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

net/publication/221216157

Smooth Reference Tracking of a Mobile Robot using Nonlinear Model

Predictive Control.

Conference Paper · January 2009

Source: DBLP

CITATIONS READS

23 486

3 authors, including:

Kiattisin Kanjanawanishkul
Mahasarakham University
40 PUBLICATIONS 413 CITATIONS

SEE PROFILE

All content following this page was uploaded by Kiattisin Kanjanawanishkul on 19 February 2016.

The user has requested enhancement of the downloaded file.

 1

Smooth Reference Tracking of a Mobile Robot

using Nonlinear Model Predictive Control
Kiattisin Kanjanawanishkul Marius Hofmeister Andreas Zell
Computer Science Department, University of Tübingen, Tübingen, Germany

Abstract— In this paper, path following control and trajectory smoother convergence to a path is achieved compared to
tracking control of a mobile robot have been studied. Reference trajectory tracking controllers, and the control signals are less
convergence in a path following problem and time convergence likely pushed to saturation. The solutions of this path following
in a trajectory tracking problem are considered in the cost
function of the nonlinear model predictive control framework. problem have been applied in a wide range of applications.
The benefit of path following control is that the path following For example, Samson [5] described a path following problem
controller eliminates aggressiveness of the tracking controller for a car pulling several trailers. In [6], Altafini addressed a
by forcing convergence to the desired path in a smooth way. path following controller for an n trailer vehicle. Furthermore,
Thus, we incorporate this benefit to the trajectory tracking path following controllers for aircraft and marine vehicles have
problem to achieve smooth convergence to the reference and
to achieve time convergence of trajectory tracking. Furthermore, been reported in [7] and [8], respectively.
by using nonlinear model predictive control, input constraints Let Γ(s) ∈ R2 be a desired geometric path parameterized by
can be handled straightforwardly in the optimization problem the curvilinear abscissa s(t) ∈ R. We then have the freedom
so that the robot can travel safely. Our controller was validated to select a temporal specification for s(t). In particular, the
by simulation and real-world experiments with a unicycle-type
rate of progression (ṡ) of a virtual vehicle, considered as an
mobile robot were also conducted.
addition control input, has been controlled explicitly (e.g. [9,
Index Terms— Path following, trajectory tracking, mobile 10, 11, 12]). Stringent initial condition constraints that are
robot, nonlinear model predictive control
present in a number of path following control strategies have
been overcome, as stated in [9].
I. I NTRODUCTION In this paper, we wish to achieve smooth spatial conver-
gence to the trajectory as well as time convergence using the
Three generic problems of motion control of a vehicle
advantage of path following control. This is accomplished by
addressed in the literature can be described below [1]:
modifying the cost function of the model predictive control
• point stabilization, where the objective is to stabilize a
framework through the addition of a time dependent penalty
vehicle at a desired robot posture, term. Based on this concept, our controller is able to opti-
• trajectory tracking, where the vehicle is required to track
mize the reference point between the virtual vehicle (path-
a time-parameterized reference, and parameterized) and the trajectory point (time-parameterized)
• path following, where the vehicle is required to converge
and also takes into account input constraints. Furthermore,
to and follow a desired path-parameterized reference, in the presence of obstacles, the controller deviates from the
without any temporal specifications. reference by incorporating obstacle information from range
Point stabilization is very different from the problems of sensors into the optimization, while respecting motion con-
path following and trajectory tracking. A central aspect of the straints.
problem, which triggered much of the subsequent research This paper is organized as follows: Section II describes the
on the control of nonholonomic systems, is that asymptotic mathematical model of a mobile robot and explains the basic
stabilization of fixed points cannot be achieved by using principle in path following and trajectory tracking. The control
continuous feedbacks which depend on the state only. This law based on nonlinear model predictive control (NMPC) is
is a consequence of an important result due to Brockett in developed in Section III. In Section IV, simulation results are
1983 [2]. shown, and then real-world experiments with a unicycle-type
The trajectory tracking problem for fully actuated systems is mobile robot are given in Section V. Finally, our conclusions
now well studied. However, when it comes to underactuated and future work are drawn in Section VI.
vehicles, i.e., when the vehicle has less actuators than state
variables to be tracked, the problem is still a very active topic
of research. II. P ROBLEM F ORMULATION
In path following control, a path following controller should A simple kinematic model of a unicycle-type mobile robot
look at (i) the distance from the vehicle to the reference path is the following:
and (ii) the angle between the vehicle’s velocity vector and    
the tangent to the path, and then reduce both to zero, without ẋm vm cos θm
any consideration in temporal specifications. Pioneering work  ẏm  =  vm sin θm  , (1)
in this area can be found in [3] as well as [4]. Typically, θ̇m ωm
 2

where xm (t) = [xm , ym , θm ]T is the state vector in the world problem is solved, the system control input signals ṡ, am , ωm
frame. vm and ωm stand of the linear and angular velocities, can be obtained by (5).
respectively. Subsequently, the error state dynamic model becomes
We first consider the path following control problem of a   
ẋe 0 ṡκ 0 0
  
xe u1

mobile robot. We wish to find control law ṡ and ωm such  ẏe  −ṡκ 0 0 0  ye  vm sin θe 
that the robot follows a virtual vehicle with position xd =  θ̇e  =  0
ẋe =      +   .
0 0 0  θe   u2 
[xd , yd , θd ]T . The kinematic model of a mobile robot can be η̇e 0 0 0 0 ηe u3
formulated with respect to a Serret-Frenet frame moving along (6)
the reference path. This frame plays the role of the body frame
of a virtual vehicle that must be followed by the real robot III. N ONLINEAR M ODEL P REDICTIVE C ONTROL D ESIGN
as depicted in Figure 1(b), together with a spatial path Γ. Nonlinear model predictive control (NMPC) is based on
In the path following problem, we normally let the forward a finite-horizon continuous time minimization of nonlinear
velocity vm track a desired velocity profile vd , while the rate predicted tracking errors with constraints on the control inputs
of progression of a virtual vehicle ṡ converges to vm . The and the state variables. It predicts system outputs based on
error state vector xe between the robot state vector xm and a current states and the system model, finds an open-loop control
virtual vehicle’s state vector xd can be expressed in the frame profile by numerical optimization, and applies the first control
of the path coordinate as follows signal in the optimized control profile to the system. However,
     due to the use of a finite horizon, control stability becomes
xe cos θd sin θd 0 xm − xd
 ye  = − sin θd cos θd 0  ym − yd  . one of the main problems. In general, the terminal region
(2)
constraint and/or the terminal penalty in the cost function are
θe 0 0 1 θm − θd
employed to enforce stability. Basically, the terminal penalty
Using (1) and (2), the error state dynamic model chosen in is assumed to be a control Lyapunov function for the system in
a rotated coordinate frame becomes the terminal region, enforcing a decrease in the value function.
ẋe = ye ṡκ − ṡ + vm cos θe The terminal region constraint is added to enforce that if the
open-loop optimal control problem is feasible once, that it will
ẏe = −xe ṡκ + vm sin θe , (3) remain feasible, and to allow establishing the decrease using
θ̇e = ωm − ṡκ the terminal penalty (see [13, 14, 15, 16] for more details).
Most model predictive controllers use a linear model of
where κ is the path curvature and ṡ is the velocity of a virtual
mobile robot kinematics to predict future system outputs. In
vehicle, bounded by 0 ≤ ṡ ≤ ṡmax .
[17, 18], a model-predictive control based on a linear, time-
However, the robot’s translation velocity vm has to be
varying description of the system was used for trajectory
controlled in order to achieve trajectory tracking. Thus, we
tracking control. Generalized predictive control (GPC) was
introduce the acceleration control input am , where am = v̇m
used to solve the path following problem in [19]. The non-
and we then obtain
linear predictive controller scheme for a trajectory tracking
η̇e = am − v̇d , (4) problem was proposed in [20, 21]. Recently, Falcone et al.
[22] implemented an MPC-based approach for active steering
where ηe = vm − vd . control design. They presented two approaches, i.e., MPC
Similar to [20], we redefine the control signals using a nonlinear vehicle model and MPC based on successive
    online linearization of the vehicle model. The differences of
u1 −ṡ + vm cos θe
this paper from other work are that this paper (i) deals with
ue = u2  =  ωm − ṡκ  . (5)
path following control, which can provide the optimal velocity
u3 am − v̇d
of a virtual vehicle to be followed along the path, (ii) achieves
The control input vector ue is used as the control input in smooth convergence to the reference with time constraints, and
our NMPC framework. When the open-loop optimal control (iii) takes into account obstacle avoidance.
A nonlinear system is normally described by the following
nonlinear differential equation:
ẋe (t) = f (xe (t), ue (t)),
(7)
subject to: xe (t) ∈ X , ue (t) ∈ U, ∀t ≥ 0
where xe (t) ∈ Rn , ue (t) ∈ Rm are the n dimensional state
vector and the m dimensional input vector of the system,
respectively. X ⊆ Rn and U ⊆ Rm denote the set of feasible
states and inputs of the system, respectively. In NMPC, the
input applied to the system is usually given by the solution
(a) (b) of the following finite horizon open-loop optimal control
problem, which is solved at every sampling instant:
Fig. 1. (a) A unicycle-type mobile robot (12 cm diameter) used in Z t+Tp
experiments and (b) a graphical representation of a unicycle mobile robot
and a reference path. min F (xe (τ ), ue (τ )) dτ + V (xe (t + Tp )), (8)
ue (·) t
 3

subject to: ẋe (τ ) = f (xe (τ ), ue (τ ))

Similar to [20], to have a negative derivative of the value
ue (τ ) ∈ U ∀τ ∈ [t, t + Tc ] function, the following conditions for the weight parameters
(9)
xe (τ ) ∈ X ∀τ ∈ [t, t + Tp ] are required:
xe (t + Tp ) ∈ Ω p11 α − q11 − α2 r11 ≥ q22
where F (xe , ue ) = xTe Qxe + uTe Rue . Tc and Tp are the p33 β − q33 − β 2 r22 ≥ 0 (17)
control horizon and the prediction horizon, respectively, with 2
p44 γ − q44 − γ r33 ≥ 0
Tc ≤ Tp . V (xe (t + Tp )) is the terminal penalty and Ω is
the terminal region. The deviation from the desired values is and the terminal-state region is defined as follows:
weighted by the positive definite matrices Q and R. Similar |xeT | ≥ |yeT |
to [20], a Lyapunov function for the terminal-state penalty can . (18)
p22 yeT vm θeT < 0
be selected as follows:
The terminal-state region is further bounded by the control
V (xe (t + Tp )) = 21 xe (t + Tp )T P xe (t + Tp ), (10) signal region. From (5) and (12), we have the following results
at the terminal state:
where P is a positive definite matrix, under the terminal-state  L    L L

controller uL u1 −αxeT −ṡ + vm cos θeT
e (t) such that the following condition is satisfied:
uL2
 =  −βθeT  =  ωmL
− ṡL κ  . (19)
V̇ (xe (t)) + F (t, xe (t), ue (t)) ≤ 0, ∀xe (t) ∈ Ω. (11) L L
u3 −γηeT am − v̇d
The terminal state feedback controller uL L L L T
e = [u1 , u2 , u3 ] Then, the system control inputs ṡL , ωm
L
, aL
m at the terminal
is defined as follows: state can be obtained by
uL
 L  
1 = −αxeT ṡ αxeT + vm cos θeT
uL
2 = −βθeT , (12)  aLm =
  βθeT + ṡκ  , (20)
uL
3 = −γηeT ωm L
−γηeT + v̇d
where α, β, γ > 0, and xe (t + Tp ) = [xeT , yeT , θeT , ηeT ]T . with the control input constraints
The subscript T denotes the terminal state. All weight param- 
ωmin
  L 
ωm ω

eters have to be selected such that (11) is satisfied. ≤ L ≤ max . (21)
amin am amax
The positive definite weight matrices of the F function in
(8) are selected as follows:
A. Time Parameterized Reference
Q = diag(q11 , q22 , q33 , q44 ), R = diag(r11 , r22 , r33 ) , In the trajectory tracking problem, the vehicle is required
(13) to track a time parameterized reference. We normally feed a
and the positive definite weight matrix of the V function in desired posture xd,t to a tracking controller. In this work, we
(10) is given by wish to combine trajectory tracking behaviors in a path fol-
lowing control law, thus achieving smooth spatial convergence
P = diag(p11 , p22 , p33 , p44 ) . (14)
to the trajectory as well as time convergence. We penalize the
Then, the stability condition becomes cost function with

V̇ (xe (t + Tp )) + F (t + Tp ) Ft = (xm (Tp ) − xd,t (Tp ))T Kt (xm (Tp ) − xd,t (Tp )), (22)
= p11 xeT ẋeT + p22 yeT ẏeT + p33 θeT θ̇eT + p44 ηeT η̇eT where Kt is a positive definite matrix. This matrix weighs
+ F (t + Tp ) the relative importance of convergence in time over spatial
convergence to the path. If Kt = 0 is chosen, pure path
= p11 xeT uL L L
1 + p22 yeT vm sin θeT + p33 θeT u2 + p44 ηeT u3
following is achieved.
+ F (t + Tp )
= p11 xeT uL L L
1 + p22 yeT vm sin θeT + p33 θeT u2 + p44 ηeT u3 B. Obstacle Avoidance
2
+ q11 x2eT + q22 yeT 2
+ q33 θeT 2
+ q44 ηeT Typically, the desired reference is generated by a plan-
2 2 2
+ r11 uL
1 + r22 uL
2 + r33 uL
3 .
ning algorithm based on a map of the environment and this
(15) reference is assumed to be collision-free. During the actual
motions it is possible that obstacles appear in the vehicle’s
Substituting the terminal state feedback controller (12) into path, which had not been present in the planning phase. This
(15), we get may also happen because of imprecision in the map, or vehicle
V̇ (xe (t + Tp )) + F (t + Tp ) =x2eT (−p11 α + q11 + α2 r11 ) localization errors. In this work, we assume that the simulated
2 sensors mimic infra-red sensors placed in a ring around the
+ θeT (−p33 β + q33 + β 2 r22 )
robot, spaced by 30◦ and they have a distance range of 50
2
+ ηeT (−p44 γ + q44 + γ 2 r33 ) cm. The obstacle information is then incorporated into the
2
+ p22 yeT vm sin θeT + q22 yeT . cost function, so that the computed control follows the desired
(16) reference, while staying away from the obstacles. In case of
 4

moving obstacles, the information such as their position and the terminal region in finite time with less effort than in case of
velocity can be used to predict the information over the next trajectory tracking control because the trajectory is the time-
Tp horizon and then the cost function can be computed. It has parameterized reference and the conditions in (18), depending
to be noted that we consider only convex polygonal obstacles. on time, need to be satisfied.
The obstacle points detected by sensors contribute to the Next, a convex polygonal obstacle was introduced in a
cost function with a term which penalizes states as follows position which prohibited path following. As it is shown in
Np Ns Figure 5, the controller deviated from the desired reference
X X e−c1 |θobs,ij | in order to safely avoid the obstacle and time convergence
Fobs = Kobs , (23)
i=1 j=1
ec2 dobs,ij could still be achieved. In Figure 6, two moving obstacles
were present. The velocity of the first obstacle was 0.2 m/s at
where Kobs , c1 , and c2 are positive constants. Ns is the number −135◦ , while the velocity of the second obstacle was 0.6 m/s
of range sensors. Np is the number of predictive steps, given at 150◦ . In the simulation results, the robot moved backward
by Np = Tp /δ, where δ is the sampling time. θobs is the angle to avoid the collision and waited until it could find a way to
of the obstacle with respect to the robot frame and dobs is the stay away from the obstacles and to follow the reference.
distance between the robot and the obstacle.

IV. S IMULATION R ESULTS V. R EAL -W ORLD E XPERIMENTS

Our NMPC controller was first implemented in Matlab and A unicycle-type mobile robot, shown in Figure 1(a) was
numerous simulations were performed. All the elements of our used in real-world experiments in this paper. The robot con-
NMPC framework were set as follows: troller is an ATMEGA644 microprocessor with 64 KB flash
Q = diag(0.2, 2, 0.01, 0.01), program memory, 16MHz clock frequency and 4 KB SRAM.
R = diag(0.0001, 0.0001, 0.0001), The robot orientation was measured by a Devantech CMPS03
P = diag(1, 1, 0.015, 0.015), compass. The localization was given by a camera looking
Kt = diag(1, 2, 0.01, 0.01), down upon the robot’s workplace and a PC was used to
Kobs = 1.5, c1 = 0.01, c2 = 10, compute the control inputs and then sent these inputs to the
Np = 3, Tc = Tp = 0.15 s, δ = 0.05 s, s(0) = 0 m, robot via WLAN. The same reference used in simulation
α = 2.5, β = 1, γ = 1. was employed in this experiment, but with vd = 0.2 m/s,
The circle reference is ωd = 0.2 rad/s and δ = 0.1 s. The free package DONLP2
[23] was used to solve the online optimization problem.
xd (t) = R cos s(t) s(t)
R , yd (t) = R sin R , However, incorporating the terminal penalty and the terminal-
where R = 1 m and the desired translation and rotation were state constraints degrades performance because of high com-
vd = 0.5 m/s and ωd = 0.5 rad/s, respectively. The maximum putational time. Therefore we did not include them in the
and minimum control inputs were set to vmax = 2 m/s, vmin = cost function in the real-world experiments. However, we still
−2 m/s, ωmax = 2 rad/s, ωmin = −2 rad/s, amax = 2 m/s2 and penalized the cost function with the time dependent penalty
amin = −2 m/s2 .
The performance achieved with pure path following, pure
trajectory tracking (see [20] for details), and for combined
trajectory tracking and path following was assessed. Fig-
ure 2(a) and Figure 2(b) show the simulation results of pure
path following control and pure trajectory tracking control,
respectively, with four different initial postures. The velocities
and the posture errors of pure path following are depicted in
Figure 3(a) and Figure 4(a), respectively, while those of pure
trajectory tracking are plotted in Figure 3(b) and Figure 4(b),
respectively when the initial posture of both cases was set to
(1.5, −0.5, π). Obviously, the control signals of path following
control were less likely pushed to saturation and motions were Fig. 5. The simulation results when a static polygonal obstacle was present.
less aggressive. However, time constraints were not achieved
in the path following control. Figure 2(c) shows the simulation
results of the combination of path following control and trajec-
tory tracking control, and the velocities and the posture errors
are shown in Figure 3(c) and Figure 4(c), respectively when
the initial posture was set to (1.5, −0.5, π). This controller can
achieve both reference convergence and time convergence with
smooth motions. As seen in the results, the robot converged
smoothly to the desired path and then it reacted to achieve
zero trajectory tracking error. Interesting enough, in case of Fig. 6. The simulation results when two moving polygonal obstacles were
path following control, the predicted system states can reach present.
 5

1 1 1

0.5 0.5 0.5

(0.5,0.5,0) (1.5,0.5,π) (0.5,0.5,0) (1.5,0.5,π) (0.5,0.5,0) (1.5,0.5,π)

y (m)

y (m)

y (m)
0 0 0

−0.5 −0.5 −0.5

(0.5,−0.5,0) (1.5,−0.5,π) (0.5,−0.5,0) (1.5,−0.5,π) (0.5,−0.5,0) (1.5,−0.5,π)

−1 −1 −1

−1 −0.5 0 0.5 1 1.5 2 −1 −0.5 0 0.5 1 1.5 2 −1 −0.5 0 0.5 1 1.5 2

x (m) x (m) x (m)

(a) (b) (c)

Fig. 2. The simulation results with four different initial postures: (a) pure path following, (b) pure trajectory tracking, and (c) the combination of path
following and trajectory tracking.

translation translation translation

2 2 2

1 1 1
(m/s)

(m/s)

(m/s)
0 0 0

−1 act. velocity −1 act. velocity act. velocity

−1
ref. velocity ref. velocity ref. velocity
−2 −2 −2
0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8
time(s) time(s) time(s)
rotation rotation rotation
2 2 2

1 1 1
(rad/s)

(rad/s)

(rad/s)
0 0 0

−1 act. velocity −1 act. velocity −1 act. velocity

ref. velocity ref. velocity ref. velocity
−2 −2 −2
0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8
time(s) time(s) time(s)

(a) (b) (c)

Fig. 3. The robot velocities when the initial posture was set to (1.5, −0.5, π): (a) pure path following, (b) pure trajectory tracking, and (c) the combination
of path following and trajectory tracking.

x error x error x error

2 2 2
(m)

(m)

(m)

0 0 0

−2 −2 −2
0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8
time(s) time(s) time(s)
y error y error y error
1 1 1
(m)

(m)

(m)

0 0 0

−1 −1 −1
0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8
time(s) time(s) time(s)
θ error θ error θ error
2 2 2
(rad)

(rad)

(rad)

1 1 1

0 0 0
0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8
time(s) time(s) time(s)

(a) (b) (c)

Fig. 4. The posture errors when the initial posture was set to (1.5, −0.5, π): (a) pure path following, (b) pure trajectory tracking, and (c) the combination
of path following and trajectory tracking.
 6

Translation velocity
x error 0.4
0.2

0.3

(m)
0

(m/s)
0.2
−0.2
0 10 20 30 40 50 60
actual position time (s) 0.1
1 desired position y error act. velocity ref. velocity
0.2
0.8 0
0 10 20 30 40 50 60
0.6 time (s)

(m)
0
0.4
Rotation velocity
0.2 0.6
−0.2
y (m)

0 10 20 30 40 50 60
0
time (s) 0.4
−0.2 θ error

(rad/s)
−0.4 0.2
0.2

(rad)
−0.6
0
−0.8
0
−0.2
act. velocity ref. velocity
−1 −0.4
0 10 20 30 40 50 60 −0.2
0 10 20 30 40 50 60
−1 −0.5 0 0.5 1 1.5 time (s)
time (s)
x (m)

(a) (b) (c)

Fig. 7. The experimental results by using our NMPC law: (a) the robot positions and its reference, (b) the posture errors, and (c) the robot’s velocities.

term in order to satisfy time constraints. Furthermore, obstacle [8] P. Encarnação and A. Pascoal, “3D path following for autonomous
avoidance will be considered in our future work because of underwater vehicle,” in Proc. of 39th IEEE Conference on Decision
and Control CDC2000, Sydney, Australia, Dec. 2000, pp. 2977-2982.
high computational time demand under real-time constraints. [9] D. Soeanto, L. Lapierre, and A. Pascoal, “Adaptive non-singular path-
The results are shown in Figure 7. following, control of dynamic wheeled robots,” in Proc. of International
Conference on Advanced Robotics, Coimbra, Portugal, June 30 - July 3,
2003, pp. 1387-1392.
VI. C ONCLUSIONS AND F UTURE W ORK [10] M. Egerstedt, X. Hu, and A. Stotsky, “Control of mobile platforms using
a virtual vehicle approach,” IEEE Trans. on Automatic Control, vol. 46,
In this paper, we presented a solution to the problem of no. 11, pp. 1777-1782, Nov. 2001.
[11] L. Lapierre, R. Zapata, and P. Lepinay, “Combined path-following and
combined trajectory tracking and path following for a mobile obstacle avoidance control of a wheeled robot,” International Journal
robot. Our approach based on the NMPC framework can of Robotics Research, vol. 26, no. 4, pp. 361-376, 2007.
control a mobile robot to smoothly converge to a reference [12] M. Aicardi, G. Casalino, A. Bicchi, and A. Balestrino, “Closed loop
steering of unicycle-like vehicles via Lyapunov techniques,” IEEE
with time and control input constraints. Robotics and Automation Magazine, pp. 27-35, March 1995.
However, the computation is one of the problems to use [13] H. Chen and F. Allgöwer, “A quasi-infinite horizon nonlinear model
NMPC in real-time systems. Improving the computation ef- predictive. control scheme with guaranteed stability,” Automatica, vol.
34, no. 10, pp. 1205-1218, 1998.
ficiency is still under our investigation. Since the initial [14] H. Michalska and D.Q. Mayne, “Robust receding horizon control of
feasibility to the optimization has been assumed in order constrained nonlinear systems,” IEEE Trans. on Automatic Control, vol.
that subsequent feasibility can be implied, feasibility analysis 38, no. 11, pp. 1623-1633, Nov. 1993.
[15] D. Q. Mayne, J. B. Rawlings, C. V. Rao, and P. O. M. Scokaert, “Con-
is one of our further research. In addition, we will extend strained model predictive control: stability and optimality,” Automatica,
our controller to accomplish coordination tasks with a larger vol. 36, pp. 789-814, 2000.
number of mobile robots in a complex environment. [16] A. Jadbabaie, J. Yu, and J. Hauser, “Stabilization receding horizon
control of nonlinear systems: A control Lyapunov function approach,”
in Proc. of American Control Conference, San Diego, CA, 1999, pp.
1535-1539.
R EFERENCES [17] W. F. Lages and J. A. V. Alves, “Real-time control of a mobile
robot using linearized model predictive control,” in Proc. of 4th IFAC
[1] P. Morin and C. Samson, “Motion control of wheeled mobile robot,” in
Symposium on Mechatronic Systems, Heidelberg, Germany, Sep. 2006,
Springer Handbook of Robotics (B. Siciliano and O. Khatib, eds.), pp.
pp. 968-973.
799-826, Springer Berlin Heidelberg, 2008.
[18] G. Klančar and I. Škrjanc, “Tracking-error model-based predictive con-
[2] R. W. Brockett, “Asymptotic stability and feedback stabilization,” in
trol for mobile robots in real time,” Robotics and Autonomous Systems,
Differential Geometric Control Theory (R. W. Brockett, R. S. Millman,
vol. 55, no. 6, pp. 460-469, 2007.
and H. J. Sussmann, eds.), pp. 181-191, Birkhauser, Boston, 1983.
[19] A. Ollero and O. Amidi, “Predictive path tracking of mobile robots.
[3] A. Micaelli and C. Samson, “Trajectory-tracking for unicycle-type and Application to the CMU Navlab,” in Proc. of International Conference
two-steering-wheels mobile robots,” Technical Report No. 2097, INRIA, on Advanced Robotics, Pisa, Italy, Jun. 1991, pp. 1081-1086.
Sophia-Antipolis, Nov. 1993. [20] D. Gu and H. Hu, “Receding horizon tracking control of wheeled mobile
[4] C. Canudas de Wit, C. Samson, H. Khennouf, and O. J. Sørdalen, robots,” IEEE Transactions on Control Systems Technology, vol. 14, no.
“Nonlinear control design for mobile robots,” in Recent trends in 4, pp. 743-749, Jul. 2006.
mobile robots (Y. F. Zheng, ed.), vol. 11, pp. 121-156, World Scientific [21] R. Hedjar, R. Toumi, P. Boucher, and D. Dumur, “Finite horizon
Publishing, 1993. nonlinear predictive control by the Taylor approximation: application
[5] C. Samson, “Control of chained systems: Application to path-following to robot tracking trajectory,” Int. Journal of Applied Mathematics and
and time-varying point stabilization of mobile robots,” IEEE Trans. on Computer Science, vol. 15, no. 4, pp. 527-540, 2005.
Automatic Control, vol. 40, no. 1, pp. 64-77, Jan. 1995. [22] P Falcone, F. Borrelli, J. Asgari, H. E. Tseng, and D. Hrovat, “Predictive
[6] C. Altafini, “Following a path of varying curvature as an output active steering control for autonomous vehicle systems,” IEEE Trans. on
regulation problem,” IEEE Trans. on Automatic Control, vol. 47, no. Control Systems Technology, vol. 15, no. 3, pp. 566-580, May 2007.
9, pp. 1551-1556, Sep. 2002. [23] P. Spellucci, “An SQP method for general nonlinear programs using only
[7] S. A. Al-Hiddabi and N. H. McClamroch, “Tracking and maneuver reg- equality constrained subproblems,” Mathematical Programming, vol. 82,
ulation control for nonlinear non-minimum phase systems: application no. 3, pp. 413-448, 1998.
to flight control,” IEEE Trans. on Control Systems Technology, vol. 10,
no. 6, pp. 780-792, 2002.

View publication stats