Beruflich Dokumente
Kultur Dokumente
Pendulum
Christoph Gruber
and Michael Hofbaur
Institute of Automation
and Control Engineering
UMIT
6060 Hall in Tyrol, Austria
Email: christoph.gruber@umit.at
michael.hofbaur@umit.at
Abstract The reaction wheel pendulum is a mechanism consisting of a planar freely rotating rotating bar with an actuated
reaction wheel at its tip. Acceleration of the reaction wheel
provides the momentum to swing the pendulum. We apply a
modern method from nonlinear control for shaping periodic
motions and designing a dedicated stabilizing controller. Specific
periodic motions are planned by introducing virtual holonomic
constraints and design of a stabilizing controller based on a
thorough analysis of the dynamics transversal to the resulting
limit cycle. Compared to previous applications, we focus on a
concise formulation of the control law and minimize the resulting
design and implementation effort.
I. I NTRODUCTION
The Reaction Wheel Pendulum (RWP) or also known as
inertia wheel pendulum is a typical setup for education and
demonstration in the field of nonlinear control systems. It
is an introducing example for design and implementation of
local linearization-based state feedback controllers to locally
stabilize equilibria. More advanced controllers also allow to
switch between equilibria. Because of its interesting model
properties, many different approaches were evaluated on the
RWP.
An energy-based swing-up controller which is able to swing
the pendulum up even with saturated control input was developed in [1], [2]. For stabilizing the upward equilibrium they
compared the afore-mentioned state feedback controller to a
partial feedback linearization design. Another approach that
uses switching between multiple controllers was presented by
[3], who use qualitative methods to design a controller that
robustly stabilizes the upright equilibrium. The same problem
was solved by [4], [5] with a single output feedback controller,
which was designed using a passivity based method.
Another interesting and challenging task with the RWP is
the stabilization of periodic motions. A two-relays controller
which was parametrized by analysis of the closed-loop dynamics using the describing function method was presented in
[6]. Another approach for stabilizing periodic motions around
the upper equilibrium of the RWP was presented in [7]. There,
a method developed in [8], [9], [10] was applied for planning
periodic motions and designing of a stabilizing controller. This
procedure follows three steps: 1) motion planning using virtual
TABLE I
M ODEL PARAMETERS
Symbol
J
Ja
m
g
l
r
g
d
k
Ra
Value [Unit]
0.0135 [kg m2 ]
0.0012 [kg m2 ]
0.416 [kg]
9.81 [ms2 ]
0.175 [m]
3.71 [1]
0.79 [1]
0.8 [1]
0.0346 [NmA1 ]
5.78 []
with
M=
Meaning
inertia of the pendulum assembly
inertia of the reaction wheel
mass of the pendulum assembly
gravity constant
length of the pendulum bar
gearbox transmission ratio
gearbox efficiency
DC-Motor efficiency
DC-Motor torque constant
DC-Motor armature resistance
G(q) =
(1)
M =
v kr( )
(2)
Ra
with v the motor voltage. We assume that the electrical is much
faster than the mechanical dynamics. Thus, we neglect the
armature inductance. The equations of motion can equivalently
be rewritten in matrix form:
M q + C q + G(q) = B v
0
mgl sin()
k 2 r2 d g k 2 r2 d g
,
0
0
krd g
B=
, q=
.
0
C=
,
= f (x, v)
(4)
= mgl sin(x1 )
Ja =
0
J
or in state-space representation:
mgl sin() + J + Ja =
J a Ra
Ja
(3)
x3
x4
(5)
with x = [ ]
T . Since there are two degrees of freedom
(DoFs) but only one actuator, the reaction wheel pendulum is
an underactuated system with underactuation degree one. The
angles and are measured using encoders with a resolution
of 4096 [cnt/rev] and r2000 [cnt/rev], respectively1 . System
(4) has four state variables but can be reduced to three
),
(, ,
since x2 = does not occur in the other equations.
The voltage v is an input and the equilibria of the reduced
system are found at (E = k, E = 0, E = 0), k Z,
where equilibria with k odd correspond to the pendulum
bar pointing upwards and k even downwards. The upward
equilibria are unstable, while the downward are stable, which
is easily verified by computing the eigenvalues of the Jacobian
at equilibrium. These properties - configuration manifold T2 ,
underactuation and existence of stable and unstable equilibria
- make the reaction wheel pendulum an interesting model for
control engineers.
III. M OTION P LANNING
The motions that we want to shape need to respect the
dynamics of the reaction wheel pendulum. There exist multiple
approaches to solve such a motion planning problem [11]: One
way would be to parametrize the desired motion and involve
the dynamics as constraint. Another method is employing virtual constraints [12]: Relations between DoF are parametrized
and introduced into the dynamics. The parameters of these
relations (the virtual constraints) show up in expressions for
the resulting motions and can be tuned to arrive at a desired
motion. Introducing a rather general virtual constraint in the
form
= f () = F () + 0
(6)
with constant 0 and F () C 2 (S1 ) allows the reduction of
the model (1) to a single differential equation
d2
(F () + 0 ) = 0
(7)
dt2
which represents the dynamics of the RWP when the virtual
constraint (6) is perfectly met. The solutions of this ordinary
differential equation are motions that are consistent with the
dynamics of the RWP and the virtual constraint. Thus, by
mgl sin() + J + Ja
1 The
with
()
()
= Ja
()
J + Ja
dF
,
d
d2 F
,
d2
= mgl sin().
(9)
(11)
e(t)
= (t) f ((t))
e(t)
= (t)
d1 (t)
e(t)
= (t)
d1 (t).
(14)
(15)
q = L
(17)
(19)
v = [ R(, ,
K
with
e
e)
1 0 L1 M 1 (CL G())
R(, ,
=
(21)
+ kr e + kr(d1 1) .
(22)
To find coordinates for a (Poincare) section in TM a further
coordinate is required. Since the equation is integrable,
an expression for 2 can be found. The difference between
this expression and the actual 2 corresponds to a so-called
integral of motion I which preserves a zero-value on the limit
cycle and can thus be used as third coordinate. This integral
of motion is [9, Theorem 1]2
Z
2(s)
I = 2 ? (0)2
ds
? (0) (s)
2mgl (cos() cos(? (0)))
= 2 ? (0)2
. (23)
J + Ja d1
2 In
our case, k = 0, 1 = 0, 2 = .
Furthermore, the time derivative of I evaluated along a solution of (20) is given by [9, Lemma 2]
I
=
=
2
u()
()
2Ja
=: kb .
J + Ja d1
(24)
with
I
x = e
e
0
A= 0
0
0
0
0
0
1
0
kb ? (t)
.
b(t) =
0
1
(t, x ) = k1 (t)
(26)
k2 k3 x .
A =
(27)
with
2
k1 kb (t)
0
A(t) =
k1 (t)
k2 kb (t)
k3 kb (t)
.
0
1
k2
k3
(28)
(29)
k2 + k3
sup
? (t)
(30)
t[0,T ]
k2 , k3
> 0, small
(31)
1
T
A|=
? ()d
k1 kb ?
=
0
0
0
0
k2
0
1
k3
(33)
where
2
?
1
T
? ()2 d.
The zeros in the first and last row of A follow from the
integration of a symmetric periodic signal over one period.
The condition from (30) is a trade-off between ensuring that
the first entry is dominant in the first row of A(t) and keeping
the first entry of the last row small. However, the first entry
in the last row can be interpreted as input (or perturbation)
to the filter4 x,2 , x,3 with low-pass characteristics and
coefficients k2 , k3 . Thus, with k2 , k3 sufficiently small, the
cut-off frequency of the filter can be placed low enough to
With k1 , k2 , k3 > 0, A is Hurwitz, since kb > 0 when
filter .
(12) is satisfied. Take t = , small and positive, then
dx
d
d
x
d
(32)
A( )|=
? x
(34)
A x
(35)
2 2
k
k
k
k
k
k
b 2
b 3
1 b ?
,
=
0
0
0
k1
0
0
h(t, x) = A (t) x,
(37)
The advantages of the controller (26) compared to the LQRbased controller are obvious. For LQR-design, the periodic
Riccati differential equation needs to be solved, which requires
tremendous computational and implementational effort (see
e.g. [14]). However, this can be done offline. Result of such a
computation is a time-varying control gain matrix, that must
somehow be implemented on real-time control hardware. This
is typically done either by memory-intensive look-up tables
or by computation-intensive interpolation methods. A further
challenge when implementing the LQR-based approach is the
which is able to
computation of a projection operator T (, )
recover the time (in the sense of a measure of progress) on
the limit cycle [10].
In comparison, the controller we propose is parametrized
by simply choosing three constants k1 , k2 and k3 according to
Lemma 1. There is no need to generate a time-varying signal
in the controller, which makes implementation almost trivial.
Instead of a controller-generated time-varying signal, we use
as an estimate from an observer in the control law and thereby
also bypass computing a projection operator. Indeed, this is
a fascinating result: A planned periodic motion5 is stabilized
without the need to generate a periodic signal in the controller.
the function
(36)
V. S IMULATION R ESULTS
u(t, x)
h(, x)d
(38)
m N : ||(t)
? (0)|| t > mT.
This implies that the approximation by the average system
is valid on the infinite interval and gets even better with
large time. Hence, the origin is a locally exponentially stable
equilibrium of the transverse dynamics and w1 converges to
the orbit exponentially.
5 around
15
1.5
10
1
5
0.5
0
5
10
0.5
15
20
25
1.5
I(t)
e(t)
d/dt e(t)
0
4
t [s]
30
Fig. 2. Evolution of x,1 (t) = I(t) (blue, solid), x,2 (t) = e(t) (green,
dash-dotted) and x,3 (t) = e(t)
35
4
t [s]
Fig. 4. Evolution of the control input (t) (blue, solid) of our controller,
and (red, dashed) of the LQR.
TABLE II
P ERIODIC M OTION C ONTROLLER PARAMETERS
Symbol
d1
? (0)
? (0)
k1
k2
k3
Ts
Value [Unit]
20
187 [ ]
0 [s1 ]
225.8
3
3.5
0.001 [s]
Meaning
virtual constraint parameter
target limit cycle initial angle
target limit cycle initial rate
controller gain 1
controller gain 2
controller gain 3
sample time
I(t)
e(t)
d/dt e(t)
0
4
t [s]
Fig. 3. Evolution of x,1 (t) = I(t) (blue, solid), x,2 (t) = e(t) (green,
dash-dotted) and x,3 (t) = e(t)
3.3
3.3
3.2
[rad]
[rad]
3.2
3.1
3.1
3
3
0
0.5
1.5
2.5
t [s]
3.5
4.5
0.5
1.5
2.5
t [s]
3.5
4.5
0.5
1.5
2.5
t [s]
3.5
4.5
d / d t [rad/s]
d / d t [rad/s]
2
1
0
1
2
0.5
1.5
2.5
t [s]
3.5
1
0
1
2
4.5
3.35
3.3
3.3
3.25
3.25
3.2
3.2
[rad]
[rad]
3.15
3.15
3.1
3.1
3.05
3.05
3
measured motion
initial state
target limit cycle
3
2.95
1.5
0.5
0
d / d t [rad/s]
0.5
measured motion
initial state
target limit cycle
2.95
1.5
1.5
0.5
0
d / d t [rad/s]
0.5
1.5
VII. C ONCLUSION
ACKNOWLEDGMENT
Press, 2006.
[12] C. Chevallereau, G. Abba, Y. Aoustin, F. Plestan, E. Westervelt,
C. Canudas-de Wit, and J. Grizzle, RABBIT: A testbed for advanced
control theory, IEEE Control Systems Magazine, vol. 23, no. 5, pp.
5779, Oct. 2003.
[13] H. K. Khalil, Nonlinear Systems, 3rd ed. Prentice-Hall, 2002.
[14] S. Gusev, S. Johansson, B. Kagstrom, A. Shiriaev, and A. Varga, A
numerical evaluation of solvers for the periodic Riccati differential
equation, BIT Numerical Mathematics, vol. 50, no. 2, pp. 301329,
2010.
[15] I. Houtzager, DPRE - Discrete-time periodic Riccati equation solver
for periodic LQ state-feedback design, 2008. [Online]. Available:
http://www.mathworks.co.kr/matlabcentral/fileexchange/21379-dpre
[16] W. Hahn, Stability of Motion. Springer, 1967.