Documente Academic
Documente Profesional
Documente Cultură
Introduction
G. Jezic et al. (Eds.): KES-AMSTA 2012, LNAI 7327, pp. 554565, 2012.
c Springer-Verlag Berlin Heidelberg 2012
555
556
I. Prodan et al.
concluding remarks are drawn in Section 6 and the illustrative examples are presented in Section 5.
The following notations will be used throughout
the paper. Minkowskis
addition of two sets X and Y is dened as X Y = A + B : A X , B Y . Let
xk+1|k denote the value of x at time instant k+1, predicted upon the information
available at time k N.
2.1
System Description
i = 1, . . . , Na ,
(1)
where xikd Rn are the state variables, uikd Rm is the control input and
wki Rn represents a bounded disturbance for the agent i. Henceforth we assume
the following:
1. The pair (Ai , Bi ) is stabilizable, with Ai Rnn , Bi Rnm .
2. The disturbance wi is bounded, i.e. wi W i , where W i is a convex and
compact set containing the origin.
Theoretically, formulation (1) suces for solving any typical multi-agent control
problem (e.g., formation stability, trajectory tracking and so forth). However,
the presence of additive noises makes the numerical computation dicult and
severely limits the practical implementability. This is particularly true for centralized schemes where the computations are to be made into an extended space.
The solution followed here is based on the ideas in [10]. As a rst step, we
consider the nominal systems associated to (1):
xik+1 = Ai xik + Bi uik ,
i = 1, . . . , Na .
(2)
By linking the control laws associated to dynamics (1) and (2), respectively,
through the relation
(3)
uikd = uik + Ki (xikd xik ),
we observe that the tracking error of the ith system, dened as zki xikd xik ,
is given by:
i
= (Ai + Bi Ki )zki + wki .
(4)
zk+1
Assuming that Ki makes the closed-loop state matrix Ai + Bi Ki to be Schur1 ,
it follows that an RPI (see Denition 1) set Si can be determined and the real
1
The stabilizability hypothesis on the pair (Ai , Bi ) implies the existence of an optimal
control law for each agent i, Ki Rnm such that the matrices Ai + Bi Ki are
stable, where the controller Ki , i = 1, . . . , Na is constructed either by a Linear
Quadratic (LQ) design using the solution of the discrete algebraic Riccati equation
or alternatively by pole placement technique.
557
trajectory generated by (1) will reside in a tube centered along the nominal
trajectory generated by (2):
xikd xik Si xikd {xik } Si
(5)
(7)
This condition takes explicitly (through the use of the sets Si , Sj ) into account
the uncertainties introduced by the bounded perturbation in (1).
Let us recall that for any two convex sets A, B the following equivalence is
true:
A B = 0
/ A {B} .
(8)
Using (8), we obtain the equivalent formulation for (7):
/ Sj (Si ) , i, j = 1, . . . , Na , i = j.
xi xj
2
(9)
The assumption that the tracking error starts inside the set is made for simplication
reasons. As long as the set is contractive, after a nite number of steps, any trajectory
starting outside will enter inside the set.
558
I. Prodan et al.
(10)
Remark 2. Note that, a solution using parametrized polyhedra (see, for instance,
[8]) to describe the safety regions of the agents is presented in [12]. For guaranteeing that two (or more) agents do not superpose, the parametrized intersections of
the invariant sets are considered and then, the domain for which the intersections
are void is described. However, we note that this approach is computationally
demanding.
Remark 3. The main technical diculty encountered in this paper is the fact
that, often, the feasible regions are non-convex. This problem rises naturally
from separation conditions (see condition (9) and Remark 1). The solution is to
use the mixed-integer programming techniques [6]. This allows us to express the
original non-convex feasible region as a convex set in an extended space. Such
an approach leads to a signicant number of binary variables in the problem
formulation, thus leading to unrealistic computational times (in the worst-case
scenarios, an exponential increase dependent of the number of binary variables).
A method for reducing the computational time is detailed in [15], where we
propose a technique for making the time of computation P-hard in the number
of Linear/Quadratic Programming (LP/QP) subproblems that have to be solved.
3
3.1
The goal of clustering the agents as close as possible to the origin is realized
through a minimal conguration for the group of agents (2). We pose the problem
as an optimization problem where the cost function is the sum of the square
distances of each agent from the origin and the constraints are the ones imposing
collision avoidance (9):
min
(xi ,ui ),
i=1,...,Na
Na
xi 22 ,
subject to:
xi xj
/ Sj {Si } ,
xi = Ai xi + Bi ui , i, j = 1, . . . , Na , i = j.
i=1
(11)
Solving the mixed-integer optimization problem (11), a set of target positions
and the associated control laws are obtained:
Na
a
,
u
)
T = (x1f , u1f ), (x2f , u2f ), . . . , (xN
,
(12)
f
f
where every pair (xif , uif ) is a xed state/input of the ith agent.
559
Remark 4. Note that the second constraint in (11) is a xed point condition.
That is, the optimization problem will nd only pairs (xif , uif ), i = 1, . . . , Na
which are also a xed point for the considered dynamics, (2). Geometrically,
this means that the points xif will nd themselves on the associated subspaces
spanned by (In Ai )1 Bi . In particular, if the agents have the same dynamics
(i.e., homogeneous agents), they will have a common subspace over which to
select the xed points xif .
3.2
N
a
ij = 1,
Na
Na
i=1
N
a
cij ij , subject to:
min
(13)
ij = 1,
ij , i,j=1,...,Na
i=1 j=1
j=1
{0, 1},
ij
where ij are the decision variables: 1 if target xif is assigned to agent j and
0 otherwise. These binary variables ensure that each agent is assigned to one
unique target position.
The problem is dened by the choice of the cost weights cij , the simplest
way is to choose it as the distance between the actual position of agent j and
the desired target position in the formation. Hence, the problem would be to
determine the minimal distance that an agent has to travel to establish the
optimal assignment in the specied formation. A more insightful way is to use
the unconstrained dynamics (2) of the agents to describe the cost of reaching
from the initial position to the desired position. Then, cij can be described by a
weighted norm:
cij = (xj xif )T P (xj xif ),
i, j = 1, . . . , Na
(14)
with the matrix P = P T 0 given by the Lyapunov function or the innite time
cost-to-go, as long as the agents follow the unconstrained optimum through the
control action:
uj = Kj (xj xif ) + uj , i, j = 1, . . . , Na , i = j,
3
(15)
In the heterogeneous case the reassignment of the nal destination points is no longer
feasible since the swapping of the safety regions will result in collisions of the agents.
560
I. Prodan et al.
where u
j is chosen such that xjf = Aj xif + Bj uj , with u
j = Bj1 (I Aj )xif , if
the matrix Bj is invertible (or the alternative pseudo-inverse which allows the
denition of a xed point for the nominal trajectory). This optimization problem
can be reduced to a simple LP problem, hence it can be eciently computed.
The goal is to drive the agents to a minimal conguration (12) (if possible,
applying optimization (13)) while in the same time avoiding collisions along
their evolution towards the formation.
To this end, we will consider the set of Na constrained systems as a global
system dened as:
k ,
(16)
xk+1 = Ag xk + Bg u
with the corresponding vectors which collects the states and the inputs of
T T
T
a
each individual nominal system (2) at time k, i.e., x
k = [x1k | |xN
] ,
k
T
a
u
= [u1k | |uN
]T and the matrices which describe the model: Ag =
k
diag[A1 , . . . , ANa ], Bg = diag[B1 , . . . , BNa ].
We consider an optimal control problem for the global system where the
cost function and the constraints couple the dynamic behavior of the individual
agents. Also, perfect knowledge of each agent dynamics described by equation (2)
is available to all the other agents. Consequently, the global model will be used
in a predictive control context which permits the use of non-convex constraints
for collision avoidance behavior.
A nite receding horizon implementation is typically based on the solution
of an open-loop optimization problem.
An optimal control action
u
is obtained
k|k , u
k+1|k , . . . , u
k+N 1|k as a result of the
from the control sequence u
u
optimization problem:
xk+1|k , . . . , x
k+N |k , u
k|k , . . . , uk+N 1|k ),
u
= arg minVN (
u
x
k+l|k = Ag xk+l1|k + Bg uk+l1|k , l = 1, . . . , N
subject to:
/ Sj {Si } , i, j = 1, . . . , Na , i = j,
Hij xk+l|k
(17)
rewrite the collision avoidance between two agents i and j in the notation of the
centralized system (16)4 .
In order to assure that the target positions (12) we require a cost function
which is minimized in the destination points (and not the origin):
Using the elements provided in Remark 3, the computational complexity of (17) can
be assessed to a polynomial number of QP problems.
561
N 1
i(k) T
VN (
xk|k , u
k|k ) = (
xk+N |k xf
i(k)
) P (k+N |k xf )+
i(k) T
(
xk+N |k xf
xk+N |k xi(k) )+
) Q(
f
l=1
N 1
i(k)
(
uT
k+l|k uf
i(k)
uk+l|k u
)R(
f
),
(18)
l=0
i(k)
i(k)
with (xf , uf ) represents the optimal target positions and the associated
= Q
T 0, R
> 0 are
control laws at current time k, i = 1, . . . , Na . Here Q
the weighting matrices with appropriate dimensions, P = P T 0 denes the
terminal cost and N denotes the length of the prediction horizon.
First let us summarize in the following algorithm the receding horizon strategy
together with task assignment mechanism:
Algorithm 1. Centralized scheme strategy for a group of agents
1
2
3
4
5
6
7
8
9
10
Due to the fact that we use invariant sets, steps 1, 2 and 3 can be executed in
an o-line procedure. In the on-line part of the algorithm, we apply a nite horizon trajectory optimization: in step 6 we execute a task assignment if possible
(only if the safety regions are identical) and then proceed with the actual computation of the receding horizon control (step 8). Finally, the rst component of
the resulting control sequence is eectively applied to the global system (step 9)
and the optimization procedure is reiterated using the available measurements
based on the receding horizon principle [9].
Remark 5. Although functional, this scheme will not scale favorably with an
increased number of agents or a large prediction horizon due to the numerical
diculties. In particular, the mixed programming algorithms are very sensitive
to the number of binary auxiliary variables. In this case a decentralized approach
is to be envisaged in order to minimize the numerical computations.
Remark 6. Note that although desirable, an increase in the length of the prediction horizon is not always practical, especially when using mixed-integer programming. We observed that a two-stage MPC, where in the rst stage a task
assignment procedure is carried and in the second, the usual optimization problem is solved oers good performances with a reduced computational eort.
562
I. Prodan et al.
Illustrative Example
For the illustrative example we consider that each of the agent is described by
the following dynamics and disturbances:
00 1
0
0 0
0.5
0 0 0
0 0
1
0.3
i
i
i
Ai =
=
=
w
:
|w
|
,
B
,
W
i
1
i
m 0
0.5 , (19)
0 0 mi 0
i
1
0 0 0 mi
0.2
0 mi
where [xi y i vxi vyi ]T , [uix uiy ]T are the state and the input of each system. The
components of the state are the position (xi , y i ) and the velocity (vxi , vyi ) of the
ith agent, i = 1, . . . , Na . The parameters mi , i are the mass and the damping
factor, respectively.
10
20
15
6
10
4
5
x2
x2
2
0
4
10
6
15
8
10
10
(a)
0
x1
10
20
20
20
(b)
15
10
10
15
x1
50
50
40
40
15
30
step 20
20
20
10
0
5
10
step 10
x2
x2
x2
step 1
30
10
10
10
20
20
step 10
step 20
10
30
15
20
10
step 1
30
40
0
x1
10
50
50
40
40
30
20
10
0
x1
10
20
30
40
50
50
50
40
30
20
10
0
x1
10
20
30
40
50
(e)
563
15
50
40
10
step 1
30
20
5
x2
x2
10
0
step 20
10
20
5
30
40
10
6
(a)
x1
50
50
(b)
40
30
20
10
0
x1
10
20
30
40
50
564
I. Prodan et al.
Conclusions
In this paper, we rst present several tools in order to provide a systematic oline procedure for the control of a group of agents towards a minimal conguration. Second, in real-time a two stage receding horizon control design is adopted
for driving the agents to the predened formation. Also, we provide several remarks, leading to computational improvements of the mixed-integer techniques
used to assure a collision free behavior along the evolution of the agents. The
results are presented through some illustrative simulations of several examples.
The current research is to develop software-in-the-loop simulations and subsequent ight tests for the control of small Unmanned Aerial Vehicles (UAVs). The
vehicle dynamics is simulated by a Piccolo software and then, the control algorithm is transmitted in real ight simulations through a communication routine
running on a PC on the ground.
References
1. Balch, T., Arkin, R.: Behavior-based formation control for multirobot teams. IEEE
Transactions on Robotics and Automation 14(6), 926939 (1998)
2. Girard, A., de Sousa, J., Hedrick, J.: Coordinated Control of Agent Formations
in Uncertain, Dynamic Environments. In: Proceedings of the European Control
Conference, Porto, Portugal (2001)
3. Grundel, D., Murphey, R., P.M., P.: Cooperative systems, Control and optimization, vol. 588. Springer (2007)
4. Hallefjord Kurt, O.: Solving large scale generalized assignment problemsAn aggregation/disaggregation approach. European Journal of Operational Research 64(1),
103114 (1993)
5. Jadbabaie, A., Lin, J., Morse, A.: Coordination of groups of mobile autonomous
agents using nearest neighbor rules. IEEE Transactions on Automatic Control 48(6), 9881001 (2003)
6. Jnger, M., Liebling, T., Naddef, D., Nemhauser, G., Pulleyblank, W.: 50 Years
of Integer Programming 1958-2008: From the Early Years to the State-of-the-Art.
Springer (2009)
7. Laerriere, G., Caughman, J., Williams, A.: Graph theoretic methods in the stability of vehicle formations. In: Proceedings of the 2004 American Control Conference,
vol. 4, pp. 37293734 (2005)
8. Loechner, V., Wilde, D.: Parameterized polyhedra and their vertices. International
Journal of Parallel Programming 25(6), 525549 (1997)
9. Mayne, D., Rawlings, J., Rao, C., Scokaert, P.O.: Constrained model predictive
control: Stability and optimality. Automatica 36, 789814 (2000)
565
10. Mayne, D., Seron, M., Rakovi, S.: Robust model predictive control of constrained
linear systems with bounded disturbances. Automatica 41(2), 219224 (2005)
11. Palekar, U., Karwan, M., Zionts, S.: A branch-and-bound method for the xed
charge transportation problem. Management Science 36(9), 10921105 (1990)
12. Prodan, I., Olaru, S., Stoica, C., Niculescu, S.-I.: Predictive control for tight
group formation of multi-agent systems. In: Proceedings of the 18th IFAC World
Congress, Milano, Italy, pp. 138143 (2011)
13. Rakovi, S., Kerrigan, E., Kouramas, K., Mayne, D.: Invariant approximations
of the minimal robust positively invariant set. IEEE Transactions on Automatic
Control 50(3), 406410 (2005)
14. Richards, A., How, J.: Aircraft trajectory planning with collision avoidance using
mixed integer linear programming. In: Proceedings of the 21th American Control
Conference, Anchorage, Alaska, USA, pp. 19361941 (2002)
15. Stoican, F., Prodan, I., Olaru, S.: On the hyperplanes arrangements in mixedinteger techniques. In: Proceedings of the 30th American Control Conference, San
Francisco, CA, USA, pp. 18981903 (2011)
16. Tanner, H., Jadbabaie, A., Pappas, G.: Flocking in xed and switching networks.
IEEE Transactions on Automatic Control 52(5), 863868 (2007)