Lionel Lapierre

Rene Zapata
Combined
Pascal Lepinay Path-following and
Université Montpellier II
161, rue Ada 34392 Montpellier Cedex 5 Obstacle Avoidance
{lapierre, zapata, lepinay}@lirmm.fr
Control of a Wheeled
Robot

Abstract 2 path-following control of nonholonomic systems,

2 obstacle avoidance strategy,

This paper proposes an algorithm that drives a unicycle type robot
to a desired path, including obstacle avoidance capabilities. The 2 coupling between the above two areas.
path-following control design relies on Lyapunov theory, backstep-
ping techniques and deals explicitly with vehicle dynamics. Further-
more, it overcomes the initial condition constraints present in a num-
1.1. Path-following control of a nonholonomic system
ber of path-following control strategies described in the literature.
This is done by controlling explicitly the rate of progression of a “vir-
The general underlying assumption in path-following control
tual target” to be tracked along the path1 thus bypassing the problems
is that the vehicle’s forward velocity tracks a desired speed
that arise when the position of the path target point is simply defined
profile, while the controller acts on the vehicle orientation to
as the closest point on the path. The obstacle avoidance part uses
drive it to the path. See Micaeli and Samson (1993), and Sam-
the Deformable Virtual Zone (DVZ) principle. This principle defines
son and Ait-Abderrahim (1991) for pioneering work in the area
a safety zone around the vehicle in which the presence of an obstacle
as well as Canudas de Wit et al. (1993), Jiang and Nimeijer
induces an “intrusion of information” that drives the vehicle reac-
(1999), and Soetanto et al. (2003) and references therein. Path-
tion. The overall algorithm is combined with a guidance solution that
following systems for marine vehicles have been reported by
embeds the path-following requirements in a desired intrusion infor-
Encarnação et al. (2000), Encarnação and Pascoal (2000), and
mation function, which steers the vehicle to the desired path while the
Lapierre et al. (2003). The main contribution of the method in
DVZ ensures minimal contact with the obstacle, implicitly bypassing
Lapierre et al. (2003) is threefold:
it. Simulation and experimental results illustrate the performance of i) it extends the results obtained by Micaeli and Samson
the control system proposed. (1993) and Soetanto et al. (2003) – for kinematic wheeled
KEY WORDS—nonlinear control, underwater robotics, un- robots – to a more general setting, in order to deal with vehicle
deractuated systems, path-following dynamics and parameter uncertainty,
ii) it overcomes stringent initial condition constraints that
are present in a number of path-following control strategies in
1. Introduction the literature. This is done by controlling explicitly the rate
of progression of a virtual target to be tracked along the path,
Real-time obstacle avoidance coupled with accurate path- fol- thus bypassing the problems that arise when the position of the
lowing control is one of the major issues in the field of mobile target point on the path is simply defined by the projection of
robotics (Ogren 1989, 2003). The underlying problem to be the actual vehicle on the path. This procedure avoids the sin-
solved can be divided into three areas: gularities that occur when the vehicle is located at the current
center of curvature of the path virtual target location (where
The International Journal of Robotics Research the closest point is not unique), and allows for global conver-
Vol. 26, No. 4, April 2007, pp. 361–375
DOI: 10.1177/0278364907076790
gence of the vehicle to the desired path. The path is a general
1
c 2007 SAGE Publications curve, described by its curvature as a function of the curvilin-
Figures 4–7 appear in color online: http://ijr.sagepub.com ear abscissa1 the singularity occurs when the robot is located at

361
362 THE INTERNATIONAL JOURNAL OF ROBOTICS RESEARCH / April 2007

the centre of the path curvature where the virtual target is lo- which repels the robot, and an attractive field located on the
cated. This is in contrast with the results described by Micaeli goal. The main difficulty with this method is the design of
and Samson (1993) for example, where only local convergence an artificial potential function without undesired local min-
is proven. To the best of the authors’ knowledge, the idea of ima. Elnagar and Hussein (2002) propose to model the poten-
exploring the extra degree of freedom that comes from control- tial field using Maxwell’s equations that completely eliminate
ling the motion of a virtual target along a path appeared for the the local minima problem, with the condition that an a pri-
first time in the work of Aicardi et al. (1995) for the control ori knowledge of the environment is available. These methods
of wheeled robots. This idea was later extended to the control are generally computationally intensive. Iniguez and Rossel
of marine craft in the work of Aicardi et al. (2001). However, (2002) propose a hierarchical and dynamic method that works
none of these addresses the issues of vehicle dynamics and pa- on a no regular grid decomposition, simple and computation-
rameter uncertainty. Furthermore, the methodologies adopted ally efficient, both in time and memory. Ge and Cui (2000)
by Aicardi et al. (1995, 2001) for control system design build address the problem of the adaptation of the potential field
on entirely different techniques that required the introduction magnitude, in order to avoid the problematic situation where
of a nonsingular transformation in the original error space. the goal is too close to an obstacle, inducing an inefficient
iii) It combines path-following control and obstacle avoid- attractive field, termed Goals Non-reachable with Obstacles
ance. Nearby. The work of Louste (1999) tackles the problem of
In this paper, controller design builds on the work reported coupling a path-planning method, based on viscous fluid prop-
in Micaeli and Samson (1993) on path-following control and agation, with nonholonomic robot kinematics restrictions. An-
relies heavily on Lyapunov-based and backstepping techniques other approach, based on a reflex behavior reaction, uses the
(Krstić et al. 19951 Fierro and Lewis 1997) in order to extend Deformable Virtual Zone (DVZ) concept, in which a robot
kinematic control laws to a dynamic setting. In Soetanto et kinematic dependent risk zone is located surrounding the ro-
al. (2003) parameter uncertainty is considered within a frame- bot. Deformation of this zone is due to the intrusion of prox-
work by augmenting Lyapunov function candidates with terms imity information. The system reaction is made in order to
that are quadratic in the parameter error. For the sake of clar- reform the risk zone to its nominal shape, implicitly moving
ity, in this paper, we do not extend the dynamic control with away from obstacles (Zapata 2004).
the parameter uncertainty robust scheme.

1.3. Coupling Obstacle Avoidance and Path Following

1.2. Obstacle Avoidance
Two different methods are generally proposed in the literature
The obstacle avoidance strategy is another major issue affect- for obstacle avoidance algorithms, coupled with a nominal ob-
ing applications in the field of mobile robotics, and comprises jective, defined as path following, trajectory tracking or point
two different issues: stabilization:
2 obstacle detection, 2 path replanning,
2 computation of the system reaction. 2 reactive path following.
Obstacle detection requires the system to be equipped with The path replanning strategy requires the system to design
sensor devices able to estimate the distance between the robot a safe path between obstacles that ensures the vehicle avoids
and the obstacles. Two different types of sensors are generally obstacles and finally reaches the desired goal. The advantage
used: of this method is that, under the assumption that an accurate
map of the environment is available, it designs a global solu-
2 A belt of proximity sensors (ultrasound, infrared, ...) tion that guarantees the system reaches the goal, if this solu-
mounted on the vehicle allowing the acquisition of a dis- tion exists. Moreover, since the system is necessarily equipped
crete estimation of the free space around the robot. with proximity sensors, the requirements for an environmen-
2 A rotating laser beam coupled with a vision system, or tal map construction are met. This allows the use of SLAM
a stereo vision system, permitting continuous estimation (Simultaneous Localization and Mapping) or CML (Concur-
of the polar signature of the free space region around the rent Localization and Mapping) techniques for accurate nav-
vehicle. igation systems, in which the path replanning method natu-
rally takes place. The drawback is the necessary computational
System reaction can be at the navigation system level (path time, which may cause the system to stop in front of an un-
replanning) or directly in the controller as a reflex behavior. charted obstacle. Finally the control objective is to follow the
Reaction quantification is generally made according to an ar- computed path (Elnagar and Hussein 20021 Iniguez and Rossel
bitrary positive potential field function attached to obstacles, (2002)1 Rimon and Koditschek 1992).
 Lapierre, Zapata, and Lepinay / Combined Path-following and Obstacle Avoidance Control of a Wheeled Robot 363

Reactive path following, which acts directly at the control

level, is generally considered as a behavior-based approach,
and includes fuzzy or neural systems, Althaus and Chris-
tensen (2002) and Arkin (1998). Several behaviors are defined
(avoid-obstacle, move to goal, ...) and a switching scheme is
designed to switch between these behaviors.

1.4. Outline of this paper

We have investigated the coupling between the path-following

algorithm described in Soetanto et al. (2003) and a DVZ-based
reactive obstacle avoidance control. The proposed method
consists in a guidance solution that embeds the path-following Fig. 1. Path following: frames definition and problem pose
requirements in a desired proximity function (with respect to description.
the obstacles) that drives the robot to contour the obstacles
while guaranteeing the path-following convergence require-
ments when there is no obstacle. This approach is based on the
derivation of a Lyapunov function that guarantees asymptotic wheels. The wheels control provides a forward force F and
convergence to the path without obstacles, and the bounded- an angular torque N applied at the vehicle’s center of mass.
ness of a variable called the intrusion ratio, that captures the The vehicle mass and moment of inertia are denoted m and
surrounding obstacles proximity and the current robot situa- I , respectively. Let u and r denote the forward and rotational
tion with respect to the path. The combination of path follow- velocity of the vehicle.
ing with a reactive – local – obstacle avoidance strategy has
a natural limitation in the situation that both controllers yield
antagonistic system reactions. In the proposed method, this 2.1.1. Kinematic equations of motion. The Serret–Frenet
situation leads to a local minimum called the corner situation, Frame
where a heuristic switch between controllers is necessary. The
advantage of the method is that outside this very specific cor- The solution to the problem of path following derived in Mi-
ner situation, no switch is required. caelli and Samson (1995) admits an intuitive explanation: a
The paper is organized as follows: Sections 2 and 3 review path-following controller should look at (i) the distance from
the basic algorithms we use for path following and obstacle the vehicle to the path and (ii) the angle between the vehicle
avoidance. Section 4 shows our solution combining both con- velocity vector and the tangent to the path, and reduce both to
trollers and Section 5 illustrates the results obtained. Finally, zero. This motivates the development of the kinematic model
Section 6 presents conclusions and suggests further work. of the vehicle in terms of a Serret–Frenet frame 3F4 that moves
along the path, with abscissa denoted s1 and ordinate y1 1 3F4
plays the role of the body axis of a “virtual target vehicle” that
2. Path-following Algorithm should be tracked by the “real vehicle”. Using this set-up, the
abovementioned distance and angle become the coordinates of
2.1. Path Following. Problem Formulation the error space in which the control problem is formulated and
solved. In this paper, motivated by the work in Micaelli and
This section reviews a solution to the problem of steering a Samson (1995), a Frenet frame 3F4 that moves along the path
unicycle type vehicle along a desired path. to be followed is used with a significant difference: the Frenet
The following assumptions are made regarding the robot frame is not attached to the point on the path that is closest
(see Figure 1). The vehicle, of width 2L, has two identical par- to the vehicle. Instead, the origin of 3F4 along the path is
allel, non-deformable rear wheels of radius R, which are con- made to evolve according to a conveniently defined function
trolled by two independent motors, and a passive front wheel. of time, effectively yielding an extra controller design parame-
It is assumed that the plane of each wheel is perpendicular to ter. As will be seen, this seemingly simple procedure allows
the ground and that the contact between the wheels and the one to lift the stringent initial condition constraints that arise
ground is pure rolling and non-slipping, i.e. the velocity of the with the path-following controller described in Micaelli and
center of mass of the robot is orthogonal to the rear wheels Samson (1993). The notation that follows is now standard.
axis. It is further assumed that the masses and inertias of the Consider Figure 1, where P is an arbitrary point on the
wheels are negligible and that the center of mass of the mobile path to be followed and Q is the center of mass of the mov-
robot is located in the middle of the axis connecting the rear ing vehicle. Associated with P, consider the corresponding
364 THE INTERNATIONAL JOURNAL OF ROBOTICS RESEARCH / April 2007

1 4 1 4
Serret–Frenet frame 3F4. The signed curvilinear abscissa of P 0 s1
2 5 2
along the path is denoted s. Clearly, Q can either be expressed 2 5 2 y 5
as q = 1X2 Y2 03 in a selected inertial reference frame 3I4 or as 1c 9 r 5 2 0 593 1 5
6
3 6
1s1 2 y1 2 03 in 3F4. Stated equivalently, Q can be given in 1X2 Y 3 74 5 cc 1s37s 0
or 1s1 2 y1 3 coordinates (see Figure 1). Let the position of point
P in 3I4 be vector p. Let 1 4
6cc 1s37s y1
1 4 2 5
2 5
cos 4 c sin 4 c 0 5 2 cc 1s37s s1 52
2 5 3 6
2
R 5 3 6 sin 4 c cos 4 c 0 5 0
6
0 0 1
equation (2) can be rewritten as
be the rotation matrix from 3I4 to 3F4, parameterized locally by 1 4 1 4
X7 s7 11 6 cc 1s3y1 3 8 s71
the angle 4 c . Define 1 c 5 47 c . Then, 2 5 2 5
2 5 2 5
7 R 2 Y7 5 5 2 y71 8 cc 1s37s s1 55
3 6 3 6
8 1c 5 47 c 5 cc 1s37s
(1) 0 0
9 c7c 1s3 5 gc 1s37s
Solving for s71 and y71 yields
where cc 1s3 and gc 1s3 5 dcdsc 1s3
denote the path curvature and 7 1 4
its derivative, respectively. The velocity of P with respect to
 X7
s71 5 cos 4 c sin 4 c 3 6 6 s7 11 6 cc y1 3
3I4 can be expressed in 3F4 to yield 8 Y7
1 4 1 4 (2)
s7  X7
dp 2 5 y71 5 6 sin 4 c cos 4 c 3 6 6 cc s7 s1 5
52 5
3 0 65
9
Y7
dt F
0
At this point it is important to notice that in Micaelli and Sam-
It is also straightforward to compute the velocity of Q in 3I4 as son (1993), s1 5 0 for all t, since the location of point P is
defined by the projection of Q on the path, assuming the pro-
dq dp dr jection is well defined. One is then forced to solve for s7 in the
5 8 R61 8 R61 11c 9 r3 equation above. However, by doing so 1 6 cc y1 appears in the
dt dt dt
denominator, thus creating a singularity at y1 5 c1c . As a result,
I I F

where r is the vector from P to Q. Multiplying the above equa- the control law derived in Micaelli and Samson (1993) requires
tion on the left by R gives the velocity of Q in 3I4 expressed in that the initial position of Q be restricted to a tube around
1
3F4 as the path, the radius of which must be less than cc2max , where
dq dp dr
cc2max denotes the maximum curvature of the path. Clearly,
R 5 8 8 1c 9 r this constraint is very conservative since the occurrence of a
dt I dt F dt F
large cc2max in only a very small section of the path will im-
Using the relations pose a rather strict constraint on the initial vehicle position, no
matter where it starts with respect to that path.
1 4
X7 By making s1 not necessarily equal to zero, a virtual target
2 5 that is not coincident with the projection of the vehicle on the
dq 2 5
5 2 Y7 5 2 path is created, thus introducing an extra degree of freedom for
dt I 3 6
controller design. By specifying how fast the newly defined
0 target moves, the occurrence of a singularity at y1 5 c1c is
1 4 removed. The velocity of the unicycle in the 3I4 frame satisfies
s71 the equation
dr 2 5
5 2 5
3 y71 6 2 1 4 1 4
dt F X7 cos 4 m
0 3 65u3 6 (3)
Y7 sin 4 m
and
 Lapierre, Zapata, and Lepinay / Combined Path-following and Obstacle Avoidance Control of a Wheeled Robot 365

where 4 m and u denote the yaw angle of the vehicle and its A.1. 91y1 2 u3 is a bounded differentiable function with respect
body-axis speed, respectively. Substituting (3) in (2) and in- to y1 and 9102 u3 5 0. This function will be called in the
troducing the variable 4 5 4 m 6 4 c gives the kinematic model sequel ‘approach angle’.
of the unicycle in 1s2 y3 coordinates as
7 A.2. y1 u sin 91y1 2 u3 02 y u5
s71 5 67s 11 6 cc y1 3 8 u cos 4
8 A.3. lim t  u1t3 5 05
y71 5 6cc s7 s1 8 u sin 4 (4)
In the V1 Lyapunov function adopted, the first term 12 1s12 8
9 47 5 6 6 c s7 y12 3captures the distance between the vehicle and the path,
m c
which must be reduced to 0. The second term aims to shape
where 1m = 47 m . the approach angle 4 5 4 m 6 4 c of the vehicle to the path as a
function of the ‘lateral’ distance y1 and speed u, by forcing it
to follow a desired orientation profile embodied in the function
2.1.2. Dynamics. Problem Formulation
9. See Samson and Ait-Abderrahim (1991), where the use of a
The dynamical model of the unicycle is obtained by augment- 9 function of this kind was first proposed.
ing (4) with the equations Assumption A515 specifies that the desired relative heading
7 vanishes as y1 goes to zero, thus imposing the condition that
8 u7 5 F the vehicle’s main axis must be tangential to the path when the
1
(5) lateral distance y1 is 0. Assumption A525 provides an adequate
9 67 5 67 6 c s 6 g s7 2
m c c reference sign definition in order to drive the vehicle to the
path (turn left when the vehicle is on the right side of the path,
where 67 m 5 N2 and 1 and 2 are the mass and the moment and turn right in the other situation). Finally, Assumption A535
of inertia of the unicycle, respectively. Let FP F and N P F be states that the vehicle does not tend to a state of rest. The need
the path-following control inputs, the forward force and the for these conditions will become apparent in the development
angular torque, respectively. that follows. The derivative of V1 can be easily computed to
With the above notation, the problem under study can be give
formulated as follows:
1
Given a desired speed profile u d 1t3 7 u min 7 0 for the vehicle V71 5 s1 s71 8 y1 y71 8 14 6 93147 6 93
7
speed u, derive a feedback control law for FP F and N P F to 8
drive y1 , 4 , and u 6 u d asymptotically to zero. 5 s1 1u cos 4 6 s7 11 6 cc y1 3 6 s7 cc y1 3
1
2.2. Nonlinear Controller Design 8 y1 u sin 4 8 14 6 93147 6 97 3
8
This section introduces a nonlinear closed loop control law to
1
steer the dynamic model of a wheeled robot described by (4) 5 s1 1u cos 4 6 s7 3 8 y1 u sin 9 8
and (5) along a desired path. Controller design builds on pre- 8
vious work by Micaeli and samson (1993) on path-following sin 4 6 sin 9
control and relies heavily on backstepping techniques.  14 6 93 47 6 97 8 8 y1 u 5
4 69
Let the ideal (also called virtual) “kinematic control laws” for
2.2.1. Kinematic Controller design
s and 4 be defined as
The analysis that follows is inspired by the work in Samson 7
s7 5 u cos 4 8 k1 s1 
and Ait-Abderrahim (1991) and Micaelli and Samson (1993) 8
on path-following control for kinematic models of wheeled ro- 47 5 97 6 8 y1 u sin 44 69
6sin 9 (7)
bots. Recall from the problem definition in the previous sec-
9 6k 14 6 935
tion that the main objective of the path-following control law 2
is to drive y1 and 4 to zero. Starting at the kinematic level,
these objectives can be embodied in the Lyapunov function where k1 and k2 are positive gains. Then,
candidate (Micaelli and Samson 1993). 14 6 932
V71 5 6k1 s12 8 y1 u sin 9 6 k2 05 (8)
1 2 1 8
V1 5 1s1 8 y12 3 8 14 6 91y1 2 u332 (6)
2 28
Note the presence of the term y1 u sin 9 in the previous equa-
where 8 is a positive gain, and it is assumed that tion and how assumption A525 is justified.
366 THE INTERNATIONAL JOURNAL OF ROBOTICS RESEARCH / April 2007

2.2.2. Backstepping the Dynamics where

7
The above feedback control law applies to the kinematic model 8 f 1 13 5 7 6 1 14 6 93 8 cc s 8 gc s7 2
8
of the wheeled robot only. In what follows, using backstepping 9 f 13 5 u7 5
techniques, that control law is extended to deal with the vehicle 2 d

dynamics. Notice how in the kinematic design the speed of the

The above control law makes V72 negative semi-definite. This
robot u1t3 was assumed to follow a desired speed profile, say
fact plays an important role in the proof of convergence of the
u d 1t3. In the dynamic design this assumption is dropped, and
robot to the path.
a feedback control law must be designed so that the tracking
error u1t3 6 u d 1t3 approaches zero. Notice also that the robot’s
angular speed 6m was assumed to be a control input. This 2.2.3. Choice of the Approach Angle 91y1 2 u3
assumption will be lifted by taking into account the vehicle
dynamics. Following Krstić et al. (1995) define the virtual As explained before, the choice of the 91y1 2 u3 approach an-
control law for 47 (desired behaviour of 47 in (7)) as gle is instrumental in shaping the transient maneuvers during
sin 4 6 sin 9 the path approach phase. Indeed, the approach angle is ex-
5 97 6 8 y1 u 6 k2 14 6 93 (9) pressing the desired angle to approach the path. In Micaelli
4 69
and Samsion (1993), the authors propose to use 91y1 2 u3 5
and let 5 47 6 be the difference between actual and desired 6sign1u3 tanh1y1 3. This choice is natural in the sense that a
values of 47 . Replacing 47 by 8 in the computation of V71 large positive lateral distance between the robot and the path,
gives will imply the desired relative heading robot/path to be  2,
that is the approach angle. As the robot approaches the path,
14 6 932
V71 5 6k1 s12 8 y1 u sin 9 6 k2 and y1 diminishes, the approach angle decreases as well. The
8 previous choice is intuitively justified, however, it raises some
14 6 93 subtle mathematical difficulties because 91y1 2 u3 is not dif-
8 (10) ferentiable with respect to u at u 5 0. Another choice is
8 possible, 91y1 2 u3 5 6 tanh1y1 u3 for instance. This com-
Augment now the candidate Lyapunov function V1 with the plicates the control derivation, and reduces the system per-
terms 2 2 and 1u 6 u d 32 2 to obtain formances in terms of convergence time. In this study, we
avoid this problematic situation by imposing a forward veloc-
1 
V2 5 V1 8 2
8 1u 6 u d 32 (11) ity u d 7 u min 7 0 (cf. problem statement of Section 2.1.1, and
2 the implicit respect of the assumption A.3). As we will see
with derivative later, this condition will also be justified in the next section on
1 obstacle avoidance (Section 3).
V72 5 6k1 s12 8 y1 u sin 9 6 14 6 932
8

1 2.2.4. Control Expression

8 14 6 93 8 7 8 1u 6 u d 31u7 6 u7 d 35
8 We now state the proposed solution for the problem exposed
Simple computations show that if in Section 2.1.2.
7
8 7 5 6 1 14 6 93 6 k3 Proposition 1: Consider the kinematic and dynamic models
8
(12) described in (4) and (5). Let the approach angle 91y1 2 u3 be
9 u7 5 u7 d 6 k4 1u 6 u d 32
91y1 2 u3 5 6sign1u34 a tanh1k9 y1 3 (15)
where k3 and k4 are positive gains. Then
where 0  4 a   2 defines the asymptotic desired approach,
1 with k9 an arbitrary positive gain. Assume that u d 1t3 7 u min 7
V72 5 6k1 s12 8 y1 u sin 9 6 14 6 932
8 0 is a C 2 function. Suppose the path to be followed is parame-
terized by its curvilinear abscissa s and assume that for each
6 k3 2
6 k4 1u 6 u d 32 05 (13) s the variables 4 , s1 , y1 , cc and gc are well defined (cf. the
Remarks section below). Then, the dynamic control law
It is now straightforward to compute the control inputs F and
N by solving the dynamics equation (5) to obtain 7
NP F 5 2 1 f 1 13 6 k3 3
7 8
8 N 5 2 1 f 1 13 6 k3 3 FP F 5 1 1 f 2 13 6 k4 1 6  d 33 (16)
(14) 9
9 F 5 1 1 f 2 13 6 k4 1u 6 u d 33 s7 5 u cos 4 8 k1 s1
 Lapierre, Zapata, and Lepinay / Combined Path-following and Obstacle Avoidance Control of a Wheeled Robot 367

where the f 1 153 and f 2 153 are defines in equation (14) and k1 , as a DVZ surrounding the vehicle. Deformation of this risk
k2 , k3 , k4 and k9 are arbitrary positive gains, drives y1 , s1 and zone  is due to the intrusion of proximity information and
4 asymptotically to zero. thus controls the robot reactions. This DVZ characterizes the
deformable zone geometry, and depends on the robot veloci-
Proof: The key steps in the proof can be briefly described as
ties (forward and rotational velocities, u and r). Briefly, the
follows: let V2 be the Lyapunov function candidate expressed
risk zone, disturbed by obstacle intrusion, can be reformed by
in equation (11). The kinematic control expression (16) yields
acting on the robot velocities. For a complete exposition of the
V72 0. Since V72 is negative semi-definite and bounded be-
DVZ principle, the reader is referred to Zapata (2004).
low, V2 is bounded and has a well defined limit. Therefore,
s1 , y1 , 4 and u are bounded, since 9 and u d are assumed to be
bounded. Considering the previous equations, it is straightfor- 3.1. DVZ Principle
ward to show that s71 , y71 , 47 and u7 are bounded as well. Direct
derivation allows one to compute the second derivative of V2 , General statement
and prove it is bounded. Therefore, V72 is uniformly continuous A rigid body R representing a controlled robot is moving
and application of Barbalat’s lemma implies that V72 tends to 0 among obstacles in 12 . The 2-dimensional vector U 5 [u r]T
as t tends to . Consider now the expression for V72 in (13). is the generalized momenta consisting of the translational and
Since V72 is a sum of negative terms and tends to 0 as t goes to rotational velocities of the robot. It characterizes the motion
, we conclude that s1 , y1 , 14 6 93, , and 1u 6 u d 3 tend to of R. Furthermore, we will assume that the robot R can be
zero as well. That is, the robot asymptotically converges to the controlled by the derivative  5 U7 of this vector.
path. Informally, we define a controlled DVZ, h , as any DVZ
Remarks: The proposed solution consists in a ‘fox–rabbit’ which depends on the vector U characterizing the motion of
problem where the rabbit is really cooperative since it adjusts R. This functional dependence will appear more clearly in
its own velocity to help the convergence, through the expres- Section 3.2, after parameterization of the DVZ:
sion: h 5 1U 35 (17)
s7 5 u cos 4 8 k1 s1 5
If we let 3 5 1h 2 3 be an ordered pair of two DVZ of
The first term u cos 4 is the projection of the vehicle (fox) onto
R (the first one being a controlled DVZ), we define the defor-
the path where the virtual target (rabbit) is located. The sec-
mation  of the DVZ h with respect to  as the functional
ond term is necessary to ensure the convergence of s1 to 0,
difference of  and h :
making the virtual target asymptotically converge to the clos-
est point with respect to the vehicle current position. In the  5  6 h 5 (18)
problem statement, the condition u d u min 7 0 has been im-
posed (implicitly covering the assumption A.3). Then the ve- We assume that the robot can perceive distances in all di-
hicle will not stay at rest and will effectively converge to the rections of space. We also assume that the set of maximum
path. The consequence of this is that the virtual target also distances that can be perceived by R and the set of actually
will not stay at rest, since 4 and s1 are guaranteed to converge perceived distances are also two DVZ surrounding R respec-
to zero, and u 7 0, for all t. Then, the path-following prob- tively named Sensor boundary and Information boundary (and
lem, as described here can never degenerate to a pose regula- respectively denoted  and ). The deformation I of  with
tion problem, for which Brockett’s limitations are inevitable. respect to  is given by I 5  6 .
The assumption that the variables 4 , s1 , y1 , cc and gc are well Let  I , be a DVZ which depends on the Sensor boundary
defined implies that there is a path parameterization that allows deformation I :
one to compute the path curvature cc 1s3 and its spatial deriv-  I 5 1I 35 (19)
ative gc 1s3, for all s denoting the current curvilinear abscissa
The deformation  of the DVZ h with respect to  I can
of the virtual target. Moreover, this implies that the system is
be written:
equipped with a sensor suite that allows computation of the ro-
bot situation in the Serret–Frenet frame, i.e. the variables 4 , s1  5  I 6 h 5 1I 3 6 1U 35 (20)
and y1 .
For a given point M on R, the deformation vector 1M3
depends on the intrusion of proximity information I 1M3, in
3. Obstacle Avoidance Algorithm the rigid body workspace, and on the controlled DVZ h .
Figure 2 shows these different DVZ.
This section presents an obstacle avoidance algorithm based Parameterization
on the use of a continuous Deformable Virtual Zone (DVZ). Equation (20) is a functional equation that has to be parame-
The main idea is to define the robot/environment interaction terized in order to be differentiable and to lead to the “usable”
368 THE INTERNATIONAL JOURNAL OF ROBOTICS RESEARCH / April 2007

Fig. 2. DVZ principle: the different DVZ for collision avoid-

ance.

equation (22). For instance, it can be spatially sampled along Fig. 3. Implementation of the function  (and projection).
the directions of the real sensors. In this case, it becomes an s-
dimensional vectorial equation where s is the number of prox-
imity sensors. It can also be evaluated as a polar signature and 3.2. The Controlled DVZ
therefore parameterized. In the case of simple mobile robots,
the DVZ can be parameterized by simple shapes like ellipsoids In order to acquire an analytical expression for the polar sig-
(see 3.2). nature of the controlled DVZ, expressed in the robot frame
centered on R, we consider an elliptic shape. Let P 5 [x y]T
Differentiation be a point on the ellipse with axis cx and c y . If we assume
By differentiating Equation (20) with respect to time, we get: that the proper reference frame of the DVZ is translated from
its center by a vector a 5 [ax a y ]T , its equation in the robot
7 5 6 U []  8  I [] 
 (21) frame is given by:
where  U and  I are the differentiation operators with respect x 8 ax 2
y 8 ay 2

to the vectorial variables U and I . We note  5 U7 and  5 I7. 8 5 15 (24)

cx cy
This equation can be rewritten as:
The coefficients ax , a y , cx and c y are chosen heuristically
7 5 A 8 B5
 (22) and depend on the first component of the control vector, i.e.
the general momentum u, the translational velocity of the ro-
Variations in  are controlled by a 2-fold input vector bot. We have chosen:
u 5 [ ]T . The first control vector , due to the robot con-
troller, tends to minimize deformation of the DVZ. The second cx 5 cx u 2 8 cmin
x
one,  5 I7, is unknown and induced by the environment it- 
self, the relative motion of obstacles with respect to the robot 5
cy 5 cx
and their shapes. It can be seen as an uncontrolled input. 3
In the rest of the paper, we will assume that the function 
ax 5 612 33cx
relating the intrusion of information to the deformed DVZ  I
is defined as follows: ay 5 05 (25)
7
8  8 I i f  8 I  h In the direction  with respect to the main axis of the el-
1I 3 5 (23) lipse, let dh 13 be the length of the controlled DVZ h , i.e.
9 0 otherise 6
the norm of the vector R P. We can write:
where the sign  means the comparison of the sensor infor-   1 41 4
6 x cos14 m 3 sin14 m 3 X
mation ( 5  8 I ) and the undeformed DVZ h in each RP 5 53 63 6 (26)
direction of measure (see Figure 3). This restriction does not y 6 sin14 m 3 cos14 m 3 Y
change the principle of the DVZ and is just a simplification
leading to confusion of the intrusion I and the deformation  where [X Y ]T are the coordinates of point P in the absolute
(their derivatives are now identical). frame.
 Lapierre, Zapata, and Lepinay / Combined Path-following and Obstacle Avoidance Control of a Wheeled Robot 369


Using the definition of dh 13 5 x 2 8 y 2 , we can write:
   
x 1
5 dh 13 (27)
y 2

where [ 1  2 ]T 5 [cos13 sin13]T is the unit vector in the di-

rection , expressed in the robot frame. Substituting equation
(27) and equation (26) in equation (24) leads to the quadratic
equation:
Adh 132 8 Bdh 13 8 C 5 0 (28)
with:
A 5  12 c2y cos2 13 8  22 c2x sin2 13

B 5 2ax  1 c2y cos13 8 2a y  2 c2x sin13

Fig. 4. Frames definition and problem pose description.
C 5 1ax c y 32 8 1a y cx 32 6 c2x c2y 5 (29)
The solution of equation 28 is:
 where  is the deformation of the DVZ in the direction .
6B 8 B 2 6 4AC This quantifies the global amount of intrusion, normalized
dh 13 5 (30) by the factor d13. Note that a null distance robot/obstacle
2A
yields an infinite value of . Thus a control that guarantees
This distance clearly depends on the form parameters of
a bounded intrusion ratio at any time, ensures that the system
the controlled DVZ, i.e. the coefficients ax , a y , cx and c y .
avoids any obstacle.
Less explicitly, it also depends on the orientation of the DVZ,
The intrusion ratio  is a function of u (through the co-
i.e. on the attitude 4 m of the robot. In order to compute the
efficients ax , a y , cx and c y ), of 4 (through equation 31) and of
derivative of dh 13 with respect to this angle 4 m (as we will
the absolute position of the detected obstacle ([x E y E ]). We
need below), we have to consider that the point P is fixed in
can write:
the absolute frame, leading to the dependence of vector  on
 5 1u2 4 m 2 x E 2 y E 35 (34)
4 m . We have: 1 4
d sin13
53 6 (31) 3.5. System Jacobian Functions
d4 m 6 cos13
The time derivation of the expression of the intrusion ratio
Another way to say this is that the obstacle in this direction
yields
(if it exists) is ground-referenced (Figure 4).
7 5 Ju u7 8 J4 r 8 F Rob5el u 8 F Obst5el (35)
3.3. The Deformed DVZ with
7  
Let c13 be the distance between the sensor and the obstacle  2 B J Bu 62C J Au 62A JCu
Ju 5 1 1
6J Bu 8 
in the direction , and d13 the “distance” between the sensor 0 d13 2A B 2 64AC
and the deformed DVZ  I . d13 is obtained by saturation of 
6B8 B 2 64AC
c13 according to the following equation: 6 2A2
J Au d
d13 5 c13 i f c13  dh 2  
 2 61 8 B J 4 62C J A4
4
8 J 5 0 d13
1
2A
6J B 8  B 2
5 dh 13 elsehere5 (32) B 64AC

6B8 B 2 64AC
6 2A2
J A4 d
3.4. The Deformation
 2 dh 13
In order to use the DVZ paradigm to control the robot collision F Rob5el 5 0 d 2 13
cos d
avoidance and to combine this method with our path-following  2 dh 13
approach, we introduce the intrusion ratio: F Obst5el 5 0 d 2 13
16[x7 E cos 14 m 8 3
 2  2 9
 dh 13 6 d13 8 y7 E sin 14 m 8 3]3 d
5 d 5 d (33) (36)
50 d13 50 d13
370 THE INTERNATIONAL JOURNAL OF ROBOTICS RESEARCH / April 2007

where 4. Combining Path-following and Obstacle

7
8 J Au 5
c
21c y uy cos2 13 8 cx c x
sin2 133
Avoidance
u
(37)
9 J 4 5 2 cos 13 sin 131c2 6 c2 3 4.1. Mathematical Inspiration
A y x
7  The combination of the two algorithms is solved as a guidance
c
JBu 5 2 cos 13[c2y au
x
8 2ax c y uy ]
problem. The requirements are:
8 
a
8 sin 13[c2x uy 8 2a y cx c
u
x
] (38) 2 the vehicle should remain far from the obstacle, i.e. in
9 8  2  the presence of obstacles  has to be bounded at any
JB 5 2 ax c y sin 13 6 a y c2x cos 13 time,
and 2 when there is no obstacle, the vehicle has to asymptoti-
 
ax c y cally converge to the desired path.
JCu 5 2 ax c y c y 8 ax
u u
To do so, we rewrite the obstacle avoidance control, adding a
  desired intrusion ratio d 5 K d tanh 1d 14 6 933, where K d
a y cx
8 a y cx cx 8 ay and d are arbitrary positive gains, 4 and 9 are path-following
u u
variables defined in the previous section (equations (4) and
  2
c y cx (15)). Let V2 5 16 d3
be a Lyapunov function candidate.
6 cx c y cx 8 cy (39) 2
u u It is straightforward to see that the choice
7
and x7 E 2 y7 E define the obstacle absolute velocity. The assump- 8 r 5 6K r 1 6 d 31J4 6 d 3 8 cc s7 8 97
tion of static obstacles yields F Obst5el 5 0. For the sake of (43)
simplicity, we consider only static obstacles. 9 u7 5 6K J u 1 6  3 6 F Rob5el u
u  d Ju 

implies that  6d asymptotically converges to a bounded set

3.6. Obstacle Avoidance Control Design
defined as:
I2 7
Consider the following Lyapunov function candidate V 5 . J4 1cc s7 8 93
The derivation yields
2
 6 d t   4  2
(44)
K r 1J 6 d 3 8 K u 1Ju 32
V7 5 1Ju u7 8 J4 r 8 F Rob5el u35 (40) 2K d d
where d 5 18d 146932
.
It is straightforward to see that the choice Note that the previous expression is bounded since 1Ju 32 7
7 0 if u 5 0, d is bounded and if the quantity cc s7 8 97 is assumed
8 u7 5 6K u Ju  6 F J u u
Rob5el

 to be bounded, which is covered by the assumption that the free

(41)
9 r 5 6K J 4  space is connected.
r 
Then, following the obstacle, the vehicle cannot be driven
where K u 2 K r are arbitrary positive gains yields V7 0 t. infinitely far from the desired path, i.e. cc s7 8 97 remains
Note that tedious but straightforward computation (using for bounded. Moreover, if there is no obstacle, the Lyapunov func-
2
the sake of simplicity the guidance functions in (25)) shows tion candidate degenerates to V50 5 2d , and the previous
control choice yields V750 5 6K r d2 d 2 0, which induces
Rob5el
that the term F J u is a positive function that acts as a non-

linear damping term. This confirms the well posedness of the the path-following asymptotic convergence requirements. The
expression (41). dynamic control corresponding to the previous solution is:
The control (41) is a hybrid kinematic/dynamic solution. 7  
8 FO A 5 1 6K u Ju 1 6 d 3 6 F Rob5el u
A backstepping step is necessary for the torque control. Let u
J
(45)
VO A 5 12 1rd 6 r 32 be a Lyapunov function candidate, where 9 N 5 2 1r7 6 K 1r 6 r 33
rd 5 6K r J4 . The choice r7 5 r7d 6 K r 1r 6 rd 3 yields V7 O A
OA d r d

0. Then the dynamic obstacle avoidance control is written as where rd 5 6K r 1 6 d 31J4 6 d 3 8 cc s7 8 97 , K u and K r
7   are arbitrary positive gains. At this stage one should note that
8 FO A 5 1 6K u Ju  6 F Rob5el u u
J this solution does not prevent the forward velocity u from be-
(42)
9 N 5 2 17r 6 K 1r 6 r 33 ing null, and thus a poor convergence rate around the path.
OA d r d
We therefore propose another control version that avoids these
where K r and K u are arbitrary positive gains, 1 and 2 are the drawbacks, but also loses any mathematical proof of conver-
mass and moment of inertia of the vehicle, and rd 5 6K r J4 I . gence.
 Lapierre, Zapata, and Lepinay / Combined Path-following and Obstacle Avoidance Control of a Wheeled Robot 371

Table 1. Corner situation.

a) Corner situation b) Extract. from Corner Sit.

c) Evolut. forward vel. d) Evolut. of the ‘rabbit’.

4.2. Practical Solution The reaction to this situation is to:

The combination of path following and obstacle avoidance ca- (i) reduce the forward velocity,
pabilities has a natural limitation in situations where both cri-
teria induce antagonist system reactions. (ii) rotate until the obstacle is present only on one side, us-
For instance, the obstacle avoidance algorithm may demand ing the following controllers.
that the forward velocity be reduced to zero or a negative value. u7 C O R N E R 5 6K u u
This situation occurs when the deformed DVZ  is sym-
metric with respect to the forward velocity direction. Then the rC O R N E R 5 rC sign1d 3 (47)
robot will orient itself to minimize the intrusion, and regulate
its forward velocity to reshape the nominal DVZ h in order where K u is a positive gain, and rC the chosen rotational
to respect the intrusion requirement, i.e. lim t   5 d . velocity to operate the corner extraction.
This situation is called a corner situation, as described in Ta-
The overall control algorithm is written
ble 1. To avoid this problematic local minimum, the following
switching scheme was designed. Let l and r be the intru- i f 1C O R N E R 5 03
sion ratios on the left (0    ) and right (    2)
side, respectively, and u min 7 0 the lowest admissible for- 3N 5 N O A 8 f 13N P F  F 5 FO A 8 f 13FP F 4
ward velocity. The chosen corner situation detection is made
with the following switching condition. The Boolean variable else
C O R N E R is initialized to zero, then
3N 5 2 16K r 1r 6 rC O R N E R 33  F 5 1u7 C O R N E R 4 (48)
i f [1l r 7 03&1u  u min 3] 3C O R N E R 5 14
where the selection function is f 13 5 18K1  , with K I a pos-
i f 1l r 5 03 3C O R N E R 5 045 (46) itive gain. Note that this algorithm causes the robot to travel
372 THE INTERNATIONAL JOURNAL OF ROBOTICS RESEARCH / April 2007

with a positive forward velocity u 7 u min 7 0, outside the Table 2. The chosen control parameters.
corner situation. This guarantees the well posedness of the
expressions (43), (44) and (45). Then the global control (48) is K rP F 5 051 K uP F 5 1 K rO A 5 0501 K uO A 5 051
well posed in any situation. Table 1(c) describes the evolution K d 5 10 d 5 0501 K I 5 100 ud 5 1
of the forward velocity while encountering a corner situation. cx 5 100 cmin 5 10 K uC O R N E R 5 051 rc 5 1
x
The increase in the intrusion ratio induces a decreasing for-
ward velocity, down to u min 5 052m5s 61 , while the obstacle
is present on the left and right side of the robot (l r 7 0).
The robot is then controlled according to (47), where rcorner is
driven by the sign of d . This allows the system to be driven 5.1. Simulation Results
away from the path, thus guaranteeing the system skirts the
obstacle. Table 1(d) shows the evolution of the virtual target We have implemented the algorithm on a unicycle simulator
(‘rabbit’) along the path, where its backward movement to- developed with Matlab. The control parameters are chosen
wards the closest point on the path is seen, while the robot is according to Table 2. The results are displayed in Table 3.
skirting the obstacle. Table 2(a) shows the path and obstacles definitions. The ro-
bot is of the unicycle type, on which a 32-proximity-sensors
belt is mounted, displayed as surrounding radial rays of length
defined by the nominal DVZ. Table 2(b) indicates a corner sit-
5. Results uation, in which we see the reduction of the DVZ due to the
decreased forward velocity. Table 2(c) shows the system after
The previous solution implicitly requires an estimation of 7 the switch from corner situation to nominal control. Finally,
in the computation of FO A . The numerical derivation of the Table 3(d) displays the global trajectory the system has made.
sensor information induces noise amplification and will not
provide an accurate estimation of . 7 Therefore, we degrade
the previous solution at a kinematic level, that does not require
7 Moreover, the considered robot (cf. Fig-
the estimation of . 5.2. Experimental Results
ure 5) accepts as control inputs the desired velocities u and r,
and its rapid actuator response (combined with its light mass
and inertia) allows implementation of a kinematic controller We have implemented the control algorithm of equations (49)
without degrading the overall performance. on the Pekee robot from the Wany company. This robot is
7 driven by two independent wheels and carries a 16-infrared-
8 r 5 r O A 8 f 13r p f proximity- sensors belt, see figure 5, and is controlled via a
(49) Mitsubishi M16C family micro-controller. The complete algo-
9 u 5 u O A 8 f 13u p f rithm, equations (49), was first implemented on an embedded
PC daughterboard, and the test results are displayed in Figure
where 6 describing the odometric trajectory of the robot. Since the
7 navigation system is based only on odometric information, the
r P F 5 97 6 K rP F 14 6 93 8 cc s7 real trajectory is not significant in the demonstration of the va-
t lidity of our approach. The robot clearly avoids the obstacle
u P F 5 0 1u7 d 6 K uP F 1u 6 u d 33dt and returns to the nominal path without requiring switching
8
s7 5 u cos 4 8 ks s1 (50) control. This is valid in this particular situation where the cor-
ner situation does not occur. Some chattering behavior has
r O A 5 6K rO A 1 6 d 31Jr 6 d 3 8 cc s7 8 97 to be filtered out by gain tuning according to the Pekee ro-
9 u 5  t 16K O A J  1 6  3 6 F r ob5el u3dt5 bot capabilities. Figure 7 shows the evolution of the forward
OA 0 u u d J u
velocity. The black dots denote the desired velocity when no
obstacle is detected, the stars indicates the evolution of the for-
The DVZ guidance functions are chosen according to (25). ward velocity in the presence of obstacles. A second stage
The other functions are chosen according to the following consisted in simplifying the algorithm complexity (including
definitions: trigonometric look-up tables) in order to reduce the compu-
7 tational burden, and implementation of the solution directly
8 d 5 K d tanh d 14 6 93 on the M16C micro-controller onboard the vehicle. The tests
(51)
9 f 13 5 1 displayed similar vehicle behavior, since the sensor (US and
18K   odometry) accuracy did not allow demonstration of the reac-
tivity benefits of the low level implementation.
 Lapierre, Zapata, and Lepinay / Combined Path-following and Obstacle Avoidance Control of a Wheeled Robot 373

Table 3. Simulation results.

a) Path and Obst. Def. b) Corner Situation

c) Extract. from Corner Sit. d) Final Trajectory

portant forward velocity. With this system, we combine the

reactivity of the DVZ principle with a path-following control
without requiring any path replanning. The next step in this
study is to intrinsically attribute the reactivity of the guidance
system to the path-following virtual target, explicitly control-
ling the evolution of an added virtual state of the virtual target,
along the y1 direction (see Figure 1). Then the main robot
control could be designed as a tracker of a cooperative virtual
target1 note that the system reactivity is then in charge of a new
dynamic guidance system.

Fig. 5. The Wany robot “Pekee”.

Acknowledgements

6. Conclusions This work is supported by a post doctoral grant from the

French national RNTL project WACIF. For further infor-
We have designed a combined path following and obstacle mation on this project, please contact http://intranet.wacif.
avoidance control law for a unicycle type robot based on the net/default.aspx .
use of the DVZ concept and on a Lyapunov and backstepping
design. Implementation of this solution on the Wany robot
“Pekee” illustrates an interesting performance, avoiding un-
necessary hot switches when the vehicle travels with an im-
374 THE INTERNATIONAL JOURNAL OF ROBOTICS RESEARCH / April 2007

Fig. 6. Experimental results, trajectories.

Fig. 7. Experimental results, forward velocity.

 Lapierre, Zapata, and Lepinay / Combined Path-following and Obstacle Avoidance Control of a Wheeled Robot 375

References Jiang, Z. P. and Nijmeijer, H. (1999). A recursive technique

for tracking control of nonholonomic systems in chained
Aicardi, M., Casalino, G., Bicchi, A., and Balestrino, A. form. IEEE Transactions on Robotics and Automation,
(1995). A closed loop steering of unicycle-like vehicles 44(2): 265–279.
via Lyapunov techniques. IEEE Robotics and Automation Krstić, M. I., Kanellakopoulos, and Kokotovic, P. (1995). Non-
Magazine, March, 27–35. linear and Adaptive Control Design. John Wiley & Sons,
Aicardi, M., Casalino, G., Indiveri, G., Aguiar, P., Encarnacao, Inc., New York.
P. and Pascoal, A. (201). A planar path following con- Lapierre, L., Soetanto, D., and Pascoal, A. (2003). Nonlin-
troller for underactuated marine vehicles. Proceedings of ear path following with application to the control of au-
MED’2001, Dubrovnik, Croatia, June. tonomous underwater vehicle. CDC 2003, 42nd IEEE Con-
Althaus, P. and Christensen, H. (2002). Behavior coordination ference on Decision and Control, Maui, Hawaii, USA, De-
for navigation in office environment, IEEE/RSJ Interna- cember.
tional Conference on Intelligent Robots and Systems, Lau- Louste, C. (1999). Conception d’une methode de planification
sanne, Switzerland, October. pour robot mobile selon la methode des milieux continus
Arkin, R. C. (1998). Behavior Based Robotics. MIT Press, appliquee aux fluides visqueux. PhD thesis no. 6701,
Cambridge, MA. LIRMM, Montpellier, France.
Canudas de Wit, C., Khennouf, H., Samson, C., and Sordalen, Micaelli, A. and Samson, C. (1993). Trajectory-tracking
O. J. (1993). Nonlinear control design for mobile robots. for unicycle-type and two-steering-wheels mobile robots.
In Recent Trends in Mobile Robotics (eds Y. F. Zheng), Vol Technical Report No. 2097, INRIA, Sophia-Antipolis, No-
11, pp.121–156, World Scientific Series in Robotics and vember.
Automated Systems. 1993 Ogren, P. (1989). Real-time obstacle avoidance for fast mobile
Elnagar, A. and Hussein, A. (2002). Motion planning using robots. IEEE Transactions on Systems, Man and Cybernet-
Maxwell’s equations. IEEE/RSJ International Conference ics, 19(5): 1179–1187.
on Intelligent Robots and Systems, Lausanne, Switzerland, Ogren, P. and Leonard, N. (2002). A provable convergent dy-
October. namic window approach to obstacle avoidance. IFAC World
Encarnação, P., Pascoal, A., and Arcak, M. (2000). Path fol- Congress, Barcelona, Spain, July.
lowing for marine vehicles in the presence of unknown cur- Ogren, P. (2003). Formation and Obstacle Avoidance in Mo-
rents. Proceedings of SYROCO’2000, 6th IFAC Symposium bile Robot Control. PhD thesis, Stockholm, Norway, June.
on Robot Control, Vienna, Austria. Rimon, E. and Koditschek, D. (1992). Exact robot navigation
Encarnação, P. and Pascoal, A. (2000). 3D path follow- using artificial potential functions. IEEE Transactions on
ing for autonomous underwater vehicle. Proceedings of Robotics and Automation, 8(5): 501–518.
CDC’2000, 39th IEEE Conference on Decision and Con- Samson, C. and Ait-Abderrahim, K. (1991). Mobile robot con-
trol, Sydney, Australia. trol Part 1: Feedback control of a non-holonomic mobile ro-
Fierro, R. and Lewis, F. (1997). Control of a nonholo- bot. Technical Report No. 1281, INRIA, Sophia-Antipolis,
nomic mobile robot: backstepping kinematics into dynam- France, June.
ics. Journal of Robotic Systems, 14(3): 149–163. Soetanto, D., Lapierre, L., and Pascoal, A. (2003). Adaptive,
Ge, S. and Cui, Y. (2000). Path planning for mobile robots nonsingular path following control of dynamic wheeled ro-
using new potential functions. Proceedings of the 3rd Asian bot. CDC’03, 42nd IEEE Conference on Decision and Con-
Control Conference, Shangai, China, July. trol, Maui, Hawaii, USA, December.
Iniguez, P. and Rossel, J. (2002). A hierarchical and dynamic Zapata, R., Cacitti, A., and Lepinay, P. (2004). DVZ-based
method to compute harmonic functions for constrained mo- collision avoidance control of non-holonomic mobile ma-
tion planning. IEEE/RSJ International Conference on Intel- nipulators. European Journal of Automated Systems, 38(5):
ligent Robots and Systems, Lausanne,Switzerland, October. 559–588.