Advanced Robotics 23 (2009) 1999–2014

www.brill.nl/ar

Full paper

Control of the Underactuated Inertia Wheel Inverted

Pendulum for Stable Limit Cycle Generation

Sébastien Andary, Ahmed Chemori ∗ and Sébastien Krut

LIRMM, University of Montpellier 2, CNRS, 161 rue Ada, 34392 Montpellier, France
Received 17 February 2009; accepted 3 July 2009

Abstract
This paper deals with a control approach dedicated to stable limit cycle generation for underactuated me-
chanical systems. The proposed approach is based on partial nonlinear feedback linearization and dynamic
control for optimal periodic reference trajectories tracking. The computation of the reference trajectories is
performed in order to optimize the behavior of the whole dynamics of the system and especially its zero
dynamics at the end of each cycle. Simulation results as well as experiments show the performance and the
efficiency of the proposed control scheme.
© Koninklijke Brill NV, Leiden and The Robotics Society of Japan, 2009

Keywords
Limit cycle, underactuated systems, optimization, partial feedback linearization, inertia wheel inverted pen-
dulum

1. Introduction

Underactuated mechanical systems [1] are those systems that have less control in-
puts than generalized coordinates (degrees of freedom), i.e., they have generalized
coordinates that are not actuated. In such systems the unactuated generalized coor-
dinates may indirectly be controlled by the actuated coordinates through a dynamic
coupling. This coupling is often nonlinear, hence the resulting dynamic constraints
are generally nonintegrable and, therefore, second-order nonholonomic.
Many examples of such systems exist; mainly in robotics, they include, among
others, the underactuated robot manipulators, the gymnast robots, particularly the
acrobot [2], the pendubot [3], planar vertical takeoff and landing aircraft [4], under-
sea vehicles [5] and other mobile robots [6].

* To whom correspondence should be addressed. E-mail: chemori@lirmm.fr

© Koninklijke Brill NV, Leiden and The Robotics Society of Japan, 2009 DOI:10.1163/016918609X12529279062438
2000 S. Andary et al. / Advanced Robotics 23 (2009) 1999–2014

Underactuation in these mechanical systems is generally introduced intentionally

at the design level to reduce the manufacturing cost, weight and/or failure rate, so
the obtained systems may be able to perform complex tasks with a reduced number
of actuators. However, such systems may require new approaches to design effec-
tive control strategies, therefore, they constitute a good framework of nonlinear
control problems of both theoretical and practical interests. For these reasons they
are attracting more and more attention from researchers from the nonlinear control
community and robotics.
In the literature many research efforts have been made on control aspects [2, 3,
7–11]; however, the field of control of such systems is still open to develop other
control strategies.
In Ref. [9], a technique design combining partial feedback linearization and Lya-
punov methods based on saturation functions, switching and energy shaping has
been proposed. In Ref. [11], the author discusses a survey on some existing results
such as geometric nonlinear control, and passivity based control for stabilizing and
tracking control of such systems. Three control algorithms have been proposed in
Ref. [8] for an underactuated two-link robot, i.e., an optimal LQ controller, a partial
feedback linearizing regulator and a sliding mode controller.
In Ref. [3], the testbed is also a two-link underactuated robot, but the actuator
is located at the base (in contrast to that proposed in Ref. [8], where the actuator
is located at the joint between the two links) and the authors propose a method
based on partial feedback linearization, nilpotent approximation and an iterative
stabilization procedure. Roughly speaking they propose an open-loop controller that
is able to steer the system closer to the desired equilibrium point in finite time
and apply it iteratively, which provides a robust exponential convergence to the
equilibrium.
In Ref. [10], the authors discuss the nonlinear behavior of the same robot pro-
posed in Ref. [3] through a graphical tool based on the Poincaré map. The study in
Ref. [7] concerns underactuated mechanical systems with several actuated degrees
of freedom and a single unactuated degree of freedom, where the authors propose a
discontinuous nonlinear feedback controller that allows the closed-loop equilibrium
at the origin to be globally attractive. In Ref. [12], a nonlinear predictive controller
is proposed for both stabilization and stable limit cycle generation with an applica-
tion to the ECP 505 inverted pendulum.
In this paper, a control approach based on collocated partial feedback lineariza-
tion [11] and optimal reference trajectories is proposed, where the objective is the
generation of stable limite cycles on both actuated and unactuated coordinates. The
chosen application is the inertia wheel inverted pendulum. Simulation results show
the effectiveness of the proposed control scheme; furthermore, experiments confirm
the simulation results and show the robustness of the controller towards external
disturbances.
This paper is organized as follows. In Section 2, the inertia inverted pendulum
is described and its dynamic model is given. The generation of the optimal refer-
 S. Andary et al. / Advanced Robotics 23 (2009) 1999–2014 2001

ence trajectories as well as the control law are discussed in Section 3. Section 4 is
dedicated to simulation results, while experiments are presented and discussed in
Section 5. Finally, concluding remarks are drawn in Section 6.

2. Inertia Wheel Inverted Pendulum

The underactuated mechanical system studied in this paper is an inertia wheel in-
verted pendulum (Fig. 1) which consists of an inverted pendulum equipped with
a rotating wheel. Its mechanical structure is sketched in Fig. 2. The motor torque
produces an angular acceleration of the rotating wheel, which generates, thanks to
the dynamic coupling, a torque acting on the pendulum’s passive joint (as shown
in Fig. 3); therefore, this passive joint can be controlled through the acceleration of
the inertia wheel.

Figure 1. Inertia wheel inverted pendulum.

Figure 2. Schematic view of the system’s mechanical structure: the joint between the frame and the
pendulum body is unactuated (passive), while the joint between the body and the inertia wheel is
actuated (active).
2002 S. Andary et al. / Advanced Robotics 23 (2009) 1999–2014

Figure 3. Mechanical principle: the inertia wheel is assumed to be equivalent to two punctual masses
situated at the acting points of F + and F − . To straighten up the pendulum, the torque acting at O
generated by F + must be greater than that generated by both gravity and F − forces.

2.1. Dynamic Modelling of the Plant

The dynamic model of the system is obtained by application of the Lagrange for-
mulation [13] and is described by:
      
i1 + i2 i2 θ̈1 −mlg sin(θ1 ) C1
+ = , (1)
i2 i2 θ̈2 0 C2
where θ1 and θ2 are, respectively, angular positions of the pendulum body and the
inertia wheel (Fig. 2), and θ̇i and θ̈i , i = 1, . . . , 2, represent their corresponding
velocities and accelerations. i1 and i2 are, respectively, their moments of inertia.
C1 is the external disturbing torque applied to the pendulum while C2 is the torque
exerted by the actuator. ml = m1 l1 + m2 l2 , where m1 and m2 are the masses of the
pendulum and the inertia wheel, respectively, and l1 and l2 are distances from origin
O to the gravity center of the pendulum body and the inertia wheel, respectively.
The proposed control approach is detailed in the next section.

3. Proposed Control Scheme

The objective of the control approach is to generate stable limit cycles on both actu-
ated and unactuated coordinates. The system is underactuated since it has only one
control input and 2 d.o.f. The control input can be used to track reference trajecto-
ries on only one coordinate. The question that may be asked is the following: which
coordinate should be used to track the reference trajectories?
The choice of the coordinate on which the reference trajectories will be defined
and tracked depends on the fixed objective. In our case, since the aim is to make
the pendulum act like a metronome, the angle of the pendulum body θ1 has been
chosen to be tracked.
The application of collocated partial feedback linearization allows us to linearize
completely the dynamics of the unactuated coordinate θ1 , whereas the remaining
nonlinear dynamics of θ2 constitutes what is called the internal dynamics of the
 S. Andary et al. / Advanced Robotics 23 (2009) 1999–2014 2003

system. Since the system is nonminimum phase, the reference trajectories have to
be optimized in order to ensure the stability of the internal dynamics. The optimiza-
tion criterion is based on the rotation speed of the inertia wheel, which has to be
minimized at the end of each cycle. Therefore, the idea is to propose parameterized
reference trajectories and to compute their parameters by solving an optimization
problem. The generation of the reference trajectories is discussed next.
3.1. Reference Trajectories Generation
Reference trajectories are to be established with the aim of generating stable limit
cycles. Their shapes have to be chosen taking into account the following consider-
ations:
(i) Since the desired dynamic behavior of the system is symmetrical with respect
to the vertical, a periodic symmetrical reference trajectory should be chosen for
θ1 in order to obtain stable periodic evolution on both actuated and unactuated
coordinates. Indeed, a perfect tracking on θ1 of the nominal system leads to a
periodic evolution of θ̇2 . Conversely, when the trajectory is asymmetrical with
respect to the vertical, the pendulum rod will spend more time on one side and
consequently the velocity of the inertia wheel at the end of the cycle will be
different from zero.
(ii) The tracking of a symmetrical trajectory when the system is subject to un-
certainties (e.g., on model parameters, external disturbances, etc.) leads to an
aperiodic evolution of θ̇2 , i.e., its final value will be different from the initial
value (zero value). For this reason, it will be clever to deform the original sym-
metrical reference trajectory in order to spend more time on one side to correct
the asymmetrical evolution of θ̇2 . Doing that, the aim is to steer θ̇2 to regain
zero at the end of the cycle.
(iii) In our case, for smooth reference trajectories, periodic functions twice con-
tinuously differentiable are chosen to avoid discontinuities on velocities and
accelerations.
Trajectories that fulfill the above considerations are designed as follows. They are
split up in half-periods and each part is expressed by a parameterized polynomial
function such that the reference trajectories on the whole cycle can be written as:
⎧   
⎪
⎪ 2t τ
⎪
⎨ AP τ , p if t ∈ 0,
2
θ1 (t, p) =
ref
   (2)
⎪
⎪ 2t τ
⎪
⎩ −AP − 1, 1 − p if t ∈ ,τ ,
τ 2
where A and τ are, respectively, the amplitude and the period of oscillations, and
p is the optimization parameter. P (t, p) is the normalized polynomial in t that
varies from 1 (at t = 0) to −1 (at t = 1) for all p. In order to join smoothly suc-
cessive parts of the reference trajectories, the first and second time derivatives of
2004 S. Andary et al. / Advanced Robotics 23 (2009) 1999–2014

P (t, p) are constrained to be zero at boundaries (cf. consideration (iii) above). Fi-
nally, p is such that P (t = p, p) = 0.
These considerations are mathematically expressed by the following constraints:
⎧
⎪
⎪ P (0, p) = 1; P (t = p, p) = 0; P (1, p) = −1
⎪
⎪
⎪
⎨ ∂P ∂P
(0, p) = 0; (1, p) = 0
∂t ∂t (3)
⎪
⎪
⎪
⎪ ∂ 2P ∂ 2P
⎪
⎩ (0, p) = 0; (1, p) = 0.
∂t 2 ∂t 2
Having seven constraints, then P (t, p) can be defined as a six-degree p-paramete-
rized polynomial in t, such that:

6
P (t, p) = αi (p)t i . (4)
i=0
The identification of coefficients αi (p) of the polynomial (4) taking into account
constraint (iii) leads to the following:
(2p − 1)(6p4 − 12p3 + 4p2 + 2p + 1)
α6 (p) =
p3 (p − 1)3
−3(4p6 − 18p4 + 16p3 − 1)
α5 (p) =
p3 (p − 1)3
3(10p6 − 18p5 + 10p3 − 1)
α4 (p) =
p3 (p − 1)3
−(20p6 − 48p5 + 30p4 − 1)
α3 (p) =
p3 (p − 1)3
α2 (p) = 0
α1 (p) = 0
α0 (p) = 1.
To show the effect of the parameter p on the reference trajectory, Fig. 4 shows
the evolution of the polynomial P (t, p) for different values of p. Note that the
pendulum spends the same amount of time on both sides of the equilibrium point
when p = 0.5.
For practical reasons (e.g., avoid touching mechanical stops) the passive joint θ1
of the pendulum is constrained to remain within interval [−A, A]; therefore, only a
subinterval [pmin , pmax ] of [0, 1] is admissible for p values. This requirement can
be written for the normalized trajectories as:
∀p ∈ [pmin , pmax ], ∀t ∈ [0, 1], |P (t, p)|  1, (5)
where pmin and pmax can be computed as follows. Assume for some p that:
∃t1 ∈ [0, 1] s.t. |P (t1 , p)| > 1. (6)
 S. Andary et al. / Advanced Robotics 23 (2009) 1999–2014 2005

Figure 4. Evolution of the polynomial P (t, p) for different values of p. Note that this corresponds to
the normalized reference trajectory of θ1 during half a period.

Consider P (0, p) = 1, P (1, p) = −1 and (6), P must admit at least one null tan-
gency point. It means that at least one root of ∂P (t, p)/∂t is within [0, 1]. The
computation of the roots of ∂P (t, p)/∂t leads to the following unique real solution:
∂P 20p6 − 48p5 + 304 − 1
(t1 , p) = 0 ⇒ t1 = . (7)
∂t 24p5 − 60p4 + 40p3 − 2
The two extremal trajectories P (t, pmin ) and P (t, pmax ) correspond, respectively,
to values 0 and 1 for t1 . This leads us to solve the following two equations:
20p6 − 48p5 + 30p4 − 1
=0 (8)
24p5 − 60p4 + 40p3 − 2
20p6 − 48p5 + 30p4 − 1
= 1. (9)
24p5 − 60p4 + 40p3 − 2
Equation (8) admits a unique real solution p = pmin , and (9) admits p = pmax as
solution, both in interval [0, 1].

Remark 1. Numerical solution of (8) is pmin  0.42, and that of (9) is pmax  0.58,
where it can be seen, due to the complementary feature between pmin and pmax , that
pmin + pmax = 1.

3.2. Tracking Control Law

The dynamic model of the system, expressed by (1), can be rewritten as:
1
θ̈1 = (C1 − C2 + mlg sin θ1 ) (10)
i1
−1 (i1 + i2 ) −mlg
θ̈2 = C1 + C2 + sin θ1 . (11)
i1 i1 i2 i1
2006 S. Andary et al. / Advanced Robotics 23 (2009) 1999–2014

The above equations describe the system dynamics in open loop. According to col-
located partial feedback linearization [9], consider the following control law (the
torque C1 is considered to be zero since the nominal system is not subject to exter-
nal disturbing torque):
   
C2 = mlg sin(θ1 ) − i1 θ̈1ref + i1 kp θ1 − θ1ref + i1 kv θ̇1 − θ̇1ref , (12)
where kp and kv are, respectively, the proportional and derivative feedback gains,
and are used to tune the dynamics of the closed-loop system. Injecting this con-
trol law in (10) leads to the following linear closed-loop subsystem (completely
linearized part):
ë + kv ė + kp e = 0, (13)
where e = θ1 − θ1ref , ė = θ̇1 − θ̇1ref and ë = θ̈1 − θ̈1ref . By choosing the feedback gains
kp and kv as positive definite matrices (scalar strictly positive gains in our case),
the resulting linear system is asymptotically stable [14]. Therefore, limt→∞ e =
limt→∞ ė = 0 and consequently θ1 → θ1ref and θ̇1 → θ̇1ref as t → ∞. The pendulum
passive joint then follows the reference trajectory θ1ref (t).
3.3. Trajectory Optimization
The aim of trajectory optimization is to minimize the motor rotating speed at the
end of each cycle. In other words, for a given initial value θ̇2 (0), the goal is to
choose the optimal value of p that minimizes a quadratic cost-function of θ̇2 (τ ).
Injecting the control law (12) in (10) and (11) leads to the following closed-loop
system:
   
θ̈1 = θ̈1ref − kv θ̇1 − θ̇1ref − kp θ1 − θ1ref (14)
i1 + i2  ref     mlg
θ̈2 = − θ̈1 − kv θ̇1 − θ̇1ref − kp θ1 − θ1ref + sin(θ1 ). (15)
i2 i2
Knowing the p-parametrized reference trajectory (2), numerical integration of
(14) and (15), starting from initial condition (θ1 (0), θ̇1 (0), θ̇2 (0)), enables us to ob-
tain the inertia wheel velocity at the end of the cycle θ̇2 (τ ). Hence, the quadratic
function to be minimized can be expressed by g(p, θ1 (0), θ̇1 (0), θ̇2 (0)). The opti-
mization problem can then be formulated as:
p̂ = Arg min g(p, θ1 (0), θ̇1 (0), θ̇2 (0)). (16)
p
This problem is to be solved at the beginning of each cycle. Its solution allows us
to define completely the reference trajectory to be tracked on the next oscillation
cycle.

4. Simulation Results
Consider the dynamic model of system (1) with dynamical parameters described in
Table 1. These parameters have been identified thanks to the real prototype of the
system.
 S. Andary et al. / Advanced Robotics 23 (2009) 1999–2014 2007

Table 1.
Description of dynamical parameters of the inverted pendulum

Parameter Description Value Unit

m1 Body mass 3.228 kg

m2 Wheel mass 0.86422 kg
i1 Body inertia 3.042 × 10−2 kgm2
i2 Wheel inertia 7.986 × 10−4 kgm2
l1 Body center of mass position 6.354 × 10−4 m
l2 Wheel center of mass position 52 × 10−3 m

Table 2.
Description of the reference trajectories and control parameters

Parameter Description Value Unit

A Oscillation amplitude 3 deg

τ Oscillation period 2 s
kp Proportional gain 300 1/s2
kv Derivative gain 20 1/s

Simulation results, obtained using Matlab software, are presented and discussed
in the following. They attest to the feasibility of the proposed control scheme. Ta-
ble 2 lists the parameters of the reference trajectories and the control approach.
The proposed simulation is started from the initial condition (θ1 (0), θ̇1 (0),
θ̇2 (0)) = (10◦ , 0 rad/s, 0 rad/s). The choice of θ1 (0) is motivated by the physi-
cal properties of the inverted pendulum (pendulum angle value is approximatively
±10◦ at standstill) since simulation is started from rest.
Figure 5 displays the whole simulation results. Figure 5a and 5b shows the
pendulum body joint position and velocity versus time in solid lines, while the
dashed lines represent their corresponding reference position and velocity. This
shows clearly the convergence of position and velocity to their reference trajec-
tories. The inertia wheel velocity versus time is displayed in Fig. 5c where it can
be observed that within two periods, the motor velocity reaches a limit cycle and
oscillates around zero. Figure 5e represents the control input that consists of the
motor driver voltage (proportional to the motor torque), where it can be noticed that
it remains within the admissible limits (±10 V).
The phase portrait (θ1 , θ̇1 ) is depicted in Fig. 5d, showing clearly the conver-
gence from the initial condition to a stable limit cycle. Figure 5f represents the
absolute value of motor velocity versus absolute motor torque and the admissible
region of the actuator, given by the manufacturer of the DC motor; it can be seen
that the trajectories remain inside the admissible region. Finally, in Fig. 5g, the evo-
2008 S. Andary et al. / Advanced Robotics 23 (2009) 1999–2014

Figure 5. Simulation results with A = 3◦ , τ = 2 s.

 S. Andary et al. / Advanced Robotics 23 (2009) 1999–2014 2009

lution of the normalized optimal parameter (which evolves within interval [−1, 1])
is displayed, where it converges after two cycles to a null value.

5. Experimental Results
In this section, the experimental results are represented with implementation issues.
It starts with a description of the inverted pendulum testbed, and then two experi-
ments are presented. The first one concerns the control of the disturbance-free real
system, whereas in the second one the system is subject to external perturbations.
5.1. Testbed and Implementation Issues
Real-time experiments are performed on an inertia wheel inverted pendulum
(Fig. 1) designed and developed at our laboratory. The pendulum angle θ1 is con-
strained to remain within the interval [−10◦ , 10◦ ] due to mechanical stops. The
actuator of the plant is a Maxon EC-powermax 30 DC motor, equipped with an
incremental encoder, allowing us to measure the inertia wheel angular position and
to compute its velocity in real-time. In order to measure the angle of the pendulum
with respect to the vertical, the system is equipped with a Micro strain FAS-G incli-
nometer. The choice of this solution, other than a classical encoder, has been made
since it does not require us to set the frame horizontally. The system is controlled
with a computer equipped with a 2.4-GHz Intel processor. The control approach is
implemented using C language and the whole system runs under the Ardence RTX
real-time OS.
The proposed control scheme requires that an optimization problem be solved
in real-time; however, to overcome the problem of computation time, online imple-
mentation is performed using a lookup table. For that a grid is defined for different
initial values θ̇2 (0). Since the reference trajectories are periodic and the proposed
controller is performant for tracking them, the influence of θ1 (0) and θ̇1 (0) on the
optimization parameter p is minor. For that reason, the lookup table was defined
only on θ̇2 (0).
The optimization parameter (16) is computed offline for each value of θ̇2 (0) on
the grid and stored in a table. Then, real-time implementation of the proposed con-
trol scheme uses this lookup table as well as an interpolation technique to obtain
instantaneously the optimization parameter.
5.2. Real-Time Experiments
In this section, real-time experiments are presented and discussed. They are per-
formed on the inverted pendulum testbed presented in the previous section. Two ex-
perimental scenarios are presented and discussed. In the first one the proposed con-
trol scheme is implemented without considering any external disturbance, whereas
in the second one, to show the robustness of the proposed approach, the system is
subject to external disturbances.
2010 S. Andary et al. / Advanced Robotics 23 (2009) 1999–2014

5.2.1. Scenario 1: Without External Disturbances

Figure 6 shows the whole results of this scenario. The pendulum joint position and
velocity are displayed in Fig. 6a and 6b, respectively. Noise can be observed on
the pendulum body velocity θ̇1 since this last one is computed using a numerical
derivation of the measured angular position θ1 .
Figure 6c represents motor velocity versus time and shows how the optimization
of the reference trajectories brings the motor velocity oscillation around zero in a
few periods. DC motor input voltage can be observed in Fig. 6e where it is shown
that it remains within the admissible range ±10 V. The phase portrait (θ1 , θ̇1 ) is
displayed in Fig. 6d. The limit cycle is reached even in the presence of noise. Fig-
ure 6f shows the admissible region of the motor power. Finally, Fig. 6g represents
the evolution of the normalized optimization parameter.
This experiment is carried out with nonzero initial conditions; however, despite
this fact the controller is able to bring the system to the reference trajectory while
bringing back and keeping the inertia wheel angular velocity oscillation around
zero.
5.2.2. Scenario 2: With External Disturbances
In this scenario, disturbances are introduced by pushing the pendulum, which gen-
erates external punctual torques applied to the pendulum joint at approximately
t = 16 and 33 s.
Figure 7 represents the whole experimental results obtained with this second
scenario. In Fig. 7a, the pendulum joint position and its corresponding reference
trajectory are plotted versus time. Figure 7b displays the curve of the pendulum
joint velocity as well as its reference trajectory. On this curve, the effect of external
disturbances can be observed as peaks that appear at instants of application of the
external disturbing torques.
External disturbance compensation can be observed in Fig. 7c and 7e where, re-
spectively, inertia wheel velocity and motor input voltage are displayed. One clearly
observes peaks at the instant of disturbance and how the system brings back the mo-
tor velocity oscillation around zero after each peak.
The phase portrait and the motor’s admissible power are displayed in Fig. 7d
and 7f, respectively. The limit cycle is rejoined after each perturbation. Figure 7g
represents the evolution of the obtained normalized optimization parameter versus
time.
The controller is able to keep the system around the reference trajectories and
reject external disturbances introduced as unpredicted torques applied on the pen-
dulum axis in a punctual fashion.

6. Conclusions and Future Work

In this paper, a control approach is proposed for generating stable limit cycles for
underactuated mechanical systems. The control scheme is designed in the special
case of the inertia wheel inverted pendulum; however, it can be easily generalized
 S. Andary et al. / Advanced Robotics 23 (2009) 1999–2014 2011

Figure 6. Experimental results with A = 3◦ , τ = 2 s (without external disturbances).

2012 S. Andary et al. / Advanced Robotics 23 (2009) 1999–2014

Figure 7. Experimental results with A = 3◦ , τ = 2 s (with external disturbances).

 S. Andary et al. / Advanced Robotics 23 (2009) 1999–2014 2013

to the more general case of underactuated mechanical systems. It is based on col-

located partial feedback linearization and optimal reference trajectories tracking.
The obtained results in simulation and real-time experiments are presented and dis-
cussed to show the effectiveness and the performance of the proposed scheme.
Currently, the work is ongoing trying to find an analytical solution to the opti-
mization problem. Furthermore, the robustness of the control scheme can be sig-
nificantly improved by adopting more flexible reference trajectories and updating
them many times during a cycle (in the actual version of the controller this is done
only once a period).

References
1. I. Fantoni and R. Lozano, Non-linear Control for Underactuated Mechanical Systems. Springer,
New York, NY (2001).
2. M. Spong, The swing up control problem for the acrobot, IEEE Control Syst. Mag. 15, 49–55
(1995).
3. A. De Luca, R. Mattone and G. Oriolo, Stabilization of underactuated planar 2r manipulator, Int.
J. Robust Nonlinear Control 10, 181–198 (2000).
4. G. Poulin, A. Chemori and N. Marchand, Minimum energy oriented global stabilizing control of
the PVTOL aircraft, Int. J. Control 80, 430–442 (2007).
5. T. I. Fossen, Nonlinear modelling and control of underwater vehicles, PhD Thesis, Norwegian
Institute of Technology (1991).
6. S. Ge and F. Lewis, Autonomous Mobile Robots. CRC Press, New York, NY (2006).
7. M. Reyhanoglu, S. Ludvigsen and O. Egeland, Discontinuous feedback control of a special class
of underactuated mechanical systems, Int. J. Robust Nonlinear Control 10, 265–282 (2000).
8. K. Lee and V. Coverstone-Carroll, Control algorithms for stabilizing underactuated robots, J. Ro-
botic Syst. 15, 681–697 (1998).
9. M. Spong, Energy based control of a class of underactuated mechanical system, in: Proc. 13th
IFAC World Congr., San Francisco, CA, pp. 431–436 (1996).
10. Y. Nakamura, T. Suzuki and M. Koinuma, Nonlinear behavior and control of nonholonomic free-
joint manipulator, IEEE Trans. Robotics Automat. 13, 853–862 (1997).
11. M. Spong, Underactuated mechanical systems, in: Control Problems in Robotics and Automation,
B. Siciliano and K. P. Valavanis (Eds), pp. 135–150. Springer, New York, NY (1998).
12. A. Chemori and M. Alamir, Nonlinear predictive control of under-actuated mechanical systems,
application: the ECP 505 inverted pendulum, in: Proc. Mathematical Theory of Networks and
Systems, Leuven (2004).
13. M. Spong and M. Vidyasagar, Robot Dynamics and Control. Wiley, New York, NY (1989).
14. H. Khalil. Nonlinear Systems, 2nd edn. Prentice Hall, Upper Saddle River, NJ (1996).
2014 S. Andary et al. / Advanced Robotics 23 (2009) 1999–2014

About the Authors

Sébastien Andary received the BS and MS degrees in Computer Science from
the University of Sciences, Montpellier, France, in 2004 and 2006, respectively.
He is now a PhD student at the University of Sciences, Montpellier, working for
the Montpellier Laboratory of Computer Science, Microelectronics and Robotics.
His research interests include robotics, nonlinear control and artificial learning.

Ahmed Chemori received the MS degree in Automatic Control, in 2001, and

the PhD degree in Automatic Control, in 2005, from Institut National Polytech-
nique de Grenoble, France. He was a Post-doctoral fellow with the Automatic
Control Laboratory of Grenoble, France, in 2006. He is currently a tenured Re-
search Scientist in Automatic Control and Robotics for the French National Center
for Scientific Research, at the Montpellier Laboratory of Computer Science, Mi-
croelectronics and Robotics, Montpellier, France. His research interests include
nonlinear control, adaptive control, predictive control and their applications in
robotics.

Sébastien Krut received the MS degree in Mechanical Engineering from the

Pierre and Marie Curie University, Paris, France, in 2000, and the PhD degree
in Automatic Control from the Montpellier University of Sciences, Montpellier,
France, in 2003. He was a Post-doctoral fellow with the Joint Japanese–French
Robotics Laboratory, Tsukuba, Japan, in 2004. He is currently a tenured research
scientist in Robotics for the French National Center for Scientific Research, at
the Montpellier Laboratory of Computer Science, Microelectronics and Robotics,
Montpellier, France. His research interests include design and control of robotic
systems.