Recommended by Frederic Mazenc

Accepted Manuscript

Aerial monitoring of slow moving convoys using elliptical orbits

Aseem V. Borkar, Vivek S. Borkar, Arpita Sinha

PII: S0947-3580(18)30171-7
DOI: https://doi.org/10.1016/j.ejcon.2018.09.010
Reference: EJCON 292

To appear in: European Journal of Control

Received date: 19 April 2018

Revised date: 2 August 2018
Accepted date: 20 September 2018

Please cite this article as: Aseem V. Borkar, Vivek S. Borkar, Arpita Sinha, Aerial monitoring
of slow moving convoys using elliptical orbits, European Journal of Control (2018), doi:
https://doi.org/10.1016/j.ejcon.2018.09.010

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service
to our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and
all legal disclaimers that apply to the journal pertain.
 ACCEPTED MANUSCRIPT

Aerial monitoring of slow moving convoys using elliptical orbits

Aseem V. Borkara,∗, Vivek S. Borkarb , Arpita Sinhaa
a Systems and Control Engineering, Indian Institute of Technology Bombay, Mumbai, Maharashtra 400076, India
b Department of Electrical Engineering, Indian Institute of Technology Bombay, Mumbai, Maharashtra 400076, India

Abstract

T
This article proposes a novel vector field based guidance scheme for tracking and surveillance by an aerial agent of a convoy moving
along a possibly non-linear trajectory on the ground. The scheme first computes a time varying ellipse that encompasses all the

IP
targets in the convoy using a simple regression based algorithm. It then ensures convergence of the aerial agent to a trajectory
that repeatedly traverses this moving ellipse. The correct tracking behaviour of the scheme is rigorously established along with
error bounds, based on the perturbation theory of ordinary differential equations. It is supported by MATLAB® simulations and a

CR
hardware implementation, the details of which are provided along with a link to the video of an actual experiment.
Keywords: Convoy monitoring, Vector field guidance, Perturbation analysis, Alekseev formula, Floquet theory
INSPEC Codes: C10; C30

1. Introduction US In particular, by exploiting the fact that we are tracking

AN
We consider the problem of tracking and surveillance of a
slowly moving convoy of ground targets using aerial agents
such as UAVs or quadrotors. For simplicity, we consider a
single monitoring agent modelled by unicycle kinematics, a
common model used in literature for such purposes. We de-
M
a slowly varying periodic trajectory, we are able to get a
good handle on the analysis of the perturbations by ex-
ploiting Floquet theory applied to the linearized dynamics
along the periodic trajectory. Compare this with the fact
that in general the stability of time-dependent lineariza-
tions is difficult to analyze and has been done only under

velop a guidance law based on this model whereby it latches stringent conditions on the time scale separation (see, e.g.,
on to a slowly moving ellipse that encircles the convoy mov- Solo (1994)). Our analysis uses in a key manner an ex-
ing along a possibly nonlinear trajectory and moves with it. A tension of Alekseev nonlinear variation of constants for-
ED

convoy, by definition, is a group of targets trailing one after the
other on a common path. As compared to the commonly studied
circular orbits, elliptical orbits offer economy of distance trav-
elled. The twofold task this entails is: to compute the moving
PT
mula that handles differing initial conditions for perturbed
and unperturbed systems, which is another aspect that was
missing in Borkar et al. (2017).
2. We supplement the MATLAB® simulations of an ideal aer-

ellipse around the convoy and to follow a trajectory that remains
faithful to traversal of this elliptical orbit as the convoy slowly
traverses a possibly nonlinear but continuously differentiable
trajectory. We assume that the aerial agent moves with much
CE
ial agent having unicycle kinematics presented in (Borkar
et al. (2017)) with a more realistic validation of the pro-
posed theory by implementing it on a quad-rotor for mon-
itoring a moving convoy of differential drive robots on the

faster speed than the convoy. We then leverage the assumed
time scale separation between the motions of the target(s) and
the aerial agent to exploit some facts from perturbation theory
of differential equations.
AC
ground. Full details of the hardware implementation are
provided, with a link to the video of actual experiment.
We also provide an additional simulation experiment with
a convoy trajectory that is only piecewise smooth.

An earlier version of this paper was presented in Borkar et al.
(2017). The present work significantly improves upon ibid. in
that:
1. The perturbation analysis in Borkar et al. (2017) was tent-
ative and incomplete. Here we present a thorough ana-
lysis which may be of independent interest as a method-
ology for problems of tracking time-varying trajectories.
∗ Corresponding
author
Email addresses: aseem@sc.iitb.ac.in (Aseem V. Borkar),
borkar@ee.iitb.ac.in (Vivek S. Borkar), asinha@sc.iitb.ac.in
(Arpita Sinha)
Preprint submitted to European Journal of Control
3. Some technical details omitted in Borkar et al. (2017) due
to page limitations are included here, making this article
self-contained.
Some potential applications are: military and police surveil-
lance, wildlife swarm tracking in forest reserves, tracking res-
cue vehicles in disaster areas, etc.
The article is organized as follows: The next section gives a
brief account of related work. Section 3 describes our choice of
the elliptical orbit for convoy encirclement. Section 4 describes
the vector field guidance strategy to guide the agent to this or-
bit. The treatment of these two is similar to that in Borkar et al.
(2017). Section 5 gives the detailed error analysis based on per-
22nd September 2018
 ACCEPTED MANUSCRIPT
turbation theory. Section 6 validates the guidance law and the
convoy encirclement strategy through MATLAB® simulations
as well as a hardware implementation. We describe the latter in
some detail and provide a link to a video of an actual experi-
ment. The Appendix details some technical results used in the
main text.
2. Prior work
This is a short overview of some of the related work, see
Robin and Lacroix (2016) for an extended survey. Among the
works within the popular paradigm of guidance strategies for
following circular orbits around stationary and/or moving tar-
gets, the closest in spirit to our work are the guidance laws
based on appropriately designing the relevant vector fields for
target tracking applications. These can be either for tracking a
single target (Frew and Lawrence (2005)) or close groups of tar-
gets (Oh et al. (2015)) with multiple UAVs while maintaining a
minimum stand-off distance from the targets. In both cases the
tracking UAVs achieve a phase separation on the circular orbit
by controlling the linear velocity with a phase error term. The
former work is extended in Frew et al. (2008) to a race-track
like path for tracking a convoy moving on a straight line. In
contrast, Frew (2007) transforms the guidance vector field de-
veloped in Frew and Lawrence (2005) to track fixed elliptical
US
AN
orbits whose parameters depend on estimation uncertainties of
the target states for a target moving in a straight line. Vector
field based guidance laws for tracking circular orbits have been
used also for atmospheric sensing (Lawrence (2003)) and path
M
following (Nelson et al. (2007)). The main difference with our
work is that barring (i) Frew et al. (2008) where the drone al-
ternates between a circle segment and a straight line, and, (ii)
Frew (2007) where the drone tracks a moving constant para-
ED
meter ellipse, the drone tracks a moving circular orbit. In our
work, the drone tracks a flexible ellipse that adapts both its
shape and position to the convoy trajectory. Another import-
ant feature of our work is that the target path is not assumed to
PT
be known a priori.
Another popular strand of research is cyclic pursuit, where
multiple agents cooperatively converge to a circular orbit with
CE
an equi-spaced formation, for both stationary (Galloway and
Dey (2015)) and moving (Ma and Hovakimyan (2013)) targets.
In Ma and Hovakimyan (2015), the latter is extended for circu-
lar orbits of varying radii around the target. In a related theme of
AC
formation control, Zhang and Liu (2016) propose agent forma-
tion strategies where the agents follow circular orbits centered
at target in order to loiter around slow moving targets and move
with a fixed equi-spaced formation on the orbit when the target
moves faster. In Paley et al. (2004), authors propose steering
control laws for cooperating agents to perform and transition
between two stably controlled group parallel and circular mo-
tions. It is also shown that this method can be used to track
a point moving on a piecewise linear path. Unlike the present
work, these schemes track a single target, not a convoy. Also,
the comments at the end of the preceding paragraph apply.
Some recent works on tracking moving objects are Zhou
et al. (2017), Kapitanyuk et al. (2017). In Zhou et al. (2017),
2
the classical linear, circular etc. orbits are considered for the tar-
get and the emphasis is on analysing the error caused by wind.
In Kapitanyuk et al. (2017), a problem similar to ours is con-
sidered, but the target is assumed to follow a smooth trajectory
specified a priori as the level curve of a known function.
For the problem of convoy protection, one approach is to
track lemniscate like orbits, e.g., Oliveira et al. (2016) pro-
pose a strategy where a single UAV tracks lemniscate like orbits
centered at the convoy centroid, while Spry et al. (2005) use a
combination of lemniscate like lateral orbits and parameterised
asymmetric longitudinal orbits to follow a convoy moving on a
straight line. Our scheme based on flexible moving ellipses is
T
much simpler and does not require any trajectory switching.
In turn, Ding et al. (2010) treat the UAVs as Dubins vehicles
IP
and design time-optimal paths for protection of stationary
ground vehicles. They propose a control strategy in order to
use these paths to monitor a convoy moving in a straight line.
CR
Our work in contrast handles a very broad class of a priori un-
known convoy paths.
3. Convoy Encirclement using Elliptical Orbits
3.1. Notation
This section introduces our strategy for continuous en-
circlement of all the targets in the convoy as they trace a certain
trajectory. For purposes of this work, we assume that the posi-
tions of the targets constituting the convoy are always known to
the monitoring agent. This could be either through sensing or
through cooperation. The proposed strategy for encirclement is
implemented by an algorithm that runs in each iteration of the
guidance loop and defines an ellipse around the positions of the
targets at each instant of time.
It is assumed that the speed VA of the tracking agent is within
a bounded interval [VAmin , VAmax ]. Furthermore, the speed Vi of
any target i is bounded above by VTmax where VTmax << VAmin .
We follow the convention that the targets are numbered 1, ..., N
along the direction of travel for the convoy, i.e., the leading
agent in the convoy is agent N. We denote the set of real num-
bers as R. Denote the rotation matrix from the right handed
global reference frame to a tilted frame with tilt angle θ(t) as
" #
cos(θ(t)) sin(θ(t))
Rθ (t) = . (1)
− sin(θ(t)) cos(θ(t))
3.2. Convoy-centric regression and projection
The algorithm fits a linear regression line to the target posi-
tions so as to define a bounding rectangle l1 (t) × l2 (t) that con-
tains all target points either inside or on it as illustrated in Fig.
1. In order to encompass all targets, we consider ellipses that
either contain or circumscribe this rectangle as candidate paths
for the monitoring agents.
The equations for computing slope m and intercept c of a
regression line y = mx + c fit to data points (xi , yi ) are obtained
by minimizing the empirical mean square error
N
1 X
(yi − mxi − c)2
N i=1
 ACCEPTED MANUSCRIPT

l2 (k) Algorithm 1 initializes the θE (k) as follows: If the agents lie

Convoy Path on a vertical line in the global reference frame, then from (2),
Target positions the numerator and denominator for slope m of the regression
at time k-1
line are both zero and the algorithm sets θE (0) = π2 . Algorithm
Average target position at
Bx time k-1 2 computes l1 (0) as length of the line segment joining the first
δθ (k) Regression line and last projections of the targets on the regression line, and
l1 (k)
By at time k-1 l2 (0) as twice the maximum normal projection error dmax from
(x̄, ȳ) Bounding rectangle the target positions to the regression line (see Fig. 1). Then
at time k-1
l2 (0) > l1 (0) implies a large projection error and the regression
Target positions
at time k
angle is set to θE (0) = π2 − θE (k).
Average target position
To avoid the problem discussed in Remark 1 in the sub-
at time k sequent iterations of the guidance loop (indexed by iteration

T
Regression line count k), Algorithm 1 selects a convoy-centric coordinate frame
θE (k) at time k
B(k) centered at the average target position ( x̄, ȳ) with the x axis

IP
θE (k − 1) Bounding rectangle
at time k
aligned along θE (k−1), and a regression line y = m0 x (solid blue
line segment in Fig. 1) is fit to the agent positions at the instant
k relative to the frame B(k). This regression line always passes

CR
Figure 1: Successive regression relative to the local frame B through the origin of the coordinate frame B(k), as the average
value ( x̄, ȳ) of the data points always lies on the regression
h iline
by virtue of (2). The change in tilt angle δθ (k) ∈ −π , π
2 for
over m and c. Thus the estimates are given by: 2
each iteration is computed as

c=
ȳ
N
P
i=1
N
P
xi2 − x̄

xi2 −
N
P
i=1

N x̄2
xi yi
, m=
N
P
i=1
N
P
xi yi − N x̄ȳ

xi2 − N x̄2
US
(2)
δθ (k) = arctan(m0 (k)), θE (k) = θE (k − 1) + δθ (k).

Algorithm 2 recomputes θE (k) as the inclination angle of the ray

AN
i=1 i=1 joining ( x̄, ȳ) to the projected position (x pN , y pN ) of the target N
N
P N
P on this regression line. As a result, θE (k) ∈ (−π, π].
xi yi
with x̄ = N and ȳ = N , where (xi , yi ) represent the positions
i=1 i=1

Algorithm 1 Convoy centric regression

of convoy target i in the global reference frame.
M

Inputs: (xi (k), yi (k)) ∀i ∈ {1, ..., N}, k

Functions: Projection
2
1 Outputs: l1 (k), l2 (k), θE (k),
! xo (k), yo (k)
1 N N
1 P P
ED

2 1 1: ( x̄, ȳ) = N xi (k), yi (k)

i=1 i=1
3
2 2: xarr = {xi (k), ∀i ∈ {1, ..., N}}
3 3: yarr = {yi (k), ∀i ∈ {1, ..., N}}
4 3 4: if k = 0 then
N
P N
P
PT

4 5: mn = xi (k)yi (k) − N x̄ȳ, md = xi2 (k) − N x̄2

5 4 i=1 i=1
5 6: if mn = 0 and md = 0 then
5
Time t = t1 Time t = t2 Time t = t3 7: θE (k) = π2
8: else m = mmdn ,
CE

Figure 2: The red points shown are the targets moving on the black 9: θE (k) = arctan(m)
path at three instants of time t1 < t2 < t3 . The green line segment is 10: [l1 (k), l2 (k), xo (k), yo (k), θE (k)]=
the regression line relative to global frame. The blue line segment is Projection(xarr , yarr , x̄, ȳ, θE (k))
the result of Algorithm 1 11: if l1 < l2 then θE (k) = π2 − θE (k)
AC

12: else
13: for "i ∈ {1, #..., N} do " #
Remark 1. A problem with linear regression is that when fit- xiB (k) xi (k) − x̄
ting line y = mx + c to the data points stacked close to the y 14: = RθE (k − 1)
yBi (k) yi (k) − ȳ
axis, linear regression yields a nearly horizontal line with large N
P B PN 2 m0n
projection errors. Hence if the regression line is fit relative to 15: mn = xi (k)yi (k), md = xiB (k), m0 =
0 B 0
m0d
i=1 i=1
a fixed inertial frame, e.g. when the convoy moves on a path in 16: δθ (k) = arctan(m0 )
the vicinity of the inertial y axis, a sudden change in inclination 17: θE (k) = θE (k − 1) + δθ (k)
angle θE of the regression line occurs as shown in Fig. 2 (green 18: [l1 (k), l2 (k), xo (k), yo (k), θE (k)]=
segment). This is clearly not desirable, because a segment of Projection(xarr , yarr , x̄, ȳ, θE (k))
this line is to be used later for defining the major axis of the
encircling ellipse. We describe below a way to circumvent this Since the motion of the agents is continuous, slope m0 of the
problem. regression line in the frame B(k) also varies continuously with
3
 ACCEPTED MANUSCRIPT

Algorithm 2 Projection
l1s
Inputs: xarr , yarr , x̄, ȳ, θE (k) ds l1
Outputs: l1 (k), l2 (k), xo (k), yo (k), θE (k)
1: xmin = 0, xmax = 0, dmax = 0 3
2: for "i ∈ #{1, ..., |xarr |}" do # 2 Fy 4 b
l2
xr x [i] − x̄
3: = RθE (k) arr (xo , yo ) Fx
yr yarr [i] − ȳ l2s
1
4: if dmax ≤ |yr | then ds 5
5: dmax = |yr |
6: if xmin ≥ xr then a
7: xmin = xr
8: if xmax ≤ xr then Figure 3: Ellipse centered frame F with origin (xo , yo ). The yel-
9: xmax = xr low dashed line is the regression segment l1 . The blue ellipse is the

T
10: if i = N then minimum area ellipse circumscribing the bounding rectangle l1 × l2
B (shown with green dashed line). The black ellipse circumscribing the
11:
# xN = xr "

IP
" # " #
x pmin xmin x̄ red dashed rectangle guarantees a minimum stand off distance d s from
12: = R−1 (k) + all the targets in the convoy.
y pmin θE 0 ȳ
" # " # " #

CR
x pmax xmax x̄
13: = R−1 (k) +
y pmax θ E 0 ȳ
" # " B# " # We assume that the aerial agent is described by the unicycle
x pN −1 x x̄
14: = RθE (k) N + kinematic model with state space representation
y pN 0 ȳ
x p +x p
15: xo (k) = min 2 max
16:
17:
18:
19:
yo (k) =
y pmin +y pmax
p 2
l1 (k) = (x pmin − x pmax )2 + (y pmin − y pmax )2
l2 (k) = 2dmax
θE (k) = arctan 2(y pN − ȳ, x pN − x̄)
US ẋA (t) = VA cos(ψA (t)), ẏA (t) = VA sin(ψA (t)),
ψ̇A (t) = ωA (t)

where (xA (t), yA (t)) are the agent position coordiantes, ψA (t) the
(3)
AN
heading angle, VA ∈ [VAmin , VAmax ] is the commanded constant
linear speed, and ωA (t) the commanded angular velocity of the
monitoring agent satisfying |ωA (t)| ≤ ωmax .
time k. As a result, δθ (k) in the local reference frame B(k) is
a small angle with respect to the local x axis and the issue dis- By Lemma 2 of Appendix, the minimum radius of curvature
M

cussed in Remark 1 is avoided. l1 (k), l2 (k) are computed in the of the ellipse E is Rmin = ba . To ensure that ωA < ωmax while
same manner as in the initialisation step using Algorithm 2. following the elliptical path, Rmin and the minimum turn radius
V max
This regression scheme is simple enough for on board real of the agent RA = ωAmax must satisfy Rmin ≥ RA , which implies
q
ED

V max aVAmax
time computation when possible, though in our experimental and b ≥
a ≥ ωAmax ωmax . Assuming that δθE is small, the
results reported later, it was computed in a central computer. maximum relative velocity between the target and the agent is
Details follow in section 6. VRmax = VAmax + VTmax . Thus we select
( )
PT

3.3. Parameter selection l1 (k) VRmax

a(k) = max √ , ,
2 ωmax
From Algorithm 1 we have the tilt angle θE (k) of the major  s 

 
axis relative to the global reference frame. The lengths of the  l2 (k)
 a(k)VRmax 


b(k) = max  (4)
CE

 √ ,  ,
semi-major and minor axes can be selected according to the ob- 
 2 ωmax  

jective of the monitoring mission. Henceforth we consider the
minimum area ellipse circumscribing the bounding rectangle which ensures that the circumscribing elliptical orbit of min-
around the convoy. By Lemma 1 of the Appendix, the minimum imum area is selected as long as it does not violate the min-
AC

2
area ellipse E : ax2 + by2 = 1 with a > b > 0 that circumscribes
2
the bounding rectangle around the convoy has semi-major axes
a = √l12 and semi-minor axes b = √l22 (shown as the blue ellipse
in Fig. 3).
Some monitoring applications such as Frew and Lawrence
(2005) may require that the monitoring agent must maintain a
minimum stand-off distance d s from the targets. To ensure this,
instead of the l1 × l2 rectangle, the l1s × l2s rectangle centered
at (xo , yo ) is considered with l1s = l1 + 2d s and l2s = l2 + 2d s ,
whereby the same algorithm guarantees a minimum stand off
distance of the elliptical path from the convoy agents (shown as
the black ellipse in Fig. 3).
4
imum turn radius or minimum speed of the monitoring agent
on the ellipse.
4. Guidance Strategy
Since the proposed algorithm assigns an ellipse around the
convoy of interest, a guidance strategy is essential for tracking
this elliptical path around the convoy. This strategy must be
able to guide the monitoring agent from any initial pose to any
ellipse of interest defined in 2D space.
To simplify analysis, it is assumed that the speed of the mon-
itoring tracking agent is significantly greater than the convoy’s
 ACCEPTED MANUSCRIPT

maximum speed. Also, the monitoring agent is characterized By differentiating with respect to time we get
by a unicycle kinematic model described in (3).  2  *" # " #+
For the case of an ellipse having its axes aligned along the dV  x (t) y2 (t)  2xA (t)/a2 ẋA (t)
= 2  A 2 + A 2 − 1 ,
2D reference frame and centered at the origin, the equation of dt a b 2yA (t)/b2 ẏA (t)
2
the ellipse is given by ax2 + by2 = 1. Differentiating this equation
2
 2 2  * " #+
 xA (t) yA (t)  cos(ψD (t))
with respect to the x coordinate, the tangential direction at a = 2VA  2 + 2 − 1 ∇EA , (9)
a b sin(ψD (t))
point (x, y) on the ellipse in the counter-clockwise direction is
given by where ∇EA = ∇EA (t) is the gradient of the ellipse of the family
  2
ψT = arctan 2 (dy, dx) = arctan 2 b2 x, −a2 y (5)
x2
a2
+ by2 = c at (xA (t), yA (t)) in the outward normal direction.

where the term arctan 2(·) yields ψT ∈ (−π, π]. This function is Case 1: If the agent position (xA (t), yA (t)) is inside the
undefined at (0, 0) where we set it equal to zero. x2 (t) 2

T
2 desired ellipse, Aa2 + yAb(t) 2 < 1, hence γ(t) ∈ [0 1). For
Consider the family of concentric ellipses ax2 + by2 = c with
2
dV
(xA (t), yA (t)) = (0, 0), dt = 0, but (0, 0) is not an equi-

IP
2
c > 0. Any point (x p , y p ) ∈ R2 lies on a unique ellipse x2
a2
+ by2 = librium point of (7) by design (note that the vector field is
x2p y2p discontinuous at (0, 0)) and the state trajectory moves out of
c p from this family, where c p = + and ψT |(x p ,y p ) gives the
a2 b2
tangential heading angle along this ellipse as shown in Fig. 4. (0, 0). For (xA (t), yA (t)) , (0, 0) inside E, ψT (t) is the counter-

CR
All points in R2 lying on the a line y = mx for any slope m clockwise tangential direction
  along the ellipse perpendicular
result in the same value of ψT . to ∇EA and ψO (t) ∈ −π 2 , 0 , the agent velocity vector with
2
Suppose the desired ellipse to be followed is ax2 + by2 = 1
2
ψA (t) = ψD (t) = ψT (t) + ψO (t) makes an acute angle with ∇E
as shown in the Fig. 4. Thus the inner product term in (9) is
with a > b and the agent position coordinates are (xA (t), yA (t)).
x2 (t) y2 (t)
Define γ(t) = Aa2 + Ab2 . The desired heading ψD (t) for the
monitoring agent is, for kγ > 0,
ψD (t) = ψT (t) + ψO (t), (6)
US inside E except (0, 0).
x2 (t) y2 (t)
positive. Therefore, as Aa2 + Ab2 − 1 < 0, dV dt < 0 for all points
AN
Case 2: If (xA (t), yA (t)) is outside the desired
  x2 (t) y2 (t)
 
ψT (t) = ψT |(xA (t),yA (t)) , ψO = arctan kγ (γ(t) − 1) . ellipse, Aa2 + Ab2 > 1, hence γ(t) ∈ (1, ∞) and ψO (t) ∈ 0, π2 .
Thus the agent velocity vector with heading direction
ψA (t) = ψD (t) = ψT (t) + ψO (t) makes an obtuse angle with
∇EA as shown in the Fig. 4 and the inner product term in (9) is
M

Y
x2 (t) y2 (t)
ψT negative. Therefore as Aa2 + Ab2 − 1 > 0 at any point outside
∇E
(0, b)
ψT ψ
D ∇E
ψD E, dV
dt < 0.
ED

ψT ψD
∇E
x2 (t) y2 (t)
Case 3: If (xA (t), yA (t)) is on the desired ellipse Aa2 + Ab2 = 1,
then dV
(−a, 0) X
dt = 0. Since γ(t) = 1, ψO (t) = 0 and ψA (t) = ψT (t),
(a, 0)
O
which is the tangential direction along the ellipse. Thus the
x2 y2
+ <1 state trajectory always remains on E, implying that it is a
PT

a2 b2

(0, −b) E: x2
+ y2
=1 positively invariant set.
a2 b2

Thus dV 2 dV
dt ≤ 0 for all (xA (t), yA (t)) ∈ R and dt = 0 for
2
x2 y
a2 + b2 >1
CE

(xA (t), yA (t)) ∈ E, where E = E ∪ {(0, 0)}. Since E is the

Figure 4: The chosen agent heading ψd shown inside, outside, and on largest invariant subset in E, from any initial (xA (0), yA (0)),
2 2
the desired ellipse E : ax2 + bx2 = 1 (xA (t), yA (t)) approaches E asymptotically by LaSalle invari-
ance principle Khalil (1996). 2
AC

Proposition 1. If ψA (t) = ψD (t) in (3), then starting at any ini-

tial position (xA (0), yA (0)), the agent asymptotically converges For an arbitrary ellipse in R2 , if F is the ellipse centric frame
2
to the desired ellipse E : ax2 + by2 = 1.
2
with origin at ellipse center (xo (t), yo (t)) and tilt angle θE (t) ∈
(−π, π] relative to the global reference frame, then the agent
Proof: For ψA (t) = ψD (t) the idealized unicycle agent has the
position and heading in the frame F are given by
following state equations:
   " #
ẋA (t) = VA cos(ψD (t)), ẏA (t) = VA sin(ψD (t)) (7)  xE (t)   xA (t) − xo (t) 
  Rθ (t) 
 yE (t)  =  E yA (t) − yo (t)  .

(10)
Define the Lyapunov function ψE (t) ψA (t) − θE (t)
 2 2
 xA (t) y2A (t) 
V =  2 + 2 − 1
 (8) The guidance law we propose can handle both counter-
a b clockwise and clockwise paths followed along the ellipse using
5
 ACCEPTED MANUSCRIPT

the heading commands ψDccw (t) and ψDcw (t) given by
ψDccw (t) = ψTccw (t) + ψO (t),
ψDcw (t) = ψTcw (t) − ψO (t),
where
 
ψTccw (t) = arctan 2 b2 (t)xE (t), −a2 (t)yE (t) ,
 
ψTcw (t) = arctan 2 −b2 (t)xE (t), a2 (t)yE (t) ,
 
and ψO (t) = arctan kγ (γ(t) − 1) for controller gain kγ and
matrix by Dx(t). Then Dx(·) satisfies the ‘equation of variation’
(ibid.)
d
Dx(t) = D f (x(t), t)Dx(t), Dx(0) = I,
(11) dt
where D f (x, t) is the Jacobian matrix of the map x 7→ f (x, t)
and I denotes the identity matrix. Let Φ(t, t0 ; x), t ≥ t0 , denote
the fundamental matrix of this time-varying linear system, i.e.,
the unique solution to the matrix differential equation
Φ̇(t, t0 ; x0 ) = D f (x(t), t)Φ(t, t0 ; x0 ), t ≥ t0 ,
with Φ(t0 , t0 ; x0 ) = I. The Alekseev formula then states that

T
x2E (t) y2E (t) Z t
γ(t) = + . (12) y(t) = x(t) + Φ(t, s; y(s))`(y(s), s)ds.

IP
a2 (t) b2 (t) 0

In order to ensure that the agent heading ψA follows the desired We first extend this to cover the case when x0 = x(0) , y(0) =
heading ψD , we use a proportional feedback control to com-

mand the agent’s angular velocity ωA (t) defined in (3):
ωA (t) (= ψ̇A(t) ) = kψ (ψD (t) − ψE (t)).
This is the classical proportional control that acts to push ψE (t)
towards ψD (t), with kψ := a prescribed gain chosen such that
this happens on a fast time scale. This allows us to treat ψE (t) ≈
ψD (t). In what follows, we exploit this to identify the two for
CR
y0 . Let Ψε : [0, ∞) 7→ [0,R∞) for ε > 0 be smooth compactly
supported functions with Ψε (x)dx = 1 and Ψε → δ0 in the
sense of distributions as ε ↓ 0, where δ0 is the Dirac delta at 0.
(13) Consider for t ≥ t0 the differential equation
US ỹ˙ ε (t) = f (ỹε (t), t) + `(ỹε (t), t) + (y0 − x0 )Ψε (t), ỹε (t0 ) = x0 .

Using Gronwall inequality, for any T > 0, we have, for L as in

(14)-(15),
AN
the perturbation theoretic analysis, ignoring the approximation
error. We justify this further in a remark at the end of the next max kỹε (t) − y(t)k
t∈[t0 ,t0 +T ]
section. Z t0 +t
≤ eLT max k(y0 − x0 ) Ψε (s)ds + (x0 − y0 )k
M

t∈[t0 ,t0 +T ] t0
5. Perturbation Analysis
ε↓0
→ 0.
For a moving ellipse, we now analyse the idealized coupled
agent-ellipse dynamics using perturbation theory for differen- That is, ỹε (·) → y(·) uniformly on compacts. Alekseev formula
ED

tial equations, specifically a small extension of the Alekseev leads to

‘nonlinear variation of constants formula’ of Alekseev (1961) Z t
(see also Lemma 3 of Brauer (1966)). We begin by recalling ỹε (t) = x(t) + Φ(t, s; y(s))`(y(s), s)ds
this formula in the next subsection. 0
PT

Z t
+ Φ(t, s; y(s))(y0 − x0 )Ψε (s)ds.
5.1. Alekseev formula 0
Consider a differential equation Letting ε ↓ 0, we have the desired ‘generalized Alekseev for-
CE

mula for differing initial data’:

ẋ(t) = f (x(t), t), x(t0 ) = x0 ,
Z t
and its perturbation y(t) = x(t) + Φ(t, 0; y0 )(y0 − x0 ) + Φ(t, s; y(s))`(y(s), s)ds.
AC

0
ẏ(t) = f (y(t), t) + `(y(t), t), y(t0 ) = x0 , (16)

where f is continuously differentiable in its first argument and 5.2. Notation

continuous in the second, and ` is continuous in its first ar- We list here for ready reference the primary notation used in
gument and measurable in the second, satisfying the uniform the analysis to follow.
(w.r.t. t) Lipschitz property: ∀ x, y, t,
• z(t) = [zA (t)T : zT (t)T ]T with zA (t) : agent position at time
k f (x, t) − f (y, t)k ≤ Lkx − yk, (14)
t and zT (t) : ellipse parameters at time t
k f (x, t) + `(x, t) − ( f (y, t) + `(y, t))k ≤ Lkx − yk. (15)
• ẑ = [zA (0)T : zT (0)T ]T
Then both differential equations are well posed. Considering
the map x0 7→ x(t), it can be shown that this map is continuously • z̃A (t) : agent position at t if ellipse parameters are frozen
differentiable (Perko (2001), section 2.3). Denote its Jacobian as z̃T (t) ≡ z∗T
6
 ACCEPTED MANUSCRIPT

• χ(t, z) : stable limit cycle for agent dynamics when target
ellipse is stationary with constant parameters z
• w(t) := χ(t, zT (t)) : desired agent trajectory
• Φ(t, t0 , ; z) : fundamental matrix of linearized unperturbed
(i.e., with immobile target) agent-target dynamics
• Φ̂(t, t0 ; z), Φ̆(t, t0 ; z) : top 2 × 2 resp., 2 × 5 submatrix of Φ
• P(t, z), Q : the periodic, resp. constant matrices in the Flo-
quet decomposition of Φ when target is stationary and the
agent follows a periodic elliptical trajectory
In fact, it traces an ellipse with fixed center and orientation
characterized by z∗T . Then
χ̇(t) = h(χ(t), z∗T ). (21)
Let w(t) := χ(t, zT (t)). This is the equation for the desired
agent trajectory where we have made the parameter z∗T of the
above periodic solution time-varying, albeit on a slower time
scale as per (18). Letting ∇z χ := the Jacobian matrix of χ w.r.t.
the second argument, the chain rule of differentiation yields:
∂χ
ẇ(t) = (t, zT (t)) + h∇z χ(t, zT (t)), żT (t)i
∂t

T
• Θ, θ : 5 × 7 zero matrix, resp., 5 × 1 zero vector = h(χ(t, zT (t)), zT (t)) + ∇z χ(t, zT (t))g(zT (t))
= h(w(t), zT (t)) + η(t), (22)

IP
• ž(t) := [zA (t)T : θ5 ]T
• Γt (x) := the projection of vector x ∈ R2 to żA (t), written as where η(t) := ∇z χ(t, zT (t))g(zT (t)). The coupled dynamics (22)-
(18) can be viewed as a perturbation of (19)-(20), where we

CR
υ x (t)żA (t) where υ x (t) := hx, żA (t)i/kżA (t)k2 ,
impose a common initial condition zT (0) = z̃T (0) for (18) and
• P̄(·, ·) := the 2 × 5 submatrix of P(·, ·) formed by its top (20).
two rows.
5.4. Lyapunov stability

5.3. Two timescale dynamics

We now cast our dynamics in a form suitable for applica-
tion of the generalized Alekseev formula above. Let zA (t) =
US We first prove that the error remains small for all time us-
ing a variant of Theorem 1, p. 339, Hirsch (1989). In Hirsch
(1989), this result is stated for stable equilibria, but the same
proof works here in view of the explicit Lyapunov function for
AN
[xA (t), yA (t)]T and (19) exhibited earlier.
Consider the dynamics in the reference frame F centered at
zT (t) = [xo (t), yo (t), a(t), b(t), θE (t)]T (xo , yo ) for the moving ellipse (shown in Fig. 3). We assume
denote respectively the position of the agent and the vector of that
M

the ellipse parameters (center position, axis lengths and tilt) at ẋo (t), ẏo (t), θ̇E (t), ȧ(t), ḃ(t) = O() (23)
time t given by Algorithm 1 and (4). Let their respective dy- uniformly in t and furthermore, a(t), b(t) ≥ c > 0 ∀t for some
namic laws be given by c > 0. Consider the time-dependent Lyapunov function
ED

!2
żA (t) = h(zA (t), zT (t)), (17) xE (t)2 yE (t)2
e
V(zA , t) := + 2 −1 . (24)
żT (t) = g(zT (t)), (18) a2 (t) b (t)
PT

2 2
for suitably defined h and g. Here  > 0 is small, so the target Let ζ(t) = xaE2(t)
(t)
+ ybE2(t)
(t)
− 1. Then by arguments similar to those
and the convoy parameters move on a slower time scale than of Proposition 1, we have
the agent. (Note that we have allowed parameters other than * " #+
the ellipse centers to vary as well for greater generality. This d e cos(ψD )
V(zA (t), t) = 2VA ζ(t) ∇EA ,
CE

accommodates, e.g., slow variation of inter-vehicle distances in dt sin(ψD )

* " #+
the convoy.) Let D := R4 × [−π, π] where we identify the right δ1 (t)
and left endpoints of the latter interval. Note that h : R2 × D 7→ + 2VA ζ(t) ∇EA , , (25)
δ2 (t))
R2 and g : D 7→ R2 are Lipschitz, so that (17), (18) are well
AC

posed. where ∇EA = ∇EA (t) is the outward normal for the above el-
Also consider the dynamics for  = 0, given by lipse at (xA (t), yA (t)) and kδi (t)k = O(), i = 1, 2 in view of
(23). Label the two summands on the right as ξ1 (t), ξ2 (t) resp.
z̃˙A (t) = h(z̃A (t), z̃T (t)), (19) Then kξ2 (t)k ≤ K for a constant K that can be estimated in
z̃˙T (t) = 0, (20) terms of the problem parameters. Consider a point outside the
current ellipse. As in Proposition 1, ζ(t) > 0 and ξ1 (t) < 0.
i.e., dynamics with z̃T (t) ≡ z∗T (say). This will serve as the Then as long as kξ1 (t)k > K, we have dtd V(z e A (t), t) < 0, imply-
unperturbed dynamics. A variant of (17)-(18) ( given by (22)- ing not only stability (i.e., the trajectory of the agent remains
(18) below) will be viewed as a perturbation of (19)-(20), so that bounded as long as that of the target does), but by the LaSalle
it falls within the purview of the generalized Alekseev formula invariance principle, that the trajectory converges to the set
above. (  * " #+ )
From Proposition 1, we know that this has a stable limit cycle  cos(ψD ) 
(xA , yA ) : Γ := ζ ∇EA , ≤ K̃
(i.e., a periodic solution) χ(t, z∗T ), t ≥ 0, parametrized by z∗T .  sin(ψD ) 
7
 ACCEPTED MANUSCRIPT
where K̃ = 2VKA and we have suppressed the slow time variation
due to (23), treating the set as quasi-static. Since Γ vanishes
only when the agent is exactly on the slowly moving ellipse,
it follows that the agent converges to an O() neighborhood
thereof. The additional fact that while it does so, it also tracks
the moving ellipse with an O() error will follow from the ana-
lysis of the next subsection.
5.5. Error bound
In this subsection we give an explicit formula for the error
that is valid for all time, using the generalized Alekseev formula
above. Consider the linearization of (19)-(20) given by
" #
Dh(zA (t, τ; ẑ), ẑ)
Ż(t) = Z(t), t ≥ τ, (26)
Θ range(Γ s ) is invariant under eQ(zA (s))(t−s) . Thus we have
where Dh(·, ·) is the Jacobian matrix of h in its first argument,
Θ the 5 × 7 matrix of zeros, and ẑ = [zA (τ)T : zT (τ)T ]T . This
is a time-varying linear system. Let Φ(t, τ; ẑ), t ≥ τ, denote its
fundamental matrix satisfying
" #
Φ̇(t, τ; ẑ) =
Dh(Z(t, τ; ẑ))
Θ
Φ(t, τ; ẑ), t ≥ τ,
with Φ(τ, τ; ẑ) = I := the identity matrix, and Z(t) =
Φ(t, τ; ẑ)Z(τ). Then by the generalized Alekseev formula, we
have
zA (t) = w(t) − Φ̂(t, 0; w(0))(w(0) − zA (0))
Zt
M
− Φ̆(t, τ; w(τ))κ(τ)dτ, t ≥ 0.
0 We may choose the reference trajectory w(·) such that (w(0)−
where Φ̂ is the 2 × 2 submatrix of Φ formed by its top two
ED
rows and columns, Φ̆ is the 2 × 7 matrix formed by its top two
rows, and κ(τ) := [η(τ)T , g(zT (t))T ]T . This gives an explicit
expression for the error between the actual agent dynamics zA (·)
and its ideal dynamics w(·).
PT
Let T > 0 denote the period of the periodic trajectory z̃A (·).
From stability theory for periodic orbits (Perko (2001), section
3.5), we deduce that
CE
Φ(t, t0 ; ẑ) = P(t; ẑ)e Q(ẑ)(t−t0 ) −1
P (t0 ; ẑ), t ≥ t0 ,
where Q(ẑ) is defined via eQ(ẑ)T := Φ(T, 0; ẑ), and
AC
P(·, ẑ) := Φ(t, 0; ẑ)e−Q(ẑ)t , P(0; ẑ) = I,
is invertible and periodic with the same period T as z̃A (·). The
continuity of these in the respective arguments is easy to de-
duce. In particular this implies the continuity of the eigenvalues
{λi (ẑ)} of Q(ẑ) in ẑ.
Since the periodic trajectory is stable (cf. the Lyapunov func-
tion argument above), the eigenvalue spectrum of Q(ẑ) will be
as follows. It will have seven eigenvalues. Of these, six will be
at the origin, i.e., zero eigenvalue with multiplicity six. Five of
these correspond to the last five coordinate vectors, i.e., the five
constant components of (20). The sixth zero eigenvalue corres-
ponds to the tangent direction to the periodic trajectory of (19).
The remaining eigenvalue, say λ1 (ẑ), will be real and strictly
negative because the limit cycle is stable. Further, from the
definition of Q, it is clear that the last five rows of eQ(z)t will
have the form [Θ0 : I] where Θ0 is the 5 × 2 matrix of zeros.
Since we are interested in motion in a compact set, we may
confine ourselves to ẑ in a compact set C, whereby
Λ = min(−λ1 (ẑ)) > 0.
ẑ∈C
Let α(t) := υη(t) (t) where ν x (t) := hx, żA (t)i/kżA (t)k2 .
Let Γ s (·) denote the projection onto the linear span of żA (s).
Also recall P̄(·, ·) introduced above. Since żA (s) is the ei-
T
genvector corresponding to zero eigenvalue of Q(zA (s)), i.e.,
eigenvector of eQ(zA (s))(t−s) corresponding to unit eigenvalue,
IP
eQ(zA (s))(t−s) η(s) = eQ(zA (s))(t−s) Γ s (η(s)) + ζ2 (t, s)
CR
= Γ s (η(s)) + ζ2 (t, s),
where kζ2 (t, s)k ≤ K̂e−Λ(t−s) for a suitable constant K̂ ∈ (0, ∞).
Similarly,
US eQ(ẑ)t (w(0) − zA (0)) = Γ0 (w(0) − zA (0)) + ζ1 (t)
with kζ1 (t)k ≤ K̂e−Λt , where we have used the same constant K̂
without loss of generality. Hence
AN
zA (t) = w(t) − (P̄(t, ẑ)Γ0 (w(0) − zA (0)) + ζ1 (t))
Z t
− (P̄(t − s, zA (s))Γ s (η(s)) + ζ2 (t, s))ds
0
(27) for t ≥ 0.
zA (0)) ⊥ z̃˙A (0), so that Γ0 (w(0) − zA (0)) = [0 : 0]T . Thus for
t ≥ 0,
w(t) − zA (t) = ζ1 (t) +
Z t Z t
 α(s)P̄(t − s, zA (s))żA (s)ds +  ζ2 (t, s)ds.
0 0
The three terms on the right behave differently:
1. The first term captures the transient due to initial mis-
alignment and is asymptotically negligible, since kζ1 (t)k ≤
K̂e−Λt → 0. Thus it can be ignored.
Rt
2. Since k 0 ζ2 (t − s)dsk ≤  K̂/Λ = O(), the third term
has only an O() bound. This is a small bounded offset
(28) that persists due to the tracking error caused by the a priori
unknown convoy motion.
3. Note that P∗ := maxt kP(t)k < ∞ because P(·) is periodic.
Thus, since kη(·)k is uniformly bounded, |α(t)| is uniformly
bounded. Also, kżA (s)k = VA . Hence the second term has
an O()t bound. Analysis of the preceding section shows
that it does not build up, i.e. the trajectory still remains
close to the moving ellipse given by the convoy paramet-
ers.
Since we are not analyzing the perturbation with respect to
a fixed periodic trajectory but with respect to a slowly vary-
ing trajectory, the correct way to see that it indeed tracks a
8
 ACCEPTED MANUSCRIPT
time-dependent ellipse in a progressive fashion is to compare
the trajectory with an appropriate ellipse on a time window
of duration ≈ the period of the ellipse, say [t, t + T (t)] where
T (t) is the (perforce bounded) period of ellipse given by VA−1 ×
its perimeter when its semi-major and semi-minor axes are
a(t), b(t) resp. That is, we compare zA (s), s ∈ [t, t + T (t)], with
w(s), s ∈ [t, t +T (t)] defined as before, but with w(t), zA (t) repla-
cing w(0), zA (0) in the foregoing. Then the above analysis leads
to:
max kw(s) − zA (s)k = O(T (t)).
s∈[t,t+T (t)]
This captures the sense in which the convoy is tracked along
a slowly moving ellipsoidal trajectory.
5.6. Further remarks
Remark 2. Another viewpoint is facilitated by the interesting
work of Artstein et al. (2007). Note that for a stationary convoy,
Ve above would be a constant of motion for the unperturbed dy-
namics, characterizing a periodic motion along a fixed ellipse.
Thus the fact that its time derivative is O() makes it a ‘slowly
moving observable’ in the terminology of Artstein et al. (2007).
The results of ibid. then say that we have a near-periodic mo-
tion along a slowly moving ellipse. Our estimates give a quant-
itative dimension to this statement.
US
AN
Remark 3. We have exploited the stable periodic nature of the
nominal ellipsoidal trajectory to use Floquet theory to advant-
age. More generally, in order to get an all time bound using
M
Alekseev formula, one would need stability of the linearized sys-
tem associated with the trajectory. This is usually not easy to
establish unless the slow variation of parameters is sufficiently
slow and certain additional conditions are satisfied, see, e.g.,
ED
Solo (1994). On the other hand, our methodology provides a
general tool for error analysis of perturbed stable periodic tra-
jectories.
PT
Remark 4. We assumed a differential equation model for con-
voy motion, implying in particular a continuously differentiable
trajectory. The analysis can be easily extended to the situation
where it is piecewise continuously differentiable, by applying
CE
the above reasoning to each continuously differentiable piece
in succession. This covers, e.g., sharp turns. This is illustrated
by a simulation study in the next section.
AC
Remark 5. The actual scheme is implemented in discrete time,
introducing discretization error. This can be analyzed along
above lines as well and will be proportional to the discretization
interval. Note that the discretization interval δ must be such
that  = o(δ), because the time scale of a tracking algorithm
must be faster than that of the slowly varying target for tracking
to work at all. The error contributed by discretization will be
O(δ).
Remark 6. A similar comment applies to the error ψE (t) −
ψD (t), which could indeed be incorporated in the above ana-
lysis with extra effort and notation, but we ignore it as it quickly
dies down to negligible value because of the faster time scale
9
involved, a fact borne by our experiments. Other small errors
such as small bounded additive noise can be handled similarly.
In case of bounded additive noise, the error bound will be pro-
portional to the bound on the noise.
6. Simulation and Experimental Results
We have validated the proposed vector field based guidance
law through simulations in Borkar et al. (2017), first for the case
of a stationary ellipse and then for a moving convoy of targets.
We recall these results here, followed by additional simulations
for a non-smooth convoy path. All simulations were performed
T
using MATLAB® 2015b. In addition to these simulation res-
ults, we also present an experimental validation of the proposed
IP
guidance law using a convoy of differential drive robots and
a quadrotor as the aerial agent. We discuss the details of the
hardware setup used and the performance of the guidance law
CR
in tracking the moving elliptical orbit around the ground con-
voy.
6.1. Simulations for following stationary orbits
The proposed law was first validated through simulation for a
stationary ellipse having different orientations θE relative to the
inertial frame in Borkar et al. (2017). For these simulations, the
agent is initialised at an arbitrary initial pose, as shown in Fig.
5. For the first case shown in Fig. 5, the unicycle agent begins
from outside the ellipse and follows it in the counter-clockwise
direction by following ψDccw given by (11). The second case
shown in Fig. 5 has the agent starting within the ellipse, and the
agent following the desired orbit in the clockwise direction by
tracking ψDcw given by (11). In either case, the major and minor
axes are chosen such that the minimum turn radius on the ellipse
V max
Rmin satisfies Rmin > RA , where RA = ωAmax is the minimum
agent turn radius. From Fig. 6, we see that except in the initial
phase where the agent is trying to align with the desired vector
field direction, the Lyapunov function V decreases with time
in either case and agent angular velocity satisfies |ωA | < ωmax .
This confirms that the agent converges to the desired ellipse
and thereafter traces it without violating the angular velocity
constraints.
6.2. Simulations for convoys moving on smooth paths
As in Borkar et al. (2017), to simulate Algorithm 1 and the
guidance strategy for a moving convoy, we select the convoy
path to be a Lissajous curve having a parametric equation
x(s) = A cos(aL s), y(s) = B sin(bL s) (29)
where aL , bL are positive integers, and parameter s ∈ [0, 2π).
The convoy moves along this curve at a constant parametric rate
Vmax
ṡ = q . (30)
A2 a2L + B2 b2L
As a result the speed of the convoy target i is given by Vi =
q
A2 a2L sin2 (aL s) + B2 b2L cos2 (bL s) ṡ ≤ Vmax . For the simula-
tions 1 and 2 in Borkar et al. (2017), we considered the Lis-
sajous curve with A = 1500 meters, B = 1000 meters, aL = 1
 ACCEPTED MANUSCRIPT

Case 1 Case 2
500 200

400 100
1000 5
4 3 l1 (t) × l2 (t) rectangle
0 1 5
300
4 2
2 Regression segment l1 (t)

Distance (meters)
Distance (meters)

200 -100
1 Elliptical orbit at time t

meters
-200 0 3 3 t=100 sec
100
Agent position at time t
0 -300
2 Convoy Path
-400
4
-100 Target position at time t
-1000 1 5
-200 -500
t=2200 sec t=3500 sec
-300 -600
0 200 400 600 -600 -400 -200 0
-1500 0 1500
Distance (meters) Distance (meters)
meters

Figure 5: Case 1: Ellipse parameters: a = 250 meters, b = 150 Figure 7: Simulation 2 snapshots for time t = 100, 2200 and 3500
meters, θE = π4 radians, (xo , yo ) = (300, 200) (in meters) Initial

T
seconds.
agent pose: (xA (0), yA (0)) = (600, −200) (in meters), ψA = π4 radi-
ans. Case 2: Ellipse parameters: a = 350 meters, b = 170 meters,

IP
θE = − π3 radians, (xo , yo ) = (−300, −200) (in meters) Initial agent
pose: (xA (0), yA (0)) = (−300, −200) (in meters), ψA = − π2 radians. wise to ẋA (t) and ẏA (t) of the unicycle agent. For a UAV this is
Controller gains: kγ = 0.5, kψ = 1. VA = 15 m/s and |ωmax | = 0.3

CR
like a velocity disturbance, e.g., due to the wind. The snapshots
rad/sec. in time of the convoy target positions and the encircling ellipt-
ical orbits for simulation 2 are shown in Fig. 7. We observe that
22
x2A yA2
1
Lyapunov function V = a2
+ b2
−1
100
V for Case 1
V × 100 for Case 2 at t = 2200 seconds, the targets are at positions roughly along
the inertial y axis direction and the Algorithm 1 fits the correct

US
50

0
0 20 40 60 80 100 120 140 160 180 200 regression line for these target positions avoiding the regression
seconds
Agent angular velocity ωA
issue discussed in Remark 1.
ωA for Case 1
AN
ωA for Case 2
0.3
rad/sec

0
-0.3 Simulation 1
0 20 40 60 80 100 120 140 160 180 200 γA
seconds γT 1
1 γT 2
γT 3
Figure 6: The Lyapunov function and angular velocity plots for the
M

γT 4
cases shown in Fig. 5. The Lyapunov function for Case 2 is scaled by γT 5
0
a factor of 100 to plot it on the same scale as Case 1. In both cases 145 1000 2000 3000 4000 5000
seconds
|ωmax | = 0.3 rad/sec.
ED

Simulation 2
γA
and bL = 2 shown in Fig. 7. The convoy comprises of five tar- 1
γT 1
get points starting from different parameter values si (0) moving γT 2
γT 3
at a speed bounded above by Vmax = 3 m/sec. From (30) this γT 4
PT

γT 5
corresponds to a constant parametric speed of ṡ = 0.0012. 0
145 1000 2000 3000 4000 5000
Since the speeds of the targets in the convoy are not con- seconds
stant, the curve length separating the target points varies with
CE

time as seen in Fig. 7 at different instances of time. The uni- Figure 8: γA (t) and γT i (t) plots for one traversal of the Lissajous curve
cycle agent’s linear velocity is VA = 15 m/sec, the permissible for simulations 1 and 2.
range of linear velocities is assumed to be given by VAmax = 20
m/sec and VAmin = 10 m/sec, and the maximum permissible
AC

angular velocity is |ωmax | = 0.3 rad/sec. The initial posi- We see that for the target i with position coordinates
tion and orientation of the unicycle agent are assumed to be (xEi (t), yEi (t)) in the ellipse centric coordinate frame, γT i (t) =
x2Ei (t) y2 (t)
(xA (0), yA (0)) = (400, 400) (in meters) and ψA (0) = −π
2 radians. a2 (t)
+ bEi2 (t) ≤ 1 for all time t for both simulations, as shown
The semi-major and semi-minor axes a(t), b(t) of the elliptical in Fig. 8. This validates the claim that the ellipse computed
orbit are chosen according to (4). The controller gain values are based on the outputs of Algorithm 1 (as discussed in Section
kψ = 1 and kγ = 5. 3) always circumscribes the targets. From Fig. 8, for the first
For simulation 1, the targets are spaced close to each other and fifth agents in simulation 2, γT i = 1 for certain time dura-
with si (0) = (i−1)π
20 , and the agent follows the elliptical orbit tions. For example, γT 5 (t) = 1 for t ∈ [0, 145] seconds. This
counter-clockwise. For simulation 2, the targets are spaced implies that the target 5 lies on the ellipse for this time dura-
farther apart with si (0) = (i−1)π
12 , and the agent follows the el- tion. This happens when an agent is on one of the corners of
liptical orbit clockwise. Also, in simulation 2, a small constant the l1 (t) × l2 (t) rectangle and the ellipse parameters a, b selected
velocity disturbance of Vw = 3 m/sec at a heading of ψw = π4 ra- according to (4) yield the minimum area ellipse circumscribing
dians relative to the global reference frame is added component- this rectangle, as shown in Fig. 7 at t = 100 seconds.
10
 ACCEPTED MANUSCRIPT

6.3. Simulation for a convoy moving on a non-smooth path
In addition to the simulations 1 and 2 from Borkar et al.
(2017) we also present a simulation 3 of the proposed monit-
oring strategy for a convoy moving on a non-smooth piecewise
continuous path shown in Fig. 9. The convoy of five targets
moves on this path with a constant speed of VT = 10 m/sec. The
unicycle aerial agent’s linear speed is VA = 50 m/sec and the
maximum permissible angular velocity is |ωmax | = 0.8 rad/sec.
For simulation 3 presented here, the initial position and orient-
ation of the unicycle agent are assumed to be (xA (0), yA (0)) =
(300, 300) (in meters) and ψA (0) = −π 2 radians and control-
ler gain values are kψ = 1.5 and kγ = 10. For all targets
convoy of differential drive robots. The Parrot® AR.Drone 2.01
is used as the aerial agent and the ground convoy comprises
of five Firebird V differential drive robots2 . The experiments
have been conducted in a motion capture environment using a
setup of 10 VICON Vantage V5 cameras3 . Thus the drone and
the convoy robots have been fitted with the reflective markers
as shown in Fig. 11, which are detected by the motion capture
cameras.

T
x2 (t) y2 (t)
i = 1, · · · , 5, γT i (t) = aEi2 (t) + bEi2 (t) ≤ 1 as shown in Fig. 10.
Thus the ellipse computed based on the outputs of Algorithm 1

IP
(as discussed in Section 3) always circumscribes the targets.

CR
300

US
meters

0
Figure 11: Reflective marker placement on the AR.Drone 2.0 and
Firebird V robots for VICON Motion capture
AN
Level One GEP1622 16 port
-300 Wired ethernet connection Gigabit PoEPlus Switch
-400 0 400 PoweroverEthernet cables
meters

Figure 9: Non-Smooth piecewise continuous convoy path for simula-

M

tion 3
AR.Drone 2.0
Simulation 3 HP Z440 Workstation




Motion
running VICON Tracker apture
γA
γT 1 V3.4 Volume
ED

1 γT 2 WiFi
Wired LAN Firebird V Firebird V Firebird V
γT 3 link
connection
γT 4
γT 5
0 Firebird V Firebird V
0 200 400 600 800 Wireless XBee
seconds Network
PT

XBee

VICON
Figure 10: γA (t) and γT i (t) plots for one traversal of the non-smooth Lenovo Z51

Pro S1 Vantage V5
70 Running
convoy path shown in Fig. 9 for simulation 3. ROS Nodes Cameras
CE

For all three simulations, we denote (xE A (t), yE A (t)) as the Figure 12: Block diagram of the experimental setup for motion cap-
position coordinates of the aerial agent in the ellipse cent- ture and control of the AR.Drone 2.0 and Firebird V robots
ric coordinate frame. Thus from the simulation data γA (t) =
x2E (t) y2 (t) Fig. 12 illustrates the experimental setup used. The data
A
+ bE2A(t) ≈ 1 (refer (12)) as shown in Fig. 8 and Fig. 10.
AC

a2 (t)
This implies that the unicucle agent follows the moving ellipse
with a small error as discussed in Section 4.
Remark 7. A video of the simulations 1, 2 and 3 for one com-
plete traversal of the convoy path by the targets can be found at
the following web-links:
https://youtu.be/JDebheE1OR8
http://www.sc.iitb.ac.in/robotics/video/convoysim.mp4
from the VICON cameras is communicated to an HP® Z-440
Workstation via a 16 port Gigabit Ethernet Switch. This com-
puter runs the VICON Tracker Version 3.4 software4 which
processes the camera data and broadcasts the pose information
of the convoy robots and the drone on the Local Area Net-
work. The controllers for the robots and the drone are run as
nodes in a ROS Indigo Igloo5 environment in Ubuntu 14.04 on
1 https://www.parrot.com/global/drones/parrot-ardrone-20-power-edition
2 http://www.nex-robotics.com/products/fire-bird-xii/fire-bird-v-

6.4. Hardware Setup atmega2560-robotic-research-platform.html

3 https://www.vicon.com/products/camera-systems/vantage
We have experimentally validated the proposed convoy mon- 4 https://docs.vicon.com/display/Tracker34/Vicon+Tracker+User+Guide

itoring strategy by implementing it on a quadrotor to monitor a 5 http://wiki.ros.org/indigo

11
 ACCEPTED MANUSCRIPT

a Lenovo® Z51-70 laptop. This Laptop has a Intel® Core i7-
5500U CPU, clocked at 2.4GHz and 16 GB of RAM and has
a Xbee® Pro Series 1 module6 connected to the USB port via a
USB-to-Serial converter board. Each of the Firebird V robots
also has a Xbee® module and the Laptop transmits control com-
mands to these robots using communication packets with ad-
dresses. The drone hosts its own Wi-Fi network and the Laptop
transmits reference control commands to the AR.Drone 2.0 on
this wireless network. The ROS node network for controlling
the Firebird V robots and the AR.Drone 2.0 running in ROS
Indigo Igloo is shown in Fig. 13.
 
y(si )−yri (t)
Pi is given by ψci = arctan . The angular speed of the
x(si )−xir (t)
  
reference point Pi is computed as ψ̇i (t) = dtd arctan ẏẋii . The
linear and angular speeds of convoy robot i are commanded as
Vir (t) = Kvi Dri cos(eψi ) and ωri (t) = Kψi eψi + ψ̇i (t)
respectively, where eψi = ψci − ψri is the heading error, Dri is the
Euclidean distance between the robot i and its reference point,
Kv = 1.1 and Kψi = 1 are controller gains. This controller was
first proposed in Atinc et al. (2013). The bot i controller node
computes these commands at a control loop rate of 10Hz, and
publishes these commands to the f irebird XBee driver node

T
bot 1 controller

VICON
as shown in Fig. 13. This node is a Python script which obtains
bot 2 controller
the command velocity data from the controller nodes of all the

IP
motion
capture
data vicon bridge bot 3 controller f irebird X ee driver
robots and transmits API packets8 with addresses, containing
stream
from LAN bot 4 controller XBee

Network to the linear and angular velocity commands to the corresponding

CR
robot via the Xbee® Network.
command Firebird
bot 5 controller V robots

ardrone autonomy
ardrone controller
WiFi link to 6.6. Control of the Aerial Agent
command the
ROS Indigo Igloo En AR.Drone 2.0
To fly the AR.Drone 2.0, we have used the

US
data logging vironment in Ubuntu
14.04
ardrone autonomy9 ROS package which communicates
Legend with the AR.Drone 2.0 over a Wi-Fi Network link, and allows
ROS Topic for
us to directly issue commands to the quadrotor within the ROS
ROS Topics for ROS Topics for
External Device communicating communicating communicating
Interface VICON locallisa control com elliptical orbit
network framework. To implement the vector field guidance
AN
tion data mands parameters

strategy with the quadrotor we have written a ROS node

Figure 13: The ROS Network running on the Lenovo® Z51-70 Laptop in C++ named ardrone controller. This node receives the
position PA = (xA , yA , zA ) and heading ψA feedback of the
To parse the VICON motion capture data-stream broadcast Drone and the position coordinates of the robots Pri = (xir , yri )
M

on the Local Area Network, we use the vicon bridge7 node in from the vicon bridge node. The inertial velocity of the
ROS which makes pose feedback data of each object (in the drone is computed from the position feedback, using discrete
motion capture volume) available in the ROS network. differences as:
ED

xA (k) − xA (k − 1) yA (k) − yA (k − 1)
6.5. Control of the Convoy Robots V xi (k) = and Vyi (k) =
dt dt
For controlling the Firebird V robot indexed i, the
bot i controller node written in C++, running in ROS, com- where dt = 0.02 sec as the controller loop rate is 50 Hz. As-
putes a reference point moving on the convoy path. For the ex- suming pitch angle θA ≈ 0 and roll angle φA ≈ 0, the body
PT

periments discussed here, the convoy paths are chosen to be Lis- frame velocity feedback of the drone is computed as
sajous curves as they have smooth turns of different curvatures " # " #
in both left and right directions, and they are parametric curves V xb (t) V (t)
= RψA (t) xi .
CE

(with parameter s). For each robot, the bot i controller node
calculates a reference point on the convoy path (29) as
Pi (t) = (xi (t), yi (t)) = (A cos(asi (t)), B sin(bsi (t))) (31)
AC
Vyb (t) Vyi (t)
The ardrone autonomy package allows us to command the
pitch angle θA , roll angle φA , yaw rate ψ̇A and vertical velo-
city żA of the drone as illustrated in Fig. 14. Thus to fly the

moving with constant parametric rate ṡi given by (30) with AR.Drone 2.0 as a unicycle aerial agent described by (3), at a
Vmax = 0.1 m/sec. The bot i controller node also obtains the fixed altitude, the inputs to the AR.Drone 2.0 are commanded
position Pri (t) = (xir (t), yri (t)) and heading ψri (t) feedback from as:
the vicon bridge node. We assume that the robot is modelled
by unicycle kinematics, i.e., θA = Kv (VA − V xb ) (32)
ẋir (t) = Vir (t) cos(ψri (t)), ẏri = Vir (t) sin(ψri (t)) and ψ̇ri = ωri (t), φA = Kv (0 − Vyb ) (33)
ψ̇A = ωA (34)
where Vir (t) and ωri (t) are the linear and angular velocities of
żA = Kz (Zcmd − ZA ) (35)
the robot i. A heading command towards the reference point

6 https://www.digi.com/products/models/xbp24-awi-001j 8 http://cms.digi.com/resources/documentation/Digidocs/90001942-13/
7 http://wiki.ros.org/vicon bridge 9 http://wiki.ros.org/ardrone autonomy

12
 ACCEPTED MANUSCRIPT

zb z_A
Experiment 1

γA
_A γT 1
1 γT 2
γT 3
γT 4
γT 5
0
0 50 100 150 200 250 300
seconds
Experiment 2

yb γA
γT 1
1
θA γT 2
γT 3
γT 4
γT 5
0
xb φA 0 50 100 150 200 250
seconds
300 350 400 450

T
Figure 14: The body fixed reference frame and inputs of the Figure 15: γA and γT i plots for one traversal of the Lissajous curve for

IP
AR.Drone 2.0 both experiments.

where VA is the commanded linear velocity of the unicycle aer-
ial agent, ωA is the commanded angular velocity of the unicycle
agent given by (13), and zcmd is the constant commanded alti-
tude for the drone, Kv = 0.8 and Kz = 0.5 are proportional con-
troller gains. The ardrone controller sends these commands
to the ardrone autonomy node, which in turn transmits these
commands to the AR.Drone 2.0 via the Wi-Fi network link, as
illustrated in Fig. 13.
US
AN
6.7. Experimental Validation
We have conducted two experiments to validate the proposed
guidance law with AR.Drone 2.0 and the Firebird V robots. For
M
CR
followed the moving elliptical orbit around the convoy with a
small error, thus validating the proposed guidance strategy.
7. Conclusions
We have proposed a novel scheme for protection and surveil-
lance by an aerial agent of a convoy moving on an arbitrary
trajectory with minimal regularity assumptions. The scheme
is based on computing a moving ellipse that circumscribes a
bounding rectangle encompassing the convoy. This rectangle is
computed using a simple regression scheme. Then by modulat-
ing the driving vector field of the agent appropriately, it prov-

both experiments the Lissajous curve used for the convoy path
has the parametric equation (29). The convoy of Firebird V
robots moves along the curve with speed bounded above by
Vmax = 0.1 m/sec. For Experiment 1, we choose the Lissa-
ED
ably converges to a trajectory that traverses the moving ellipt-
ical orbit repeatedly. The elliptical orbits prove to be more eco-
nomical than circular ones in terms of coverage. The scheme is
very simple to implement and has been given a rigorous justi-

jous curve having A = 2 meters, B = 1.5 meters, aL = 1 and
bL = 2 (refer (29)), and the robots have initial parameter values
si (0) = (i−1)π
12 and ṡ = 0.02774 (from (30)). For Experiment 2,
we choose the Lissajous curve having A = 1.5 meters, B = 2
PT
fication inclusive of error analysis using Alekseev’s non-linear
variation of constants formula for regular perturbations of or-
dinary differential equations. The supporting simulations also
show performance that matches the theoretical predictions. In

meters, aL = 3 and bL = 2 (refer (29)), and the robots have addition, a hardware implementation is reported, which shows
initial parameter values si (0) = (i−1)π
24 and ṡ = 0.01661 (from a behavior matching the theoretical expectations.
(30)). The AR.Drone 2.0 is commanded to fly with a constant The present analysis is restricted to a single agent. We are in
CE

linear speed of VA = 0.4 m/sec and |ωmax | = 0.7 rad/sec, flying the process of implementing a multi-agent version which will
at an altitude of zcmd = 0.75 meters. be reported in a sequel. Another future direction is to incorpor-
ate realistic noise models and ensure robustness vis-a-vis these.
Remark 8. A video of experiments 1 and 2 for one complete
traversal of the Lissajous curve by the convoy can be found at
AC

the following web-links: Acknowledgements

https://youtu.be/HjORqSJoN9s
http://www.sc.iitb.ac.in/robotics/video/convoyexpt.mp4
As shown in Fig. 15, for the convoy robot i with position co-
ordinates (xEi (t), yEi (t)) in the ellipse-centric coordinate frame,
x2 (t) y2 (t)
γT i (t) = aEi2 (t) + bEi2 (t) ≤ 1 for all time t for both experiments.
This implies that the ellipse computed using Algorithm 1 al-
ways contains all the targets. Fig. 15 shows that the value of
x2 (t) y2 (t)
γ(t) = aE2A(t) + bE2A(t) remains close to 1, where (xE A (t), yE A (t)) are
position coordinates of the AR.Drone 2.0 in the ellipse cent-
ric coordinate frame. This indicates that the AR.Drone 2.0 has
13
The authors thank G. K. Arunkumar and Dr. Pranjal Vyas
for discussions, and Vraj Parikh for developing the ROS driver
for controlling the Firebird V robots. Work done by VSB was
supported in part by a J. C. Bose Fellowship.
Appendix
2
Lemma 1. The area of an ellipse E : ax2 + by2 = 1 circumscribing
2
a rectangle of dimensions l1 × l2 with l1 > l2 > 0 is minimized
by a = √l12 , b = √l22 .
 ACCEPTED MANUSCRIPT

2 2 
 2 2
Proof: For an ellipse√the semi major axis a and minor axis b q ∈ Z+ . Evaluating ddsR2 at s∗ gives ddsR2  ∗ = 3(a a−b ) > 0 and
 s =qπ
are related as: b = a 1 − e2 , wheree ∈ [0, 1) is the eccentri- d2 R  −3(a2 −b2 )
city. The vertices of the rectangle at ±l21 , ±l22 lie on the ellipse. 2 
ds s∗ = (2q+1)π
= b < 0 as a > b > 0. Thus from second
  2
l2 l2 l2 order sufficient conditions of minimization, s∗ = qπ minimizes
Thus 4a12 + 4b22 = 1, which simplifies to a2 = 14 l12 + 1−e2 2 As
√ R and the minimum radius of curvature for the ellipse is Rmin =
2
a result, the area of the ellipse is A = πab = πa2 1 − e2 = R| s∗ =qπ = ba
π 2
√ l2
2 2
4 l1 1 − e + 1−e2 . The first two derivatives of A w.r.t. e
√ 2
are:
dA π e(l22 − l12 (1 − e2 ))
= , References
de 4 (1 − e2 ) 2
3

Alekseev, V. M., 1961. An estimate for the perturbations of the solutions of

d2 A π l22 (1 + 2e2 ) − l12 (1 − e2 )

T
= . ordinary differential equations (in Russian). Vestnik Moskov. Uni. Series I
de2 4 (1 − e2 ) 2
5
Mat. Mech. 2, 28–36.
Artstein, Z., Kevrekidis, I. G., Slemrod, M., Titi, E. S., 2007. Slow observables

IP
From first order necessary conditions for minimization, dA de =
of singularly perturbed differential equations. Nonlinearity 20 (11), 2463–
r 2482.
l 2 2
0, extremizer value e∗ = 0, ± 1 − l22 . Evaluating ddeA2 at e∗ , Atinc, G. M., Stipanovic, D. M., Voulgaris, P. G., Karkoub, M., 2013. Collision-

CR
1 free trajectory tracking while preserving connectivity in unicycle multi-
 agent systems. In: 2013 American Control Conference. pp. 5392–5397.
d2 A  π(l12 − l22 ) Borkar, A. V., Borkar, V. S., Sinha, A., 2017. Vector field guidance for convoy
 = − < 0, monitoring using elliptical orbits. In: Proceedings of 57th IEEE Conference
de2 e∗ =0 4 on Decision and Control, Melbourne. pp. 918–924.

d2 A 
 s
de2 e∗ =± 1− l22
2
l
=
πl13 2 2
(l − l ) > 0
2l23 1 2
1
US Brauer, F., 1966. Perturbations of nonlinear systems of differential equations.
Journal of Mathematical Analysis and Applications 14 (2), 198–206.
Ding, X. C., Rahmani, A. R., Egerstedt, M., 2010. Multi-UAV convoy protec-
tion: An optimal approach to path planning and coordination. IEEE Trans-
actions on Robotics 26 (2), 256–268.
AN
∗ Frew, E., Lawrence, D., 2005. Cooperative stand-off tracking of moving targets
as l1 > l2 > 0. As e ∈ [0, 1), rfrom second order sufficient
by a team of autonomous aircraft. In: AIAA Guidance, Navigation, and
l2
conditions of minimization, e∗ = 1 − l22 minimizes A. Using Control Conference and Exhibit. p. 6363.
1 Frew, E. W., 2007. Cooperative standoff tracking of uncertain moving targets
b2 2 using active robot networks. In: Robotics and Automation, 2007 IEEE In-
the ellipse relation a2
= 1 − e , for the minimum area ellipse,
ternational Conference on. IEEE, pp. 3277–3282.
M

l1 l2 l12
a = b = η for a positive constant η. Substituting this in 4a2
+ Frew, E. W., Lawrence, D. A., Morris, S., 2008. Coordinated standoff tracking
l22 √ of moving targets using Lyapunov guidance vector fields. Journal of Guid-
4b2
= 1, we get η = 2. Hence a = √l12 and b = √l22 . 2 ance, Control, and Dynamics 31 (2), 290–306.
Galloway, K. S., Dey, B., 2015. Station keeping through beacon-referenced
ED

cyclic pursuit. In: American Control Conference (ACC), 2015. IEEE, pp.
2
4765–4770.
Lemma 2. For an ellipse E : ax2 + by2 = 1 with a > b > 0, the
2

2
Hirsch, M. W., 1989. Convergent activation dynamics in continuous time net-
minimum radius of curvature Rmin = ba . works. Neural networks 2 (5), 331–349.
Kapitanyuk, Y. A., Garcia de Marina, H., Proskurnikov, A. V., 2017. Guid-
PT

Proof: For the parametric representation of an ellipse ing vector field algorithm for a moving path following problem. IFAC-
(x(s), y(s)) = (a cos(s), b sin(s)), radius of curvature is PapersOnLine 50 (1), 6983–6988.
Khalil, H. K., 1996. Noninear Systems, 2nd Edition. Prentice-Hall, New Jersey.
 3 Lawrence, D., 2003. Lyapunov vector fields for UAV flock coordination. In:
( ẋ2 + ẏ2 ) 2
3
a2 sin2 (s) + b2 cos2 (s) 2 2nd AIAA Unmanned Unlimited Conference (Workshop, and Exhibit),. pp.
CE

R= = . (36) 1–8.
ÿ ẋ − ẍẏ ab Ma, L., Hovakimyan, N., 2013. Cooperative target tracking in balanced circular
formation: Multiple UAVs tracking a ground vehicle. In: American Control
The first two derivatives of R with respect to s are Conference (ACC), 2013. IEEE, pp. 5386–5391.
q Ma, L., Hovakimyan, N., 2015. Cooperative target tracking with time-varying
AC

2 2 2 2 2 2 formation radius. In: European Control Conference (ECC), 2015. IEEE, pp.
dR 3(a − b ) sin(2s) a sin (s) + b cos (s) 1699–1704.
= ,
ds 2ab Nelson, D. R., Barber, D. B., McLain, T. W., Beard, R. W., 2007. Vector field
path following for miniature air vehicles. IEEE Transactions on Robotics
d2 R 23 (3), 519–529.
= Oh, H., Kim, S., Shin, H.-S., Tsourdos, A., 2015. Coordinated standoff track-
ds2 ing of moving target groups using multiple UAVs. IEEE Transactions on

3(a2 − b2 ) b2 cos4 (s) − a2 sin4 (s) + Aerospace and Electronic Systems 51 (2), 1501–1514.
Oliveira, T., Aguiar, A. P., Encarnação, P., 2016. A convoy protection strategy


2(a2 − b2 ) cos2 (s) sin2 (s) using the moving path following method. In: Unmanned Aircraft Systems
(ICUAS), 2016 International Conference on. IEEE, pp. 521–530.
 q −1 Paley, D., Leonard, N. E., Sepulchre, R., 2004. Collective motion: Bistability
ab a2 sin2 (s) + b2 cos2 (s) . and trajectory tracking. In: 43rd IEEE Conference on Decision and Control
(CDC). Vol. 2. IEEE, pp. 1932–1937.
From the first order necessary conditions for minimization, Perko, L., 2001. Differential Equations and Dynamical Systems, 3rd Edition.
dR
ds = 0, the extremizer value is s∗ = qπ, (2q + 1) π2 for some Springer Verlag, New York.

14
 ACCEPTED MANUSCRIPT

Robin, C., Lacroix, S., 2016. Multi-robot target detection and tracking: tax-
onomy and survey. Autonomous Robots 40 (4), 729–760.
Solo, V., 1994. On the stability of slowly time-varying linear systems. Mathem-
atics of Control, Signals, and Systems (MCSS) 7 (4), 331–350.
Spry, S. C., Girard, A. R., Hedrick, J. K., 2005. Convoy protection using mul-
tiple unmanned aerial vehicles: organization and coordination. In: American
Control Conference, 2005. Proceedings of the 2005. IEEE, pp. 3524–3529.
Zhang, M., Liu, H. H., 2016. Cooperative tracking of a moving target using
multiple fixed-wing UAVs. Journal of Intelligent & Robotic Systems 81 (3-
4), 505–529.
Zhou, B., Satyavada, H., Simone, B., 2017. Adaptive path following for un-
manned aerial vehicles in time-varying unknown wind environments. In:
2017 American Control Conference. pp. 1127–1132.

T
IP
CR
US
AN
M
ED
PT
CE
AC

15