Download as pdf or txt
Download as pdf or txt
You are on page 1of 20

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 23, NO.

5, SEPTEMBER 2015 1807

Broadcast Control Mechanism for Positional


Consensus in Multiagent Systems
Kaushik Das and Debasish Ghose

Abstract— In this paper, a strategy for controlling a group of shown that when there are only two agents, there exists a
agents to achieve positional consensus is presented. The problem broadcast control command [i.e., both agents receive identical
is constrained by the requirement that every agent must be instructions from the central controller (CC)] using which
given the same control input through a broadcast communication
mechanism. Although the control command is computed using both agents can meet at the same location at the same time for
state information in a global framework, the control input almost all initial conditions. However, if the number of agents
is implemented by the agents in a local coordinate frame. is more than two, then the best that the agents can achieve is to
We propose a novel linear programming (LP) formulation that come close within a certain distance of each other (measured
is computationally less intensive than earlier proposed methods. by the radius of the smallest disc that contains all the agents’
Moreover, a random perturbation input in the control command
that helps the agents to come close to each other even for positions). Bretl [4] formulates this problem as an optimization
a large number of agents, which was not possible with an problem that minimizes the radius of the disc, and proposes
existing strategy in the literature, is introduced. The method a solution using the second-order cone programming (SOCP)
is extended to achieve positional consensus at a prespecified technique [5]–[7]. However, using Bretl’s strategy the agents
location. The effectiveness of the approach is illustrated through cannot be made to converge to a point. Once the solution of
simulation results. A comparison between the LP approach and
the existing second-order cone programming-based approach is the SOCP is implemented, no further improvement is possible.
also presented. The algorithm was successfully implemented on Bretl’s paper was in turn motivated by an interesting paper
a robotic platform with three robots. by Donald et al. [8] on the development of untethered,
Index Terms— Broadcast control, consensus, ensemble control, steerable microrobots, where every robot receives the same
LEGO robots, linear programming (LP), multiagent system. power and control signal through an underlying electrical
grid. These robots mainly have two components. One is an
I. I NTRODUCTION untethered scratch drive actuator used for forward locomotion.

T HE principle of using multiple agents is motivated by


the idea that instead of using highly sophisticated and
expensive robots, it may be advantageous in certain situations
The other component is a cantilevered steering arm used
for turning through a frictional contact with the substrate.
These microrobots are simple in construction and can operate
to use a group of small, simple, and relatively cheap robots. only in a local coordinate frame. They do not have sensory
One of the problems that is of paramount importance in mul- capabilities to determine their position and orientation in the
tiagent systems is that of achieving consensus, i.e., achieving global coordinate frame neither do they have the capability to
identical values for some specified subset of the states of the localize themselves in relation to their neighboring objects or
agents. An important consensus problem is when the agents other robots.
may want to converge to a point [1]–[3]. Some applications There are other applications, where these concepts can be
of this class of problem are in midair refueling of the aircraft useful. The microscale and nanoscale robots have very little
or unmanned aerial vehicles (UAVs), exchange of information onboard computation capability. They are mainly controlled by
between agents, targeting a geographical location, and so on. global signals. One can find examples in the literature, where
The problem addressed in this paper was motivated by robots are controlled through global signals. Tung et al. [9]
a paper by Bretl [4], where a control strategy for a group present a class of wireless resonant magnetic microactuator
of microrobots is developed to perform a useful task even that integrates a polymer spring/body structure with electro-
when every robot receives the same control signal. This paper plated ferromagnetic masses. Schürle et al. [10] describe a
considers point robots with simple kinematics. It was holonomic manipulation system consisting of eight individual
Manuscript received May 9, 2013; revised January 22, 2014 and July electromagnets for the generation of magnetic fields and gradi-
16, 2014; accepted November 26, 2014. Date of publication January 23, ents to control objects at the nanoscale. Diller et al. [11], [12]
2015; date of current version August 7, 2015. Manuscript received in develop methods for controlling multiple untethered magnetic
final form December 31, 2014. This work was supported by the Asian
Office of Aerospace Research and Development—Air Force Office of microrobots (called Mag-μ Bots) having all dimensions under
Scientific Research, Arlington, VA, USA. Recommended by Associate 1 mm, without the need for a specialized surface where
Editor M. Mesbahi. Mag-μ Bots are geometrically and magnetically designed
The authors are with the Guidance, Control, and Decision Systems
Laboratory, Department of Aerospace Engineering, Indian Institute to respond uniquely to the same magnetic field.
of Science, Bangalore 560012, India (e-mail: kshikdas@gmail.com; Hasegawa et al. [13] present a novel real-time visual feedback
dghose@aero.iisc.ernet.in). control system for electronic-field controlled paramecium.
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. Steager et al. [14] present electrokinetic and optically
Digital Object Identifier 10.1109/TCST.2015.2388732 controlled microbiorobots (MBRs), where bacteria are
1063-6536 © 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
1808 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 23, NO. 5, SEPTEMBER 2015

attached to the surface of the robots through microfabrication in earlier work), which leads to a simple but elegant linear
process. Felfoul et al. [15] report the successful navigation of programming (LP) formulation. This allows standard com-
magnetotactic bacteria toward regions located inside a solid putational algorithms to be used for obtaining the solution
tumor using a computer controlled set of magnetic coils. and reduces the computation time significantly. Moreover,
Ou et al. [16] present a model predictive control algorithm this formulation also retains the property that the number of
to control the magnetotactic Tetrahymena pyriformis as a decision variables, whose values are to be communicated to
MBR based on a discrete-time plant model. Shirai et al. [17] agents, remains unchanged even when the number of agents
present directional control of thermally driven single-molecule increases. Finally, we also identify some interesting properties
nanocars. Chiang et al. [18] present the development of a related to the positional consensus of agents when the
light driven nanocar. LP-based strategy is applied iteratively.
The above examples are of micro and nanorobots being used It is worth noting that the randomization feature in the
for various applications and some of them are designed to algorithm has some similarity with the random perturbation
respond to identical signals available to all the robots (for used in [26], which proposes a simple discrete time model of n
instance, magnetic fields). This is similar to the broadcast autonomous agents (points or particles) all moving on a plane
mechanism used in [4] and in this paper. with constant speeds but with different headings. Each agent
Broadcast-based control of multiagent systems is also sim- updates its heading using a local rule based on the average of
ilar to ensemble control [19]. There are several works related the headings of its neighbors plus some random perturbation.
to ensemble control that can be found in the literature. Some They show that this nearest neighbor rule can cause all agents
of the works are listed below. Becker et al. [20] present to move eventually in the same direction although each agent’s
a globally asymptotically stabilizing feedback control policy set of nearest neighbors changes with time as the system
for a group of differential-drive robots based on ensemble evolves. Subsequent work (for instance, [27]) attempt to
control. This paper addresses the problem of consensus to a modify the basic control scheme to obtain faster convergence.
specified rendezvous point using innovative control policies. Some of our previous works [28]–[30] present some
Becker and Bretl [21] present a problem of steering a nonholo- preliminary results on the broadcast-based consensus
nomic unicycle in the presence of model perturbation based problem. In [28], an LP-based broadcast mechanisms was
on ensemble control. Becker and Bretl [22] present a steering proposed to achieve positional consensus. In [29], a flag
algorithm of a modified plate-ball system under model pertur- setting-based modified broadcast algorithm was proposed to
bation using ensemble control. Bretl [23] presents a problem of achieve positional consensus for a system of multiagents.
minimum time optimal control problem for a group of robots In [30], an LP-based decentralized algorithm for a group
that move in the same direction but with different speeds. of multiple autonomous agents was proposed to achieve
If we have a centralized system with perfect information positional consensus. Some of our other works [31]–[34] used
then achieving consensus is a trivial matter, since the CC the idea of rectilinear decision domain, which also led to an
can instruct each agent to reach a common consensus point. LP formulation, to achieve rendezvous for multiagent systems.
However, if the communication system is simple and can only This paper is organized as follows. In Section II, we
send messages by broadcast then one may opt for a broadcast consider two agents and show that it is possible to move
protocol, where the CC communicates simple and identical two agents to a common location using identical control.
instructions to all the agents through a broadcast mechanism. In Section III, we formulate an LP problem for minimizing the
It may be further required to impose the additional con- proximity between agents using identical broadcast control.
straint that each agent can interpret the control command only In Section IV, we introduce the notion of iterative solution
in its local coordinate frame or local state space. Only the CC of the problem by repeated use of the LP algorithm and
has access to the global states of the system. Some of these show that introducing a random perturbation in the broad-
constraints are common to other problems of a similar nature cast mechanism leads to point convergence of the agents.
(see [2], [24], [25]). In Section V, we present a modification of the algorithm
Our paper makes several specific contributions. The first to ensure that the swarm of agents converge to a prespeci-
is to propose a strategy that uses the basic Bretl’s model fied point. In Section VI, we present some broadcast-based
with an additional randomization feature that helps in dislodg- algorithms in conjunction with selective enabling of agents.
ing the solution from local minima, allowing a reasonably In Section VII, we show several simulation results that illus-
large number of agents to achieve positional consensus or trate some interesting features of the proposed algorithm.
point convergence on repeated application of the algorithm In Section VIII, we demonstrate the applicability of the
without compromising the broadcast constraint on the control algorithm on a testbed made by LEGO Mindstorms NXT
command. Although exact positional consensus may not be Robots. Section IX concludes this paper with a discussion of
possible when the number of agents is large, the method possible future directions of research.
allows such agents to come very close to each other. The
second contribution is that our method can be extended to II. F ORMULATION AND S OLUTION FOR T WO AGENTS
the case, where the agents can be made to converge to any As in Bretl’s paper [4], point robots with simple kinematics
prespecified point. The third contribution is to propose a are considered. Here, the problem is first formulated in a
rectilinear bounding box approach for the optimization general framework, followed by a specific two agents case
problem (as against the circular bounding box approach used which is used to clarify the assumptions and concepts.
DAS AND GHOSE: BROADCAST CONTROL MECHANISM 1809

Fig. 1. (a) Basic configuration. (b) Configuration with two agents.

Assume that n agents are located on an obstacle-free plane. Letting x 1 f = x 2 f we obtain


A CC has access to the global state of the system which, in  
d
this case, consists of the position (x i ∈ R2 ) and orientation R(θ ) = [R(θ20 ) − R(θ10 )]−1 (x 10 − x 20 ). (3)
0
(θi ∈ (−π, π]) of all the agents, i = 1, . . . , n. The CC
computes a common local control and broadcasts it to all the Equation (3) can be solved easily for d and θ if the
agents for implementation. The local control is in the form inverse on the right-hand side exists, which is true when
of a tuple (θ, d), which is interpreted by each agent in its θ10 = θ20 .
local frame of reference. Here, θ refers to the angle by which This is illustrated in Fig. 1(b). Note that this result is also
each agent changes its orientation, and d is a scalar that refers available in [4] in a slightly modified form.
to the distance by which each agent moves after effecting the The above theorem shows that it is possible to use a
orientation change. Note that due to the broadcast mechanism, broadcast control command to make two agents meet at the
(θ, d) is the same for all the agents. In addition, the local same location simultaneously for almost all initial conditions.
frame of reference for each agent is centered at the agent’s However, the solution is unique and hence the location of the
location and its reference axis is oriented along its current meeting point cannot be chosen arbitrarily.
orientation. As an illustration, see Fig. 1(a), where agents are One can also interpret this result by noting that the final
shown located initially at x i0 with initial orientation θi0 in the meeting point is on the Voronoi edge (equidistant line)
global reference frame. between the two initial positions of the agents [Fig. 2(a)].
If the control command broadcast to all the agents is (θ, d), It can be shown that only one unique point on the Voronoi
then the agents implement it in their local coordinate frame edge satisfies the requirement that the orientation angle change
by each of them changing their orientation by the same angle is the same for both the agents [Fig. 2(b)]. In Fig. 2(a), the
θ and advancing by the same distance d to reach the final point p moves on the equidistant line from −∞ to +∞
destination x i f . Even in this figure it can be seen that by doing and the corresponding orientation angle change θ is plotted
this the agents have come closer to each other. Our objective for the two agents in Fig. 2(b). The two curves intersect
is to determine a tuple (θ, d) such that the agents can achieve at ( p∗ , θ ∗ ) which represents the unique control command
the closest proximity with each other. point.
Theorem 1: For two agents, for all initial conditions of Note that when the number of agents is more than two,
the agents except when θ10 = θ20 , there exists a unique each pair gives rise to a different unique meeting point.
control (θ, d) that makes the agents converge to a point. In general, there does not exist a common control command
Proof: Assuming only two agents, we have to be broadcast, so that all the agents meet at a point. In the
absence of such a command, the best that can be done is to
 
d determine a (θ, d) that brings the agents in closest proximity
x i f = x i0 + R(θi0 )R(θ ) , i = 1, 2 (1) with each other. Note that in this case (θ, d) may not be
0
unique.
where R(α) is the rotation matrix Bretl [4] defines the radius of the smallest circle that
  contains all the final points as the performance measure for
cos α − sin α the algorithm. One of the problems that Bretl addressed in [4]
R(α) = . (2)
sin α cos α was to determine (θ, d) that minimizes the radius of the circle.
1810 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 23, NO. 5, SEPTEMBER 2015

Fig. 2. Voronoi interpretation. (a) Equidistant from two points. (b) Unique point with same orientation change angle.

Bretl solves the problem numerically by formulating it as an


SOCP. This approach has two drawbacks.
1) The equations are nonlinear and consequently the SOCP
may become prohibitively compute-time intensive when
a large number of agents are involved.
2) Even when a solution is obtained, the agents can at best
come as close as possible to each other but they do not
converge to a point, thus defeating the ultimate goal of
achieving perfect consensus.
In the following section, we will propose a LP-based
solution to overcome both the drawbacks without compromis-
ing on the broadcast-based control mechanism.

III. L INEAR P ROGRAMMING F ORMULATION


AND S OME R ESULTS
Consider n points in R2 denoted by (x i , yi ), i = 1, . . . , n.
Fig. 3. Smallest rectangle and the smallest squares containing all the n points.
Define
x max = max{| x i − x j |: i, j = 1, . . . , n} (4)
ymax = max{| yi − y j |: i, j = 1, . . . , n}. (5) smallest square oriented along the global coordinate frame,
and containing all the final positions of the agents. Let this
These are the dimensions of the smallest rectangle (aligned square be centered at z = (z 1 , z 2 ) ∈ R2 . As pointed out earlier,
with the global coordinate frame) containing all the the square need not be unique, and hence, the center of the
points. square also need not be unique.
Define Assuming that all the agents execute the command (θ, d),
their final positions, given by pi, f = [qi1 qi2 θ f ] will be
max{x max , ymax } = 2r (6)
⎡ ⎤
which is also the length of the side of the smallest square cos θi0 − sin θi0 0
(aligned with the global coordinate frame) containing all the pi, f = pi,0 + ⎣ sin θi0 cos θi0 0 ⎦
n points. Note that the smallest rectangle is unique, whereas 0 0 1
⎡ ⎤⎡ ⎤
the smallest square is not unique. This is illustrated by Fig. 3. cos θ − sin θ 0 d
It is also clear that the smallest rectangle is a subset of the × ⎣ sin θ cos θ 0 ⎦ ⎣ 0 ⎦. (7)
smallest square. 0 0 1 θ
Let the initial position and initial orientation of the n
That is
agents be pi,0 = [ pi1 pi2 θi0 ] where θi0 ∈ (−π, π], for all ⎡ ⎤ ⎡ ⎤ ⎡ ⎤⎡ ⎤
i ∈ {1, . . . , n}. As before, we define the control command to qi1 pi1 cos θi0 − sin θi0 0 u1
be broadcast as (θ, d). We define our performance measure ⎣ qi2 ⎦ = ⎣ pi2 ⎦ + ⎣ sin θi0 cos θi0 0 ⎦ ⎣ u 2 ⎦ (8)
as the half length, denoted by r > 0, of the side of the θf θi0 0 0 1 θ
DAS AND GHOSE: BROADCAST CONTROL MECHANISM 1811

Fig. 5. Generalization of the LP algorithm to multisteps.

However, note that starting from an arbitrary set of initial


coordinates the LP may not be able to produce an improved
solution. As an example, consider four agents with positions
and orientations as [(1, 5), 0°], [(1, 2), −262°], [(4, 3), 40°],
and [(4, 3.5), 180°]. The LP will not be able to produce an
improved solution.
Fig. 4. Illustration of how the square reduces in size using the LP solution.
The solution of the LP problem will yield control
instructions that can be broadcast to all the agents using
where u 1 = d cos θ and u 2 = d sin θ are the control variables which the agents will move to a new position or within a new
that replace (θ, d). Note that (8) is a linear equation. Here square region of smaller area, i.e., a square with a reduced
 side length from 2r0 to 2r ∗f , as shown in Fig. 4.
d = u12 + u22 (9) It can be shown that no further improvement of the
−1 u 2 performance (i.e., reduction in r ) can be achieved by repeated
θ = tan . (10)
u1 use of the algorithm. In other words, repeated application of
Now, we formulate the LP problem as the LP algorithm with the new final positions will not reduce
the value of r any further.
min r Suppose, the LP algorithm is represented as an operation L
s.t. − r ≤ pi1 + u 1 cos θi0 − u 2 sin θi0 − z 1 ≤ r (11) on the initial conditions that yields the solution as
−r ≤ pi2 + u 1 sin θi0 + u 2 cos θi0 − z 2 ≤ r,

L( pi0 |i = 1, . . . , n) = u ∗1 , u ∗2 , r ∗f , pi f (16)
i = 1, . . . , n (12)
r ≥ 0. (13) then

The above is an LP problem with the decision vector as L( pi f |i = 1, . . . , n) = 0, 0, r ∗f , pi f . (17)


(r, z 1 , z 2 , u 1 , u 2 ). Note that the decision vector remains the
same irrespective of the number of agents. Only the number That is, there will be no further change in the performance
of inequality constraints increases with the number of agents. measure r . In other words, pi f is a stationary point so far as
In addition, note that z 1 , z 2 , u 1 , and u 2 are free variables the LP algorithm is concerned.
and can take both positive and negative values. Finally, after We can generalize this process by assuming that each step
obtaining u 1 and u 2 from the LP solution, the (θ, d) command in the iteration is denoted by the index k, with the first step
can be obtained using (9) and (10). An illustration of this in the iteration as k = 1. Then, we can represent the process,
process is shown in Fig. 4. as shown in Fig. 5.
After executing the movement command the distance Theorem 2: For k ≥ 2, u ∗1,k = u ∗2,k = 0 and pi,k+1 = pi,k ,
between agents ai and a j along the x- and y-axis will be i.e., repeated use of the LP solution on subsequent positions
will not reduce r .
dxi j = |( pi1 − p j 1) + d(cos(θi + θ ) − cos(θ j + θ ))| Proof: We will prove this by contradiction. Suppose for
= |( pi1 − p j 1) + dC| (14) a given initial condition pi01 we have r = r as the measure
0
d yi j = |( pi2 − p j 2 ) + d(sin(θi + θ ) − sin(θ j + θ ))| of proximity of the agents. Applying the LP algorithm, we
= |( pi2 − p j 2 ) + d S| (15) obtain u 1∗ 1∗ 1
1 , u 2 , r1 , and the final positions as pi f . Now,
considering the initial conditions as pi0 = pi f and applying
2 1
where C = [cos(θi + θ ) − cos(θ j + θ )], S = [sin(θi + θ ) − the LP algorithm let us assume that we get u 2∗1  = 0, u 2  = 0,
2∗
sin(θ j + θ )], and i, j ∈ {1, . . . , n}. Thus, the resulting r2 < r1 , and the final positions as pi f . Then, let us define
2
distance along the x-axis and y-axis are dictated by the value
of C and S. û i = u 1∗
i + û i , i = 1, 2, where û i is u i expressed in
2∗ 2∗ 2∗

If (x i,k , yi,k ) and (x i,k+1 , yi,k+1 ), i = 1, . . . , n are two the initial local reference frame in which u 1∗
i is represented.
sets of n points obtained at the kth and k + 1th iteration, It can be shown that if in the first step û i is used it would
respectively (where each of these sets of points are obtained yield a r = r2 < r1 , which implies that r1 was not a solution
as the result of executing LP). Then, ri,k+1 ≤ ri,k . to the LP. This leads to a contradiction.
1812 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 23, NO. 5, SEPTEMBER 2015

Fig. 7. Perturbed case.

The proof for more than one interior agent is an easy


extension, where we can define d < (1/4) min{ηi1 , ηi2 ,
ηi3 , ηi4 }, ∀i , where i defines an interior agent.
This theorem shows that if there are only two agents on the
boundary of the smallest rectangle then the LP will certainly
yield a better solution.

Fig. 6. Illustration for the proof of Theorem 3. IV. ACHIEVING C ONSENSUS U SING P ERTURBATIONS
The solution of the LP problem will yield control instruc-
Before proving some general results on formation of agents tions that can be broadcast to all the agents, using which
after the LP is executed we will state a lemma that will be the agents will move to a new position or within a new
useful to prove the main results. square region of a smaller area, i.e., a square with a reduced
Lemma 1: If C = cos(θi + θ ) − cos(θ j + θ ) then there side length. It has been shown (Theorem 2) that no further
exists a θ such that C < 0 where θi , θ j ∈ (−π, π] and θi = θ j . improvement of the performance (i.e., reduction in r ) can be
Proof: Let θ = π − (θi − θ j /2) − φ. After replacing θ achieved by repeated use of the algorithm. In other words,
in C we get C = −2 sin(θi − θ j /2) sin(φ). It is clear from repeated application of the LP algorithm with the new final
the expression that we can make C < 0 by choosing φ positions will not reduce the value of r any further.
appropriately. Now, consider a case where the agents receive a broadcast
An interesting observation is as follows. command containing the LP solution and a command to
Theorem 3: After executing the LP, the number of agents randomly perturb the final orientation angle after the LP
on the boundary of the smallest rectangle containing all the solution has been implemented.
agents is more than two. This will be referred to as the perturbed case as against
Proof: We will prove this by contradiction. Let the the algorithm discussed in the previous section, which can be
number of agents on the smallest rectangle be two. Thus, referred to as the unperturbed case. The multistep process for
the two agents will be located at diagonally opposite corners. the perturbed case is shown in Fig. 7, where the orientation
The other agents will be in the interior of the rectangle. For the angle, after implementing the LP solution, is perturbed by each
sake of simplicity, we will consider only three agents. This is agent as follows:
illustrated in Fig. 6, where M N is the Voronoi edge (equidis-
θ̂i,k+1 = θi,k+1 + νi,k+1 . (18)
tant line) between agents at A and B. According to Theorem 1,
there always exists (θ, d) such that the agents can meet at As an example, consider three agents with positions and
an unique point on the Voronoi edge (M N). Let the unique orientations [(1, 1), 45°], [(5, 4), 135°], and [(2, 6), −45°].
meeting point be F. Let us define a small positive quantity d After executing the LP, the generated broadcast com-
such that d < (1/4) min{η1 , η2 , η3 , η4 }, where η1 , . . . , η4 mand will be (θ = 15.92°, d = 2.58). Therefore,
are as shown in Fig. 6. Consider that the agents use a common the agents will move to new positions [(2.25, 3.25), 60.92°],
(θ, d) and move from their current positions A and B such [(2.75, 5.25), 150.92°], and [(4.25, 4.75), −29.08°]. The center
that the new position of agents are C and D, respectively. of the square is at (3.25, 4.25) and the length of its side is 2.
Similarly, the interior agent will also move according to the Further application of the LP algorithm will give zero solution
same (θ, d). Note that all the agents will remain within using these updated positions and orientations. After introduc-
the circles of radius d centered at their previous positions. ing perturbation of 50°, 100°, and 70° in the orientations of
In particular, it can be seen that there exists a d such that the the agents, and using these perturbed orientations, the LP gen-
interior agent will remain inside the interior of the smallest erates the broadcast command (θ = 273.5°, d = 0.2602). The
rectangle formed from the next positions of the three agents. agents will now move to new positions [(2.49, 3.36), 25.68°],
Now, AB F ∼ C D F and so C D AB and C D < AB. [(2.49, 5.32), 165.7°], and [(4.44, 4.56), 315.6°]. The center
This implies that further reduction in size of the rectangle is of the new square is at (3.46, 4.34) and the length of its side
possible or the rectangle formation with only two agents at the is 1.95. This numerical example has been illustrated in Fig. 8.
boundaries is not optimal. This theorem can also be proved This example shows that introducing a perturbation helps in
using Lemma 1 and (14), selecting a d as given above. moving the solution away from a local minima.
DAS AND GHOSE: BROADCAST CONTROL MECHANISM 1813

Fig. 8. Example of the perturbed case.

Any perturbation in one robot has the potential to move


the agents away from the local minimum and subsequent
Fig. 9. Illustration of agents gathering toward a specific point (z 1 , z 2 ).
application of LP will yield a solution which is at least as
good as, or better than, the previous local minimum. This is define the smallest square that contains all the agents and has
true for perturbations to any number of robots. It is easy to its center at (z 1 , z 2 ). Note that if we consider only the square
see that the above statement is true since the perturbations that contains all the agents then its center will most likely be
are applied only to the directions and not to the positions of different from (z 1 , z 2 ). Here, the input to the LP are the initial
the agents, which implies that the solution (0, 0), which will positions, the initial orientation, and the meeting point. We can
keep the agents at the same positions, is a valid solution to formulate the modified LP problem as
the perturbed problem.
To generalize this process, the perturbation angle νi,k+1 is min r g
represented by s.t. − r g ≤ pi1 + u 1 cos θi0 − u 2 sin θi0 − z 1 ≤ r g (20)
νi,k+1 = ηi β (19) −r g ≤ pi2 + u 1 sin θi0 + u 2 cos θi0 − z 2 ≤ r g ,
i = 1, . . . , n (21)
where ηi is a random number generated by each agent inde-
pendently and β is a scaling angle which is common to all r g ≥ 0. (22)
the agents. The scaling angle can be set manually and η can The above is an LP problem with the decision vector as
be generated through various distributions. To ascertain if the (r g , u 1 , u 2 ). Note that the number of decision variables has
nature of the distribution effects the results, we consider two reduced over the previous formulation, where the rendezvous
cases. In one, η is generated as a uniformly distributed random point was not specified. The above formulation will try to min-
number and in the other, it is generated through a normal imize the square centered at the (z 1 , z 2 ) with side length 2r g .
distribution. However, in both the cases, we ensure that the The above formulation is equivalent to
variance remains the same.
min r g

V. ACHIEVING P OSITIONAL C ONSENSUS s.t. − r g ≤ pi1 + u 1 cos θi0 − u 2 sin θi0 ≤ r g (23)
AT A D ESIRED P OINT
−r g ≤ pi2 + u 1 sin θi0 + u 2 cos θi0 ≤ r g ,
In the previous section, we considered the problem of i = 1, . . . , n (24)
positional consensus, but did not have control over the position
rg ≥ 0 (25)
at which the agents meet.
= p
Suppose we have a prespecified meeting point defined where pi1 i1 − z 1 and pi2 = pi2 − z 2 . It will try
as (z 1 , z 2 ). This can be achieved as a consensus point by to minimize the square centered at (0, 0) in the modified
modifying the previous formulation. In the modified formu- coordinate system with side length 2r g . It is clear that,
lation given below, we define the performance measure to be |, | p |}. For further analysis, we will use this
r g ≤ max{| pi1 i2
minimized as the length of the side of the smallest square, formulation.
centered at the specified meeting point (z 1 , z 2 ), which contains The following interesting properties of the algorithm are
all the agents. This is different from the problem formulated relevant as they are in contrast to the results for the previous
in the previous section, where we considered the smallest case where the final point was not specified.
square that contained all the agents. Here, (z 1 , z 2 ) is fixed, Property 1: For two agents, for the specified final point
unlike in the previous cases, where (z 1 , z 2 ) was also a decision case, even when θ10 = θ20 , there may not exist a control
variable. The illustration of the process can be observed in (θ, d) using which point convergence can be achieved.
Fig. 9, where initial positions of the agents are ai , a j , ak , Proof: From Fig. 2(a), it is clear that the meeting point for
and al . The final positions of the agents are ai , a j , ak , and al , two agents is on the Voronoi edge, which may not go through
respectively. The specified rendezvous point (z 1 , z 2 ) is used to the origin (0, 0). Now, in the specified consensus point case,
1814 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 23, NO. 5, SEPTEMBER 2015

VI. S ELECTIVE E NABLING OF AGENTS


In the formulation given in the previous section, when the
number of agents is high, LP will take a longer time to
bring the agents close. However, in case of smaller number
of agents the convergence time is also small. Hence, in case
of higher number of agents, for faster convergence, if one
can choose a group of smaller number of agents at each step
and execute LP for that group, then it may be possible to
bring a large number of agents to a point in a small time.
This is the motivation behind the selective enabling approach
in which a selected subset of agents are enabled to execute
the LP command. The choice of the agents that belong to
the subset at each step was made by a flag setting approach.
Diller et al. [11], [12], [35] deal with remotely addressable
magnetically actuated microrobots that have their magnetic
Fig. 10. Example of the specified meeting point.
moments selectively turned ON or OFF by application of a large
magnetic field pulse. This is conceptually similar to the idea
used here. Although the objective of switching the magnetic
field in their work is different from what we propose here.
the meeting point is always the origin. Therefore, the two The main idea behind this approach is that each agent ai
agents cannot meet at a point for all initial conditions. randomly generates a flag f i , that can take one of two values.
Property 2: For the specified final point case, the center of One is high ( f h ), denoting an active agent, and the other is
the square (0, 0) need not be within the smallest rectangle low ( fl ), denoting a passive agent. The set of agents that have
made by the agents along the reference frame after the LP high flags (active) is denoted by Ra and the set of agents that
solution is implemented. have low flags (passive) denoted by R p .
Proof: This can be prove through an example. Let us con- At time t let the set of active agents be Ra (t) = {i | f i = fh }
sider three agents with positions and orientations [(1, 1), 45°], and the set of passive agents at time t be R p (t) = {i | fi = fl },
[(5, 4), 135°], and [(2, 6), −45°] and the specified meeting where i ∈ {1, . . . , n}. An active agent will implement the
point is (0, 0). After executing the LP, the broadcast command broadcast command and change its state, whereas a passive
is (θ = 1.37°, d = 1.44). Therefore, the agents will move agent will ignore the command and remain stationary. The
to new positions [(2, 2.04), 46.37°], [(3.95, 5), 136.37°], and flag value for each agent may or may not be observed by
[(3.04, 5), −43.63°]. The center of the square is at (0, 0) and the CC. The CC has information of position and orientation
the length of its side is 5 (Fig. 10). Note that the origin (0, 0) of all the agents (active and passive) at each step. The control
is outside of the smallest rectangle containing all the agents command (u ∗1 , u ∗2 ) computed by the CC will be broadcast to
and aligned along the reference frame. all the agents, but only the active agents will move.
Property 3: The side length of the square (2r g ), in case of Let the position and orientation of the active and passive
specified rendezvous point, is greater than or equal to the side agents be ( pa,k , θa,k ) and ( p p,k , θ p,k ), respectively. After
length of the square in case of nonspecified rendezvous point, receiving the broadcast command the new positions of the
i.e., r g ≥ r for the same initial positions and orientations of active and passive agents are
the agents. ⎡ ⎤
Proof: This can be easily proved by observing that the cos θi,a,k − sin θi,a,k 0
solution of the LP with the specified rendezvous point will sat- pi,a,k+1 = pi,a,k + ⎣ sin θi,a,k cos θi,a,k 0 ⎦
isfy all the constraints of the LP with nonspecified rendezvous 0 0 1
⎡ ⎤⎡ ⎤
point. This can be seen from (11)–(13) to (23)–(25). Thus, the cos θ − sin θ 0 d
solution of the LP for the nonspecified rendezvous point will × ⎣ sin θ cos θ 0 ⎦ ⎣ 0 ⎦ (26)
yield a solution which is at least as good as the solution for 0 0 1 θ
the specified rendezvous point. pi, p,k+1 = pi, p,k (27)
In the specified final point case also, no further improvement
of the performance (i.e., reduction in r g ) can be achieved by where (θ, d) is the broadcast control command.
repeated use of the LP algorithm when the angles are not
perturbed. A. Algorithms
Theorem 4: For k ≥ 2, u ∗1,k = u ∗2,k = 0 and pi,k+1 = pi,k . In this section, we propose five different algorithms obtained
Proof: The proof is similar to Theorem 2. by combining the two randomization mechanisms described in
To reduce r g further a perturbation in the orientation of the previous section in different ways. Their relations with
the agents is required. Therefore, the agents receive broadcast each other are shown in Fig. 11 and the algorithms are
command containing the LP solution and a command to described as follows.
randomly perturb the final orientation angle after the LP 1) Perturbation (P): This is the basic perturbation
solution is implemented. algorithm described in the previous section.
DAS AND GHOSE: BROADCAST CONTROL MECHANISM 1815

Fig. 11. Taxonomy of the randomization mechanism-based algorithm.


TABLE I
M AIN F EATURES OF A LL THE A LGORITHMS (Y = Y ES ; N = N O )

2) Flag Observed (FO ): In this case, the flags of the agents 4) Perturbation With Flag Observed (P, FO ): In this case,
will be set randomly but the orientation will not be perturbed. the active agents in Ra (t), whose flag is high, will perturb its
The CC will observe the flags as well as the positions orientation randomly. Then, the CC will observe the positions,
and orientations of all the agents and compute the control orientations, and flags of all the agents. However, the LP will
command for the active agents only. The CC will broadcast be executed using the position and orientation of the active
control command and all active agents will move from their agents only. The control command will be broadcast to all the
respective positions. Thus agents. After receiving the control command only the active

agents will change their state. The passive agents will remain
L( pi1,a,k , pi2,a,k , θi,a,k ) = u ∗1 , u ∗2 . (28)
stationary. Thus
Note that in this case, there is no perturbation in the agent

orientation. Even then, the LP may give nonzero solution L( pi1,a,k , pi2,a,k , θ̂i,a,k ) = u ∗1 , u ∗2 (30)
at each step. This is because the group of active agents is
where θ̂i,a,k = θi,a,k + νa,k .
different at each step due to random flag setting.
5) Perturbation With Flag Not Observed (P, FN ): In this
3) Algorithm (FN ): In this case too, the flags of the agents
case also, the active agents will perturb their orientations
will be set randomly and orientation will not be perturbed.
randomly. However, the CC will not observe the flag infor-
However, unlike the (FO ) case, the CC will not observe the
mation and will compute the control command (θ, d) using
flags and will only observe positions and orientations of all
positions and orientations of both the active agents and passive
the agents and compute the control command for all of them.
agents. The control command will be broadcast and only
The CC will broadcast the control command to all the agents,
the active agents will move from their respective positions.
but only the active agents will implement these and move.
Thus
Thus

L(( pi1,a,k , pi2,a,k , θi,a,k ), ( pi1, p,k , pi2, p,k , θ p,k )) = u ∗1 , u ∗2 . L(( pi1,a,k , pi2,a,k , θ̂i,a,k )( pi1, p,k , pi2, p,k , θ̂i, p,k )) = u ∗1 , u ∗2
(29) (31)
Even here, although there is no perturbation in the orientation, where θ̂i,a,k = θi,a,k + νa,k .
the LP may give nonzero solution at each step as the group The main features of the above algorithms are summarized
of active agents is different due to random flag setting. in a tabular form in Table I.
1816 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 23, NO. 5, SEPTEMBER 2015

Fig. 12. Consensus with three agents. (a) Trajectory of the agents. (b) Zoomed-in view of the ending locations of the agents. (c) r versus iteration.
(d) Convergence in the X-coordinate. (e) Convergence in the Y -coordinate. (f) θ versus d (number shows the iteration number).

VII. S IMULATION R ESULTS both the cases. These results are given for three agents and
This section gives simulation result for all the cases obtained by averaging over 200 trials. From these results,
discussed previously and provides very important insights we conclude that using larger scaling angle reduces r faster
into the broadcast-based algorithm. All the simulations have and yields shorter path lengths. It can also be observed
been carried out in a computer with Intel Core 2 CPU, that the performance of the normal distribution and uniform
1.86 GHz, and 3-GB RAM. distribution are similar. In the rest of this paper, we will use
only normal distribution.
It is obvious that the computation time and the number
A. Nonspecified Point Consensus of iterations should rise with the number of agents. This is
In the first set of simulations, we consider three agents expected since the complexity of the algorithm is the same as
with initial conditions [(1, 1), 45°], [(5, 4), 135°], and that of the LP algorithm. This can also be observed in Fig. 14,
[(2, 6), −45°] as the initial position and orientation angle where the half length of the square after 300 iterations has been
of the three agents. Using the perturbation technique, with plotted against the number of agents. As the number of agents
normal distribution for η and a scaling angle of β = 120°, the increases, the rate of convergence of the algorithm decreases
agents converge to a point after a few iterations (Fig. 12). The substantially.
variation in the X and Y coordinates of the three agents against In Table II, the effort required by the agents as they move
the number of iterations are also shown. The convergence toward the rendezvous point is given. The results are given
criterion for terminating the simulation was the value of r for a three agents case. The first column gives the sum of the
becomes <0.01 (denoted by rt ). The variations of d against distances traveled by each robot up to the kth iteration. This is
θ are also shown in Fig. 12(e), where the agents are seen a measure of the rectilinear motion effort by the robots. Note
to move by infinitesimally small distances but changing their that multiplication by the number of robots n(= 3) is done
heading drastically. because the distance moved by each robot at each iteration is
In Fig. 13, we plot the average number of iterations needed, the same for each robot. Similarly, the second column gives
and the average length of path traveled by each agent, as a the sum of the angular rotations in degrees executed by each
function of the scaling angle β for the two cases when the robot up to the kth iteration and represents the rotational effort
random number η is generated by a uniform distribution or by the robots. The third column represents the sum (summed
by a normal distribution using the same variance (0.33) for over the number of robots) of the Euclidean distance covered
DAS AND GHOSE: BROADCAST CONTROL MECHANISM 1817

Fig. 13. Convergence results for three agents. (a) Average number of iterations. (b) Standard deviations for number of iterations: uniform distribution.
(c) Standard deviations for number of iterations: normal distribution. (d) Average length of path. (e) Standard deviations for path length: uniform distribution.
(f) Standard deviations for path length: normal distribution.

TABLE II
R ECTILINEAR AND ROTATIONAL E FFORT BY ROBOTS (AVERAGE H AS
TAKEN OVER 200 S IMULATIONS )

Fig. 14. r versus number of agents after 300 iterations.

by each robot from its initial position at the kth iteration


and is given by D = ni=1 Dik , where Dik is the Euclidean
distance of the agent i at the kth iteration, from its initial
position. This is not a measure of the effort by the robots
but can be used as benchmark to ascertain how much actual
displacement from their initial positions is achieved by the
robots. The table shows that after the first few iterations,
the displacement achieved, denoted by D, and the rectilinear
motion effort retains a linear relationship, while the angular or implemented by a randomly selected subset containing
effort keeps rising at a rapid rate. This shows that as the robots increasing number of agents. For example, in the first instance,
get closer to the rendezvous point there is a tendency by them only one randomly selected agent executes the LP at each
to rotate by large angles. iteration till rendezvous is achieved. The agent, which executes
Next, a simulation is done with 10 agents in which the LP, is selected randomly at each iteration. Similarly, in
although the LP is solved for all the agents, it is executed the next simulation, two agents are randomly selected at each
1818 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 23, NO. 5, SEPTEMBER 2015

determination of the next position of the agents was decided


by LP in one case and by SOCP in another. Initially, the
number of agents considered was three. The perturbation was
done using normal distribution and the average was taken over
200 simulation runs. Table III and Fig. 18 give a comparison
between the computation time of LP and SOCP for different
scaling angles used for implementing the perturbation. It can
be seen that the SOCP-based algorithm takes a considerably
longer time than the LP-based algorithm.
The comparison between LP and SOCP in terms of number
of iterations taken to converge has been done using a scaling
angle of 120°. The number of iterations taken to converge as
Fig. 15. Number of iterations versus fixed number of randomly chosen active
agents. the number of agents varies and the corresponding time taken
to converge is given in Table IV. The number of iterations
and computation times were averaged over 200 simulations.
iteration. In Fig. 15, the number of iterations to converge for Fig. 19 gives a plot of the average computation time and
each number of randomly selected agents is given. It can be the standard deviations. It shows that the LP-based broadcast
observed that the required number of iterations decreases up algorithm performs better than SOCP.
to a certain number of agents, after which it increases. Hence,
it appears that the number of agents implementing the LP
solution is an important factor in determining the speed of D. Selective Enabling Approach
convergence of the algorithm. We will show other results with The simulation results for the basic perturbation-based
similar ideas when we address the selective enabling class of broadcast algorithm (which was denoted by P) was discussed
problems later in this section. However, it would be prudent in the previous section.
to note that the number of agents chosen to be in the subset We now consider Algorithm FO , in which the CC observes
of active agents is fixed here, although the agents themselves the flag values and there is no perturbation in orientation. Here,
are selected randomly. In the selective enabling case, this five agents were considered to demonstrate the result. The
number cannot be enforced as each agent chooses to be active convergence criterion for terminating the simulation is r ≤ 0.1.
or inactive based upon its own randomized algorithm. The simulation results are given in Fig. 20 and show that the
agents do not converge to a point. In fact, they seem to be
B. Specified Consensus Point sometimes moving away from each other. The reason for this is
that the CC computes the control command based on the active
To demonstrate that the agents can meet at any prespecified
agents’ information only and the computation is independent
point, we consider the initial positions and orientation angles
of the passive agents’ information. This implies that the active
as [(1, 1), 45°], [(5, 4), 135°], and [(2, 6), −45°] for the three
agents move to a point defined by the LP solution without
agents. The meeting point is set as the origin (0, 0). The results
taking into consideration the presence of the passive agents.
are given in Fig. 16. Comparing with Fig. 12(b), which had the
This leads to the active agents moving away from the passive
same initial conditions and did not have a specified rendezvous
agents. Therefore, after repeatedly using this algorithm, the
point, we can observe that the initial square in the specified
agents do not converge to a point.
point case is expectedly larger and also takes a larger number
Next, Algorithm FN is considered, in which the CC does
of iterations to converge. The simulation results with 20 agents
not observe the flags and there is no perturbation of angles.
is given in Fig. 17. It is seen that even after 3000 iterations the
Here, 5, 10, 15, and 20 agents have been considered. The
agents do not converge to a point but the value of r reduces
simulation results are given in Fig. 21. The log scale plot
from 7.5 to 3 in less than 1000 iterations.
shows the details for the smaller number of iterations with
greater clarity and shows that the value of r need not decrease
C. Performance Comparison Between LP and SOCP monotonically. The agents converge to a point for up to
Bretl [4] formulates the robot gathering problem as an 15 agents. The simulation results also show that this algorithm
optimization problem that minimizes the radius of the disc. does not work for 20 agents and higher but it still gives better
The proposed algorithm was solved using the SOCP tech- result than algorithm P and substantially better result than FO
nique [5]–[7], which is a nonlinear programming technique since it moves the passive agents in a way that keeps them
and is computationally time intensive. On the other hand, the close to the active agents.
robot gathering problem in our work was formulated using Next, Algorithm (P, FO ) is considered. Note that in this
an LP approach, which has better convergence properties in case, the control command is computed by considering only
standard problems, especially with smaller number of robots. the active agents. For simulation, five agents are considered.
In this section, a comparison between algorithms based on The simulation result is given in Fig. 22. The simulation results
SOCP and LP technique is presented for this problem for show that the agents do not converge to a point even for small
the nonspecified rendezvous point case. The problem formu- number of agents. The reason is the same, as discussed in the
lation with perturbation was used in both cases. Only the FO algorithm.
DAS AND GHOSE: BROADCAST CONTROL MECHANISM 1819

Fig. 16. Consensus with three agents at specified point. (a) Trajectory of the agents. (b) r versus iteration.

Fig. 17. LP with 20 agents. (a) Trajectory of the agents. (b) Zoomed-in view of the ending locations of the agents. (c) r versus iteration.

TABLE III
C OMPUTATION T IMES OF LP- AND SOCP-BASED B ROADCAST
A LGORITHM W ITH T HREE A GENTS FOR D IFFERENT S CALING
A NGLES (AVERAGE H AS TAKEN OVER 200 S IMULATIONS )

Fig. 18. Comparison of computation times of LP- and SOCP-based broadcast


algorithm with three agents for different scaling angles.

Next, Algorithm (P, FN ) is considered. The CC does not


observe the flags and considers all the agents, while computing
the control command. Here, 5, 10, 15, and 20 agents were
considered for the simulation. The results are given in Fig. 23.
We also observe that this algorithm is able to bring higher agents in the LP computation keeps the agent swarm together.
number of agents to a point. The reason why this algorithm Thus, not having access to the flag information (by the CC)
works better than (P, FO ) is that consideration of the passive actually helps in better convergence.
1820 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 23, NO. 5, SEPTEMBER 2015

TABLE IV
C OMPARISON B ETWEEN LP- AND SOCP-BASED B ROADCAST A LGORITHM IN T ERMS OF N UMBER OF I TERATIONS AND
C OMPUTATION T IMES (AVERAGE H AS TAKEN OVER 200 S IMULATIONS )

Fig. 19. Comparison based on computation time and number of interactions between the LP- and SOCP-based algorithms.

Fig. 20. Algorithm FO . (a) Zoomed-in view of the starting positions. (b) Trajectory of the agents. (c) r versus iterations.

Fig. 21. Algorithm FN . (a) r versus number of iterations in log scale. (b) r versus number of iterations in linear scale. (c) Trajectory for five agents.

It is quite evident that not all the algorithms are able to bring terms of the number of iterations and computation times for
the agents to a point. Table V shows performance comparison 5, 10, 15, 20, 50, and 100 agents. Here, we consider rt = 0.1.
of all the algorithms, for unspecified rendezvous point, in Note that NC denotes those cases, where r is not less than
DAS AND GHOSE: BROADCAST CONTROL MECHANISM 1821

Fig. 22. Algorithm (P, FO ). (a) r versus number of iteration. (b) Trajectory of the agents.

Fig. 23. Algorithm (P, FN ). (a) r versus number of iterations in log scale. (b) r versus number of iterations in linear scale. (c) Trajectory for five agents.

TABLE V
C OMPARISON B ETWEEN A LGORITHMS FOR D IFFERENT N UMBER OF A GENTS W ITH U NSPECIFIED R ENDEZVOUS P OINT (AVERAGE WAS
TAKEN OVER 200 S IMULATIONS , rt = 0.1, I NITIALLY, THE ROBOTS W ERE D ISTRIBUTED IN 10 × 10 W INDOW )

rt after 10 000 iterations. From the above results one can Algorithms P and FN . When the number of agents is 50 or
conclude that the algorithms P, FN , and (P, FN ) are able more, none of the algorithms seem to converge.
to bring the agents to a point. However, for higher number Next, to demonstrate the algorithm’s performance for the
of agents, for instance 20, Algorithm (P, FN ) is superior to case when the agents are required to meet at a prespecified
1822 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 23, NO. 5, SEPTEMBER 2015

Fig. 24. Final point is specified. (a) r versus number of iterations in linear scale. (b) r versus number of iterations in log scale. (c)–(g) Trajectories for the
five agents case. (c) Algorithm P. (d) Algorithm (P, FO ). (e) Algorithm (P, FN ). (f) Algorithm FO . (g) Algorithm FN .

TABLE VI
C OMPARISON B ETWEEN A LGORITHMS FOR D IFFERENT N UMBER A GENTS W ITH S PECIFIED R ENDEZVOUS P OINT (AVERAGE WAS
TAKEN OVER 200 S IMULATIONS , rt = 0.1, I NITIALLY, THE ROBOTS W ERE D ISTRIBUTED IN 10 × 10 W INDOW )

point, five agents were considered. Here, the meeting point is agents, Algorithms (P, FO ) and FO take fewer iterations to
specified as (0, 0). The trajectories are given in Fig. 24. converge. This is in contrast to the results when the final
Table VI gives comparison of the number of iterations and rendezvous point was not specified. The reason for this is that
of the computations time between the algorithms for 5, 10, since the final point is specified, even a subset of the agents
15, and 20 agents, respectively, when the meeting point is for which the LP is executed tends to always move toward
specified. The results show that even for large number of the specified point. Since this point is fixed and is always
DAS AND GHOSE: BROADCAST CONTROL MECHANISM 1823

with the laptop, one of which is used for robot detection


and the other is used to provide online streaming video of
the experiment. The software architecture of the experiment
has various components, such as high-level robot control,
low-level robot control, and robot detection and robot com-
munication [36]. The high-level control is performed in the
laptop for all the computations related to the algorithms.
The algorithm was implemented using MATLAB. The
MATLAB Image Acquisition Toolbox was used to capture
and process the images [37]. The captured images were
processed to extract the rectangular markers on the top of the
robots. Before starting the experiments, offline camera cal-
ibration was performed using MATLAB Camera Calibration
Toolbox [38], [39]. The low-level robot control was done using
Not eXactly C (NXC) [40] programming. The NXC codes are
placed in the microprocessor of each vehicle. The low-level
control was decomposed into two tasks, which run in parallel.
Fig. 25. Hardware architecture.
The first of which checks whether there are any incoming
messages from the CC (the laptop) via Bluetooth connection.
If there are any messages, then the messages will pass to
the second task. The second task controls the rotation of the
two wheels. The communication between the robots (NXT)
and CC has been done through the Bluetooth adapter. The
Rheinsch-Westfälisehe Techniche Hosehsehule—Mindstorms
NXT Toolbox for MATLAB [41] has been used for wireless
communication. This communication has been used for send-
ing commands for rotation of the wheel and receiving health
monitoring data from NXT robots (for instance, battery-level
Fig. 26. LEGO NXT mobile robot. (a) Without markers. (b) With markers. information).
In Fig. 27, some snapshots of the experiment are shown.
This experiment was carried out for the broadcast consensus
included in the LP formulation at each iteration, it serves to algorithm to bring the vehicles to a desired point. In Fig. 27(a),
make the agents move toward it, thus avoiding the phenomena the robots are at their initial position and orientation. In
of spreading in arbitrary directions. On the other hand, when Fig. 27(b), the robots have turned by the command angle
Algorithm FN is used, the LP is always executed for all the θ after receiving the broadcast command. In Fig. 27(c), the
agents (i.e., a larger number of agents, as compared with the robots have advanced by the distance d, thus coming closer
FO case, which executes the LP only for the active agents) to each other. Although the experiment shows only one step
thus leading to a significant slowing down of the convergence of the algorithm it validates the implementability of the
rate. algorithm. In Fig. 28, two step of an experiment is shown.
One of the challenges in implementing the broadcast-based
VIII. E XPERIMENTAL VALIDATION consensus algorithm with nonspecified rendezvous point is that
The hardware architecture for implementing the broadcast- for small number of agents the LP solution may come up
based consensus algorithm is shown in Fig. 25. The hardware with a value of d that would be large and can take the robots
consists of mainly four parts: a team of three LEGO out of the controlled laboratory space. One could to either
Mindstorms NXT mobile robots, a laptop, a Bluetooth adapter, ignore the solution and execute the LP again after perturbing
and two cameras. The core of all the robots is NXT brick the orientations a second time, or one could implement the
that consist of a microprocessor equipped with computation algorithm using a fixed fraction of the computed value of d.
capabilities, motor actuation, and Bluetooth communication. The second approach is suboptimal but serves to reduce the
The left and right wheel of the robots are driven by two size of the square. This can be proved using the reasoning
servomotors independently. A third wheel acts as a third employed in the proof of Theorem 3. Of course, this problem
support and it rotates freely. This implies that the robots are is unlikely to appear when the number of agents is large.
of differential-drive kinematics type. Two rectangular markers Other challenges such as communication between CC and
(different in size) are put on the top of the robots (Fig. 26). robots, obtaining the coordinates (position and orientation) of
This allows visual detection using cameras. the agents also need to be solved. The above experiment imple-
A laptop provided the necessary processing power and ments only the basic algorithm P. There are some additional
communication capabilities. MATLAB was used for all technical and implementation challenges that will arise when
the computations. The laptop used a Bluetooth adapter to implementing the others algorithms. For example, identifying
communicate to the LEGO robots. Two cameras are attached if the agent is active or passive, dynamics of the agents, and
1824 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 23, NO. 5, SEPTEMBER 2015

Fig. 27. Snapshot from experiment results. (a) Initial positions of the robot. (b) After implementing θ . (c) After implementing d.

Fig. 28. Snapshot from experiment results. (a) Initial positions of the robot. (b) After implementing θ1 . (c) After implementing d1 . (d) After perturbation
in the orientation. (e) After implementing θ2 . (f) After implementing d2 .

noise in measurements are important challenges. The first can (i.e., identical control) for all the agents has been considered.
be implemented using some LED lights on the top of each The results show that it is possible for a group of agents
robot to indicate their flag values, which are identified using to meet at a location, or come very close to each other,
standard image processing techniques. A point of concern is by sending them a sequence of simple and exactly identical
that, in practical implementations, the crisscrossing behavior of instructions. The location point may also be prespecified.
the agent paths can lead to wasteful power/fuel consumption, We were able to show that introducing a perturbation in
but given the basic premise of broadcast control, this seems the orientations of the agents helps to dislodge the agents
to be inevitable. from a local minimum point and allows the algorithm to
look for a better solution. Repetition of this process helps
IX. C ONCLUSION in achieving consensus. We also proposed a novel LP-based
In this paper, the problem of controlling a group of agents solution approach that is computationally less intensive than
to converge at a location using only broadcast control input the SOCP technique proposed in the literature. The algorithms
DAS AND GHOSE: BROADCAST CONTROL MECHANISM 1825

were tested successfully on a real multirobot testbed made [12] S. Floyd, E. Diller, C. Pawashe, and M. Sitti, “Control methodologies
using LEGO Mindstorms NXT mobile robots. for a heterogeneous group of untethered magnetic micro-robots,” Int. J.
Robot. Res., vol. 30, no. 13, pp. 1553–1565, 2011.
There are several opportunities for future work in this [13] T. Hasegawa, N. Ogawa, H. Oku, and M. Ishikawa, “A new framework
direction. An interesting theoretical work could be to esti- for microrobotic control of motile cells based on high-speed tracking
mate the expected value of the number of iterations or rate and focusing,” in Proc. IEEE Int. Conf. Robot. Autom. (ICRA), Pasadena,
CA, USA, May 2008, pp. 3964–3969.
of convergence of the various algorithms using Lyapunov [14] E. B. Steager, M. S. Sakar, D. H. Kim, V. Kumar, G. J. Pappas, and
analysis (similar to [20]). As can be seen, these algorithms M. J. Kim, “Electrokinetic and optical control of bacterial microrobots,”
yield almost point convergence for up to about 20 agents, J. Micromech. Microeng., vol. 21, no. 3, p. 035001, 2011.
[15] O. Felfoul, M. Mohammadi, L. Gaboury, and S. Martel, “Tumor
but for larger number of agents the proximity measure r targeting by computer controlled guidance of magnetotactic bacteria
does not reduce beyond a certain value. Establishing lower acting like autonomous microrobots,” in Proc. IEEE/RSJ Int. Conf.
bounds on r as a function of the number of agents could Intell. Robots Syst. (IROS), San Francisco, CA, USA, Sep. 2011,
pp. 1304–1308.
be another interesting theoretical work. It seems possible to [16] Y. Ou, D. H. Kim, P. Kim, M. J. Kim, and A. A. Julius, “Motion control
extend this paper to consider process noise, sensor uncer- of Tetrahymena pyriformis cells with artificial magnetotaxis: Model
tainty, the presence of obstacles in the environment, and predictive control (MPC) approach,” in Proc. IEEE Int. Conf. Robot.
Autom. (ICRA), Saint Paul, MN, USA, May 2012, pp. 2492–2497.
incorporating collision avoidance. The collision avoidance [17] Y. Shirai, A. J. Osgood, Y. Zhao, K. F. Kelly, and J. M. Tour, “Directional
can possibly be implemented in the robotic platform using control in thermally driven single-molecule nanocars,” Nano Lett.,
proximity sensors on the robots so that robots can be made vol. 11, no. 5, pp. 2330–2334, Feb. 2005.
[18] P.-T. Chiang et al., “Toward a light-driven motorized nanocar: Synthesis
to execute a collision avoidance maneuver when they come and initial imaging of single molecules,” ACS Nano, vol. 6, no. 1,
closer than a certain prespecified distance of other robots. It pp. 592–597, 2012.
is also possible to extend this paper to higher dimensions. [19] J.-S. Li and N. Khaneja, “Ensemble control of Bloch equations,” IEEE
Trans. Autom. Control, vol. 54, no. 3, pp. 528–536, Mar. 2009.
This paper can be implemented in Nanorobots, UAVs, and [20] A. Becker, C. Onyuksel, and T. Bretl, “Feedback control of many
so on. However, the details of the implementations could be differential-drive robots with uniform control inputs,” in Proc. IEEE/RSJ
interesting areas for future research. Again, it would be a very Int. Conf. Intell. Robots Syst. (IROS), Vilamoura, Portugal, Oct. 2012,
pp. 2256–2262.
interesting future research work to compare the convergence [21] A. Becker and T. Bretl, “Approximate steering of a unicycle under
rate between the constrained case of having a fixed meeting bounded model perturbation using ensemble control,” IEEE Trans.
point and the unconstrained case of an arbitrary meeting Robot., vol. 28, no. 3, pp. 580–591, Jun. 2012.
[22] A. Becker and T. Bretl, “Approximate steering of a plate-ball system
point. under bounded model perturbation using ensemble control,” in Proc.
IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS), Vilamoura, Portugal,
Oct. 2012, pp. 5353–5359.
ACKNOWLEDGMENT [23] T. Bretl, “Minimum-time optimal control of many robots that move in
the same direction at different speeds,” IEEE Trans. Robot., vol. 28,
The authors would like to thank the reviewers for their no. 2, pp. 351–363, Apr. 2012.
perceptive comments which have helped to improve this [24] G. Antonelli and S. Chiaverini, “Kinematic control of platoons
paper. of autonomous vehicles,” IEEE Trans. Robot., vol. 22, no. 6,
pp. 1285–1292, Dec. 2006.
[25] V. Gervasi and G. Prencipe, “Coordination without communication: The
R EFERENCES case of the flocking problem,” Discrete Appl. Math., vol. 144, no. 3,
pp. 324–344, Dec. 2004.
[1] M. Cieliebak, P. Flocchini, G. Prencipe, and N. Santoro, “Solving the [26] T. Vicsek, A. Czirók, E. Ben-Jacob, I. Cohen, and O. Shochet, “Novel
robots gathering problem,” in Proc. 30th Int. Colloq. Autom., Lang. type of phase transition in a system of self-driven particles,” Phys. Rev.
Program., 2003, pp. 1181–1196. Lett., vol. 75, pp. 1226–1229, Aug. 1995.
[2] I. Suzuki and M. Yamashita, “Distributed anonymous mobile robots: [27] M. George and D. Ghose, “Reducing convergence times of self-propelled
Formation of geometric patterns,” SIAM J. Comput., vol. 28, no. 4, swarms via modified nearest neighbor rules,” Phys. A, vol. 391, no. 16,
pp. 1347–1363, 1999. pp. 4121–4127, Aug. 2012.
[3] H. Ando, Y. Oasa, I. Suzuki, and M. Yamashita, “Distributed mem- [28] K. Das and D. Ghose, “Positional consensus in multi-agent systems
oryless point convergence algorithm for mobile robots with limited using a broadcast control mechanism,” in Proc. Amer. Control Conf.,
visibility,” IEEE Trans. Robot. Autom., vol. 15, no. 5, pp. 818–828, St. Louis, MO, USA, Jun. 2009, pp. 5731–5736.
Oct. 1999. [29] K. Das and D. Ghose, “Randomization mechanisms based positional
[4] T. Bretl, “Control of many agents using few instructions,” in Proc. 3rd consensus in homogenous multi-agent system,” in Proc. Int. Conf.
Robot. Sci. Syst. Conf., Atlanta, GA, USA, 2007. Aerosp. Eng., Bangalore, India, 2009, pp. 1327–1335.
[5] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, U.K.: [30] K. Das and D. Ghose, “Positional consensus of multi-agent systems
Cambridge Univ. Press, 2004. using linear programming based decentralized control with rectilinear
[6] D. Bertsimas and J. Tsitsiklis, Introduction to Linear Optimization. decision domain,” in Trends in Intelligent Robotics, Communications
Belmont, MA, USA: Athena Scientific, 1997. in Computer and Information Science, vol. 103. Berlin, Germany:
[7] D. G. Luenberger, Introduction to Linear and Nonlinear Programming. Springer-Verlag, 2010, pp. 178–185.
Upper Saddle River, NJ, USA: Pearson, 1973. [31] K. Das and D. Ghose, “Multi-agent rendezvous algorithm with rectilin-
[8] B. R. Donald, C. G. Levey, C. D. McGray, I. Paprotny, and D. Rus, ear decision domain,” in Trends in Intelligent Robotics, Communications
“An untethered, electrostatic, globally controllable MEMS micro-robot,” in Computer and Information Science, vol. 103. Berlin, Germany:
J. Microelectromech. Syst., vol. 15, no. 1, pp. 1–15, Feb. 2006. Springer-Verlag, 2010, pp. 162–169.
[9] H. W. Tung, D. R. Frutiger, S. Pane, and B. J. Nelson, “Polymer-based [32] K. Das and D. Ghose, “Multi-agent rendezvous algorithms under various
wireless resonant magnetic microrobots,” in Proc. IEEE Int. Conf. Robot. information paradigms,” in Proc. Int. Conf. Intell. Unmanned Syst., Bali,
Autom. (ICRA), Saint Paul, MN, USA, May 2012, pp. 715–720. Indonesia, 2010.
[10] S. Schürle, K. E. Peyer, B. E. Kratochvil, and B. J. Nelson, “Holonomic [33] K. Das and D. Ghose, “Convergence results for multi-agent rendezvous
5-DOF magnetic control of 1D nanostructures,” in Proc. IEEE Int. Conf. algorithm with rectilinear decision domain,” in Proc. IEEE Multi-Conf.
Robot. Autom. (ICRA), Saint Paul, MN, USA, May 2012, pp. 1081–1086. Syst. Control, Denver, CO, USA, Sep. 2011, pp. 804–809.
[11] E. Diller, S. Floyd, C. Pawashe, and M. Sitti, “Control of multiple het- [34] K. Das and D. Ghose, “Multi-agent rendezvous algorithms with reduced
erogeneous magnetic microrobots in two dimensions on nonspecialized communication,” Int. J. Artif. Intell., vol. 7, no. 11, pp. 329–346,
surfaces,” IEEE Trans. Robot., vol. 28, no. 1, pp. 172–182, Feb. 2012. Oct. 2011.
1826 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 23, NO. 5, SEPTEMBER 2015

[35] E. Diller, S. Miyashita, and M. Sitti, “Magnetic hysteresis for Debasish Ghose received the B.Sc. (Eng.) degree
multi-state addressable magnetic microrobotic control,” in Proc. from the National Institute of Technology, Rourkela,
IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS), Vilamoura, Portugal, India, in 1982, and the M.E. and Ph.D. degrees from
Oct. 2012, pp. 2325–2331. the Indian Institute of Science (IISc), Bangalore,
[36] D. Benedettelli, M. Casini, A. Garulli, A. Giannitrapani, and A. Vicino, India, in 1984 and 1990, respectively.
“A LEGO mindstorms experimental setup for multi-agent systems,” in He has held visiting positions at the University
Proc. Int. Conf. Control Appl. (CCA), Saint Petersburg, Russia, Jul. 2009, of California at Los Angeles, Los Angeles, CA,
pp. 1230–1235. USA, and several other universities. He is currently a
[37] R. C. Gonzalez and R. E. Woods, Digital Image Processing. New Delhi, Professor with the Department of Aerospace Engi-
India: Prentice-Hall, 2006. neering, IISc. He has authored a book entitled
[38] J.-Y. Bouguet. (2008). Camera Calibration Toolbox for MATLAB. Scheduling Divisible Loads in Parallel and Distrib-
[Online]. Available: http://www.vision.caltech.edu/bouguetj/calib-doc/ uted Systems (Wiley, 1996). His current research interests include guidance
[39] R. Hartley and A. Zisserman, Multiple View Geometry in Computer and control of aerospace vehicles, collective robotics, multiple agent decision-
Vision. Cambridge, U.K.: Cambridge Univ. Press, 2003. makings, distributed decision-making systems, and scheduling problems in
[40] J. Hansen. (2007). Not eXactly C. (NXC)—Programmer’s Guide. distributed computing systems.
[Online]. Available: http://bricxcc.sourceforge.net/nbc/nxcdoc/NXC- Dr. Ghose is a fellow of the Indian National Academy of Engineering.
Guide.pdf He is also on the Editorial Board of the IEEE T RANSACTIONS ON
[41] Institute of Imaging and Computer Vision—RWTH Aachen University. A EROSPACE and E LECTRONIC S YSTEMS .
(2009). RWTH—Mindstorms NXT Toolbox. [Online]. Available:
http://www.mindstorms.rwth-aachen.de
[42] The LEGO Group. (2010). The Lego Mindstorms. [Online]. Available:
http://mindtorms.lego.com

Kaushik Das received the B.E. degree in electrical


engineering from the Indian Institute of Engineering
Science and Technology, Shibpur, India, in 2003,
the M.E. degree from Jadavpur University, Kolkata,
India, in 2006, and the Ph.D. degree from the Indian
Institute of Science, Bangalore, India, in 2014.
He was a Post-Doctoral Fellow with the Uni-
versity of Nebraska Omaha, Omaha, NE, USA,
from 2013 to 2014. His current research interests
include distributed algorithm, multirobot system, and
autonomous system and control theory.

You might also like