturation margin, ,
w.r.t. free parameters in real-time.
An equivalent objective follows immediately from part (ii)
of the corollary: Minimize tendon force deviation from t
d
(50%-yield) w.r.t. free parameters in real-time,

:= argmin

|t() t
d
|

= argmin
,
s.t. 0,
e G

+G
+
t
d
e
where t() is dened by (2) with the G

notation replacing
G
R
hereafter. Alternatively, the free parameters can be com-
puted as

2
:= argmin

|t() t
d
|
2
= argmin

|G

+G
+
t
d
|
2
= G
T
t
d
where G

G
T
t
d
is the orthonormal projection of t
d
into
null(G). The least-squares minimizer,

2
, can be expressed
5246
Proceedings of the American Control Conference
Denver, Colorado June 4-6, 2003
analytically, and computed independent of the controller. In
contrast, the min-max minimizer,

, requires a linear pro-

gram, but is more effective for saturation avoidance.
How effective is

at saturation avoidance? For a given

torque loading, , and robot geometry, q, the percent satura-
tion, S, can be computed as a function of the free parameters,
, as
S() := (2/ f
y
)|G

(q)+G
+
(q) t
d
|

(3)
In the example below, we illustrate how effective

is at
saturation avoidance in the presence of torque loading and a
variable structure geometry.
Example. In Fig. (2) we are given two tendon actuation
Figure 2: Benchmark comparison.
systems for a single link manipulator. The coupling matrices
that map t R
2
into R
1
for systems 1 and 2 are given
respectively by
G
1
= r

1 1

G
2
= a

esin

c
1
esin

c
2

where c
1
= 1+e
2
+2ecos, c
2
= c
1
4ecos and e = b/a.
We let a =

3/2, b = 1/2, r = 1/10, and plot in Fig. (3)

the level curves of percent saturation, S(

p
), where

p
is a
function of and . For instance,

= argmin

c
1

c
2

0.5c
2

c
1
(e+e
3
)sin
0.5c
1

c
2
(e+e
3
)sin

f
y
2
f
y
2

yields tendon forces t = G

+
2
+G

for system 2. Figure

(3b,c) shows the least-squares minimizer is slightly inferior
to the minmax minimizer for saturation avoidance. Figure
30 90 150
40
0
40
, (deg)

,

(
N
m
)
(a)
100
60
20
100
60
20
30 90 150
40
0
40
, (deg)
(b)
1
0
0
6
0
2
0
1
0
0
6
0
2
0
30 90 150
40
0
40
, (deg)
(c)
1
0
0
6
0
2
0
1
0
0 6
0
2
0
Figure 3: Level curves of percent saturation, S(

p
), as a function
of torque and geometry, and . (a) system 1 with p =
, (b) system 2 with p = , (c) system 2 with p = 2.
(3a,b) shows that system two is guaranteed to sustain greater
torque than system one.
The apparent leverage advantage of system two extends to
most tensegrity systems as follows. Let S
1
denote S(

) for
pulley-based robots, Fig. (1a,b). Let S
2
denote S(

) for
tensegrity-based robots, Fig. (1c-h). The upper bounds on
percent saturation,
S
1
= (2/ f
y
)|G
+
|

||

2
f
y
S
2
= (2/ f
y
)|G

t
d
+G
+
|

k
1
(q) +k
2
(q)
||

2
f
y
validate tensegritys leverage advantage for most congura-
tions where k
1
0, k
2
1 and the link lengths are suf-
ciently greater than the pulley radii, i.e. a r. Increasing
the sustainable torque means point-to-point maneuvering can
occur in less time.
5 Minimum-time control along a path
Given a rigid-body system, a tendon network and a path, we
pose the following problem statement:
What admissible tendon force inputs will move plant from
point A to point B along a prescribed path in minimum time?
The solution is broken down into three tasks: First, convert
the free dynamics to the path-following dynamics. Second,
substitute the path-following dynamics into the tendon satu-
ration constraint to get a path-acceleration constraint. Third,
construct the minimum-time solution by maximizing the path
velocity at each point on path [2].
Path-following dynamics. The rigid-body free dynamics
are expressed as, M(q) q +h(q, q) = . For kinematically-
invertible robots/tensegrities, the tip of the manipulator can
be written as an explicit function of the joint angles or the
distance along the path:
r(q) = r(s)
Differentiating w.r.t. time yields
r
q
q = r
s
s
r
q
q+ r
q
q = r
s
s + r
ss
s
2
where r
q
= r/q is the Jacobian. We now can compute
q = q(s), q = q(s, s), q = q(s, s, s) and the path-following
dynamics as
(s, s, s) = u(s) s +v(s, s) (4)
u(s) = M(q)r
1
q
r
s
v(s, s) = M(q)r
1
q
( r
ss
s
2
r
q
q) +h(q, q)
5247
Proceedings of the American Control Conference
Denver, Colorado June 4-6, 2003
Path-acceleration constraint. Recall the linear algebra
problem: (s, s, s) = G(s)t whose solution t = G
+
+G

exists iff [I GG
+
] =0. Assuming solution exists (which is
true automatically when G is full row rank), substitute it into
the tendon saturation constraint of theorem 1, |Dt e|

,
to get
|a(s, s, ) + sc(s)|

where a = DG

+DG
+
v e and c = DG
+
u. Equivalently,
we get a
i
(s, s, ) + sc
i
(s) for i = 1 to m. Rear-
rangement yields path-acceleration constraint,
f (s, s, ) s g(s, s, )
f = max
i
((sign(c
i
)a
i
)/c
i
)
g = min
i
((sign(c
i
)a
i
)/c
i
)
Minimum-time solution. The minimum-time solution is
obtained by choosing the acceleration s to make the veloc-
ity s as large as possible at every point s without violating
f (s, s, ) g(s, s, ). This follows by minimizing the cost,
J =

T
0
dt =

B
A
1
s(s)
ds
In [2], it was shown that J is minimized if and only if s al-
ways takes either its largest or its smallest admissible value.
In summary, it is easy to show the following.
Theorem 2 The path-following minimum time solution sub-
ject to t A is obtained by switching between maximum ac-
celeration, s = g(s, s,

), and maximum deceleration s =

f (s, s,

) where

= argmin

|A(s)+b(s, s) + sc(s)|

and A = DG

, b = DG
+
v e, c = DG
+
u.
Corollary 2 The path-following minimum time solution
subject to t A and min

|Dt() e|
2
is obtained by
switching between s = g(s, s,

2
) and s = f (s, s,

2
) where

2
= G
T
(s)D
1
e.
The remaining task is to determine when to switch between
f and g. Techniques for locating switching points are de-
scribed in [2, 11] and applied to an example at the end of
the paper. The minimum time solution yields the desired
open-loop trajectory, q
d
. Feedback can be used to reduce the
tracking error, z = qq
d
, as discussed next.
6 Model uncertainty and feedback control
For the uncertain plant, M(q) q+V(q, q) q+g(q) +d
w
= G t,
we compute t =

G

+

G
+
+u
r
as the feedback law where
u
r
is a robust control input, d
w
is an external disturbance
and G = G(q) is assumed. Three sources of parametric
uncertainty are considered here, namely, the standard er-
ror, = , that occurs in all robotic systems, and two
tensegrity-based errors,

G
+
= G
+

G
+
psuedoinverse error

= G

nullspace error
Assuming q is measured exactly and G is full row rank, the
closed loop system becomes M(q) q+V(q, q) q+g(q)+d
t
=
+Gu
r
, where the total disturbance is d
t
= d
w
+G

G
+
+
G

+G

G
+
.
Lyapunov design. Given time-optimal reference trajecto-
ries how do we stay on path using feedback? This can be
solved by using a Lyapunov function: V
1
=
1
2
r
T
M(q)r where
r =z+ z. First, we can rewrite dynamics in terms of ltered
tracking error r as, M(q) r =Y()V(q, q)r d
t
++Gu
r
,
where Y() = M(q)( z q
d
) +V(q, q)(z q
d
) g(q).
Then we choose nominal input as = Y() K
r
r with
error = Y()

. The robust input, u

r
= k
d

G
+
sgn(r) can
be implemented without error since we assume the state
is known exactly. The psuedoinverse error is bounded as
gI G

G
+
gI, where 0 g 1. The time-derivative
of the Lyapunov function becomes,

V
1
= r
T
M(q) r +r
T

M(q)r/2
= r
T
(Y()d
t
+ +Gu
r
)
+r
T
(

M(q)/2V(q, q))r
= r
T
(Y()d
t
+ +Gu
r
)
= r
T
K
r
r r
T
d
t
+r
T
Gu
r
= r
T
K
r
r r
T
d
t
+(k
d
+ g[k
d
[)
n

i=1
[r
i
[
r
T
K
r
r +
n

i=1
[r
i
[[d
ti
[ +(k
d
+ g[k
d
[)
n

i=1
[r
i
[
r
T
K
r
r +
n

i=1
[r
i
[

max
i
[d
ti
[ +k
d
+ g[k
d
[

r
T
K
r
r if k
d
=[k
d
[, [k
d
[
[[d
t
[[

(1 g)
(5)
If the robust control gain requirement in (5) holds, then

V
1
0 and a La Salles argument given in [7] can be used
to show that the tracking errors are asymptotically stable.
(The skew-symmetric property is used to simplify equality
2 above. For equality 4, use r
T
Gu
r
= k
d
r
T
G

G
+
sgn(r) =
k
d
r
T
sgn(r)k
d
r
T
G

G
+
sgn(r) (k
d
+ g[k
d
[)
n
i=1
[r
i
[ to sim-
plify.)
The gain requirement on [k
d
[ can be reduced with the adap-
tive inertial-related control [7]. Finite-bandwidth robust con-
trol is possible by replacing the sign function with a satura-
tion function in u
r
[9]. For standard non-tensegrity robots,
(5) reduces to [k
d
[ [[d
w
+ [[

.
5248
Proceedings of the American Control Conference
Denver, Colorado June 4-6, 2003
7 Tensegrity Model
In this section, a family of dynamical system models is gen-
erated for a generalized class of tendon-driven robots that
includes tensegrity structures. First, the rigid body dynamic
models for serial-link and free-link systems are developed.
Second, tendon actuation model building is conveyed by an
illustrative example.
Figure 4: (a) Free-link. (b) Serial-link.
Serial-link model. The geometry of a serial-link rigid-body
system with b bars (links) is shown in Fig. (4b) for the two-
bar case. For convenience, its n degrees of freedom are or-
ganized as
q =

q
T
1
q
T
2
. . . q
T
b

T
R
n
, q
i
=

i

i

T
The position of the i
th
bars tip, p
i
, and center of mass,
p
ci
, are expressed in terms of q as follows
p
i
= p
o
+
i

k=1
a
k

k
p
ci
= p
o
+
i1

k=1
a
k

k
+a
ci

i
=

c
i
c
i
s
i
c
i
s
i

where c() = cos(), s() = sin(), a

i
is the length of the i
th
bar, and a
ci
is the distance from node to center of mass.
Time-derivative of position yields velocities,
p
i
=
i

k=1
a
k

k
q
k
q
k
=:
i
(q) q
p
ci
=
i1

k=1
a
k

k
q
k
q
k
+a
ci

i
q
i
q
i
=:
ci
(q) q
where (3 n) matrices,
i
(q) =

J
1
J
2
J
i
O
ni

and

ci
(q) =

J
1
J
2
J
i1
J
ci
O
ni

, consist of Jacobians, J
i
=
a
i

i
/q
i
and J
ci
= a
ci

i
/q
i
, and a matrix of zeros with
n 2i columns, O
ni
. The kinetic energy for the rigid body
system becomes
K =
1
2
b

i=1
(m
i
p
T
ci
p
ci
+ q
T
i
I
i
(q
i
) q
i
) (6)
=
1
2
q
T

i=1
m
i

T
ci
(q)
ci
(q) +I (q)

q =:
1
2
q
T
M(q) q
where I = diag

I
1
I
2
I
b

R
nn
, consists of inertial
blocks, I
i
= diag

m
i
12
(a
i
c
i
)
2 m
i
12
a
2
i

. Notice how the kinetic

energy derivation yields the mass matrix as a freebee! The
potential energy for the rigid body system is
V =
b

i=1
p
T
ci
f
gi
, f
gi
=

0 m
i
g 0

T
The kinetic and potential energies are differentiated in La-
granges equations of motion,
d
dt
K
q

K
q
+
V
q
=
and rearranged into the following matrix form [12],
M(q) q+V(q, q) q+g(q) =
Specically, if we denote the mass and coriolis/centripetal
matrices elementwise as M = [
k j
] and V = [
k j
], then V
can be completely determined from M using the following
Christoffel parameters [12],

k j
=
1
2
n

i=1

k j
q
i
+

ki
q
j

i j
q
k

q
i
(7)
The gravity-induced torque vector becomes
g(q) :=
V
q
=
b

i=1

p
ci
q

T
f
gi
=
b

i=1

T
ci
f
gi
Nodal forces, f =

f
T
1
f
T
2
f
T
b

T
, in cartesian space,
f
i
R
3
, are induced by the tendon actuation system. Using
the principle of virtual work [12], the mapping from nodal
forces, f R
3b
, to torques, R
n
, is
=
b

i=1

T
i
f
i
=

T
1

T
2

T
b

f =: H(q) f
Free-link model. For tensegrity structures designed with
inertially-isolated rigid bars as depicted in Fig. (4a), a single
bar has ve degrees of freedomthree translations of its mass
center, p
c1
R
3
, and two Euler angles, q
1
R
2
. Using (6)
with b = 1, the kinetic energy of a single free link is
K =
1
2
(m
1
p
T
c1
p
c1
+ q
T
1
I
1
(q
1
) q
1
)
=
1
2

p
T
c1
q
T
1

m
1
I 0
0 I
1
(q
1
)

p
c1
q
1

=:
1
2
q
T
M(q) q
Notice that q is now a 5-vector, not a 2-vector as before.
Use (7) again to get V(q, q) from M(q). The gravity vector,
g(q) =

I 0

T
f
g1
is trivial. Using the principle of virtual
work, this nodal force to torque mapping becomes
=

I I
J
T
c1
J
T
c1

f
1
f
2

=: H(q) f
5249
Proceedings of the American Control Conference
Denver, Colorado June 4-6, 2003
8 Illustrative Examples
In this section, models for two different plants are generated
and a minimum-time control simulation is given to illustrate
the simplicity, scope, and potential of light and agile robotic
tensegrity structures.
Plant 1. Two inertially-isolated double-link manipulators in-
terconnected with exible tendons are shown in gure (5).
This system can be described by two copies of the serial-link
Figure 5: Tensegrity robot joining two workpieces.
model where b = 2, n = 2 and
i
= 0 for each copy. Com-
bining the copies yields
M(q) q+V(q, q) q+g(q) = H(q) f (8)
as the global model where
M = diag

M
12
M
34

, V = diag

V
12
V
34

H =

H
12
H
34

, q =

q
12
q
34

, f =

f
12
f
34

M
i j
=

m
i
a
2
ci
+m
j
a
2
c j
+I
i
m
j
a
c j
a
i
(c
i j
)
m
j
a
c j
a
i
(c
i j
) m
j
a
2
c j
+I
j

V
i j
=

0 m
j
a
c j
a
i
(s
i j
)

j
m
j
a
c j
a
i
(s
i j
)

i
0

H
i j
=

J
T
i
J
T
i
0 J
T
j

, J
T
i
= a
i

s
i
c
i

and
i j
=
j

i
, q
i j
=

i

j

T
, f
i j
=

f
T
i
f
T
j

T
, s
i
=
sin
i
, c
i
= cos
i
. Gravity acts orthogonal to the plane of
motion, and therefore g = 0. The nodal points, p
i
, are com-
puted using forward kinematics,
p
i
= p
oi
+a
i

c
i
s
i

, p
i+1
= p
i
+a
i+1

c
i+1
s
i+1

where i = 1 for the rst 2-link arm and i = 3 for the second
2-link arm. The tendon orientation vectors,
i
, are dened in
terms of the nodal points as

I
I
I
I
I I
I I
I I
I I

p
o1
p
o3
p
1
p
2
p
3
p
4

=C
o
p
o
+Cp
which is used to compute the direction cosine matrix,
D = diag(

d
1
d
2
d
3
d
4
d
5
d
6

) where d
i
=
i
(
T
i

i
)
1/2
.
The principle of virtual work yields f =C
T
D(q)t. The rigid-
body and tendon-actuation systems are joined as
M(q) q+V(q, q) q+g(q) = G(q)t, G := HC
T
D
G =

0 J
T
1
d
2
J
T
1
d
3
J
T
1
d
4
J
T
1
d
5
J
T
1
d
6
0 0 0 J
T
2
d
4
0 J
T
2
d
6
J
T
3
d
1
0 J
T
3
d
3
J
T
3
d
4
J
T
3
d
5
J
T
3
d
6
0 0 0 0 J
T
4
d
5
J
T
4
d
6

It is easy to see when G R

46
drops rank, since J
T
i
d
j
= 0
if and only if tendon j is unattached or colinear to bar i.
Plant 2. A two-stage tensegrity-based manipulator shown in
gure (6). This system can be described by three copies of
p311
p211
p111
p221
p121
p212
p312
p222
p322
p112
p122
p321
Figure 6: 3D tensegrity robot xed at p
111
, p
211
and p
311
.
the serial-link model with b = 1 and n = 2, and three copies
of the free-link model. Assuming zero-gravity yields g = 0
and the global model (8) as
M = diag

M
11
M
21
M
31
M
12
M
22
M
32

V = diag

V
11
V
21
V
31
V
12
V
22
V
32

H = diag

H
11
H
21
H
31
H
12
H
22
H
32

q =

q
T
11
q
T
21
q
T
31
q
T
12
q
T
22
q
T
32

T
, q
i1
=
i1
f =

f
T
11
f
T
21
f
T
31
f
T
12
f
T
22
f
T
32

T
, f
i1
= f
i12
q
i2
=

p
T
ci

T
i2

T
, f
i2
=

f
T
i21
f
T
i22

T
M
i1
= 4I
i1
, V
i1
= 4C
i1
, H
i1
= 2J
i1
M
i2
= diag

m
i j
I I
i2

, V
i2
= diag

0 C
i2

H
i2
=

I I
J
i2
J
i2

, I
i j
=
m
i j
a
2
i j
12

(c
i j
)
2
0
0 1

C
i j
=
m
i j
a
2
i j
s
i j
c
i j
12

i j

i j

i j
0

J
i j
=
a
i j
2

s
i j
c
i j
c
i j
c
i j
0
c
i j
c
i j
s
i j
s
i j
c
i j

Euler-angles are denoted by

i j
=

i j

i j

T
. Nodal force
f
i jk
R
3
is applied to node p
i jk
R
3
where
p
i12
= p
i11
+a
i1

i1
p
i21
= p
ci
(a
i2
/2)
i2
p
i22
= p
ci
+(a
i2
/2)
i2

i j
=

c
i j
c
i j
s
i j
c
i j
s
i j

5250
Proceedings of the American Control Conference
Denver, Colorado June 4-6, 2003
for i = 1, 2 and 3. The nodal points, tendon orientation
vectors, direction cosines and G are computed in the same
manner as in the previous example.
Simulation. The minimum-time tendon forces are designed
for the two-link serial manipulator pictured in (1a). The ten-
don forces illustrated in gure (7a) are designed to move the
manipulator tip along a specic path, i.e. x =4(s0.5)
3
and
y = s +5, in minimal time according to corollary 2. Four
tendons are used, but only two of them are illustrated in the
gure, since t
2
= f
y
t
1
and t
4
= f
y
t
3
due to symmetry in
the tendon network. The phase plane trajectory, i.e. s versus
0 0.05 0.1 0.15 0.2
0
10
20
30
40
50
60
70
80
90
100
t
e
n
s
i
o
n
,

(
N
)
time, (seconds)
t
t
t
t
1
3
1
3
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
s, (m)
d
s
/
d
t
,

(
m
/
s
e
c
)
MVC
Max Deceleration Max
Accel.
Switch
Point
MVC
Figure 7: (a) Time-optimal tendon force trajectories for 5 t
95N with switching point at 0.14 sec. (b) Maximum
Velocity Curve (MVC) and optimal path trajectory.
s, for this design is illustrated in gure (7b). Clearly, the con-
trol input is feasible since it is bounded above by the maxi-
mum velocity curve. The switching point indicates where on
the path the maximum path-acceleration, g, is switched to
the maximum path-deceleration, f. See [2, 11] for algorithm
details. This procedure can be repeated for plants 1 and 2 by
prescribing a feasible path for all degrees of freedom.
Tension rate constraint. Notice in gure (7a) the ten-
don forces are discontinuous at the switching point. Since
tendons have nite bandwidth, the control law must be
smoothened for implementation. Smooth minimum-time tra-
jectories are computed by using the algorithm given in [3]
to ensure tension rates, t, instead of torque rates, , are
bounded. This substitution is straightforward once we ex-
press t in exactly the same form as as follows
t = (s)
...
s +(s, s, s), t R
m
= (s)
...
s +(s, s, s), R
n
. (9)
To derive , , and , we rst dene
i
as the i
th
column of
G
+
R
mn
,
i
as the i
th
column of G

R
m
, and note that
[3] shows how in (4) can be differentiated to get and
in (9). Then, =
n
i=1
(
i

i
q
q+
i

i
) +

i=1
(
i

i
q
q+
i

i
)
and =
n
i=1

i
, where the elements of , , , and are
denoted by subscript i. To keep independent of
...
s , we use
the least-squares minimizer, = ( f
y
/2)G
T
e R

, which
yields
i
= ( f
y
/2)
m
j=1

i j
and
i
= ( f
y
/2)
m
j=1

i j
q
q. All
trajectories are smooth because [3] uses cubic splines to pa-
rameterize s(s).
9 Conclusion
This paper describes a feedback linearization control lawthat
uses the parameters in the nullspace of the control distribu-
tion matrix, G, to minimize the norm of the tendon force
tracking error, |t t
d
|, while avoiding saturation of the con-
trol signals. This paper has also shown that when the free
parameters in the tendon control law are optimized in real-
time, control synthesis for a generalized class of light and
agile robotic tensegrity structures is possible. Future work
will focus on integrating structure and control design so that
optimal candidates within this new robotics paradigm can be
identied on a task-by-task basis.
References
[1] A. Bicchi and D. Prattichizzo. Analysis and opti-
mization of tendinous actuation for biomorphically designed
robotic systems. Robotic systems, 2000-1.
[2] J.E. Bobrow, S. Dubowsky, and J.S. Gibson. Time-
Optimal Control of Robotic Manipulators Along Specied
Paths. Int J of Robot Res, 4(3), 1985.
[3] D. Constantinescu and E.A. Croft. Smooth and
Time-Optimal Trajectory Planning for Industrial Manipula-
tors along Specied Paths. J Robot Sys, 2000.
[4] V. Hayward and J. Cruz-Hernandez. Parameter sen-
sitivity analysis for design and control of tendon transmis-
sion. Experimental Robotics, 1995.
[5] C. Johnstun and C. Smith. Modelling and design of
a mechanical tendon actuation system. ASME Trans J Dyn
Sys Meas Control, 1992.
[6] H. Kobayashi, K. Hyodo, and D. Ogane. On intel-
legent control of tendon-driven robotic mechanisms with re-
dundant tendons. Int J Robot Res, 17(5):561571, 1998.
[7] F.L. Lewis, C.T. Abdallah, and D.M. Dawson. Control
of Robotic Manipulators. Macmillian, 1993.
[8] R. Murray, Z. Li, and S. Sastry. A Mathematical In-
troduction to Robotic Manipulation. CRC Press, 1994.
[9] Z. Qu and D.M. Dawson. Robust Tracking Control of
Robot Manipulators. IEEE Press, 1996.
[10] R.E. Skelton, J.W. Helton, R. Adhikari, J. Pinaud, and
W. Chan. An introduction to the mechanics of tensegrity
structures. In Conference on Decision and Control, vol-
ume 5, pages 42549, 2001.
[11] J.-J. E. Slotine and H. Yang. Improving the Ef-
ciency of Time-Optimal Path-Following Algorithms. IEEE
Trans Robot Automat, 5(1):118124, 1989.
[12] M. Spong and M. Vidyasagar. Robot Dynamics and
Control. Wiley, 1989.
[13] W.T. Townsend and J.A. Guertin. Teleoperator slave -
W.A.M. design methodology. Industrial Robot, 26(3):167
177, 1999.
5251
Proceedings of the American Control Conference
Denver, Colorado June 4-6, 2003