3208
4, OCTOBER 2018
Impulsive Dynamics and Control of the
Inertia-Wheel Pendulum
Nilay Kant and Ranjan Mukherjee
Abstract—The dynamics of the inertia-wheel pendulum, when
subjected to impulsive inputs, can be described by algebraic
equations. Optimal sequences of these inputs, that minimize their
infinity norm, are designed for rest-to-rest maneuvers. The results
are applied to the well-studied swing-up problem, and high-gain
feedback is used for continuous approximation of the inputs and
simulation of the impulsive dynamics. Analytical and simulation
results establish a direct link between high wheel velocities during
swing-up and control strategies that take the pendulum directly
to the upright configuration. They also indicate that optimal
trajectories resemble those of energy-based controllers and can be
designed to satisfy the torque constraint of the actuator.
Index Terms—Dynamics, optimization and optimal control, un-
deractuated robots.
NOMENCLATURE
1 , 2 distance of center-of-mass of pendulum from the pas-
sive joint, and length of the pendulum (m).
m1 , m2 mass of the pendulum, and combined mass of the
motor and the wheel, (kg).
t0 initial time, (s).
t i , tf time instant at which the i-th impulsive torque is
applied, i = 1, 2, . . . , N , and final time (s).
I1 , I2 mass moment of inertia of pendulum and wheel about
their center of mass, (kg.m2 ).
N number of impulsive torques applied.
θ angular position of the pendulum, measured CCW
with respect to the x axis, (rad).
φ angular position of the wheel, measured CCW with
respect to the pendulum, (rad).
θ̇, φ̇ angular velocity of pendulum and wheel, (rad/s).
θi value of θ at time ti .
θ̇i− , θ̇i+ value of θ̇ immediately before and after application
of impulsive torque at time ti .
φ̇− + the angular displacement of the wheel is not included in the
i , φ̇i value of φ̇ immediately before and after application
of impulsive torque at time ti .
τ torque applied by the motor on the wheel, (Nm).
τi impulsive torque applied at time ti , (Nm).
Manuscript received February 22, 2018; accepted June 21, 2018. Date of
publication June 27, 2018; date of current version July 13, 2018. This letter
was recommended for publication by Associate Editor E. Papadopoulos and
Editor D. Song upon evaluation of the reviewers’ comments. This work was
supported by the National Science Foundation under Grant CMMI-1462118.
(Corresponding author: Ranjan Mukherjee.)
The authors are with the Department of Mechanical Engineering, Michigan
State University, East Lansing, MI 48824 USA (e-mail:,kantnila@egr.msu.edu;
mukherji@egr.msu.edu).
Digital Object Identifier 10.1109/LRA.2018.2851029
2377-3766 © 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
IEEE ROBOTICS AND AUTOMATION LETTERS, VOL. 3, NO.
τhg continuous-time implementation of impulsive torque
using high-gain feedback, (Nm).
Ii angular impulse of impulsive torque τi applied at time
ti , (Nm-s).
IN the N -dimensional vector [I1 I2 · · · IN ].
I. INTRODUCTION
MPULSIVE control of dynamical systems has been widely
I investigated but a majority of the investigations have been
theoretical in nature - see [1]–[3] and the references therein.
Recently, however, impulsive control has been experimentally
demonstrated in underactuated systems [4]–[9], and singular
perturbation theory has been used to justify approximation of
the impulsive inputs by high-gain feedback [8], [9]. Impulsive
inputs are very useful in applications where a sudden change
in the system configuration is desired. For instance, impulsive
inputs can be used to regain the stability of an equilibrium from
configurations outside their region of attraction [8], [9]. These
and other demonstrations, which have been performed using
standard actuators, are significant as they dispel the notion that
large actuators are required for implementing impulsive control.
In reality, actuators such as electric motors can apply signifi-
cantly larger forces over short intervals of time compared to the
maximum force that they can apply continuously [10]. Having
established the feasibility of implementing impulsive control in
standard physical systems, we focus on the theoretical problem
of impulsive control of the simplest underactuated system, the
inertia-wheel pendulum (IWP).
The IWP is comprised of a simple pendulum and a motor
mounted at its distal end; the motor drives a wheel - see Fig. 1.
The IWP is similar to the acrobot [11], where the motor drives
a link instead of a wheel. For the IWP, the torque produced by
the motor can be used to accelerate the wheel in both CW and
CCW directions and the resulting reaction torque can be used
to control the pendulum angle. Since the wheel is symmetric,
state-space representation and the only source of nonlinearity
in the equation of motion is the gravity term. The IWP is an
ideal candidate for impulsive control since the dynamics of
the system can be described by simple algebraic equations: the
effect of impulsive forces can be described by changes in the
velocities of the system along the impulse manifold [8], and
conservation of energy and conservation of wheel momentum
describe the dynamics when no torque is applied by the
motor.
For the IWP, stabilization of its upright posture and swing-
up to this configuration have been investigated in the literature,
[12]–[18], for example. Some of the early work on the IWP can
be found in [12] where a hybrid controller comprised of separate
swing-up and balancing controllers was designed using energy
KANT AND MUKHERJEE: IMPULSIVE DYNAMICS AND CONTROL OF THE INERTIA-WHEEL PENDULUM 3209

B. Effect of Impulsive Input and Constants of Free Motion

The application of an ideal impulsive torque τi by the motor
at time ti results in discontinuous jumps in the velocities of
both the pendulum and the wheel while their angular positions
remain unchanged. By integrating (2a) with respect to time over
+
the infinitesimal interval [t−
i , ti ], it can be shown [8] that these
jumps satisfy the relation:
m22
(θ̇i+ − θ̇i− ) = −C(φ̇+ −
i − φ̇i ), C (3)
(m11 + m22 )
Using (2b), the jump in the pendulum velocity can be shown to
be related to the angular impulse as follows:
Fig. 1. The inertia wheel pendulum is shown in an arbitrary configuration.
 t +i
Ii  τi dt = 2 m22 [(θ̇i+ − θ̇i− ) + (φ̇+ −
i − φ̇i )] (4a)
methods. Swing-up and stabilization was achieved using a single t−
i
controller using the IDA-PBC method [13], [14] and a global
change of coordinates that converted the dynamics of the IWP ⇒ Ii = −2 m11 (θ̇i+ − θ̇i− ) (4b)
into strict feedback form [15]. The IDA-PBC based controllers
[13], [14] and the globally stabilizing controller [15] result in where (4b) was obtained from (4a) using (3).
high wheel velocities and actuator torques. If τ = 0, the dynamics of the IWP described by (2a) or (2b)
The swing-up problem of the IWP is a rest-to-rest maneuver simplifies to the form
between its lowest and highest potential energy configurations.
2 m11 θ̈ + β cos θ = 0 (5a)
In this letter, we address the general problem of designing rest-
to-rest maneuvers using a sequence of impulsive inputs. The θ̈ + φ̈ = 0 (5b)
use of impulsive inputs simplifies the dynamics of the IWP and
permits the design of optimal sequences under some general The above equations are the differential forms for conservation
assumptions. It is found that a sequence with an odd number of energy of the IWP and conservation of angular momentum
of inputs is less optimal than the two adjacent sequences with of the wheel, namely
even number of inputs. The high wheel velocities and/or actu-
ator torques associated with some results in the literature [13]– m11 θ̇2 + β sin θ = constant (6a)
[15], is explained using the analysis based on two impulsive
inputs. An optimal sequence of impulsive inputs, implemented θ̇ + φ̇ = constant (6b)
using high-gain feedback [8], is used for swing-up of the IWP; It can be seen from (6a) and (6b) that the dynamics of the
the simulations results show similarities with the energy-based pendulum and wheel are decoupled and simplified when τ = 0.
methods in [12]. It is shown that an optimal input sequence can For controlling the IWP, this motivates the use of impulsive
be designed to accommodate torque constraint of the actuator; torques at discrete instants of time.
this has not been discussed in earlier works.
C. Problem Statement: Rest-to-Rest Maneuvers
II. MATHEMATICAL MODEL AND PROBLEM STATEMENT
We consider the problem of designing a set of discrete im-
A. Equations of Motion pulsive inputs over the time interval [t0 , tf ] for rest-to-rest ma-
The Inertia-Wheel Pendulum (IWP), shown in Fig. 1, is an neuvers of the IWP between two configurations where the final
underactuated system comprised of a simple pendulum and an configuration has higher potential energy than the initial config-
inertia wheel. The inertia wheel, also referred to as a reaction uration, i.e.,
wheel, is driven by a motor mounted at the distal end of the θ̇0 = φ˙0 = 0 (7a)
pendulum. The kinetic energy T and the potential energy V of
the IWP are given by the relations: θ̇f = φ˙f = 0 (7b)
2
T = m11 θ̇2 + m22 (θ̇ + φ̇) , V = β sin θ (1) β (sin θf − sin θ0 ) > 0 (7c)
  and θ lies in the range (−3π/2, π/2]. For a given value of N ,
m11 = m1 l12 + m2 l22 + I1 /2, m22 = I2 /2
the task is to design the vector I N = [I1 I2 · · · IN ] where
β = (m1 l1 + m2 l2 )g
sign(Ik ) = −sign(Ik −1 ), k = 2, 3, . . . , N (8)
Using (1), the Euler-Lagrange equation of motion of the IWP
can be written as: and t0 < t1 < t2 < · · · < tN < tf . The rationale for imposing
the constraint in (8) will be provided later, at the beginning of
2 (m11 + m22 )θ̈ + 2 m22 φ̈ + β cos θ = 0 (2a) Section IV. For the sake of simplicity, it is assumed that the time
instants ti are such that the angular velocity of the pendulum is
2 m22 θ̈ + 2 m22 φ̈ = τ (2b)
momentarily zero, i.e.,
Note that the angle φ does not appear in (2a) or (2b) - this is
because the wheel is symmetric about its axis of rotation. θ̇i−  θ̇(t−
i ) = 0, i = 1, 2, . . . , N (9)
3210 IEEE ROBOTICS AND AUTOMATION LETTERS, VOL. 3, NO. 4, OCTOBER 2018

Without loss of generality, it is assumed that the first impulsive For [t+ − +
1 , t2 ] and [t2 , tf ], the conservation laws in (6), together
torque is applied at time t1 , where t−
1 = t0 . Therefore, with (7b) and (11), give
θ̇1− = θ̇0 = 0, φ̇−
1 = φ̇0 = 0 (10) 2 2
m11 θ˙1+ + β sin θ0 = m11 θ˙2− + β sin θ2 (16a)
Since the angular positions of the pendulum and the wheel do
θ̇2− + φ̇−
2 = θ̇1+ + φ̇+
1 (16b)
not change during application of an impulsive torque, we have
2
θ1 = θ0 , φ1 = φ0 (11) m11 θ˙2+ + β sin θ2 = β sin θf (16c)
θ̇2+ + φ̇+
2 = θ̇f + φ̇f = 0 (16d)
III. REST-TO-REST MANEUVERS:
CASE OF ONE AND TWO IMPULSIVE INPUTS For the second impulse, the relationship between the velocity
jumps can be obtained from (3) as:
A. Case of One Impulsive Input (N = 1)
Using (3) and (10), we can write (θ̇2+ − θ̇2− ) = −C(φ̇+ −
2 − φ̇2 ) (17)

θ̇1+ = −C φ̇+ (12) By substituting the relations φ̇+ + −

2 = −θ̇2 from (16d) and φ̇2 =
1
θ̇1+ + φ̇+ −
1 − θ̇2 from (16b) into the right-hand side of (17) and
The main result can now be stated as follows: simplifying using (12), we get
Result 1: (One Impulsive Input) The rest-to-rest maneuver
described by (7) cannot be accomplished using I 1 = [I1 ]. (θ̇2+ − θ̇2− ) = −θ̇1+ (18)
Discussion: Since τ = 0 for [t+1 , tf ], we get from (6) and
(11): By combining (16a) and (16c) and substituting (18), we get
 
2 2 2
β (sin θf + sin θ0 − 2 sin θ2 ) = m11 θ˙2+ + θ˙2− − θ˙1+
2
m11 θ˙1+ + β sin θ0 = m11 θ˙f + β sin θf
2
(13a)
θ̇1+ + φ̇+
1 = θ̇f + φ̇f = 2 m11 θ˙2− θ˙2+
(13b)

Using (7b), (13b), and (12), it can be shown that θ̇1+ = 0. From
(13a) we now get β (sin θf − sin θ0 ) = 0, which violates (7c).
This establishes Result 1 by contradiction.  in the above equation and simplifying, we get
Remark 1: A single impulsive torque can always be chosen
to impart sufficient angular momentum to the pendulum such
that it reaches its desired configuration with zero angular veloc-
ity. However, this impulsive torque will also cause the wheel to
have a nonzero angular velocity in the inertial reference frame.
This implies that φ̇ = 0 when θ̇ = 0.
B. Case of Two Impulsive Inputs (N = 2)
We assume that the first impulse is applied at t1 and hence
(12) is still valid. Two results are presented next. In the first
result (Result 2), we relax the assumption in (9) and design a
more general sequence of impulses that satisfies (7). We will
show that the second result (Result 3), which is a special case
of the first, automatically satisfies the assumption in (9).
Result 2: (Two Impulsive Inputs) The rest-to-rest maneuver
described by (7) can be accomplished using I 2 = [I1 I2 ],
where I2 = −I1 .
Discussion: From (4a) and (7a) we have
I1 = 2 m22 (θ̇1+ + φ̇+
1 )
and since (6b) holds good for [t+ −
1 , t2 ], we can write
I1 = 2 m22 (θ̇1+ + φ̇+ − −
1 ) = 2 m22 (θ̇2 + φ̇2 ) (14)
From (7b) we have (θ̇f + φ̇f ) = 0 and since (6b) holds good
for [t+ + +
2 , tf ], we get (θ̇2 + φ̇2 ) = 0. Using (4a), we get
I2 = −2 m22 (θ̇2− + φ̇−
2) (15)
It is clear from (14) and (15) that the conditions in (7b) require
I2 = −I1 .
Substituting the expressions for θ˙2+ and θ˙2− from (16a) and (16c)

+ 1 β (sin θf − sin θ0 )
| θ̇1 |=  (19)
2 m11 sin θf − sin θ2
From (7c) we know that (sin θf − sin θ0 ) > 0 and it can be
seen from (16c) that (sin θf − sin θ2 ) > 0; therefore, the above
equation is well-defined. For a given pair {θ0 , θf }, (19) provides
a functional relationship between the initial angular velocity θ̇1+
(resulting from application of the first impulse I1 ) and the con-
figuration θ2 where the second impulse (I2 = −I1 ) is applied.
There are infinite solutions given by the pair {θ̇1+ , θ2 }; for each
solution, the value of θ̇1+ can be used to compute the impulses
I1 and I2 (I2 = −I1 ), using (12), (14), and (15). An ex-
ample showing the initial, intermediate, and final configuration
of the IWP for the case with two impulsive inputs is shown in
Fig. 2. 
The next result pertains to the particular solution that mini-
mizes the magnitude of the impulses.
Result 3: (Optimal Input I 2 ) The minimum magnitude of
the impulsive inputs required for the rest-to-rest maneuver de-
scribed by (7) is
| I1 |=| I2 |= 2 m11 β (sin θf − sin θ0 ) (20)
Discussion: Using (12) and (14), I1 can be expressed as
I1 = 2 m22 (θ̇1+ + φ̇+ +
1 ) = 2 m22 (1 − C)θ̇1 (21)
Therefore, the magnitude of I1 can be minimized by minimiz-
ing the magnitude of θ̇1+ . From (16a) it can be seen that
2 2
m11 θ˙1+ = m11 θ˙2− + β (sin θ2 − sin θ0 )
KANT AND MUKHERJEE: IMPULSIVE DYNAMICS AND CONTROL OF THE INERTIA-WHEEL PENDULUM 3211

effect of these two impulses in terms of change in the velocities

+
of the pendulum and wheel over the interval [t− k , tk +1 ] can be
achieved by a single impulse I¯ at time tk , where
I¯ = Ik + Ik +1 (25)
Discussion: From (4a) we have


Ik = 2 m22 (θ̇k+ − θ̇k− ) + (φ̇+
k − φ̇ −
k )


Ik +1 = 2 m22 (θ̇k++1 − θ̇k−+1 ) + (φ̇+ −
k +1 − φ̇k +1 )

Since τ = 0 for [t+ −

k , tk +1 ], (6b) can be used to rewrite the above
equations as follows


Ik = 2 m22 (θ̇k−+1 − θ̇k− ) + (φ̇− −
k +1 − φ̇k )


Ik +1 = 2 m22 (θ̇k++1 − θ̇k−+1 ) + (φ̇+ k +1 − φ̇ −
k +1 )
Fig. 2. An example showing the initial, intermediate, and final configuration

of the IWP for the case with two impulsive inputs (N = 2).
⇒ Ik + Ik +1 = 2 m22 (θ̇k++1 − θ̇k− ) + (φ̇+ k +1 − φ̇ −
k )

If (sin θ2 − sin θ0 ) ≤ 0, θ̇2− = 0 but the minimum value of θ̇1+ The change in the velocities of the pendulum and wheel over
+ + +
is equal to zero. This implies (sin θf − sin θ0 ) = 0 from (19), the interval [t− − −
k , tk +1 ] are (θ̇k +1 − θ̇k ) and (φ̇k +1 − φ̇k ), re-
which contradicts (7c). Since (sin θ2 − sin θ0 ) must be positive, spectively. To achieve the same change, I¯ must satisfy
the minimum magnitude of θ̇1+ can be obtained by choosing

θ̇2− = 0 1 ; this magnitude is equal to I¯ = 2 m22 (θ̇k++1 − θ̇k− ) + (φ̇+ −
k +1 − φ̇k )

| θ̇1+ |= (β/m11 ) (sin θ2 − sin θ0 ) (22) ⇒ I¯ = Ik + Ik +1
By equating (19) and (22), we get This establishes Result 4. 
Result 4: clearly indicates that two consecutive impulses of
1 the same sign can be replaced by a single impulse of the same
(sin θ0 + sin θf )
sin θ2 = (23)
2 sign. This justifies the constraint imposed in our problem state-
where θ2 is the angle at which the second impulse is applied.2 ment that consecutive impulses must have opposite sign.
Substituting (23) into (21) and (22), and comparing (14) and An extension of Result 4 is now considered. For a rest-to-rest
(15) we get maneuver using two impulsive inputs (N = 2), Result 4 implies
that I¯ = I1 + I2 = 0. This is true since (6b) and (7b) implies
| θ̇1+ | = (β/2m11 ) (sin θf − sin θ0 ) (24a) (θ̇k++1 + φ̇+ + + + +
k +1 ) = (θ̇2 + φ̇2 ) = (θ̇f + φ̇f ) = 0, and (7a) and
(10) implies (θ̇k− + φ̇− − −
k ) = (θ̇1 + φ̇1 ) = 0. This is consistent
⇒ | I1 | =| I2 |= 2m11 β (sin θf − sin θ0 ) (24b) with Result 2, where it was shown that I2 = −I1 . A general-
ization of this result in stated next.
This establishes Result 3.  Result 5: (Zero Sum of Impulses) For a rest-to-rest maneuver
Remark 2: From Result 3 it can be seen that the time instant involving N impulsive inputs, N ≥ 2, the following equation
t2 is automatically known when the magnitudes of the individual must hold.
impulses are minimized. This is different from Result 2, where
the choice of t2 is not unique and each feasible choice of t2
(alternatively θ2 ) uniquely determines the magnitudes of the
impulses.
IV. REST-TO-REST MANEUVERS:
GENERALIZATION OF THE RESULTS TO N INPUTS
A. Revisiting the Problem Statement
We start this section with the result that justifies the rationale
for imposing the constraint in (8).
Result 4: (Sum of Two Consecutive Impulses) Consider two
impulses Ik and Ik +1 applied at times tk and tk +1 . The net
1 This choice automatically satisfies the assumption in (9).
2 Itis clear that (23) can have multiple solutions for θ2 . The procedure for
computing the correct solution will be discussed in Section IV-B. Knowing the
value of θ2 , it will be possible to determine the time instant t2 .

N
Ii = 0 (26)
i=1
Discussion: A sequence of N impulses, N ≥ 2, can be replaced
by two impulses by applying Result 4 iteratively. For a rest-
to-rest maneuver, the sum of these two impulses is zero - this
follows from our discussion above. 
Remark 3: It is clear from the discussion above that both
Result 4 and Result 5 are quite general and they do not require
the assumption in (9) to be satisfied.
With the motivation of investigating the minimum values of
the magnitudes of the impulsive torques, we investigate rest-to-
rest maneuver of the IWP with I 3 and I 4 . As in the cases with
I 1 and I 2 , (12) holds good.
B. Rest-to-Rest Maneuvers: Even Number of Impulsive Inputs
Theorem 1: (Optimality of Even Impulse Sequence)
3212 IEEE ROBOTICS AND AUTOMATION LETTERS, VOL. 3, NO. 4, OCTOBER 2018

For a rest-to-rest maneuver of the IWP that satisfies (7) and for (2m + 2) impulsive inputs, i.e., n = (m + 1). It has been
(9) and uses 2n impulsive inputs, n = 1, 2, . . ., I 2n ∞ is shown earlier in (20) that (27) is satisfied for N = 2 (n = 1).
minimized by the following choice of inputs: By induction we can now claim that (27) will be satisfied for
any even number of impulsive inputs.
1
| Ii |= √ 2m11 β (sin θf − sin θ0 ), ∀ i = 1, 2, . . . , 2n The values of | θ̇i+ |, i = 1, 2, . . . , 2n, can be obtained from
n
(27), (4b), and (9), namely
(27)

1 β
The angles where the impulsive inputs are applied satisfy the | θ̇i+ | = √ (sin θf − sin θ0 ) (34)
following relation n 2m11
   
2n − i + 1 i−1 Substituting θ̇i+ from (34) in the energy conservation law
sin θi = sin θ0 + sin θf 2
2n 2n for rest-to-rest maneuvers, namely, m θ˙+ + β sin θ =
11 i i
i = 1, 2, . . . , 2n (28) β sin θi+1 and solving sequentially and iteratively for each i,
where i = 1, 2, . . . , 2n, we get the relations in (28). 
Proof: We use induction to first prove (27). Assuming that Remark 4: It follows from (27) in Theorem 1 and (8) that
(27) is satisfied for 2m impulsive inputs, i.e., n = m. We express all the 2n impulses have the same magnitude and consecutive
the magnitudes of the impulses for the case with (2m + 2) inputs impulses have opposite signs. Since the pendulum and wheel
using the relation are both at rest at the initial time, it follows that each pair of
consecutive impulses (starting with I1 and I2 ) result in a rest-
ki
| Ii | = √ 2m11 β (sin θf − sin θ0 ), to-rest maneuver.
m
Remark 5: It follows from (4b), (8), and (9) that the veloc-
i = 1, 2, . . . , (2m + 2) (29) ity of the pendulum immediately after application of an im-
pulsive input will have opposite sign for two consecutive im-
where ki , i = 1, 2, . . . , (2m + 2), are arbitrary positive num-
pulses, i.e., sign(θ̇k++1 ) = −sign(θ̇k+ ), k = 1, 2, . . . , (2n − 1).
bers. Using (4b) and (9) we can show
Using this fact, the following algorithm can be constructed to

ki β determine the unique value of θi , i = 1, 2, . . . , 2n, from (28).
| θ̇i+ | = √ (sin θf − sin θ0 ),
m 2m11
i = 1, 2, . . . , (2m + 2) (30) If θf belongs to quadrant I (includes θ = π/2) or IV, then
For m = 1, 2, . . . , n, compute sin θ2m using (28)
Using (9), the conservation law in (6a) for the time intervals If sin θ2m > 0, θ2m belongs to quadrant II
[t+ − + − + − +
1 , t2 ], [t2 , t3 ], · · · [tj , tj +1 ], . . . , [t2m +2 , tf ] can be written Elseif sin θm < 0, θ2m belongs to quadrant III
as: Else θ2m = −π.
2
m11 θ˙1+ + β sin θ0 = β sin θ2 For m = 0, 1, . . . , (n − 1), compute sin θ2m +1 using
(28)
2 If sin θ2m +1 > 0, θ2m +1 belongs to quadrant I
m11 θ˙2+ + β sin θ2 = β sin θ3
If sin θ2m +1 < 0, θ2m +1 belongs to quadrant IV
.. Else θ2m +1 = 0.
. Else θf belongs to quadrant II or III, then
(31) For m = 1, 2, . . . , n, compute sin θ2m using (28)
2
m11 θ˙j+ + β sin θj = β sin θj +1 If sin θ2m > 0, θ2m belongs to quadrant I
Elseif sin θm < 0, θ2m belongs to quadrant IV
.. Else θm = 0.
.
For m = 0, 1, . . . , (n − 1), compute sin θ2m +1 using
2
+
+2 + β sin θ2m +2 = β sin θf
m11 θ̇2m (28)
If sin θ2m +1 > 0, θ2m +1 belongs to quadrant II
Addition of equations in (31) and substitution of | θ̇i+ | from (30) If sin θ2m +1 < 0, θ2m +1 belongs to quadrant III
in the resulting equation gives the following Else θ2m +1 = −π.
2m +2
Endif

ki2 = [k1 k2 · · · k2m +2 ]22 = 2m (32)
i=1
C. Rest-to-Rest Maneuvers: Odd Number of Impulsive Inputs
Using (32) and the property of norms it can be shown that We generalize the result presented in Remark 4.
√
2m + 2 [k1 k2 · · · k2m +2 ]∞ ≥ [k1 k2 · · · k2m +2 ]2 Theorem 2: (Lack of Optimality of Odd Impulse Sequence)
 It is not possible to design an odd impulse sequence for which
m the magnitudes of all the impulsive inputs are less than the opti-
⇒ [k1 k2 · · · k2m +2 ]∞ ≥ (33)
m+1 mal magnitude for the preceding and succeeding even impulse
 sequence. In other words, the following inequality holds for
It can be shown that ki = m/(m + 1), i = 1, 2, . . . , (2m + n = 1, 2, · · · .
2), satisfy (32) and minimize [k1 k2 · · · k2m +2 ]∞ . Substi-
tution of these values of ki in (29) shows that (27) is satisfied I 2n +1 ∞ > min I 2n ∞ > min I 2n +2 ∞ (35)
KANT AND MUKHERJEE: IMPULSIVE DYNAMICS AND CONTROL OF THE INERTIA-WHEEL PENDULUM 3213

Proof: It is clear from (27) that min I 2n ∞ >

min I 2n +2 ∞ . We proceed to prove the left inequality by
contradiction. To this end, we assume
| Ii | = ki min I 2n ∞ , ki ∈ (0, 1],
i = 1, 2, . . . , 2n + 1

+ ki β
⇒ | θ̇i | = √ (sin θf − sin θ0 ) (36)
n 2m11
From (8), (26), and (36) we can show
k1 + k3 + · · · k2n +1 = k2 + k4 + · · · k2n (37) Fig. 3. Simulation results for the globally stabilizing controller [15] with
controller parameter values: c0 = −π/10, c1 = 13, c2 = 16 and c3 = 8.0.
Substituting θ̇i+ from (36) in the energy conservation law
2
for rest-to-rest maneuvers, namely, m θ˙+ + β sin θ =
11 i i
β sin θi+1 for i = 1, 2, . . . , 2n + 1, and adding them, we get
k12 + k22 + · · · k2n
2
+1 = 2n (38)
Since ki ∈ (0, 1], the following inequality holds true
k12 + k22 + · · · k2n
2
+1 ≤ k1 + k2 + k3 · · · k2n +1 (39)
Substituting (37) and (38) in the above inequality, we get
k2 + k4 + · · · k2n ≥ n ⇐⇒ k2 = k4 = · · · = k2n = 1
(40)
which implies from (37) and (38) Fig. 4. Simulation results for the globally stabilizing controller [15] with
controller parameter values: c0 = −π/10, c1 = 13, c2 = 16 and c3 = 4.5.


[k1 + k3 + · · · + k2n +1 ] = n = k12 + k32 + · · · + k2n
2
+1
high-gain feedback [19] for the IWP can be obtained as
Since ki ∈ (0, 1] ∀ i, the above equation cannot be satisfied.
1
From (40), this implies τhg = [K T M −1 K]−1 [K T M −1 H + (φ̇+ − φ̇−
i )]
 i
k2 + k4 + · · · k2n > n (41)    
(m11 + m22 ) m22 β cos θ
M =2 , H= (43)
which violates our assumption ki ∈ (0, 1] ∀ i.  m22 m22 0
where the matrices M and H above were reconstructed from (2),
V. THE SWING-UP PROBLEM T
K  [ 0 1 ] , and  > 0 is a small number. Implementation of
A. Optimal Swing-Up Using Even Impulse Sequences impulsive inputs using high-gain feedback also enables us to
compare our results with those published in the literature. A
The swing-up problem is a rest-to-rest maneuver where the discussion of select results in the literature is presented next.
final pendulum angle is θf = π/2. We consider special case
where the initial angle of the pendulum is in the vertically
downward configuration, i.e., θ0 = −π/2. For swing-up using C. Discussion of Results in the Literature
an even number of impulsive inputs, the optimal solution can 1) Globally Stabilizing Controller: We implemented the
be obtained from (27). For N = 2n, n = 1, 2, . . ., the optimal globally stabilizing controller in [15] using their kinematic and
solution is given by a sequence of equal and opposite impulses dynamic parameter values, which are given below
of the following magnitude:
m11 = 4.83 × 10−3 , m22 = 32 × 10−6 , β = 37.9 × 10−2
1  (44)
| Ii |= √ I, ∀ i = 1, 2, . . . , 2n, I  2 m11 β (42)
n and controller parameters: c0 = −π/10, c1 = 13, c2 = 16 and
Since the magnitude of the impulses is inversely proportional to c3 = 8.0; the results are shown in Fig. 3. The plots are slightly
√ different from those presented in [15] but the overall trends are
n, the magnitude of each impulse in the sequence is reduced
if the number of impulsive inputs are increased by an even num- similar. It is clear from the plot of θ that the control input drives
ber. This information will be useful for designing an impulse the pendulum directly towards the desired value of θf = π/2.
sequence that takes into account actuator saturation. There is a small overshoot beyond π/2 but the wheel velocity
is extremely high, of the order of 8000 rad/s. A change in the
controller parameter c3 from 8.0 to 4.5 increases the overshoot
B. Implementation Using High-Gain Feedback slightly but reduces the maximum wheel velocity by almost 50%
Ideal impulsive inputs are Dirac-delta functions and cannot - see Figs. 3 and 4.
be generated by actuators. In real physical systems, continuous- The above observation can be explained by the analysis pre-
time implementation of impulsive inputs has be achieved using sented in Section III even though the nature of the inputs are
high-gain feedback in both theory and experiments [6]–[8]. The completely different (continuous inputs in [15] vs a pair of
3214
Fig. 5. High-gain feedback implementation of two impulsive inputs (N = 2)
for swing-up of the IWP. For the purpose of comparison with the results in
Figs. 3 and 4, the controller is designed to keep θ in the domain (−π/2, 3π/2].
impulsive inputs, N = 2). By changing the domain of θ from
(−3π/2, π/2] to (−π/2, 3π/2] 3 and using (12) and (19), we
get for θf = π/2 and θ0 = −π/2:
 It can be seen from Fig. 6 that the second impulse is applied
+ 1 β when θ2 ≈ −π rad. Similar to the results in Fig. 5, swing-up is
| φ̇1 |= (45)
C m11 (1 − sin θ2 )
It is clear from (45) that the wheel velocity immediately
after application of the first impulse depends only on the angle
where the second impulse is applied, namely θ2 , and tends to
infinity when θ2 = π/2+ , i.e., when the overshoot approaches
zero. While it is clear from (45) that θ2 = π/2+ is not a good
choice for application of the second impulse, the value of θ2
that minimizes the magnitude of the wheel velocity | φ̇+ 1 | can
be obtained using the energy conservation law in (6a). For the
IWP to cross the upright configuration, the following inequality
must be satisfied:
2 1
m11 θ˙1+ + β sin θ0 > β ⇒ | φ̇+ 1 |> 2β/m11 (46)
C
where θ0 = −π/2 and (12) were used. Comparing (45) and
(46), we can show that θ2 = (5π/6)− minimizes | φ̇+ 1 |.
A simulation was performed using the high-gain feedback law
in (43) with  = 0.01, the parameter values in (44), and θ2 ≈
3π/4 (slightly less that 5π/6); the results are shown in Fig. 5.
After the IWP reached a neighborhood of θf = π/2, a linear
controller was invoked for stabilization. The linear controller
was designed to place the poles of the closed loop system at
−4 ± 2i and −8. The simulation results indicate that swing-up
is achieved in less than 1.0 s, which is much faster than that
achieved in [15]. The maximum velocity of the wheel is still
quite high (3000 rad/s) but it is significantly lower than that in
Fig. 3. The torque required is quite high (≈ 13 Nm) but this can
be reduced significantly by simply changing the domain of θ2 ,
as we will show in the next simulation.
The simulation results presented in Fig. 5 were obtained by
assuming θ ∈ (−π/2, 3π/2]; this was motivated by the need
to generate trajectories of the IWP similar to those generated
in [15], for comparison. If we switch the domain of θ back to
(−3π/2, π/2], as defined in Section II-C, the maximum wheel
speed and the magnitude of the maximum torque can both be
reduced from their values in Fig. 5. Simulation results of high-
gain feedback implementation of the optimal impulse sequence
based on two inputs, described by (27) and (28) with n = 1,
3 This change in the domain is necessary to ensure that the trajectory of θ is
similar to that in [15] but it does not change the analysis whatsoever.
IEEE ROBOTICS AND AUTOMATION LETTERS, VOL. 3, NO. 4, OCTOBER 2018
Fig. 6. High-gain feedback implementation of the sequence of two optimal
impulsive inputs (N = 2) for swing-up of the IWP. The controller is designed
to keep θ in the domain (−3π/2, π/2].
is shown in Fig. 6. The high-gain controller was implemented
using  = 0.02 and stabilization of the equilibrium was achieved
by the same linear controller that was used in the last simulation.
achieved in less than 1.0 s, but the maximum wheel velocity is
now reduced from 3000 rad/s to 2000 rad/s and the magnitude
of the maximum torque is reduced from ≈ 13 Nm to ≈ 3 Nm.
The maximum torque of ≈ 3 Nm in Fig. 6, although larger
than those reported in the literature, is not a significant con-
cern because it is applied for a very short duration of time.
Motors can apply substantially larger torques 4 than their maxi-
mum continuous torque over short time intervals. The maximum
torque of ≈ 3 Nm also corresponds to the continuous-time im-
plementation of the optimal I 2 . The magnitude of this torque,
as well as the maximum velocity of the wheel, can be easily re-
duced if we consider continuous-time implementation of I 2n ,
n = 2, 3, · · · . This is discussed in the next section.
Remark 6: Similar to the globally stabilizing controller [15],
the IDA-PBC method also takes the pendulum directly to the
desired upright configuration and results in a large torque [13]
or large wheel velocity [14] during swing-up; these results are
not presented here because of space constraints.
2) Energy Based Controller: When the number of impulses
are increased from N = 2 to N = 8, for example, the√magni-
tude of the impulsive torques are reduced by a factor of 4 = 2;
consequently, the magnitude of the maximum high-gain torque
and the wheel velocity are reduced proportionately - see Fig. 7.
The trajectories of the state variables in Fig. 7 resemble those of
the energy based controllers [12], [20] during swing-up phase
of the IWP; the PFBLC + AL energy-based controller presented
in [12] is simulated here to show the similarities in the trajecto-
ries - see Fig. 8. It can be seen from Figs. 7 and 8, that, unlike
the globally stabilizing controller [15] (see Fig. 3) where the
pendulum is aggressively driven towards its desired configura-
tion, both controllers (presented here and in [12]) gradually add
energy to the pendulum over several cycles.
A comparison of Figs. 7 and 8 indicates that the magnitude
of the maximum torque required by our method is larger than
that required by the approach proposed in [12]. However, since
the torques are applied intermittently over very short intervals of
time, feasibility of our approach is determined by the peak torque
rating of the actuator as opposed to the maximum continuous
4 This is referred to a peak torque [10]; for different motors, the peak torque
can be twice to ten times larger than the maximum continuous torque.
KANT AND MUKHERJEE: IMPULSIVE DYNAMICS AND CONTROL OF THE INERTIA-WHEEL PENDULUM
Fig. 7. High-gain feedback implementation of the sequence of eight optimal
impulsive inputs (N = 8) for swing-up of the IWP. The high-gain controller
was implemented with  = 0.02.
Fig. 8. Simulation using the PFBLC + AL controller in [12]; the controller
parameters were chosen as k e = 3.1 × 10 7 and k v = 0.1. This choice of pa-
rameters ensured that the time required for swing-up and the magnitude of the
maximum control torque in simulations matched those of the experiments. The
initial configuration was chosen to be slightly different from θ0 = −π/2 since
the controller is unable to swing-up from this configuration.
TABLE I
SWING-UP TIME AND MAXIMUM MAGNITUDE OF HIGH-GAIN TORQUE
REQUIRED FOR DIFFERENT VALUES OF N
torque rating, which is always lower [10]. A salient feature of
our approach is that, given any actuator, an optimal impulse
sequence can be designed such that the peak torque rating of
the motor is not exceeded. A higher value of N reduces the
peak torque requirement of the motor but increases the time
required for swing-up. The swing-up time and the magnitude
of the maximum torque for several different values of N are
presented in Table I.
VI. CONCLUSION
Rest-to-rest maneuvers of the inertia-wheel pendulum was
studied in the framework of impulsive control. Assuming a set
of discrete impulsive inputs, optimal sequences were designed
to minimize their infinity-norm. It was shown analytically that a
sequence with an odd number of inputs is less optimal than the
two adjacent sequences with even number of inputs. Analytical
and simulation results with two inputs were used to explain the
high wheel velocities and large continuous torques associated
with methods that attempt to take the pendulum directly to its
desired configuration and minimize overshoot. Implementation
of impulsive control using high-gain feedback also results in
3215
large torques but they act over short intervals of time; therefore,
feasibility of impulsive control is determined by the peak torque
rating of the actuator, which is always larger than the continu-
ous torque rating. It was shown that the number of impulsive
inputs can be increased to not exceed the peak torque rating
of the actuator; this, of-course, increases the time required for
swing-up. Simulation results for swing-up showed similarities
between the optimal trajectories and the trajectories obtained
using the energy-based controllers. Future work will investigate
the possibility of extending the method presented here to other
underactuated systems.
REFERENCES
[1] R. Goebel, R. G. Sanfelice, and A. R. Teel, “Hybrid dynamical systems,”
IEEE Control Syst., vol. 29, no. 2, pp. 28–93, Apr. 2009.
[2] W. M. Haddad, V. Chellaboina, and S. G. Nersesov, Impulsive and Hybrid
Dynamical Systems (Princeton Series in Applied Mathematics). Princeton,
NJ, USA: Princeton Univ. Press, 2006.
[3] B. M. Miller and E. Y. Rubinovich, Impulsive Control in Continuous and
Discrete-Continuous Systems. New York, NY, USA: Springer, 2012.
[4] Y. Aoustin, D. T. Romero, C. Chevallereau, and S. Aubin, “Impulsive
control for a thirteen-link biped,” in Proc. 9th IEEE Int. Workshop Adv.
Motion Control, 2006, pp. 439–444.
[5] S. Weibel, G. Howell, and J. Baillieul, “Control of single-degree-of-
freedom Hamiltonian systems with impulsive inputs,” in Proc. 35th IEEE
Conf. Decis. Control, 1996, vol. 4, pp. 4661–4666.
[6] T. Albahkali, R. Mukherjee, and T. Das, “Swing-up control of the pen-
dubot: An impulse-momentum approach,” IEEE Trans. Robot., vol. 25,
no. 4, pp. 975–982, Aug. 2009.
[7] F. B. Mathis, R. Jafari, and R. Mukherjee, “Impulsive actuation in
robot manipulators: Experimental verification of pendubot swing-up,”
IEEE/ASME Trans. Mechatronics, vol. 19, no. 4, pp. 1469–1474, Aug.
2014.
[8] R. Jafari, F. B. Mathis, R. Mukherjee, and H. Khalil, “Enlarging the
region of attraction of equilibria of underactuated systems using impulsive
inputs,” IEEE Trans. Control Syst. Technol., vol. 24, no. 1, pp. 334–340,
Jan. 2016.
[9] N. Kant, D. Chowdhury, R. Mukherjee, and H. Khalil, “Estimation of the
region of attraction of underactuated systems and its enlargement using
impulsive inputs,” IEEE Trans. Robot., to be published.
[10] H. A. Toliyat and G. B. Kliman, Handbook of Electric Motors, vol. 120.
Boca Raton, FL, USA: CRC Press, 2004.
[11] M. W. Spong, “The swing up control problem for the acrobot,” IEEE
Control Syst., vol. 15, no. 1, pp. 49–55, Feb. 1995.
[12] M. W. Spong, P. Corke, and R. Lozano, “Nonlinear control of the reaction
wheel pendulum,” Automatica, vol. 37, no. 11, pp. 1845–1851, 2001.
[13] R. Ortega, M. W. Spong, F. Gómez-Estern, and G. Blankenstein, “Sta-
bilization of a class of underactuated mechanical systems via intercon-
nection and damping assignment,” IEEE Trans. Autom. Control, vol. 47,
no. 8, pp. 1218–1233, Aug. 2002.
[14] M. Ryalat and D. S. Laila, “A simplified IDA-PBC design for underac-
tuated mechanical systems with applications,” Eur. J. Control, vol. 27,
pp. 1–16, 2016.
[15] R. Olfati-Saber, “Global stabilization of a flat underactuated system: The
inertia wheel pendulum,” in Proc. IEEE Int. Conf. Decis.. Control, 2001,
vol. 4, pp. 3764–3765.
[16] N. Qaiser, N. Iqbal, and A. Hussain, “Stabilization of non-linear in-
ertia wheel pendulum system using a new dynamic surface control
based technique,” in Proc. IEEE Int. Conf. Eng. Intell. Syst., 2006,
pp. 1–6.
[17] A. Khoroshun, “Stabilization of the upper equilibrium position of a pen-
dulum by spinning an inertial flywheel,” Int. Appl. Mech., vol. 52, no. 5,
pp. 547–556, 2016.
[18] A. Zhang, C. Yang, S. Gong, and J. Qiu, “Nonlinear stabilizing control of
underactuated inertia wheel pendulum based on coordinate transformation
and time-reverse strategy,” Nonlinear Dyn., vol. 84, no. 4, pp. 2467–2476,
2016.
[19] N. Kant, R. Mukherjee, and H. K. Khalil, “Swing-up of the inertia wheel
pendulum using impulsive torques,” in Proc. IEEE 56th Annu. Conf. De-
cis.. Control, Dec. 2017, pp. 5833–5838.
[20] I. Fantoni, R. Lozano, and M. Spong, “Stabilization of the reaction wheel
pendulum using an energy approach,” in Proc. Eur. Control Conf., 2001,
pp. 2552–2557.