See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/224331947

Optimal Motion Planning for Differentially Flat Underactuated Mechanical

Systems

Conference Paper · October 2008

DOI: 10.1109/ICAL.2008.4636403 · Source: IEEE Xplore

CITATIONS READS
3 450

2 authors:

Guangping He Zhiyong Geng

North China University of Technology Peking University
85 PUBLICATIONS   397 CITATIONS    159 PUBLICATIONS   855 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Guidance and control on non-Euclidean spaces in the presence of uncertainties View project

NSFC Project: Dynamics Synthesis And Control of Robotic Systems with Nonholonomic Constraints View project

All content following this page was uploaded by Guangping He on 28 March 2015.

The user has requested enhancement of the downloaded file.

 Proceedings of the IEEE
International Conference on Automation and Logistics
Qingdao, China September 2008

Optimal Motion Planning for Differentially Flat

*
Underactuated Mechanical Systems
Guang-Ping He Zhi-Yong Geng
State Key Laboratory for Turbulence and Complex Systems, State Key Laboratory for Turbulence and Complex Systems,
Department of Mechanics and Aerospace Engineering Department of Mechanics and Aerospace Engineering
Peking University Peking University
Beijing 100871, P.R. China Beijing 100871, P.R. China
hegp55@sina.com hegp55@sina.com

Abstract - It is shown that the flat output of the single input
underactuated mechanical system can be obtained by finding a
smooth output function such that the system has relative degree
equals to the dimension of the state space. Assuming the flat
output of an underactuated system can be solved explicitly, an
optimization method is proposed for the motion planning of the
differentially flat underactuated mechanical systems by
constructing a shape adjustable curve with satisfying specific
boundary conditions in flat output space. The inertia wheel
pendulum is used to verify the proposed optimization method
and some numerical simulation results are included.
Index Terms - Differential flatness\, Underactuated system,
Motion planning, Optimization.
I. INTRODUCTION
The underactuated mechanical systems that the degrees of
freedom (DOF) are more than the numbers of independent
inputs, allow to reduce cost, weight as well as the occurrence
of failures, thus can be used in the fields such as space [1],
underwater [2], and biomechanical systems [3] etc. The
underactuated mechanical systems are generally slaved by
first-order [1] or second-order [4] nonholonomic constraints in
high nonlinear form. In some cases, the linear approximation
of the underactuated mechanical systems lose the
controllability [5], thus the motion planning and control
problem for the underactuated system is generally nonlinear in
nature. In the field of nonlinear control, so far, some effective
methods [6-8] are just developed for the nonlinear systems
necessary condition, while for multi-inputs nonlinear system,
only a necessary condition had been presented [6].
If an underactuated mechanical system can be confirmed
to be differential flat and the flat output can be expressed
explicitly, then the problem of finding a trajectory satisfying
the underactuation constraints becomes the relatively simple
algebraic problem of finding a curve to fit the start and final
constraints on flat output. As to be shown in this paper, by
constructing a polynomial with redundant design parameters,
optimizing the feasible trajectory for the underactuated
mechanical system is also possible. For instance, considering
the main application fields of underactuated mechanical
system, improving the energy efficiency of the motion of the
underactuated system is appealing because that the power of
the remote manipulation system is limited generally.
II. THE DYNAMICS OF UNDERACTUATED MECHANICAL
SYSTEMS
Consider a mechanical system with n DOF, denote by
q ∈ R n the vector of generalized coordinates of the system.
Assume there are no external constraints on the system, the
dynamics of the mechanical system can be calculated by
Euler-Lagrange equation given as
d ∂L ∂L
− = F (q ) IJ (1)
dt ∂q ∂q
where L is the Lagrangian, IJ ∈ R m is the vector of
generalized forces. If m < n , then F (q) = [0, I m ] , where
T

with specially geometric or algebraic structure such as

differentially flat [9] or nilpotent [7] systems. I m ∈ R m×m is the identity matrix. Partition the generalized
In this paper, optimizing the motion of underactuated
mechanical system with differentially flat property is
[ ]
coordinates as q = qp qa ∈ Rn−m × Rm according to F (q ) ,
investigated. Differential flatness was first defined by Fliess et where q p and qa are the unactuated and actuated generalized
al. [9]. Now it is well known that both motion planning and coordinates respectively. After partitioning the inertia matrix
control problem for a nonlinear system with differentially flat of the system accordingly, then the dynamics (1) can be
property are simple [6] since that the motion planning problem partitioned as
can be transformed to solve a algebraic equation in flat output ªm pp m pa º ªqp º ª c1 (q, q ) º ª 0 º
space, as well as the nonlinear control problem is equivalent to « » + = (2)
¬ m ap m aa ¼ «¬qa »¼ «¬c 2 (q, q )»¼ «¬IJ »¼
a linear one in flat output space. Unfortunately, so far there is
no systematic way to determine if a nonlinear system is ªm m º ªc (q, q)º
differentially flat, or what the flat outputs for a nonlinear where « pp pa» is the inertia matrix of the system, « 1 »
¬map maa ¼ ¬c2 (q, q)¼
system are, with the major exception of the single-input
system. For the single-input system, there is a sufficient and

*
This work is partially supported by NSFC with No.50475177, Beijing NSF with No.3062009 and CPSF with No. 20070410001 to G-P. He.

978-1-4244-2503-7/08/$20.00 © 2008 IEEE

1567

Authorized licensed use limited to: North China University of Technology. Downloaded on October 5, 2008 at 19:40 from IEEE Xplore. Restrictions apply.
includes centrifugal, Coriolis, gravitational and frictional outputs without integration. More rigorously, if the system has
forces. Since there are no external generalized forces, the first states x ∈ R n , and inputs u ∈ R m then the system is flat if we
n − m rows of equations (2) are given by can find outputs y ∈ R m (equal in number of inputs) with the
m pp qp + m pa qa + c1 (q, q ) = 0 (3)
form y = h( x , u, u ,!, u (α ) ) such that the states and inputs can
and it can be seen as “constraints” of the subsystem
be expressed as
m ap qp + m aa qa + c 2 (q, q ) = IJ (4)
x = ϕ ( y, y ,!, y ( β ) )
The second-order differential “constraints” (3) have first
u = Ȗ ( y, y ,!, y ( β ) )
integral if and only if the passive generalized coordinates q p
Since the algebraic structure between the states and the flat
are cyclic [4], viz., satisfying outputs, the motion planning and control design for the
∂L differentially flat nonlinear systems become simple.
=0 (5)
∂q p Unfortunately, so far there is no systematic way to determine
This means the dynamics (1) has conserved quantities if a multi-inputs nonlinear system is differentially flat, or what
∂L the flat outputs for a multi-inputs nonlinear system are. For the
= constant (6) single-input nonlinear system, the sufficient and necessary
∂q p condition was presented by the following theorem [6]:
Equations (6) are a set of first-order differential equations that
Theorem 1: A single input nonlinear system of the form
indicate the generalized momentums are constants such as
showing in the space free-floating manipulator system [1]. We x = f ( x, u ) , x ∈ R n , u ∈ R is differentially flat if and only if
consider the case that the passive generalized coordinates q p it is feedback linearizable.
are not cyclic, viz., the equations (3) absence the first integrals Exact linearization via nonlinear feedback is one of the
then showing a set of second-order “nonholonomic most important fruits of the geometric control theory in
constraints”. This terminology used in this paper emphasizes nonlinear system during the past two decades [13]. The
that the motion planning and control problem for the feedback linearization theorem for the single-input nonlinear
underactuated mechanical system can indeed be dealt with by system can be recited as follow [13]:
the same tools as for the classical nonholonomic systems.
For simplifying the dynamics (3)-(4), by the partial Theorem 2: The single-input nonlinear system of the form
feedback linearization proposed by Spong [12], the dynamics x = f ( x ) + g ( x )u , x ∈ R n , u ∈ R , is feedback linearizable if
(2) can be transformed to and only if there exist a neighborhood U of a point x 0 and a
qp = −m pp−1c1 (q, q ) − m pp−1 m pa u real-valued function h( x ) , defined on U , such that the single
(7)
qa = u input and single output (SISO) system
by the input transformation x = f ( x ) + g ( x )u
(10)
IJ = (m aa − m ap m pp−1m pa )u + c 2 (q, q ) − m ap m pp−1c1 (q, q ) (8) y = h( x )
where u = qa is the new input. Then, the system (7) can be satisfies conditions:
written to a form in state space as (a). Lg Lkf h( x ) = 0, k < n − 1
x = f ( x ) + g ( x )u , x ∈ R 2 n , u ∈ R m (9) (b). Lg Lnf−1h( x ) ≠ 0 .
where
The Theorem 2 indicates the single-input nonlinear
ª x1 º ª q p º ª q p º ª 0 º
« » « ». system is feedback linearizable if and only if the real-valued
« x » «q » , − m −1
c q q m pp m pa »
( ,  ) , − −1
2 « p «
x = « » = « » f ( x) = «
» pp 1 » g( x) = « function h( x ) defined at a point x 0 can be found such that the
« x3 » q q a » « 0 »
« » « a»
x
«
« 0
»
»
«
I
» relative degree of the SISO system satisfies r = n . Obviously,
¬ 4 ¼ ¬«q a ¼» ¬ ¼ ¬ m ¼
the relative degree of a nonlinear system depends on the
The main difficulty for motion planning and control design for
selection of the output function, and if the point x 0 can be
the underactuated system in (9) is that after partial feedback
given arbitrarily, the nonlinear system is globally linearizable
linearization, the new control u appears in the dynamics of
by nonlinear feedback. If the output function h( x ) is
both subsystems ( x1 , x 2 ) and ( x 3 , x 4 ) . It will be shown in
available, then the Theorem 2 indicates the derivative of the
next section that the underactuated system (9) can be output has relationships
transformed to a linear system if it is differentially flat.
y = L f h( x )
III. DIFFERENTIALLY FLAT UNDERACTUATED MECHANICAL #
SYSTEMS (11)
y ( n −1) = Lnf−1 h( x )
Differential flatness was first defined by Fliess et al. [9]
using the formalism of differential algebra. Roughly speaking, y ( n ) = Lnf h( x ) + L g Lnf−1 h( x )u
a system is differentially flat if one can find a set of outputs If one defines the coordinates transformation as
such that all states and inputs can be determined from these z i = φi ( x ) = Lif−1 h( x ), 1 ≤ i ≤ n (12)

1568

Authorized licensed use limited to: North China University of Technology. Downloaded on October 5, 2008 at 19:40 from IEEE Xplore. Restrictions apply.
and defines the input change as For ∀x ∈ R n satisfying Lg Lnf−1 h( x )) ≠ 0 , the matrix
1
u= (− Lnf h( x ) + v) (13) ∂
Lg L f h ( x )
n −1
ĭ (x ) has full rank [13], i.e. there exist the inverse
∂x
then the system (10) can be transformed to a linear system coordinate transformation
with form x = ĭ −1 ( z ) = ĭ −1 ( y, y ,!, y ( n−1) ) (18)
z1 = z 2
and the input change
# 1
(14) u= (− Lnf h(ĭ −1 ( z )) + v) (19)
z n −1 = z n Lg Lnf−1 h(ĭ −1 ( z ))
z n = v therefore both the states and input can be expressed as the
It is obviously that the linear system (14) is controllable. functions of flat output and its finite order derivatives.
Therefore, the main task for linearizing the nonlinear system IV. OPTIMAL MOTION PLANNING FOR THE FLAT
x = f ( x ) + g ( x )u , x ∈ R n , u ∈ R by nonlinear feedback (13) UNDERACTUATED MECHANICAL SYSTEM
is to seek an output function h( x ) that makes the system has
For the flat underactuated mechanical system, the problem
relative degree r = n . Thanks to the following theorem [13] of finding a feasible trajectory ( x(t ), u (t )) between the initial
the output function can be found for the SISO nonlinear
system. state x 0 = x (t 0 ) and the final state x 1 = x (t1 ) for the
underactuated system is changed to the problem of finding a
Theorem 3: There exists the output function h( x ) for which flat output curve y (t ) satisfying boundary conditions
the single-input nonlinear system with form x = f ( x) + g( x)u ,
z 0 = [y y ( n −1) ] (t o ) and z 1 = [y y " y ( n−1) ] (t1 )
T T
y "
x ∈ R , u ∈ R , has relative degree r = n at a point x if and
n 0
specified by x 0 = x (t 0 ) and x 1 = x (t1 ) respectively.
only if the following conditions are satisfied.
[ ]
(a) The matrix g adf g " adnf−2 g adnf−1g is full rank at x 0 ;
Therefore, the problem finding a trajectory satisfying the
underactuated constraints becomes the relatively simple
(b) The distribution Δ = span{g, ad f g,!, ad nf−2 g} is involutive algebraic problem of finding a curve to fit initial and final
conditions on y (t ) . By the inverse coordinate transformation
near x 0 .
(18), any curve y (t ) maps directly to a consistent pair of state
Note that if the point x 0 can be specified arbitrarily in the and control histories x (t ) and u (t ) . For instance, the flat
state space, then the Theorem 3 is globally effective. output can be parameterized by the polynomial
Assuming this is the case we here consider, then the following s −1
k
§t·
Theorem is easy to be proven. y = ¦ a s −1 ¨ ¸ (20)
s =1 ©T¹
where t ∈ [t 0 , t1 ] , T = t1 − t 0 , and a s −1 , s = 1,2,!, k are the
Theorem 4: The single nonlinear system with
form x = f ( x) + g(x)u, x ∈Rn ,u ∈R is differentially flat if and
design parameters. Define τ T = t T ∈ [0,1] to be a new time
only if the following conditions are satisfied.
variable, and let k > 2n , where n is the dimension of the state
[
(a) The matrix g ad f g " ad nf−2 g ad nf−1 g is full rank; ] space. Substitute the boundary conditions into (20), one has
(b) The distribution Δ = span{g, adf g,!, ad n−2
f
g} is involutive. ªz 0 º
And the flat output can be solved by « 1 » = Aa (21)
¬z ¼
∂h ( x )
= [0 0 " 0 δ ( x )] C −1 ( x ) (15) where a = [a0 a1 " ak ] , and the matrix A ∈ R 2 n×k has
T
∂x
where form
[
C ( x ) = g ad f g " ad nf−2 g ad nf−1 g ] (16) ª1 τ T (0)
«
τ T2 (0) " τ Tk −1 (0) º
»
«0 1 T 2τ T (0) T " (k - 1)τ T (0) T
k −2
is addressed as the controllability matrix of the system, and »
δ ( x) ≠ 0 . «# # # # # »
«0 0 0 " ( k − 1)!τ Tk − n (0) ((k - n)!T n -1 )»
The Theorem 4 provides a systemic way to find the flat A=« »
«1 τ T (1) τ T2 (1) " τ Tk −1 (1) »
output for the single-input underactuated mechanical system. «0 1 T 2τ T (1) T " (k - 1)τ T (1) T
k −2
»
Assume the flat output could be obtained by equation (15), the « »
coordinate transformation (12) gives «# # # # # »
z = ĭ( x) (17) «¬0 0 0 " ( k − 1)!τ Tk − n (1) ((k - n)!T n -1 ) »¼
where where τ T (0) = 0,τ T (1) = 1 . Obviously, the matrix A has full
z = [z1 z 2 ! z n ] = [y y " y ( n −1) ] rank with rank( A) = 2n . Since k > 2n , the general solution of
T T

[
ĭ ( x ) = h( x ) L f h( x ) " L h( x ) n −1
f ]
T
algebraic equation (21) can be written as

1569

Authorized licensed use limited to: North China University of Technology. Downloaded on October 5, 2008 at 19:40 from IEEE Xplore. Restrictions apply.
 ªz 0 º
a = A + « 1 » + μ ( I − A + A)İ
¬z ¼
where A + denote the Moore-Penrose generalized inverse and
is given by A + = A T ( AA T ) −1 , μ > 0 is a scale multiplier, I is
the identity matrix, İ is an arbitrary vector. The term
(I − A+ A) span the null space of A . This formulation
provides a decoupled solution. The first term is the special
solution that makes a 2 be minimized, and it gives a set of
design parameters a * such that the polynomial y (t ) satisfies
the boundary conditions. The second term μ ( I − A + A)İ can
change the shape of the curve y (t ) while not violating the
boundary conditions z 0 and z 1 since A( I − A + A)İ ≡ 0 .
Therefore, the equation (22) provides an approach to
optimizing the motion of the flat mechanical system for
specific sake.
To improve the energy efficiency is a broadly interested
task for designing the control for the mechanical system. As
an example, considering the optimization problem that
minimizes a measure about the kinetic energy. The measure
can be defined as
K = ³ ( x T Gx )dt
1 t1
(23)
2 t0
where G is a weighted matrix . With considering the equation
(16), the measure (23) can be rewritten as
generalized actuation force IJ i is not an algebraic function of
(22)
state variables, the gradient of the measure (26) cannot be
obtained by (25), the optimization problem can but be solved
by a optimization method without using the gradient of the
measure.
V. MOTION PLANNING AND CONTROL OF THE INERTIA
WHEEL PENDULUM
The inertia wheel pendulum (IWP) is a planar inverted
pendulum with a revolving wheel at the end, as shown in
Figure 1. The wheel is actuated and the joint of the pendulum
at the base is passive. The IWP was first introduced by Spong
et al. [14]. The task is to stabilize the pendulum at its upright
equilibrium point while the wheel stops to a given position.
The IWP is the first example a flat underactuated of
mechanical system with two DOF and single actuator. This is
due to the constant inertia matrix of the system [15]. In this
section, the IWP is considered to verify the feasibility of the
optimization method.
y θ2
m2 , I 2
m1 , I1
θ1 lc
l
o
x

2 t0
[
K (a ) = ³ (ĭ −1 ( z ) ) G (ĭ −1 ( z ) )dt
1 t1 T
]
(24)
Fig.1. The inertia wheel pendulum
Therefore, the kinetic energy is a function of the design
parameters of the polynomial (20). Let A. The flat output of the IWP
∂ Given the length of the link is l , the distance between
İ = − K (a ) (25) the center of mass (CM) of the link and the passive joint is lc ,
∂a
with μ > 0 , then the equation (22) gives a locally optimal the mass of the link and the wheel are m1 and m2
solution for minimizing the measure (23). If the arbitrary respectively, I1 and I 2 are the inertias of the link and wheel
vector İ can be expressed explicitly, then by adjusting the respectively. Denote the generalized coordinates of the system
scale μ , a better suboptimal solution can be obtained
to be q = [θ 1 θ 2 ] . The dynamics of the IWP is given by
T

generally. If the arbitrary vector İ cannot be obtained

explicitly, a global optimization algorithm such as the m11θ1 + m12θ2 + c1 (θ1 ) = 0
(27)
evolution methods (for instance, genetic algorithm) or the m21θ1 + m22θ2 + c2 = τ
randomized methods (for instance, probabilistic roadmap where
method) can be used to solve the problem, and the formulation
m11 = m1lc2 + m2 l 2 + I1 + I 2
(22) is also useable as long as one regards the arbitrary vector
İ and the scale multiplier μ as new design variables. Some m12 = m21 = m22 = I 2
global optimization methods permit determining the arbitrary c1 (θ1 ) = −(m1lc + m2l ) g sin θ1
vector İ without using the gradient of the measure. Therefore, c2 = 0
the energy efficiency measure can be defined by a function
without depending on the design parameters a explicitly. For Let x = [θ1 θ1 θ2 θ2 ]T , h0 =
(m1lc + m2l ) g , h1 = −I 2 ,
m1lc2 + m2l 2 + I1 + I 2 m1lc2 + m2 l 2 + I1 + I 2
instance, the measure can also be defined as
t1
the state space equation of the system (27) can be written as
K = ³ IJ T q a dt (26) x = f ( x ) + g ( x )u (28)
t0
where
f ( x ) = [x2 h0 sin x1 x4 0]
The measure (26) calculates the total works of the actuators on T

the actuation duration. This is a more direct way to evaluate

g ( x ) = [0 h1 0 1]
T
the energy efficiency of the specific motion. Since the

1570

Authorized licensed use limited to: North China University of Technology. Downloaded on October 5, 2008 at 19:40 from IEEE Xplore. Restrictions apply.
 u = x 4 ª1 0 0 0 0 0 0 0 0 º
«0 1 T 0 0 0 0 0 0 0 »
The controllability matrix of the system (28) is given by « »
[
C ( x ) = g, ad f g, ad 2f g, ad 3f g ] (29)
«0 0 2 T 2
«
0 0 0
0
6 T3
0
0
0
0
0
0
0
0
0 »
0 »
» (37)
A=«
Let δ ( x ) = −h0 h1 cos x1 , the flat output for the IWP can be «1 1
«
1 1 1 1 1 1 1 »
»
obtained by the equation (15), and gives «0 1 T 2 T 3T 4T 5T 6T 7T 8T »
«0 0 2 T 2 6 T 2 12 T 2 20 T 2 30 T 2 42 T 2 56 T 2 »
y = h( x) = x1 − h1 x3 (30) « »
¬«0 0 0 6 T3 24 T 3 60 T 3 120 T 3 210 T 3 336 T 3 ¼»
The derivatives of the flat output can be obtained as
As in section 4, the measure of the IWP system can be defined
y = x2 − h1 x4 Δ y = h0 sin x1 Δ y = h0 x2 cos x1 (31) t
to be K = 1 ³[x2 x4 ]M[x2 x4 ]Tdt , where M = ª«m11 m12 º» . Despite the
`

then, the inverse coordinate transformation can be given by 2 t0 ¬m21 m22 ¼

x1 = arcsin( y h0 )
( )
arbitrary vector in the equation (22) can be calculated as
x2 = y h0 1 − ( y h0 ) 2 ∂ t
§ ª ∂x ∂x4 º T ·
M [x 2 x4 ] ¸¸dt (38)
1

İ=− K (a ) = − ³ ¨¨ « 2
1 ∂a t © ¬ ∂a ∂a »¼
x3 = (arcsin( y h0 ) − y ) (32) 0 ¹
h1 ∂x2 ∂x4
where ∈ R 8×1 and ∈ R 8×1 are vectors, the negative
x4 =
1
h1
( (
y h0 1 − ( y h0 ) 2 − y ) ) ∂a ∂a
gradient of the measure only gives the best direction for
reducing the energy dissipation. We adopt the global
Substituted (32) into (19), the input transformation has form optimization method to optimize the motion for the IWP
u=−
1
y +
1 ª ( 4)
y +
(y)2 y º (33) system. In the simulations section, the vector İ and the scale
« »
h1 h0 h1 1 − ( y h0 ) 2 ¬ 1 − ( y h0 ) 2 ¼ multiplier μ are determined by the standard genetic
Since both the states and the input can be expressed as the algorithm, which can be found in many textbooks.
function of the output (30), as shown by equation (32) and
(33), this confirms the output (30) is indeed a flat output for C. Simulation results
the IWP system. The physical parameters of the IWP are listed in the
Refer to the equations (32) and (33), one can find that Appendix. For the motion planning task, we assume the initial
state is x 0 t =0 s = [30o 0 0 0] and the final state is
T
ʌ
there is a singularity at the point x1 = θ1 = ± kʌ, k = 0,1,2,! . 0
2
x = [0 0 0 0] , the duration of the task is
1 T
Outside the singularity region, by the nonlinear feedback t1 =10 s

u=−
1
y +
1 ª
v+
(y)2 y º (34) T = t1 − t0 = 10s . By the genetic algorithm, the approximately
« 2 »
h1 h0 h1 1 − ( y h0 ) 2 ¬ 1 − ( y h0 ) ¼ optimal values for the arbitrary vector and the scale multiplier
are obtained to be ε = −[0 0 0 0 5 1 1 0 −1] and
T
the system is equivalent to a fourth order linear system
y ( 4) = v (35) μ = 6 × 10 4 respectively. Fig. 2 and Fig. 3 show the motion
where v is the auxiliary control input defined in (34). planning (dashed line) and trajectories tracking (solid line)
simulation results, of which the former corresponds to μ = 0
B. Optimal motion planning for the IWP system
and the later corresponds to μ = 6 × 10 4 . The energy
The underactuated mechanical systems are appealing for
T
the application fields where the weight of the system has efficiency is evaluated by E = ³ τθ2 dt . For the fourth-order
rigorous limit. The energy efficiency of the underactuated 0
mechanical must be considered even if the system is carefully linear system (35), it is easy to design a closed-loop controller
designed to be flat, such as the flat biped robots [16] and the v = y( 4) d + k3 (yd − y) + k2 ( yd − y) + k2 ( y d − y ) + k0 ( y d − y) (39)
flat space manipulators [17]. For the IWP system (27), the
benchmark single input flat underactuated mechanical system where y d is the flat trajectory given by (36). The controller
with flat output (30), the trajectory in the flat output space can (39) is stabilizable by choosing the parameters k i , i = 1,2,3,4
be parameterized as a eighth-order polynomial
s −1 such that the closed loop characteristic polynomial
9
§t · λ4 + k3 λ3 + k 2 λ2 + k1λ + k 0 has all its roots in the left half of
y (t ) = ¦ as −1 ¨ ¸ (36)
s =1 ©T¹ the complex plane. For testing the performance of the closed
where a redundant design parameter is included. Assume the loop controller, the state of the system is started from a new
initial state and the final state are x 0 and x 1 respectively, and
initial state [
x 0 t =0 s = 70o 0 − (3 ×103 )
o
0 ]T
, which
the corresponding positions in flat output space are z and z 0 1 0

significantly deviates from the target trajectory. Refer to the

respectively. The matrix A of the equation (21) is given by
Fig.2 and 3, one can find that not only the controller (39) is
exponentially stable but also the actual energy dissipation is
reduced correspondingly.

1571

Authorized licensed use limited to: North China University of Technology. Downloaded on October 5, 2008 at 19:40 from IEEE Xplore. Restrictions apply.
 θ1 ( o ) θ 2 (o ) control task. Since there is no systemic way to determine if a
100 10000 multi-inputs nonlinear system is flat or what the flat outputs
θ1d θ 2d for a multi-inputs nonlinear system are, the motion planning
θ1 θ2 and control for general multi-inputs underactuated mechanical
50 5000
systems are still open problems, say nothing of the motion
optimization.
0 0
REFERENCES
-50 -5000 [1] Y. Nakamura and R. Mukherjee. Nonholonomic path planning of space
0 5 10 15 20 0 5 10 15 20
robots via a bidirectional approach, IEEE Transactions on Robotics and
time(s) time(s) Automation, 1991, 7(4): 500-514.
(a) (b) [2] K.D. Do, J. Pan and Z. P. Jiang. Robust and adaptive path following for
underactuated autonomous underwater vehicles, Ocean Engineering,
τ (Nm) E (J) 2004, 31:1967-1997.
10 600 [3] L. Birglen and C. M. Gosselin. Kinetostatic analysis of underactuated
τd Ed fingers, IEEE Transactions on Robotics and Automation, 2004, 20(3):
211-221.
τ 400 [4] G. Oriolo and Y. Nakamura. Free-joint manipulators: motion control
5
E under second-order nonholonomic constraints. IEEE/RSJ International
Workshop on Intelligent Robots and Systems, 1991, 1248-1253.
0 200 [5] A. De Luca, R. Mattone and G. Oriolo. Stabilization of an
underactuated planar 2R manipulator, International Journal of Robust
and Nonlinear Control, 2000, 24:181-198.
-5 0 [6] H. Sira-Ramirez and S. K. Agrawal. Differentially flat systems. Marcel
0 5 10 15 20 0 5 10 15 20 Dekker, Inc., New York, 2004.
time(s) time(s) [7] R. M. Murray, Z. Li and S.S. Sastry. A mathematical introduction to
(c) (d) robotic manipulation, CRC Press, 1994.
Fig.2. The trajectory tracking control results with μ = 0 [8] M. Krsti , I. Kanellakopoulos and P. Kokotovi . Nonlinear and adaptive
control design, John Wiley & Sons, Inc.,1995.
θ 2 (o ) [9] M. Fliess, J. Lévine, P. Martin, and P. Rouchon. Flatness and defect of
θ1 ( o )
nonlinear systems: Introductory theory and examples. International
1 00 6000
Journal of Control, 1995, 61(6): 1327-1361.
θ1d θ 2d [10] S. K. Agrawal and A. Fattah. Motion control of a novel planar biped
4000
θ1 θ2 with nearly linear dynamics. IEEE/ASME Transactions on
50 Mechatronics, 2006, 11(2): 162-168.
2000
[11] V. I. Arnold. Mathematical methods of classical mechanics. Springer-
0 Verlag, New York, 1989.
0
[12] M. W. Spong. The swing up control problem for the acrobot, IEEE
-2000 Control Systems Magazine, 15: 49-55, 1995.
-5 0 -4000
[13] A. Isidori. Nonlinear control systems. Springer-Verlag, 1995.
0 5 10 15 20 0 5 10 15 20 [14] M. W. Spong, P. Corke, and R. Lozano. Nonlinear control of the inertia
time(s) time(s) wheel pendulum, Automatica, 2001, 37: 1845-1851.
(a) (b) [15] Reza Olfati-Saber. Global stabilization of a flat underactuated systems:
the inertia wheel pendulum, Proceedings of the 40th IEEE Conference
on Decision and Control, Orleans, Florida, 2001, 3764-3765.
τ (Nm) E (J) [16] V. Sangwan and S. K. Agrawal. Leg-like motion with an under-
10 400 actuated two DOF linkage using differential flatness. Proceedings of
τ d
d
E the 2006 American Control Conference, Minneapolis, Minnesota, 2006,
5 τ 300 1790-1795.
[17] J. Franch and S. K. Agrawal. Planar space robots with coupled joints:
0 200 E Differentially flat design. Proceedings of the 2004 IEEE International
Conference on Robotics and Automation, New Orleans, 2004, 2116-
100
2121.
-5

-1 0 0
0 5 10 15 20 0 5 10 15 20
time(s) time(s)
(c) (d)
Fig. 3. The trajectory tracking control results with μ = 6 × 10 4

VI. CONCLUSIONS
For a flat underactuated mechanical system, it is shown
that the motion of the system can be optimized by constructing
a shape adjustable curve in the flat output space. By the
benchmark singe-input underactuated mechanical system, the
inertia wheel pendulum system, it is verified that the proposed
optimal motion planning method can effectively improve the
energy efficiency of the IWP system for a given position

1572

Authorized
View licensed
publication stats use limited to: North China University of Technology. Downloaded on October 5, 2008 at 19:40 from IEEE Xplore. Restrictions apply.