Professional Documents
Culture Documents
Abstract—This paper considers the synchronous distributed between the neighboring vehicles), and the algebraic graph
receding horizon control (RHC) for a general problem of theory in [15], [25]–[27] (where the controller is designed by
the nonholonomic multi-vehicle systems, i.e., the simultaneous guaranteeing the desired property of the Laplacian matrix).
forward/backward tracking, regulation and formation with the
collision avoidance. First, for each vehicle, a positively invariant However, these techniques either ignore the physical constraints
terminal-state region and an auxiliary controller are developed. (the saturation of input, the collision avoidance, etc.), or just
When every vehicle lies in its terminal-state region, all the dis- consider some specific constraints implicitly (e.g., [23] pun-
tributed control targets are achieved by the auxiliary controller. ishes the collision by a potential function). The receding hori-
Second, the compatibility constraint, restricting the norm of the zon control (RHC), with the advantages of explicitly handling
uncertain deviation between the assumed and actual predictive
trajectories of each vehicle, is given, which respects both the the physical constraints and optimizing the trajectories, is a
collision avoidance and convergence guarantee. Thirdly, a robust preferred approach in distributed control of multi-agent/multi-
collision avoidance constraint tolerating for the uncertain devia- vehicle systems (see, e.g., [28]–[30]). A survey work of the
tion is designed. By these designs, an overall control algorithm is distributed RHC (DRHC) can be found in [31].
proposed, by applying which all the control targets are achieved. In practice, usually a wheeled vehicle (e.g., the unicycle-
Two illustrative examples are provided to show the advantage and
effectiveness of the proposed approach. modeled vehicle in [32]–[34]) is constrained by its mechanical
dynamics to move only along its orientation (i.e., ẏ cos θ −
Index Terms—Collision avoidance, compatibility constraint, ẋ sin θ = 0), which is the so-called nonholonomic constraint.
distributed receding horizon control (DRHC), nonholonomic
multi-vehicles. For the nonholonomic vehicle, its local linear controllability
is lost when both the linear and angular speeds converge
I. I NTRODUCTION to zeros, i.e., limt→∞ (vr2 (t) + ωr2 (t)) = 0 (see the regulation
problem in [33]–[35]), which invalidates the use of linearized
I N the framework of control of multi-agent/multi-vehicle
systems there are nowadays many important fields of re-
search: flocking [1]–[3], consensus [4]–[8], cooperation [9],
model. Recently, some approaches are proved to be efficient for
the tracking and regulation problem of single nonholonomic
vehicle, e.g., the dynamic feedback linearization in [33] (for
[10], air/ground traffic control management [11], [12], tracking
the tracking and regulation, respectively), the backstepping
[13], [14] and formation [15]–[18]. This framework has broad
technique in [34] (for simultaneous tracking and regulation,
applications (e.g., battle damage assessment, reconnaissance,
with input saturation), and the RHC in [35] (for the simultane-
vegetation growth analysis, rapid assessment of topographical
ous tracking and regulation). As reported in [35], a terminal-
changes, localization of chemical sources, demining, and search
state region and an auxiliary controller for the simultaneous
and rescue missions; see [19]–[21]). For the multi-agent/multi-
tracking and regulation problem of single nonholonomic ve-
vehicle systems, the distributed/decentralized control strategy,
hicle, are developed. However, the positive invariance of the
with less cost of communication and computation, is preferred
developed terminal-state region cannot be guaranteed, which
to the centralized one. So far, a great number of techniques
may induce the loss of recursive feasibility. A similar condi-
have been applied for the distributed/decentralized control, e.g.,
tion is provided with [36] for the nonholonomic multi-vehicle
the leader-follower approach in [14], [19], [22] (where the
systems. Up to now, the development of the positively invari-
controller is designed by vanishing the relative error between
ant terminal-state region and the auxiliary controller is still a
each pair of leader and follower), the artificial potential function
challenge for the simultaneous tracking and regulation problem
in [23], [24] (where the controller is designed by shrinking
of the single nonholonomic vehicle. Aiming at the simulta-
an artificial potential function that defines interaction forces
neous forward/backward tracking and regulation problem of
nonholonomic multi-vehicle systems with collision avoidance,
Manuscript received August 4, 2012; revised January 24, 2013, July 28, we develop a positively invariant terminal-state region with
2013, and November 30, 2013; accepted December 22, 2013. Date of pub-
lication February 3, 2014; date of current version May 20, 2014. This work an auxiliary controller for each vehicle, which provides the
was supported by the National Nature Science Foundation of China under foundation for the RHC of nonholonomic vehicles. This is the
Grant 61174095. Recommended by Associate Editor R. Beard. (Corresponding first contribution of this paper.
author: B. Ding.)
The authors are with the Ministry of Education Key Lab For Intelligent Net- By employing a DRHC strategy for the multi-vehicle sys-
works and Network Security (MOE KLINNS Lab), Department of Automation, tems, each vehicle obtains its current input via solving a con-
School of Electronic and Information Engineering, Xi’an Jiaotong University, strained optimization problem at each sampling instant, taking
Xi’an 710049, China (e-mail: edison.wang@stu.xjtu.edu.cn; baocang.ding@
gmail.com). advantage of the predictive information of itself and of its
Digital Object Identifier 10.1109/TAC.2014.2304175 neighbors. The optimization problems of the vehicles can be
0018-9286 © 2014 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.
1440 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 59, NO. 6, JUNE 2014
TABLE I
N OTATIONS
Fig. 1. Different DRHC strategies. In the synchronous DRHC, all the vehicles
solve the optimization problems synchronously, and in the sequential one, solve
in the specified order. In the iterative DRHC, each vehicle solves its optimiza-
tion problem and communicates the solution with its neighbors iteratively, until
the local convergence of solutions is attained.
region (i.e., over the control horizon). Inside the terminal-
solved either synchronously (see, e.g., [16], [17], [29], [30]), state region, the auxiliary controller guarantees the collision
or sequentially (see, e.g., [11], [37], [38]), or iteratively (see, avoidance by providing a backup maneuver (this controller is
e.g., [39]–[41]). The differences of the three are illustrated in never implemented). This is the third contribution of this paper.
Fig. 1. In the synchronous DRHC which is the target of this This paper proposes the DRHC synthesis approach, as-
paper, each vehicle uses the assumed predictive trajectories of suming that the nonholonomic dynamics is nominal and the
its neighbors which deviate uncertainly from the corresponding optimization time at each sampling instant is trivial. The
actual ones. Due to the existing uncertain deviation, some rest of this paper is organized as follows. In Section II, we
constraints or conditions are desired in order to guarantee the introduce the nominal nonholonomic multi-vehicle systems
convergence of the overall multi-vehicle systems (to mention and the control objective, and formulate the individual cost
a few, the contractive constraint imposed in [42], the compat- function for each vehicle accordingly. Section III details the
ibility constraints proposed in [16], [17], the contractive-like synthesis approach. After designing the terminal, compatibility
constraint in [29], and the stability condition in [43]). In [30], and collision avoidance constraints, an individual optimization
an interacting compatibility condition is derived, and disassem- problem is formulated for each vehicle with the guarantees for
bled as the compatibility constraints which are incorporated in recursive feasibility, collision avoidance and convergence, and a
the individual optimization problem. However, these relevant synchronous DRHC algorithm is presented for implementation.
works are not applicable for the simultaneous tracking, regu- Section IV provides a numerical example to demonstrate the
lation and formation problem of nonholonomic multi-vehicle unique positive invariance of the designed terminal-state region,
systems, because their assumption of linear controllability in and a simulation example to illustrate the effectiveness of the
the terminal-state region is not satisfied. Motivated by [16], proposed DRHC. Finally, a conclusion summarizes this paper
[17], [30], we propose a time-varying compatibility constraint in Section V. Table I clarifies the symbols used in the paper.
in the norm form, which plays an important role in both the
convergence guarantee and the collision avoidance. This is the
II. P ROBLEM S TATEMENT
second contribution of this paper.
For the safety of multi-vehicle systems, the collision avoid- Consider the nonholonomic multi-vehicle system of Na ve-
ance should be considered in its control problem (see, e.g., hicles which are formulated by the following unicycle-modeled
[3], [21], [38], [41]). In the DRHC works [21], [38], [41], the dynamics (see, e.g., [32]–[34]):
collision avoidance is guaranteed by i) imposing a collision ⎡ ⎤
avoidance constraint over the control horizon in the individual vi (t) cos θi (t)
optimization problem; ii) developing an admissible protection żi (t) = f (zi (t), ui (t)) = ⎣ vi (t) sin θi (t) ⎦ , i ∈ Na . (1)
zone (i.e., admissible safety set) with positive invariance, and ωi (t)
an backup controller to perform emergency maneuver (e.g.,
stopping at the current position, loitering in a circle with a Here zi (t) = [pi (t); θi (t)] and ui (t) = [vi (t); ωi (t)] denote the
constant turn radius, and hovering). All the three works do not state and control input of vehicle i, respectively; pi (t) =
consider the nonholonomic vehicle, and there is no involvement [xi (t); yi (t)] and θi (t) denote the position (of the midpoint
of the assumed predictive trajectory. In comparison, we design of the rear axis) in the Cartesian coordinate frame and the
a collision avoidance constraint for each vehicle which tolerates orientation of vehicle i, respectively; vi (t) and ωi (t) are the
the uncertain deviation existing between the assumed and actual linear and angular speeds of vehicle i, respectively. ui (t) ∈
predictive trajectories of every other vehicle. This collision Ui = {ui (t)||vi (t)| ≤ viM , |ωi (t)| ≤ ωiM } where viM and ωiM
avoidance constraint is imposed outside of the terminal-state are the limits of linear and angular speed, respectively. The
WANG AND DING: DRHC FOR TRACKING AND FORMATION OF NONHOLONOMIC MULTI-VEHICLE SYSTEMS 1441
reference trajectory to be tracked is generated by a virtual and the terminal cost gi (zir (tk + T |tk )), which is a continuous,
reference vehicle of the same dynamics differentiable function (chosen in Section III-A), satisfying
gi (0) = 0 and gi (zir (t)) > 0 for any zir (t)
= 0.
żr (t) = f (zr (t), ur (t)) (2)
with the reference state zr (t) = [pr (t); θr (t)], and reference III. S YNCHRONOUS D ISTRIBUTED R ECEDING
input ur (t) = [vr (t); ωr (t)] satisfying |vr (t)| ≤ vrM < viM and H ORIZON C ONTROL
|ωr (t)| ≤ ωrM < ωiM .
For the tracking and formation objective, we denote dir = In the synchronous DRHC, each vehicle i does not know
[dxir ; dyir ] the desired tracking vector between the vehicle i the current actual predictive trajectories of the others. As the
and the virtual reference vehicle, dij = [dxij ; dyij ] the desired substitutions, it utilizes the assumed predictive trajectories of
formation vector between the vehicles i and j, Ni = {j ∈ other vehicles which are received before solving the optimiza-
Na | dij is desired} the formation neighbor index set (j ∈ tion problem. To distinguish the different predictive trajectories,
Ni ⇔ i ∈ Nj ), and R the safety radius of vehicles. To steer the we clarify
nonholonomic multi-vehicle systems to track the virtual refer- u∗i (τ |tk ) the optimal predictive input of vehicle i, the first
ence vehicle and keep the desired formation without collision, portion of which is implemented as the real-time
the control objective of each vehicle i is described as input;
ûi (τ |tk ) the assumed predictive input of vehicle i, which is
Tracking : as t → ∞, pr (t) − pi (t) = dir , θi (t) = θr (t),
used to evolve the assumed predictive trajectory.
(3)
The corresponding variables evolved according to u∗i (τ |tk )
Formation : as t → ∞, pj (t) − pi (t) = dij , ∀j ∈ Ni , (4) and ûi (τ |tk ) are labeled in the same manner, e.g., the optimal
Collision avoidance : pj (t) − pi (t) ≥ 2R, ∀j ∈ Na \ i, (i.e., actual) predictive trajectory p∗i (τ |tk ) and the assumed
(5) predictive trajectory p̂i (τ |tk ). By implementing the real-time
control input ui (τ ) = u∗i (τ |tk ), τ ∈ [tk , tk+1 ), the real-time
Consistency : dij = dir − djr = dil − djl ,
trajectory is identical to the optimal predictive one, i.e., pi (τ ) =
∀j, l ∈ Ni . (6) p∗i (τ |tk ), τ ∈ [tk , tk+1 ].
Here the condition (6) is required to guarantee the consistency The assumed predictive input is taken as the combination of
between the tracking and formation. For a desired tracking and the optimal predictive input and the terminal controller at the
formation objective, dir and dij are given a priori, following last sampling instant (see, e.g., [16], [17], [30]), i.e.,
which Ni is determined. To guarantee the collision avoidance
u∗i (τ |tk−1 ), τ ∈ [tk , tk−1 + T ),
condition (5), each vehicle i has to communicate with all the ûi (τ |tk ) = (9)
vehicles j ∈ Na \ i in order to plan a safe trajectory. Hence, a
κ∗
κi (zir (τ |tk−1 )) , τ ∈ [tk−1 + T, tk + T ]
global communication is desired in this paper.
According to the desired control objective, we de- where κi (zir κ∗
(τ |tk−1 )) is the terminal controller (i.e., the
fine pir (t) = pi (t) − pr (t) + dir the position tracking er- auxiliary control law designed in Section III-A) depend-
ror, θir (t) = θi (t) − θr (t) the angle tracking error, vir (t) = ing on the state tracking error zir κ∗
(τ |tk−1 ) = ziκ∗ (τ |tk−1 ) −
vi (t) − vr (t) cos θir (t) the linear speed tracking error, ωir (t) = zr (τ ) + [dir ; 0], with zi (τ |tk−1 ) being the terminal state
κ∗
ωi (t) − ωr (t) the angular speed tracking error, pij (t) = evolved along the dynamics (1) from ziκ∗ (tk−1 + T |tk−1 ) =
pi (t) − pj (t) + dij the position formation error, zir = [pir ; θir ] zi∗ (tk−1 + T |tk−1 ) by implementing the feedback control input
the state tracking error, and uir = [vir ; ωir ] the input tracking ui (τ |tk−1 ) = κi (zirκ∗
(τ |tk−1 )) in closed-loop. Accordingly, the
error. For the weighting scalars αi , ρi and βi given a priori, the assumed predictive trajectory is evolved as
ideal cost function of each vehicle i is formulated as
p∗i (τ |tk−1 ), τ ∈ [tk , tk−1 + T ],
t
k +T p̂i (τ |tk ) = (10)
i (τ |tk−1 ),
pκ∗ τ ∈ (tk−1 + T, tk + T ]
Ji (tk , zir , pi· , uir ) = Li (τ |tk , zir , pi· , uir )dτ
where pκ∗ i (τ |tk−1 ) is the associated terminal trajectory. The
tk
+ gi (zir (tk + T |tk )) (7) relationship between the different inputs and corresponding tra-
jectories are illustrated in Fig. 2. Following the assumed predic-
with the stage cost tive trajectory (10), we define the uncertain deviation between
Li (τ |tk , zir , pi· , uir ) = αi pir (τ |tk )2 + ρi vir (τ |tk )2 the assumed and actual predictive trajectories of vehicle i as
εi (τ |tk ) = p̂i (τ |tk ) − pi (τ |tk ), and denote the corresponding
position tracking predictive position formation error as p̂ij (τ |tk ) = pi (τ |tk ) −
+ αi θir (τ |tk )2 + ρi ωir (τ |tk )2 p̂j (τ |tk ) + dij .
angle tracking
Let Ωi (tk ) define the terminal-state region at the in-
stant tk , hCB i (εi (τ |tk )) ≤ 0 the compatibility constraint,
+ βi pij (τ |tk )2 (8) and hCA i (p i (τ |tk ), p̂j (τ |tk )) ≤ 0, ∀j ∈ Na \ i the collision
j∈Ni
avoidance constraint. Each vehicle i takes its current cost
position formation as Ji (tk , zir , p̂i· , uir ) instead of Ji (tk , zir , pi· , uir ) in (7), by
1442 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 59, NO. 6, JUNE 2014
Fig. 3. Terminal-state region Ωi for the case of forward tracking and reg-
ulation. For the given parameters λi , μi , θ̄ir , and Di , the boundary of this
Fig. 2. Relationship between different inputs and corresponding trajectories. region is determined by solving an optimization problem subjecting to the
Please refer to Algorithm 1 for implementation details. constraint (15).
TABLE II
WAY TO H ANDLE THE T ERMINAL C ONDITIONS
⎧
with the non-negative scalars λi , μi , θ̄ir . For simplicity, we have ⎨ 1+λ2i αi +2|Ni |βi +2 j∈N βj +αi μ2i +ρi a2i
omitted the superscript “κ” in x̃ir (τ |tk ), ỹir (τ |tk ), θ̃ir (τ |tk ) γi ≥ max
i
,
and p̃ir (τ |tk ). Hence, the terminal-state region is ⎩ ai − λi μi vrM
Ωi (tk ) = zi (τ |tk )(15) . (16) ρi 1 + λ2i bi μi vrM + ci . (19e)
In the terminal constraint (15), (15a)–(15d) are designed in Theorem 1: Let the terminal-state region and terminal cost
view of the nonholonomic property, to constrain the terminal function be given as in (16) and (18). Suppose Assumption 1
state for simultaneous tracking and regulation. The two cases holds and zi (t|tk ) ∈ Ωi (tk ). Then, for any τ ≥ t, under the
in (15d) are designed, respectively, for forward and backward auxiliary control law (17),
tracking with regulation. The constraint (15e) is designed to a) the requirements in Definition 1 are satisfied;
guarantee (12c), which follows the intuition that if each vehicle b) the tracking error system (14) is exponentially stable.
is closed enough to its desired position, then it will not collide Proof: The proof follows that, under ui (τ |tk ) =
with the others (for more details see Theorem 1 in the sequel). κi (zir (τ |tk )), for any zi (t|tk ) ∈ Ωi (tk ), the requirements
The region enclosed by dotted lines in Fig. 3 is the terminal- in Definition 1 are guaranteed by (15) and (19), and gi (·)
state region for the case of forward tracking and regulation. decreases exponentially to zero. For more details, see
The auxiliary control law is Appendix A.
κ Following Assumption 1, the auxiliary controller scalars
vi (τ |tk )
κi (zir (τ |tk )) = κ ai , bi , ci , the terminal constraint scalars λi , μi , θ̄ir , and the
ωi (τ |tk ) terminal weighting scalar γi can be chosen off-line (see
Algorithm 1).
vr (τ |tk ) cos θ̃ir (τ |tk ) + vir
κ
(τ |tk )
= (17) The brief relationship between the conceptual condition (12)
ωr (τ |tk ) + ωir κ (τ |t )
k
and the realistic conditions (15), (19) is shown in Table II.
with κ
vir (τ |tk ) = −ai x̃ir (τ |tk ) and ωirκ
(τ |tk ) = Remarks 1: Compared with the results in [35], [36], [45],
−(sgn(θ̃ir (τ |tk ))(bi |x̃ir (τ |tk )vr (τ )θ̃ir (τ |tk )| + ci x̃2ir (τ |tk ))/ the terminal-state region Ωi (tk ) in (16) has the following
(x̃2ir (τ |tk ) + ỹir
2
(τ |tk )), where ai , bi , ci are positive scalars. advantages.
Additionally, we define the terminal cost function as • It handles forward tracking (vr (τ ), vi (τ ) ≥ 0), backward
tracking (vr (τ ), vi (τ ) ≤ 0) and regulation (vr (τ ) = 0)
1 problem in a unified manner. In comparison, [35] requires
gi (zir (τ |tk )) = γi pir (τ |tk )2 + |θir (τ |tk )|
2 vr (t) ≥ 0 for the single vehicle, while [36], [45] require
vr (t)
= 0.
1 2
= γi p̃ir (τ |tk ) + θ̃ir (τ |tk ) (18) • It avoids the collision between the multi-vehicles.
2
• It guarantees the positive invariance (illustrated in
with the weighting scalar γi > 0, which will serve as a Section IV-A).
Lyapunov function for the tracking error system (14).
For the existence of the above terminal cost function, B. Design of the Compatibility and Collision
terminal-state region and auxiliary controller (which are the Avoidance Constraints
three ingredients of [44] for a usual RHC) to satisfy the require-
ments in Definition 1, the following assumption is desired. Due to the satisfaction of (12c), the collision avoidance is
Assumption 1: Given the input limits viM , ωiM , vrM , ωrM and guaranteed in the terminal-state region. When the predictive
the weighting scalars αi , βi , ρi , i ∈ Na , there exist scalars trajectory is planed over the control horizon τ ∈ [tk , tk + T ],
ai , bi , ci , λi , μi , θ̄ir ∈ [0, π/2] and γi such that the collision avoidance should also be guaranteed. Under the
ideal condition, each vehicle i prefers to consider the collision
v M − vrM avoidance constraint
max λi ωiM , λi μi vrM ≤ ai ≤ i , (19a)
Di
pj (τ |tk ) − pi (τ |tk ) ≥ 2R, ∀j ∈ Na \ i (20)
bi ≥ 1 + λ2i , (19b)
for τ ∈ [tk , tk + T ]. However, the actual predictive trajectory
ci ≥ max 1 + λ2i ωrM + ai λi , 1 + λ2i ai θ̄ir , (19c)
pj (τ |tk ) is unavailable for vehicle i at tk , so the correspond-
bi μi vrM + ci ≤ ωiM − ωrM , (19d) ing assumed predictive trajectory p̂j (τ |tk ) is utilized as the
1444 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 59, NO. 6, JUNE 2014
substitution. As a result, the influence of the uncertain deviation The constraint (26) is still valid by substituting σij (k) with
εj (τ |tk ) on the collision avoidance cannot be ignored. Addi- σj (k). We select the variable σij (k) in (26) for less commu-
tionally, the influence of the uncertain deviation on the con- nication cost (see the second paragraph after Algorithm 1 in
vergence of tracking and formation should also be considered. Section III-D). The constraint (26) is the designed collision
In this section, we propose to design a compatibility constraint avoidance constraint, i.e.,
to handle the effects of, and design the collision avoidance
i (pi (τ |tk ), p̂j (τ |tk )) = 2R+σij (k)−p̂j (τ |tk )−pi (τ |tk ).
hCA
constraint tolerating for, the uncertain deviation.
Constraining the uncertain deviation εi (τ |tk ), referred to as (27)
the compatibility constraint in [16], [17], [30], is imposed in the
individual optimization problem with the fundamental effect In summary, if the constraints (23) and (26) are satisfied for all
of enforcing a degree of consistency between what a vehicle τ ∈ [tk , tk + T ], then the implementation of ui (τ ) = u∗i (τ |tk ),
plans to do and what other vehicles believe it will do. This is τ ∈ [tk , tk+1 ) guarantees the collision avoidance condition (5)
important for guaranteeing the state trajectory to converge to the over the interval [tk , tk+1 ].
desired target (see, e.g., [30]) or a neighborhood of it (see, e.g., Example: Let us illustrate how the compatibility constraint
[16], [17]). In this paper, we endow the compatibility constraint (23) plays its role in the collision avoidance. Suppose for three
with an additional effect for the collision avoidance. To these vehicles i, j and l
ends, we design two bounds for the compatibility constraint, in
the following. p̂j (τ |tk )− p̂i (τ |tk ) = p̂l (τ |tk )− p̂i (τ |tk ) = 2R, (28)
The first bound is designed in view of the collision p̂j (τ |tk )− p̂l (τ |tk ) = 4R (29)
avoidance, as
for all τ ∈ [tk , tk+1 ]. The assumed predictive trajectory of vehi-
σi (k) = min {σij (k)}
j∈Na \i cle i is sandwiched between those of j and l. According to (21),
σi (k) = σj (k) = σl (k) = 0 is computed. Furthermore, accord-
minτ ∈[tk ,tk +T ] {p̂j (τ |tk )− p̂i (τ |tk )} ing to (23), it leads to pi (τ |tk ) = p̂i (τ |tk ), pj (τ |tk ) = p̂j (τ |tk )
= min −R .
j∈Na \i 2 and pl (τ |tk ) = p̂l (τ |tk ), τ ∈ (tk , tk+1 ], i.e., the three vehicles
(21) move along their assumed predictive trajectories, respectively,
so that the collision between these vehicles is avoided over
The second bound is designed for the tracking and formation the interval [tk , tk+1 ]. Without (23), there are many choices
convergence guarantee, as for pi (τ |tk ), pj (τ |tk ), pl (τ |tk ) satisfying (26) but incurring
collision.
ϕ αi zir (tk )2 + j∈Ni βi pij (tk )2
ηi (k) = (22)
(T − δ) j∈Ni βj (2ξij (k) + 3σij (k)) C. Main Results
where ϕ ∈ [0, 1) is a pre-specified scalar denoting the Now, by substituting the above designed terminal, compat-
desired convergence speed, and ξij (k) = maxτ ∈[tk+1 ,tk +T ] ibility and collision avoidance constraints into (11), the opti-
p̂j (τ |tk ) − p̂i (τ |tk ) + dji . Then, by incorporating both mization problem of each vehicle i to be directly solved on-line
bounds, we constrain the uncertain deviation by is formulated as
εi (τ |tk ) ≤ min {σi (k), ηi (k)} (23) min Ji (tk , zir , p̂i· , uir )
ui (τ |tk )
for all τ ∈ [tk , tk + T ]. The constraint (23) is the designed s.t. zi (tk |tk ) = zi (tk ),
compatibility constraint, i.e., for all τ ∈ [tk , tk + T ],
hCB
i (εi (τ |tk )) = εi (τ |tk ) − min {σi (k), ηi (k)} . (24) żi (τ |tk ) = fi (zi (τ |tk ), ui (τ |tk )) ,
ui (τ |tk ) ∈ Ui , (23), (26),
Following the compatibility constraint (23) and the triangle
inequality property, yields: for τ = tk + T, (15). (30)
pj (τ |tk ) − pi (τ |tk ) = p̂j (τ |tk ) − pi (τ |tk ) − εj (τ |tk ) Since the optimization problem (30) is constructed following
the common practice of RHC, the properties of the usual RHC
≥ p̂j (τ |tk ) − pi (τ |tk ) − σij (k)
(i.e., the recursive feasibility of the optimization problem, and
(25) convergence of the closed-loop state; see [44]) will be illus-
trated. The collision avoidance will be proved as an additional
by applying which the ideal constraint (20) can be property of the multi-vehicle systems.
guaranteed by Theorem 2: For the multi-vehicle system (1), each vehicle
solves the optimization problem (30) at each sampling instant
p̂j (τ |tk ) − pi (τ |tk ) ≥ 2R + σij (k), ∀j ∈ Na \ i. (26) tk , k ≥ 0. For each vehicle, if there exists a feasible solution at
WANG AND DING: DRHC FOR TRACKING AND FORMATION OF NONHOLONOMIC MULTI-VEHICLE SYSTEMS 1445
the initial instant t0 , then by implementing the real-time input (II) Over the interval t ∈ [tk , tk+1 ), k ≥ 0,
ui (τ ) = u∗i (τ |tk ), τ ∈ [tk , tk+1 ), a) implement the real-time input ui (t) = u∗i (t|tk );
b) assume the predictive trajectory p̂i (τ |tk+1 ), τ ∈
a) the optimization problem (30) is feasible at any tk , k > 0;
b) the collision between each pair of vehicles is avoided for [tk+1 , tk+1 + T ] according to (10);
all t ≥ 0; c) send p̂i (τ |tk+1 ) to j ∈ Na \ i and receive p̂j (τ |tk+1 )
from j ∈ Na \ i.
c) the tracking and formation errors of multi-vehicles con-
verge to zeros.
In the off-line stage of Algorithm 1, the determination of
Proof: The proof follows that, by applying the real-time the parameters is discussed as follows. The weighting scalars
input ui (τ ) = u∗i (τ |tk ), τ ∈ [tk , tk+1 ) for each vehicle, the αi , βi , ρi are given according to the desired tracking and for-
assumed predictive input ûi (τ |tk+1 ), τ ∈ [tk+1 , tk+1 + T ] is mation objectives. The more the tracking (formation) objective
feasible to (30) at the instant tk+1 ; the collision avoidance is emphasized, the larger the scalars αi and ρi (βi ) should be
condition (5) is guaranteed over the interval τ ∈ [tk , tk+1 ]; the given. The scalar ϕ ∈ [0, 1) is given to tune the convergence
sum of all the individual optimal cost functions (serving as a speed of tracking and formation. The faster the convergence
Lyapunov function of the multi-vehicles) decreases exponen- speed is desired, the smaller the ϕ should be given. However, a
tially to zero. For more details, see Appendix B. smaller scalar ϕ leads to a smaller ηi (k), which will force the
Now, following the proof of Theorem 2, the main results vehicle i to keep closer with its assumed trajectory, i.e., leave
of this paper are concluded as follows. For the nonholonomic less margin for improving trajectory. The scalars λi , μi and θ̄ir
multi-vehicle system (1), the reference trajectory generated by are chosen to tune the terminal-state region Ωi (tk ). According
(2), and the given control objective (3)–(6), if the optimization to the constraints (15a)–(15c), the larger the scalars λi , μi and
problem (30) of each vehicle i is feasible at the initial instant θ̄ir are chosen, the larger the region Ωi (tk ) can be determined,
t0 , then the implementation of the real-time input ui (τ ) = however, the less margin is reserved for choosing the controller
u∗i (τ |tk ), τ ∈ [tk , tk+1 ) for each vehicle can steer the multi- parameters ai , bi and ci (see the conditions (19a)–(19d) in
vehicle system to track the desired reference trajectory and Assumption 1). Thus, the terminal-state region Ωi (tk ) should
achieve the desired formation with collision avoidance. be determined properly to guarantee the existence of the auxil-
Remark 2: Compared with the compatibility constraints pro- iary control law κi (·). After the determination of Ωi (tk ) and
posed in [16], [17], [30], the constraint (23) has the following κi (·), the terminal weighting scalar γi can be easily chosen
advantages. according to (19e).
• It guarantees the tracking and formation errors to converge In the on-line stage of Algorithm 1, the trick of setting the
to zeros (while [16], [17] do not), and is in the norm form initial assumed predictive trajectory as p̂i (τ |t0 ) = pi (t0 ) is mo-
(while [30] is in a complex linear matrix inequality form). tivated by [16] and [17], to avoid the computation complexity
• It plays an important role in both the collision avoidance of centralized computation. In item a) of step (I), the calcula-
(depending on the designed bound σi (k)) and the conver- tion of σi (k) depends on the collection of the communicated
gence guarantee (depending on the designed bound ηi (k)), information p̂j (τ |tk ), j ∈ Na \ i. Then, using σj (k) in the
while the former is not considered in [16], [17], [30]. collision avoidance constraint (26) would require each vehicle
to communicate with the others once more, which is avoided
in this paper. The calculation of ηi (k) depends on the current
D. Implementation of the DRHC Algorithm state zi (tk ) and the current position pj (tk ) of vehicles j ∈ Ni ,
which are identical to zi∗ (tk |tk−1 ) and p̂j (tk |tk ), respectively.
In this section, the synchronous DRHC algorithm is pre- In item b) of step (I), the time of solving the optimization
sented for each vehicle to schedule the optimization, implemen-
problem (30) is assumed to be trivial at each sampling instant tk
tation, calculation and communication.
for immediate implementation. For the non-trivial optimization
times, a predictive version can be adopted (as referred in [17]),
Algorithm 1:
i.e., solving (30) before tk using the optimal state zi (tk ) =
Off-line stage: For each vehicle i, give the weighting scalars
zi∗ (tk |tk−1 ).
αi , ρi , βi , and the convergence speed ϕ; choose the terminal For the above proposed distributed RHC, when Na in-
constraint scalars λi , μi , θ̄ir , the terminal controller parame-
creases, the number of collision avoidance constraints, say
ters ai , bi , ci , and the terminal weighting scalar γi satisfying Na − 1, is increased, but the number of decision variables
Assumption 1. ui (τ |tk ) (τ ∈ [tk , tk + T ]) never grows. In contrast, for the
On-line stage: At the initial time t = t0 , each vehicle i sets
centralized RHC, the number of collision avoidance constraints
p̂i (τ |t0 ) = pi (t0 ), τ ∈ [t0 , t0 + T ], sends p̂i (τ |t0 ) to j ∈ Na \ is (1/2)Na (Na − 1), while the number of decision variables
i, and receives p̂j (τ |t0 ) from j ∈ Na \ i. Then, it performs the ui (τ |tk ) (τ ∈ [tk , tk + T ], i ∈ Na ) increases linearly with Na .
following procedure:
In [46], some analyses on the computational complexity of the
(I) At each sampling instant t = tk , k ≥ 0, RHC optimization problem for the continuous-time nonlinear
a) compute the bounds σij (k) and σi (k) according to system have been given, and a comparison of the complexity
(21), and ηi (k) according to (22); bounds between the centralized and distributed RHCs has been
b) solve the optimization problem (30) to obtain provided. These results similarly support the computational
u∗i (τ |tk ), τ ∈ [tk , tk + T ]; advantage of Algorithm 1.
1446 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 59, NO. 6, JUNE 2014
TABLE III
T ERMINAL C ONSTRAINTS , AUXILIARY C ONTROLLERS AND A SSOCIATED PARAMETERS FOR D IFFERENT A PPROACHES
TABLE IV
C HECK OF THE P OSITIVE I NVARIANCE OF T ERMINAL -S TATE R EGIONS
Fig. 4. Trajectories of multi-vehicles and reference in scenario 1 (without Fig. 6. Trajectories of multi-vehicles and reference in scenario 3 (with colli-
collision avoidance guarantee). At the time 0.65 s, 4.65 s and 9.45 s, the vehicles sion avoidance guarantee). At the time 12 s and 18 s, the vehicles are marked
are marked with the orientation arrows. The vehicles 1 and 3 collide during the with the orientation arrows. The collision is avoided.
interval [0.5 s, 0.7 s].
Fig. 7. Optimization time of multi-vehicles over the whole simulation Fig. 9. Linear speeds of multi-vehicles over the whole simulation horizon.
horizon.
Fig. 10. Angular speeds of multi-vehicles over the whole simulation horizon.
Fig. 8. Cost values of multi-vehicles over the whole simulation horizon.
phenomenon occurs. Over the whole simulation horizon, the
also avoided. This verifies the effectiveness of the main results linear speed of each vehicle does not exceed the range [0, viM ]
for backward tracking. before 9.5 s, and the range [−viM , 0] after 9.5 s (see Fig. 9), and
In the three scenarios, the optimization problems are always the angular speed does not exceed [−ωiM , ωiM ] (see Fig. 10). It
feasible by receding-horizon implementation, which verifies means that the input constraints are always satisfied. Thus, the
the property of recursive feasibility. In the scenarios 2 and 3, effectiveness of the main results has been verified.
the optimization time of each vehicle, say T ci (tk ), over the For a reasonable illustration on the computational complexity
whole simulation horizon [0 s, 18 s] are plotted in Fig. 7. The of the optimization problem (30) solved by IPOPT, the Monte
optimal performance costs, Ji (tk )’s, are plotted in Fig. 8, where Carlo simulations (10 times for each of the 15 sets of feasible
the sum of the individual optimal cost values decreases mono- random initial positions) are run for the above scenarios 1 and 2.
tonically for the forward and backward tracking, respectively. For scenario 1, since the Na − 1 collision avoidance constraints
It verifies the convergence guarantee. are removed from the optimization problem, it has the same
In scenario 2, Figs. 9 and 10 show that the linear and angular computational complexity as the case Na = 1. Some statistical
speeds of all vehicles converge to the reference linear and values of the optimization time T ci (tk ) for i = 1, 2, 3 are
angular speeds after the time t9 = 4.5 s. It means that the shown in Table V. It is shown that the vehicle number Na does
desired forward tracking and formation objective is achieved affect the computational complexity.
after the time t9 = 4.5 s. Note that, the inputs of each vehicle Remark 3: In this simulation, the optimization time at each
deviate a little bit from the corresponding reference inputs sampling instant is longer than the chosen sampling period
during the control horizon [7.5 s, 8 s) (see Figs. 9 and 10). It δ = 0.5 s. For the real-time applications, the computers (e.g.,
means that each vehicle takes a regulation to adjust its position the embedded computers with hardware acceleration) with
and orientation for parallel parking. In scenario 3, the similar hundreds times faster computing speed than PC are required.
WANG AND DING: DRHC FOR TRACKING AND FORMATION OF NONHOLONOMIC MULTI-VEHICLE SYSTEMS 1449
2|x̃ir ỹir |
≥ 4 ci − 1 + λ2i ωrM − ai λi x̃2ir
ỹir
V. C ONCLUSION + (bi − 1)|x̃ir vr θ̃ir |
−1
≤ 2|ỹir θ̃ir | bi |x̃ir ỹir vr θ̃ir | + ci x̃2ir ỹir According to (15a), (15b), (19a), and (19e), the condition (43) is
2 2
(1 + λi ) x̃ir guaranteed by summing the following conditions for all i ∈ Na :
⎛ ⎞
2
+ bi x̃2ir vr θ̃ir + ci x̃3ir θ̃ir
! ġi (zir )+ ⎝αi + 2|Ni |βi + 2 βj ⎠ pir 2 + αi θir 2
2 j∈Ni
+ |ωr x̃ir θ̃ir | + vr θ̃ir
+ ρi vir + ωir
2 2
⎛ ⎞
bi |x̃ir ỹir vr θ̃ir | ci |ỹir |
≤ −2|ỹir θ̃ir | + = ġi (zir ) + ⎝αi + 2|Ni |βi + 2 βj ⎠ p̃ir 2
(1 + λ2i ) x̃2ir 1 + λ2i
j∈Ni
bi 2
+ 2 − 1 vr θ̃ir + αi θ̃ir 2 + ρi vir κ 2
+ ωir κ 2
1 + λi
!
κ κ 2
ci = γi x̃ir vir + ỹir vr sin θ̃ir +sgn(θ̃ir )ωir + αi θ̃ir
+ − ωr |x̃ir θ̃ir |
M
⎛ ⎞
1 + λ2i
≤0 (37) + ⎝αi + 2|Ni |βi + 2 βj ⎠ x̃2ir + ỹir
2
j∈Ni
which guarantee the satisfaction of (15d) for τ = t+ . Applying
κ2 κ2
+ ρi vir + ωir
(15a), (15b), (15d) and (19a), yields ⎛
d
p̃ir (τ |tk )2 = 2 −ai x̃2ir + ỹir vr sin θ̃ir ≤ γi ⎝−ai x̃2ir + λi μi vrM x̃2ir
dτ
≤ − 2 ai − λi μi vrM x̃2ir ⎞
≤0
(38) sgn(θ̃ir ) bi μi vrM + ci x̃2ir
− ⎠
which guarantees the satisfaction of (15e) for τ = t+ . By x̃2ir + ỹir
2
|vir
κ
| = |ai x̃ir | ≤ |ai |p̃ir ≤ viM −vrM ⇒ |viκ | ≤ viM , (39) + αi μ2i x̃2ir
+ ρi a2i x̃2ir
2
while (15b) and (19d) lead to ρi sgn(θ̃ir ) bi μi vrM + ci x̃4ir
+ 2
(x̃2ir + ỹir
2 )
sgn(θ̃ir ) bi |x̃ir vr θ̃ir | + ci x̃2ir
|ωir
κ
|=
x̃2ir + ỹir
2
≤ − γi ai − λi μi vrM
≤ bi μi vr + ci ≤ ωi − ωrM ⇒ |ωiκ | ≤ ωiM .
M M
(40)
⎛ ⎞
Hence, the condition (12b) holds for any τ ≥ t.
− ⎝αi + 2|Ni |βi + 2 βj ⎠ 1 + λ2i
According to the constraint (15e), we have
j∈Ni
!
pir + pjr = p̃ir + p̃jr ≤ dir − djr − 2R (41)
− αi μ2i − ρi a2i x̃2ir − bi μi vrM + ci
which, by applying the triangle inequality, guarantees
γi
2R ≤ dir − djr − (pir + pjr ) × − ρi b i μ i v M
r + c i sgn( θ̃ ir )
1 + λ2i
≤ dir − djr − (pi − pr + dir ) − (pj − pr + djr ) ≤ 0. (44)
WANG AND DING: DRHC FOR TRACKING AND FORMATION OF NONHOLONOMIC MULTI-VEHICLE SYSTEMS 1451
+ Li (τ |tk , zir
κ∗ κ∗ κ∗
, pi· , κi (zir )) dτ
For the next instant tk+1 , the derivations are provided in the
tk +T
following. To distinguish from the optimal and assumed pre- !
dictive inputs, we introduce “−” as a superscript, i.e., denote + κ∗
gi (zir (tk+1 + T |tk )) − ∗
gi (zir (tk + T |tk )) .
the feasible predictive input as u− i (τ |tk+1 ); the accordingly
evolved variables are labeled with the same superscript. (51)
a) We show how the inequality constraints in (30) are
satisfied at instant tk+1 . By adopting the predictive input Following the triangle inequality property and the compatibility
u−i (τ |tk+1 ) = ûi (τ |tk+1 ), τ ∈ [tk+1 , tk+1 + T ], each vehicle constraint (23), we have:
evolves its predictive trajectory as p− i (τ |tk+1 ) = p̂i (τ |tk+1 ),
τ ∈ [tk+1 , tk+1 + T ]. Following the property of the terminal- t
k +T
& & & &
state region Ωi (tk ), u− i (τ |tk+1 ) ∈ Ui holds for all τ ∈ βi &p∗ij (τ |tk )&2 − &p̂∗ij (τ |tk )&2 dτ
[tk+1 , tk+1 + T ], and the terminal constraint (15) is satis- i∈Na j∈Ni tk+1
fied for τ = tk+1 + T . Because ε− i (τ |tk+1 ) = p̂i (τ |tk+1 ) − t
k +T
p− & &
i (τ |tk+1 ) = 0, the compatibility constraint (23) is satisfied
≤ βi 2 &p̂∗ij (τ |tk )& εj (τ |tk )
naturally for all τ ∈ [tk+1 , tk+1 + T ]. According to (46) and
i∈Na j∈Ni tk+1
the property of the auxiliary controller, it holds that
+ εj (τ |tk )2 dτ
2R ≤ p̂j (τ |tk+1 ) − p̂i (τ |tk+1 )
& &
&p∗ (τ |tk ) − p∗ (τ |tk )& ,
t
k +T
ACKNOWLEDGMENT [23] N. E. Leonard and E. Fiorelli, “Virtual leaders, artificial potentials and
coordinated control of groups,” in Proc. 40th Conf. Decision Control,
The authors would like to thank the anonymous reviewers 2001, pp. 2968–2973.
for their valuable comments and suggestions on improving [24] A. Ajorlou, A. Momeni, and A. G. Aghdam, “A class of bounded dis-
tributed control strategies for connectivity preservation in multi-agent
the work. systems,” IEEE Trans. Autom. Control, vol. 55, no. 12, pp. 2828–2833,
2010.
[25] W. J. Dong and J. A. Farrell, “Cooperative control of multiple non-
R EFERENCES holonomic mobile agents,” IEEE Trans. Autom. Control, vol. 53, no. 6,
pp. 1434–1448, 2008.
[1] H. G. Tanner, A. Jadbabaie, and G. J. Pappas, “Flocking in fixed [26] X. L. Wang, Y. G. Hong, J. Huang, and Z. P. Jiang, “A distributed control
and switching networks,” IEEE Trans. Autom. Control, vol. 52, no. 5, approach to a robust output regulation problem for multi-agent linear
pp. 863–868, 2007. systems,” IEEE Trans. Autom. Control, vol. 55, no. 12, pp. 2891–2895,
[2] H. S. Su, X. F. Wang, and Z. L. Lin, “Flocking of multi-agents with a 2010.
virtual leader,” IEEE Trans. Autom. Control, vol. 54, no. 2, pp. 293–307, [27] D. V. Dimarogonas, E. Frazzoli, and K. H. Johansson, “Distributed event-
2009. triggered control for multi-agent systems,” IEEE Trans. Autom. Control,
[3] H. T. Zhang, C. Zhai, and Z. Y. Chen, “A general alignment repul- vol. 57, no. 5, pp. 1291–1297, 2012.
sion algorithm for flocking of multi-agent systems,” IEEE Trans. Autom. [28] D. Jia and B. Krogh, “Min-max feedback model predictive control for
Control, vol. 56, no. 2, pp. 430–435, 2011. distributed control with communication,” in Proc. Amer. Control Conf.,
[4] R. Olfati-Saber and R. M. Murray, “Consensus problems in networks of 2002, pp. 4507–4512.
agents with switching topology and time-delays,” IEEE Trans. Autom. [29] C. Wang and C. J. Ong, “Distributed model predictive control of dynam-
Control, vol. 49, no. 9, pp. 1520–1533, 2004. ically decoupled systems with coupled cost,” Automatica, vol. 46, no. 12,
[5] K. Y. You and L. H. Xie, “Network topology and communication data rate pp. 2053–2058, 2010.
for consensusability of discrete-time multi-agent systems,” IEEE Trans. [30] B. C. Ding, L. H. Xie, and W. J. Cai, “Distributed model predictive control
Autom. Control, vol. 56, no. 10, pp. 2262–2275, 2011. for constrained linear systems,” Int. J. Robust Nonlin. Control, vol. 20,
[6] U. Munz, A. Papachristodoulou, and F. Allgower, “Consensus in multi- no. 11, pp. 1285–1298, 2010.
agent systems with coupling delays and switching topology,” IEEE Trans. [31] R. Scattolini, “Architectures for distributed and hierarchical model predic-
Autom. Control, vol. 56, no. 12, pp. 2976–2982, 2011. tive control—A review,” J. Process Control, vol. 19, no. 5, pp. 723–731,
[7] H. Kim, H. Shim, and J. H. Seo, “Output consensus of heterogeneous un- 2009.
certain linear multi-agent systems,” IEEE Trans. Autom. Control, vol. 56, [32] J. A. Marshall, M. E. Broucke, and B. A. Francis, “Pursuit formations of
no. 1, pp. 200–206, 2011. unicycles,” Automatica, vol. 42, no. 1, pp. 3–12, 2006.
[8] H. J. Fang, Z. H. Wu, and J. Wei, “Improvement for consensus perfor- [33] G. Oriolo, A. De Luca, and M. Vendittelli, “WMR control via dy-
mance of multi-agent systems based on weighted average prediction,” namic feedback linearization: Design, implementation, experimental val-
IEEE Trans. Autom. Control, vol. 57, no. 1, pp. 249–254, 2012. idation,” IEEE Trans. Control Syst. Technol., vol. 10, no. 6, pp. 835–852,
[9] M. G. Earl and R. D’Andrea, “A decomposition approach to multi-vehicle 2002.
cooperative control,” Robot. Auton. Syst., vol. 55, no. 4, pp. 276–291, [34] T. C. Lee, K. T. Song, C. H. Lee, and C. C. Teng, “Tracking control
2007. of unicycle-modeled mobile robots using a saturation feedback con-
[10] Y. F. Su and J. Huang, “Cooperative output regulation of linear multi-agent troller,” IEEE Trans. Control Syst. Technol., vol. 9, no. 2, pp. 305–318,
systems,” IEEE Trans. Autom. Control, vol. 57, no. 4, pp. 1062–1066, 2001.
2012. [35] D. B. Gu and H. S. Hu, “Receding horizon tracking control of wheeled
[11] G. Chaloulos, P. Hokayem, and J. Lygeros, “Distributed hierarchical MPC mobile robots,” IEEE Trans. Control Syst. Technol., vol. 14, no. 4,
for conflict resolution in air traffic control,” in Proc. Amer. Control Conf., pp. 743–749, 2006.
2010, pp. 3945–3950. [36] J. Chen, D. Sun, J. Yang, and H. Chen, “Leader-follower formation con-
[12] S. A. Reveliotis and E. Roszkowska, “On the complexity of maximally trol of multiple non-holonomic mobile robots incorporating a receding-
permissive deadlock avoidance in multi-vehicle traffic systems,” IEEE horizon scheme,” Int. J. Robot. Res., vol. 29, no. 6, pp. 727–747,
Trans. Autom. Control, vol. 55, no. 7, pp. 1646–1651, 2010. 2010.
[13] W. Ren, “Consensus Seeking, Formation Keeping, Trajectory Tracking [37] A. Richards and J. P. How, “Robust distributed model predictive control,”
in Multiple Vehicle Cooperative Control,” Ph.D. dissertation, Brigham Int. J. Control, vol. 80, no. 9, pp. 1517–1531, 2007.
Young University, Provo, UT, Aug. 2004. [38] Y. Kuwata, A. Richards, T. Schouwenaars, and J. P. How, “Dis-
[14] G. D. Shi, Y. G. Hong, and K. H. Johansson, “Connectivity and set tributed robust receding horizon control for multivehicle guidance,”
tracking of multi-agent systems guided by multiple moving leaders,” IEEE Trans. Control Syst. Technol., vol. 15, no. 4, pp. 627–641,
IEEE Trans. Autom. Control, vol. 57, no. 3, pp. 663–676, 2012. 2007.
[15] G. Lafferriere, A. Williams, J. Caughman, and J. J. P. Veerman, “Decen- [39] A. N. Venkat, I. A. Hiskens, J. B. Rawlings, and S. J. Wright, “Distributed
tralized control of vehicle formations,” Syst. Control Lett., vol. 54, no. 9, MPC strategies with application to power system automatic generation
pp. 899–910, 2005. control,” IEEE Trans. Control Syst. Technol., vol. 16, no. 6, pp. 1192–
[16] W. B. Dunbar, “Distributed Receding Horizon Control of Multiagent 1206, 2008.
Systems,” Ph.D. dissertation, California Institute of Technology, [40] E. Camponogara and L. B. de Oliveira, “Distributed optimization for
Pasadena, CA, Apr. 2004. model predictive control of linear-dynamic networks,” IEEE Trans. Syst.,
[17] W. B. Dunbar and R. M. Murray, “Distributed receding horizon control Man Cybern. A, vol. 39, no. 6, pp. 1331–1338, 2009.
for multi-vehicle formation stabilization,” Automatica, vol. 42, no. 4, [41] Y. Kuwata and J. P. How, “Cooperative distributed robust trajectory
pp. 549–558, 2006. optimization using receding horizon MILP,” IEEE Trans. Control Syst.
[18] W. Zhang and J. H. Hu, “Optimal multi-agent coordination under tree Technol., vol. 19, no. 2, pp. 423–431, 2011.
formation constraints,” IEEE Trans. Autom. Control, vol. 53, no. 3, [42] E. Camponogara, D. Jia, B. H. Krogh, and S. Talukdar, “Distributed model
pp. 692–705, 2008. predictive control,” IEEE Control Syst. Mag., vol. 22, no. 1, pp. 44–52,
[19] D. M. Stipanovic, G. Inalhan, R. Teo, and C. J. Tomlin, “Decentral- 2002.
ized overlapping control of a formation of unmanned aerial vehicles,” [43] T. Keviczky, F. Borrelli, and G. J. Balas, “Decentralized receding hori-
Automatica, vol. 40, no. 8, pp. 1285–1296, 2004. zon control for large scale dynamically decoupled systems,” Automatica,
[20] J. T. Feddema, C. Lewis, and D. A. Schoenwald, “Decentralized control of vol. 42, no. 12, pp. 2105–2115, 2006.
cooperative robotic vehicles: Theory and application,” IEEE Trans. Robot. [44] D. Q. Mayne, J. B. Rawlings, C. V. Rao, and P. O. M. Scokaert, “Con-
Autom., vol. 18, no. 5, pp. 852–864, 2002. strained model predictive control: Stability and optimality,” Automatica,
[21] T. Keviczky, F. Borrelli, K. Fregene, D. Godbole, and G. J. Balas, “De- vol. 36, no. 6.
centralized receding horizon control and coordination of autonomous [45] F. Xie and R. Fierro, “On motion coordination of multiple vehicles with
vehicle formations,” IEEE Trans. Control Syst. Technol., vol. 16, no. 1, nonholonomic constraints,” in Proc. Amer. Control Conf., 2007, vol. 1–13,
pp. 19–33, 2008. pp. 5986–5991.
[22] J. L. Wang and H. N. Wu, “Leader-following formation control of multi- [46] W. B. Dunbar, “Distributed receding horizon control of dynamically cou-
agent systems under fixed and switching topologies,” Int. J. Control, pled nonlinear systems,” IEEE Trans. Autom. Control, vol. 52, no. 7,
vol. 85, no. 6, pp. 695–705, 2012. pp. 1249–1263, 2007.
WANG AND DING: DRHC FOR TRACKING AND FORMATION OF NONHOLONOMIC MULTI-VEHICLE SYSTEMS 1453
[47] P. Falugi, E. C. Kerrigan, and E. V. Wyk, Imperial College London Opti- Baocang Ding was born in Hebei, China. He re-
mal Control Software User Guide (ICLOCS), 2010. [Online]. Available: ceived the M.S. degree from the China University
http://www.ee.ic.ac.uk/ICLOCS/ of Petroleum, Beijing, in 2000 the Ph.D. degree
[48] A. Wächter and L. T. Biegler, “On the implementation of an interior- from Shanghai Jiaotong University, Shanghai, China,
point filter line-search algorithm for large-scale nonlinear programming,” in 2003.
Mathematical Programming, vol. 106, no. 1, pp. 25–57, 2006. From September 2005 to September 2006, he was
a Postdoctoral Research Fellow in Department of
Chemical and Materials Engineering, University of
Alberta, Canada. From November 2006 to August
Peng Wang was born in Shanxi, China. He re-
2007, he was a Research Fellow in the School of
ceived the M.S. degree from Chang’an University,
EEE, Nanyang Technological University, Singapore.
Shaanxi, China, in 2009 and is currently pursuing the From September 2003 to August 2007, he was an Associate Professor in Hebei
Ph.D. degree at the Xi’an Jiaotong University, Xi’an,
University of Technology, China. From September 2007 to December 2008,
China.
he was a Professor in Chongqing University, China. He is now a Professor
His research interests include predictive control,
with Xi’an Jiaotong University, Xi’an, China. His research interests include
networked control and distributed control systems. predictive control, fuzzy control, networked control and distributed control
systems.
Dr. Ding was the recipient of the 2009 Program for New Century Excellent
Talents in University of China.