1
Roto-Translation Invariant Formation of Fixed-Wing
UAVs in 3D: Feasibility and Control
Xiaodong He, Zhongkui Li, Xiangke Wang, and Zhiyong Geng
Abstract—This paper investigates the formation of fixed-wing
arXiv:2302.11765v1 [eess.SY] 23 Feb 2023
UAVs in 3D, which communicate via a directed acyclic graph.
Different from common formation problems, we consider the
roto-translation invariant (RTI) formation, where the “roto-
translation” refers to a rigid-body motion obtained by composing
rotation and translation. Besides, the fixed-wing UAV is modelled
by a 3-D nonholonomic constrained rigid body instead of a
particle agent. The main results of this paper include proposing
the formation feasibility conditions and designing the formation
controller. Firstly, we define the RTI formation and propose
the conditions to guarantee that the formations are feasible
for the fixed-wing UAVs under the nonholonomic and input
saturation constraints. Secondly, given feasible formations, we
design a formation controller by introducing a virtual leader and
employing the compensation of rotation, followed by proving the
stability of the closed-loop system. Finally, simulation examples
are presented to verify the theoretical results.al with the motion
coupling among multiple robots. Finally, numerical simulations
are conducted to verify the theoretical results.
Index Terms—Formation control; Roto-translation; Fixed-
wing UAV; Nonholonomic constraints; Rigid body.
I. I NTRODUCTION
Fixed-wing UAVs are playing increasingly prominent roles
in defense programs and civil applications, such as border
patrol, disaster relief, and environmental monitoring [1], [2],
[3], [4], [5]. Among various coordination tasks, the most
common one for fixed-wing UAVs is the formation control
[6], which aims at driving multiple UAVs to achieve the pre-
defined constraints regarding their states, generally interpreted
as the desired geometric formation patterns. Plenty of existing
articles have investigated the formation control of fixed-wing
UAVs with diverse approaches, such as leader-follower struc-
ture [7], [8], consensus-based method [9], [10], guiding vector
field [11], [12], graph rigidity [13], [14]. Additionally, several
papers further focus on the formation reconfiguration [15],
[16] and maneuverability [17], [18]. Although having been
extensively studied, the formation of fixed-wing UAVs still
remains some implicit but attractive facts which are typically
omitted.
One fact is that the model of the fixed-wing UAV is a
rigid body in 3D, indicating that the complete 3-D rigid body
This work was supported by the National Natural Science Foundation of
China (61973006, T2121002), and the China Postdoctoral Science Foundation
(2022M720242). (Corresponding author: Zhongkui Li.)
X. He, Z. Li, and Z. Geng are with the State Key Laboratory for Turbulence
and Complex Systems, Department of Mechanics and Engineering Science,
College of Engineering, Peking University, Beijing 100871, China (e-mail:
hxdupc@pku.edu.cn; zhongkli@pku.edu.cn; zygeng@pku.edu.cn)
X. Wang is with the College of Intelligence Science and Technology,
National University of Defense Technology, Changsha 410073, China (xk-
wang@nudt.edu.cn)
kinematics should be established to capture in mathematics the
important characteristics of the physical system. However, in
a large amount of research [7], [9], [13], [15], [18], [19], [20],
the fixed-wing UAV is modelled by a planar nonholonomic
vehicle, generally known as a unicycle. Although the model
of a 2-D nonholonomic vehicle coincides with the cases of
flying at a fixed altitude, the results obtained in 2D cannot be
directly applied to the formation problems of the fixed-wing
UAVs in 3D.
Another fact in formation control is that it is more appealing
in certain real-world scenarios to focus on the overall motion
of a formation pattern. It has been demonstrated in [21] that
the concept of formation control should involve not only
controlling the relative positions and orientations of the agents
in a group, but also allowing the group to move as a whole.
Nevertheless, only a few papers [22], [23], [24] investigate
the formation problem from the perspective of overall motion.
Although providing some insightful results, these articles
omitted the formation feasibility, which is indeed the premise
of control especially under certain motion constraints.
Generally, it is said that a formation pattern is feasible, if all
of the agents in the formation satisfy the motion constraints,
such as nonholonomic constraints, input saturation constraints,
and so on. The seminal paper [25] develops a systematic
framework for studying the motion feasibility of multi-agent
formations, where feasibility conditions are proposed to main-
tain formation specifications described by equality constraints.
The work by Sun and Anderson [26] investigates the formation
feasibility of heterogeneous robot teams modelled by control
affine nonlinear systems, particularly presenting two special
motions that preserve motion constraints. The paper [27]
characterizes the mobility for formations of unicycle robots,
where the feasibility is formulated based on the distance-
bearing constraints. Colombo and Dimarogonas [28] have
studied the motion feasibility problem for multi-agent systems
on Lie groups, especially for the left-invariant fully actuated
kinematic and dynamic systems. Besides, the dissertation
[29] addresses the multi-robot planning problem with motion
constraints, and develops an algorithm for fixed-wing UAVs
in formation flight.
Motivated by the above-mentioned literature, in this paper,
we investigate the roto-translation invariant (RTI) formation
feasibility and control of fixed-wing UAVs in 3D under the
communication topology of a directed acyclic graph. The
terminology of “roto-translation invariant” is proposed in [30],
where roto-translation describes a rigid-body motion obtained
by composing rotation and translation. Thus, the RTI forma-
tion can rotate and translate simultaneously, and meanwhile
 2

the shape of the formation pattern keeps invariant, moving p = [x y z]T , while the attitude is specified by a rotation
like a single rigid body. The contributions of this paper are matrix R in the Special Orthogonal group SO(3) = {R ∈
threefold. R3×3 RT R = I3 , detR = 1}. Thus, the configuration of the
Firstly, we establish the kinematics of the fixed-wing UAV, fixed-wing UAV can be given by
which is a general 3-D rigid body under the nonholonomic  
R p
constraints, rather than a 3-D nonholonomic particle agent g= ∈ R4×4 . (1)
01×3 1
with decoupled rotation [31], [32]. Different from the 2-D
rotation which has only one degree of freedom (DOF), the Since g is uniquely defined by R and p, it can also be written
rotation kinematics in 3D is not an integrator anymore but as g = (R, p). All the configurations g constitute a Lie group
has a nonlinear formulation with the coupling of three DOFs. named the Special Euclidean group
This leads to the rigid body’s attitude being related to the 
R p
  
4×4  3

rotation sequence, which brings essential challenges than the SE(3) = ∈ R  R ∈ SO(3), p ∈ R .
01×3 1
2-D case as given in [33]. Particularly, in order to globally and
uniquely depict the rotation, the fixed-wing UAV’s kinematics Therefore, the configuration g of the fixed-wing UAV is an
is established in the Lie group SE(3) [34], [35], [36], [37], element in the Lie group SE(3).
[38], avoiding the singularity and unwinding phenomenon The fixed-wing UAV’s angular velocity and linear ve-
resulting from Euler angles. locity are denoted by ω = [ω x ω y ω z ]T ∈ R3 and
Secondly, we define the RTI formation of the fixed-wing v = [v x v y v z ]T ∈ R3 , which are both defined in F b .
UAVs, and propose the formation feasibility under the non- Regarding
 SO(3), the associated Lie algebra is so(3) = {S ∈
holonomic and input saturation constraints. Since the RTI
formation possesses special motion characteristics, not all
formation patterns are feasible to keep RTI. Thus, we propose
the formation feasibility by deriving the condition of formation
maintenance, and embedding the nonholonomic and saturation
R3×3  S T = −S}. Define a map ·∧ : R3 → so(3) by
(a∧ )b = a × b for all a, b ∈ R3 , where “×” is the vector
cross-product, and let ·∨ : so(3) → R3 denote the inverse
isomorphism. Then, the associated Lie algebra of SE(3) is
   
S v 4×4 
 3
constraints into it. It is shown that the relative positions in se(3) = ∈ R  S ∈ so(3), v ∈ R .
01×3 0
the formation can be arbitrarily specified, while the relative
orientations are decided by the nonholonomic constraints. With slight abuse of notation, we introduce the linear map
Moreover, it is derived mathematically that when the RTI for- ·∧ : R3 ⊕ R3 → se(3) defined by
 ∧ 
mation turns around, the outside fixed-wing UAVs accelerate ω v
while the inside ones decelerate. ξ∧ = ∈ R4×4 , (2)
01×3 0
Thirdly, based on the feasible formations, we design the
formation controller for the fixed-wing UAVs. Note that the for ξ = (ω, v) ∈ R3 ⊕ R3 . Similarly, let ·∨ : se(3) → R3 ⊕ R3
fixed-wing UAV is an underactuated system in the sense that represent the inverse isomorphism. The elements in se(3) are
six DOFs are controlled by only four inputs. Hence, problem, referred to as twists. Thus, the velocity ω ∈ R3 and v ∈ R3
we present the idea of feedback coupling for underactuated are formulated to be the twist ξ ∧ ∈ se(3).
systems. Specifically, the underactuated states are made cou- Having defined the configuration g and velocity ξ ∧ , we can
pled with the actuated directions by designing appropriate establish the kinematics of the fixed-wing UAV by
feedback in the controller, so that the lack of control inputs ġ = gξ ∧ , (3)
can be compensated by the additional constructed terms. More-
over, these additional feedback are designed from a physical where g is the state and ξ ∧ is the control input. It is observed
perspective, representing the rotation from the constrained that the system (3) is a global description independent of local
velocity to the standard velocity. Then, intuitively, the fixed- coordinates. In addition to the kinematics, the adjoint map
wing UAVs move along the trajectories of fully actuated Adg : se(3) → se(3) will be used in later derivation, which is
systems to achieve the formation. defined by
The paper is organized as follows. Section II provides Adg η ∧ = g(η ∧ )g −1 , ∀η ∈ se(3). (4)
the preliminaries and the problem statement, followed by
the communication topology transformation in Section III. Although having six DOFs, the fixed-wing UAV is restricted
Section IV and Section V propose the formation feasibility and by the nonholonomic constraints, where the word “nonholo-
the formation controller, respectively. Numerical simulation nomic” refers to the nonintegrability of a differential equation
examples are given in Section VI, and the paper concludes [39]. Intuitively, nonholonomic constraints of a fixed-wing
in Section VII. UAV exhibit as zero linear velocities along the y−, z−axis of
F b , i.e., v y = v z = 0, which can be expressed in a coordinate-
free formulation as
II. P RELIMINARIES AND PROBLEM STATEMENT
T
e2 e3 (ξ ∧ )e4 = 02×1 ,

A. Kinematics of fixed-wing UAV (5)
The fixed-wing UAV is modelled by a 6-DOF rigid body where ei ∈ R4 (i = 2, 3, 4) are the unit vectors with the i-th
moving in R3 . Let F e denote the earth-fixed frame and F b entry as 1. Hence, the nonholonomic constraints render the
the body-fixed frame. The position is described by a vector fixed-wing UAV to be an underactuated system in the sense

that the six DOFs of the UAV are controlled only by four
available control inputs, i.e., v x , ω x , ω y , ω z .
Thus, the kinematics of the fixed-wing UAV is described by
(3) together with the nonholonomic constraints (5). Although
it seems abstract, the kinematics can be transformed to a
common formulation by using the Euler angles. Let φ, θ, ψ
denote the roll, pitch, yaw angles, respectively, and we employ
the rotation sequence of φ − θ − ψ. Then, the rotation matrix
R can be parameterized to be
 
cψ cθ −sψ cφ + cψ sθ sφ sψ sφ + cψ sθ cφ
R = sψ cθ cψ cφ + sψ sθ sφ −cψ sφ + sψ sθ cφ  , (6)
−sθ cθ s φ cθ cφ
where c⋆ , cos ⋆ and s⋆ , sin ⋆. We take the time derivative
of R, and the rotation kinematics in the coordinates of the
Euler angles can be obtained as
     x
φ̇ 1 tan θ sin φ tan θ cos φ ω then ḡ0i is referred to as the roto-translation invariant (RTI)
 θ̇  = 0 cos φ − sin φ  ω y  , (7)
ψ̇ 0 sec θ sin φ sec θ cos φ ωz
implying that the rotation kinematics in 3D is not an integrator
like in 2D anymore [1]. Regarding the translation kinematics,
it follows from (3) that the time derivative of p is ṗ = Rv,
and by substituting (6) into it, we have
   
ẋ cos ψ cos θ
 ẏ  = v x  sin ψ cos θ  , (8)
ż − sin θ
where v y = v z = 0 is utilized. Therefore, the globally-
described kinematics (3) can be locally parameterized as the
rotation kinematics (7) and translation kinematics (8). Al-
though providing an intuitive way to represent the attitude, the
Euler angles suffer from singularity. Hence, we establish the
model in the Lie group SE(3) without local coordinates, and
employ the kinematics (3) with the nonholonomic constraints
(5) to describe the fixed-wing UAV.
B. Basic Graph Theory
A group of fixed-wing UAVs interact with each other via
communication networks, and it is convenient to model the
information exchanges by directed graphs. A directed graph
G is a pair (V, E), where V = {ν1 , · · · , νN } is a nonempty
finite node set and E ⊆ V × V is an edge set of ordered
pairs of nodes, called edges. The edge (νi , νj ) in the edge set
E denotes that the node νj can obtain information from the
node νi , but not necessarily vice versa. For an edge (νi , νj ),
the node νi is called the parent node, and νj is the child node.
A directed path from node νi1 to node νil is a sequence of
ordered edges of the form (νik , νik+1 ), k = 1, · · · , l − 1. A
cycle is a directed path that starts and ends at the same node.
A directed acyclic graph is a directed graph without any cycle.
A directed spanning tree is a directed graph in which every
node has exactly one parent except for one node, called the
root node, which has no parent and has directed paths to all
other nodes.
3
C. RTI formation in 3D
Consider N + 1 fixed-wing UAVs communicated by a
directed acyclic graph. Let gi = (Ri , pi ) ∈ SE(3) and
ξi∧ = (ωi∧ , vi ) ∈ se(3) denote the configuration and velocity
of the fixed-wing UAVs (i = 0, 1, · · · , N ), where 0 represents
the leader (or the root node of the graph), while 1, · · · , N
represent the followers.
Definition 1 (RTI formation). Given a group element ḡ0i =
(R̄0i , p̄0i ) ∈ SE(3) (i = 1, · · · , N ), if the follower i’s
configuration gi satisfies
gi = g0 ḡ0i , (9)
where g0 is the leader’s configuration, then it is called that the
follower gi achieves the desired formation pattern specified by
ḡ0i with respect to the leader g0 . Furthermore, if the desired
rotation matrix R̄0i and position vector p̄0i are both constants,
in other words, ḡ0i is a constant group element in SE(3),
formation. If only the desired position vector p̄0i is a constant,
then ḡ0i is referred to as the pseudo roto-translation invariant
(P-RTI) formation.
Remark 1. The formation pattern ḡ0i describes a relative
configuration of gi with respect to g0 , rather than an absolute
configuration. The RTI formation requires that the relative
position p̄0i should be unchanged in the leader’s body-fixed
frame, and the relative attitude R̄0i with respect to the
leader should also keep fixed. Regarding the P-RTI formation,
only fixed relative position is required, while no formation
constraint is introduced to the relative attitude. However,
the relative attitude in the P-RTI formation cannot be set
arbitrarily. Instead, it is typically decided by the nonholonomic
constraints.
Problem Statement: Regarding N + 1 fixed-wing UAVs
connected by a directed acyclic graph, the investigation of the
formation problem involves two aspects as given below.
1) Propose the RTI formation feasibility and determine
what types of formation patterns can be realized by
the fixed-wing UAVs under the nonholonomic and input
saturation constraints;
2) Design the formation controller so as to achieve feasible
RTI formation patterns for the fixed-wing UAVs.
III. C OMMUNICATION TOPOLOGY TRANSFORMATION
In this section, we propose a transformation for the com-
munication topology, which is a directed acyclic graph. As
shown in Definition 1, the formation pattern ḡ0i is given with
respect to the leader g0 . Nevertheless, in the swarm of the
fixed-wing UAVs, an arbitrary follower gi possibly does not
interact with the leader g0 directly, while gi always has parent
nodes gPi , i.e., the nodes from which gi receives information
in the communication graph. Thus, the formation of gi should
be given with respect to its parent nodes gPi . For a directed
spanning tree, each follower gi has one and only one parent
node gPi . Then, we can similarly define ḡPi i , that is, the
formation pattern of gi with respect to gPi , by
gi = gPi ḡPi i . (10)
 4

However, regarding a directed acyclic graph, which is consid- the nonholonomic constraints. According to (13), it is obtained
ered in this paper, each follower gi probably has more than that
one parent node. Then, the main challenge is how to define the T
e2 e3 (ξP1,2 )∧ e4

formation with respect to multiple parent nodes. To solve this i
T
= e2 e3 ((1 − λ1i )(ξP1 i )∧ + λ1i (ξP2 i )∧ )e4

problem, we utilize the geometric convex combination in the
Lie group SE(3) to construct a virtual parent node for each T
=(1 − λ1i ) e2 e3 ((ξP1 i )∧ )e4

follower. T
+ λ1i e2 e3 ((ξP2 i )∧ )e4 .

(15)
Definition 2 ([40], geometric convex combination). Let
gP1 i , · · · , gPMii denote the configurations of the parent nodes of Since [e2 e3 ]T ((ξP1 i )∧ )e4 = 0 and [e2 e3 ]T ((ξP2 i )∧ )e4 = 0,
the follower gi , where Mi is the number of the parent nodes of by substituting them into (15), it is further derived that
gi . Then, the geometric convex combination of gP1 i , · · · , gPMii , T
e2 e3 (ξP1,2 )∧ e4 = 0.

denoted by gGi , is iteratively defined by i
(16)
Then, we can obtain (14) by iterative computation.
gP1,2
i
= gP1 i exp(λ1i log((gP1 i )−1 gP2 i )),
gP1,2,3 = gP1,2 exp(λ2i log((gP1,2 )−1 gP3 i )), Therefore, based on Lemma 1, the virtual parent node gGi
i i i
is described by the kinematics (12) with the nonholonomic
··· constraints (14), which indicates that the virtual parent node
gGi = gP1,···
i
,Mi −1
exp(λiMi −1 log((gP1,···
i
,Mi −1 −1 Mi
) gPi )), has the same model as any of fixed-wing UAV in the swarm.
(11) Furthermore, given the fact that there is no cycle in the
where λji are the convex combination coefficients satisfying communication graph, then for each follower gi , the motion of
0 ≤ λji ≤ 1, j = 1, · · · , Mi − 1, and the exponential map its virtual parent node gGi is independent of gi ’s own motion.
exp : se(3) → SE(3) and the logarithmic map log : SE(3) → Thus, similar to (10), we define the formation pattern of gi
se(3) are given in [41]. with respect to gGi , denoted by ḡGi i , in the following equality
Regarding an arbitrary follower gi , Definition 2 provides gi = gGi ḡGi i . (17)
a virtual parent node gGi , which is defined by the geometric
convex combination of the parent nodes of gi . Since the graph This demonstrates that, with the help of the geometric convex
is acyclic, the geometric convex combination gGi does not combination in the Lie group SE(3), the leader-follower
involve the information of gi , so that there is no algebraic loop formation of the fixed-wing UAVs in a directed acyclic graph
regarding gi . Similar to every node in the graph, the geometric can be handled by the formation of the pair of gi and gGi as
convex combination gGi also possesses its own kinematics, given in (17). Note that such a case is equivalent to the case
which is given in the following lemma. of a leader with one follower; for simplicity, we let gL , gF , ḡ
denote gGi , gi , ḡGi i , respectively. Then, based on (17), the
Lemma 1. Let ξP1 i , · · · , ξPMii denote the velocities of the follower gF achieves the formation ḡ with respect to the leader
parent nodes of gi . Then, the kinematics of the geometric gL , if there holds
convex combination gGi is given by gF = gL ḡ, (18)
∧ which is referred to as the condition of formation achievement.
ġGi = gGi ξG i
, (12)
Therefore, in the subsequent sections, we shall study the
where ξGi is the convex combination of the velocities and is formation feasibility and design the formation controller based
defined by on (18).

ξP1,2
i
= (1 − λ1i )ξP1 i + λ1i ξP2 i , IV. RTI FORMATION FEASIBILITY ANALYSIS
ξP1,2,3
i
= (1 − λ2i )ξP1,2
i
+ λ2i ξP3 i , In this section, we investigate the RTI formation feasibility
(13)
··· under two kinds of constraints: the nonholonomic constraints
and the input saturation constraints. It is obvious that the
ξGi = (1 − λM
i
i −1
)ξP1,··· ,Mi −1
+ λM
i
i −1 Mi
ξPi .
i condition (18) provides a requirement for the configurations.
But in order to maintain such a formation pattern, the velocities
Furthermore, if the velocities ξP1 i , · · · , ξPMii are all restricted should be restricted as well. Thus, we give the condition of
by the nonholonomic constraints (5), then the velocity convex formation maintenance as follows, which is firstly proposed
combination ξGi defined in (13) satisfies the same nonholo- in [33].
nomic constraints
Lemma 2 ([33]). The follower gF maintains the desired
T ∧ formation pattern ḡ with respect to the leader gL , if the

e2 e3 (ξG i
)e4 = 02×1 . (14)
∧
leader’s velocity ξL and follower’s velocity ξF∧ satisfy
Proof. It has been proved in [40] that the geometric convex ξF∧ = Adḡ−1 ξL
∧
, (19)
combination gGi has the kinematics as (12). In the following,
we prove that the velocity convex combination ξGi satisfies where Adḡ−1 is the adjoint map defined in (4).
 5

Therefore, the conditions (18) and (19) guarantee that well. Thus, we embed the nonholonomic constraints (5) into
the RTI formation can be achieved and maintained, which (23), and it follows that
propose the requirements for configurations and velocities,  T ∧
e2 e3 (Adḡ−1 ξL )e4 = 02×1 , (26)
respectively. Regarding the fully actuated systems, which do
not impose any velocity constraints, the velocity condition which is further simplified to be (22) based on (23).
(19) can always be satisfied for arbitrary formation pattern ḡ.
It is observed that the constraints (22) involve the leader’s
However, the velocity of the fixed-wing UAVs is restricted by ∧ ∧
velocity ξL = (ωL , vL ) and the formation pattern ḡ = (R̄, p̄).
the nonholonomic and input saturation constraints, indicating
Generally, the leader’s velocity has been defined in advance
that the velocity condition (19) may not hold for certain types
and independent of the formation controller. In contrast, the
of ḡ. Then, we would like to investigate under what conditions
formation pattern ḡ is specified by users according to the
of ḡ, the velocity condition (19) can be guaranteed, which is
task. Then, the constraints (22) rely on the choice of (R̄, p̄).
referred to as the formation feasibility as defined below.
Moreover, the position vector p̄ in the formation can usually be
Definition 3 (formation feasibility). Let ΠE : se(3) → R and specified arbitrarily, while the rotation matrix R̄ cannot. This
ΠI : se(3) → R represent the nonholonomic constraints and is because the nonholonomic constraints restrict the direction
saturation constraints, respectively. Assume that the velocity of the linear velocity, i.e., the orientation of the fixed-wing
of the fixed-wing UAVs satisfies ΠE (ξ ∧ ) = 0 and ΠI (ξ ∧ ) ≤ 0. UAV. Hence, the rotation matrix R̄ in the formation pattern
Then, we call the formation pattern ḡ is feasible for the fixed- should possess a particular form to satisfy the nonholonomic
wing UAVs, if there holds constraints, which exhibits as (22). Then, it is implied that we
∧ can propose the formation feasibility by investigating what
ΠE (Adḡ−1 ξL ) = 0, (20) kinds of R̄ make the constraints (22) hold.
∧
ΠI (Adḡ−1 ξL ) ≤ 0. (21) It is not difficult to find out that the rotation matrix R̄
could be obtained by solving the matrix equation (22), where
Thus, in the following, we study the formation feasibility the variables except for R̄ are all known. However, it is not
under these two kinds of constraints, which lays the foundation
trivial to solve such a matrix equation directly. Moreover, when
to the subsequent formation control. defining an attitude in applications, we typically provide the
Euler angles and rotation sequence firstly, owing to the explicit
A. Feasibility under nonholonomic constraints physical interpretations, and then derive the rotation matrix
by the transformation in (6). Thus, it will be more convenient
Basically, Definition 3 provides the idea of studying the
to define the formation pattern in real-word scenarios, if the
formation feasibility, that is, embeding the nonholonomic con-
formation feasibility under the nonholonomic constraints is
straints into the velocity condition (19), and then figuring out
characterized from the perspective of Euler angles. Therefore,
what conditions the formation pattern should satisfy. Before
we reformulate the (22) by using Euler angles, and propose
the main results, a lemma of nonholonomic adjoint velocity is
the conditions when such constraints are able to hold.
given below.
Let φ̄, θ̄, ψ̄ denote the desired roll, pitch, yaw angles of the
Lemma 3. Let ḡ = (R̄, p̄) denote the desired formation follower gF with respect to the leader gL . Then, the formation
pattern of the follower gF with respect to the leader gL . feasibility is presented in the following theorem.
∧ ∧
Regarding the leader’s velocity ξL = (ωL , vL ), the formation
Theorem 1. Consider a formation pattern ḡ = (R̄, p̄), which
pattern ḡ is feasible for the follower gF under the nonholo-
describes the desired relative configuration of the follower gF
nomic constraints, if there holds
with respect to the leader gL . Let p̄ = [x̄ ȳ z̄]T be an
arbitrary constant vector in R3 , and let R̄ be decided by a
T ∧
e2 e3 R̄T (ωL

p̄ + vL ) = 02×1 . (22)
set of Euler angles (φ̄, θ̄, ψ̄), where φ̄ is arbitrarily chosen in
∧
Proof. We firstly compute the adjoint velocity Adḡ−1 ξL . (−π, π), θ̄ is given by
According to the adjoint map defined in (4), we have z
τAd
 T  ∧ θ̄ = − arctan x + sin ψ̄τ y
, (27)
−R̄T p̄
 
∧ R̄ ω L vL R̄ p̄ cos ψ̄τAd Ad
Adḡ−1 ξL =
01×3 1 01×3 0 01×3 1 and ψ̄ is given by
 T ∧ T ∧

R̄ ωL R̄ R̄ (ωL p̄ + vL ) y
= . (23) τAd
01×3 1 ψ̄ = arctan x . (28)
τAd
For the sake of illustration, we define x y z ∧
Herein, τAd , τAd , τAd are the components of ωL p̄ + vL . Then,
∧ ∧
ωAd = R̄T ωL R̄, (24) the following statements hold. 1)
vAd = R̄ T ∧
(ωL p̄ + vL ), (25) 1) The formation pattern ḡ is feasible for the fixed-wing
UAVs under the nonholonomic constraints.
∧
where ωAd and vAd are regarded as the angular velocity and 2) The formation pattern ḡ is RTI, if the leader’s velocity
∧
linear velocity of Adḡ−1 ξL , respectively. Note that ξF∧ is re- ξL∧
= (ωL ∧
, vL ) is time-invariant.
stricted by the nonholonomic constraints (5). Then, according 3) The formation pattern ḡ is P-RTI, if the leader’s velocity
∧ ∧ ∧
to Definition 3, Adḡ−1 ξL should satisfy such constraints as ξL = (ωL , vL ) is time-varying.
 6

x
y z
− sin ψ̄ cos φ̄ + cos ψ̄ sin θ̄ sin φ̄ τAd + cos ψ̄ cos φ̄ + sin ψ̄ sin θ̄ sin φ̄ τAd + cos θ̄ sin φ̄τAd =0 (30)

x
y z
sin ψ̄ sin φ̄ + cos ψ̄ sin θ̄ cos φ̄ τAd + − cos ψ̄ sin φ̄ + sin ψ̄ sin θ̄ cos φ̄ τAd + cos θ̄ cos φ̄τAd = 0. (31)

Proof. 1) Since the relative position p̄ has been predefined derived. To this end, the formulas (30) and (31) are reorganized
as constant, then based on Lemma 3, the feasibility of the to be
formation patten ḡ is equivalent to making (22) hold by choos- x y

∧ − sin ψ̄τAd + cos ψ̄τAd cos φ̄ + sin φ̄×
ing appropriate R̄. Note that if ωL p̄ + vL = 0, the rotation y

x z


matrix R̄ can be arbitrarily chosen in SO(3). However, such sin θ̄ cos ψ̄τAd + sin ψ̄τAd + cos θ̄τAd = 0, (37)
x y

a case results in vAd = 0 based on (25). According to (19), sin ψ̄τAd − cos ψ̄τAd sin φ̄ + cos φ̄×
this further leads to the follower’s linear velocity satisfying x
sin θ̄ cos ψ̄τAd y

+ sin ψ̄τAd z
+ cos θ̄τAd


= 0. (38)
vF = vAd = 0. Nevertheless, the fixed-wing UAVs cannot
hover in the space, in other words, the linear velocity cannot Substituting (35) into (37) and (38), it follows that
∧
be zero. Thus, the case of ωL p̄ + vL = 0 is not considered x
sin ψ̄τAd y

− cos ψ̄τAd cos φ̄ = 0, (39)
∧
in this paper. Regarding ωL p̄ + vL 6= 0, it is obtained that x y

sin ψ̄τAd − cos ψ̄τAd sin φ̄ = 0. (40)
 y z z y x
  x 
ωL p̄ − ωL p̄ + vL τAd According to (33), we have sin ψ̄τAd x
− y
cos ψ̄τAd = 0. Thus,
ωL ∧
p̄ + vL =  ωL z x
p̄ − ωL x z
p̄  , τ y  . (29)
x y y x
Ad
z
the formulas (39) and (40) always hold no matter what value
ωL p̄ − ωL p̄ τAd of the roll angle φ̄ is predefined. Therefore, the formation
Additionally, the rotation matrix R̄ can be computed base on pattern ḡ given by (27) and (28) satisfies the constraints (30)
φ̄, θ̄, ψ̄ according to the transformation in (6). We substitute and (31), indicating that ḡ is feasible for the formation of the
the R̄ in (6) and the ωL ∧
p̄ + vL in (29) into (22), and after a fixed-wing UAVs.
series of derivation, the constraints (22) can be reformulated 2) According to Definition 1, the formation pattern ḡ is RTI
as two algebraic equations (30) and (31). Thus, according to if the rotation matrix R̄ and position vector p̄ are both time-
Lemma 3, the formation pattern ḡ is feasible if these two equa- invariant. Note that p̄ has been specified as a constant vector
tions (30) and (31) hold. Since p̄ and ξL ∧
are both predefined, in R3 . Thus, we only have to prove R̄ is constant. Since the
∧ ∧
x y z
the variables τAd , τAd , τAd are all known. Therefore, we can leader’s velocity ξL = (ωL , vL ) and the position vector p̄ are
x y z
propose conditions to ensure formation feasibility by deriving both time-invariant, it follows from (29) that τAd , τAd , τAd
the Euler angles φ̄, θ̄, ψ̄ from (30) and (31) straightforwardly. are all constants. Then, the attitude angles θ̄ and ψ̄ given in
(27) and (28) can both keep fixed values. Plus a constant φ̄,
We firstly derive the yaw angle ψ̄. By the transformation of
the rotation matrix R̄ given by (6) is time-invariant, which
(30)× cos φ̄+(31)× sin φ̄, we cancel the terms involving θ̄ in
indicates that the formation pattern ḡ is RTI.
(30) and (31), and obtain that ∧
3) Considering the fact that the leader’s velocity ξL =
∧
(sin2 φ̄ + cos2 φ̄) sin ψ̄τAd
x y
− (sin2 φ̄ + cos2 φ̄) cos ψ̄τAd = 0. (ωL , vL ) is time-varying, we can obtain the attitude angles θ̄
(32) and ψ̄ given in (27) and (28) will vary with time as well, which
Due to sin2 φ̄ + cos2 φ̄ = 1, we have contributes to a time-varying rotation matrix R̄ accordingly.
Therefore, only the position vector p is constant, implying that
x y
sin ψ̄τAd = cos ψ̄τAd , (33) the formation pattern ḡ is P-RTI according to Definition 1.

which is further written to be Remark 2. The formulas (30) and (31) are regarded as the
y
nonholonomic constraints for the RTI formation. Note that a
τAd fixed-wing UAV is nonholonomic constrained in two directions
tan ψ̄ = x . (34)
τAd of the translational motion, i.e., swaying and heaving, while
it is interesting to observe that these two nonholonomic
So that, the yaw angle ψ̄ is obtained as in (28). Next, constraints are eventually formulated to be two constraints
the pitch angle θ̄ will be derived. By the transformation of (30) and (31) for three attitude angles φ̄, θ̄, ψ̄. This further
(30)× sin φ̄+(31)× cos φ̄, the terms involving φ̄ are cancelled implies an unconstrained rotation DOF is indeed preserved
and it follows that in the choice of the formation pattern, which is verified by
x
sin θ̄(cos ψ̄τAd y
+ sin ψ̄τAd z
) + cos θ̄τAd = 0. (35) the proof of Theorem 1 that ψ̄ and θ̄ should be given by (28)
and (27) to make the RTI formation feasible, while φ̄ is freely
Then, it is further obtained that chosen as needed.
z In Theorem 1, the desired relative attitude matrix R̄ is
τAd
tan θ̄ = − x + sin ψ̄τ y
, (36) decided by three attitude angles φ̄, θ̄, ψ̄. Then, a naturally-
cos ψ̄τAd Ad
arising question is when φ̄, θ̄, ψ̄ can all be 0, followed by the
where ψ̄ is provided in (28). Hence, the pitch angle θ̄ is given fact that R̄ degenerates to be the identity matrix I3 . Intuitively,
by (27). Lastly, the expression of the roll angle φ̄ should be this case means the follower gF has the same attitude as the

x =0
Fig. 1. Schematic diagram of the velocities in the case of ωL
leader gL . In other words, they point to the same direction
in the formation pattern, which is relatively common in real-
world scenarios. Therefore, the following theorem illustrates
the feasibility of such a particular formation pattern for the
fixed-wing UAVs.
Corollary 1. Let ḡ = (I3 , p̄) denote an RTI formation pattern,
where I3 is the identity matrix in R3×3 and p̄ is a constant
vector in R3 . Then, ḡ is feasible for the fixed-wing UAVs under
the nonholonomic constraints, if there holds ωL ∧
= 03×3 or
x x
ωL = 0, p̄ = 0.
Proof. According to Lemma 3, a formation pattern R̄ is
∧
feasible if the condition (22) holds. Once it has ωL = 03×3 ,
the condition (22) degenerates to
[e2 e3 ]T R̄T vL = 02×1 .
x T
Due to vL = [vL 0 0] , then the formula (41) naturally
holds for R̄ = I3 , which indicates the formation pattern
ḡ(I3 , p̄) is feasible for the fixed-wing UAVs. Regarding the
x probably violates the velocity saturation constraints, leading
case of ωL = 0, p̄x = 0, it follows from (29) that
∧ x
ωL p̄ + vL = [τAd 0 0]T .
By substituting (42) and R̄ = I3 into (22), we can verify
the rightness of (22), implying the feasibility of the formation
pattern ḡ = (I3 , p̄).
∧ the set of linear maps from the space U to the space V.
Remark 3. In Corollary 1, it is evident that ωL = 03×3 refers
x y
to a straight line trajectory, while the case of ωL = 0 (ωL 6=
z by [AdR ] = R.
0, ωL 6= 0) implicitly represents the leader moves along a
circle in the plane perpendicular to ωL . As shown in Figure 1,
x Now, we propose the formation feasibility under the input
once it has ωL = 0, the angular velocity ωL lies in the y − z
plane and is orthogonal to the linear velocity vL , followed
by the fact that ωL and vL defines a plane (the red plane).
Then, according to theoretical mechanics, the fixed-wing UAVs
move along a circle in the blue plane (which is perpendicular
to the red plane) with the radius r = kvL k/kωL k. Therefore,
Corollary 1 demonstrates intuitively that the fixed-wing UAVs
have the same orientations in two cases, i.e., moving along a
straight line, or a circle with parallel formation patterns.
Remark 4. It is well known that the main drawback of the
Euler angles is the singularity at certain points. However, the
formation feasibility and controller proposed in this paper do
not suffer from any singularity. The reasons are given below.
For the φ − θ − ψ sequence, the singularity appears when
7
θ = ± π2 . In such a case, one rotation DOF is lost, so that the
rest two angles cannot be well defined. This can be explicitly
viewed in mathematics. Once θ = ± π2 , the rotation matrix in
(6) degenerates to
 
0 sin(φ − ψ) cos(φ − ψ)
R = ±  0 cos(φ − ψ) − sin(φ − ψ) , (43)
−1 0 0
which is only decided by the error between φ and ψ. Therefore,
given a certain attitude (or a rotation matrix) with θ = ± π2 , we
cannot derive the specific φ and ψ from (43) exactly. In short,
the singularity arises from the fact that the transformation
from rotation matrix to Euler angles is one-to-many. However,
in this paper, we use Euler angles φ̄, θ̄, ψ̄ to derive rotation
matrix R̄ instead. Once φ̄, θ̄, ψ̄ are provided, R̄ can be
uniquely obtained without any singularity according to (6).
Given the fact that the transformation from the Euler angles
to the rotation matrix is a surjective map, R̄ is always well
defined by (6), even if in the case of θ = ± π2 . Thus, there does
not exist any singularity from the Euler angles.
B. Feasibility under input saturation constraints
For the fixed-wing UAVs, the angular velocity and linear
velocity are typically restricted by the saturation constraints.
Particularly, the linear velocity of the fixed-wing UAV should
always be positive and has a lower bound to maintain the flight
altitude. Note that the RTI formation requires the followers
(41) to rotate with the leader so as to keep the formation pattern
invariant. Then, if the leader takes sharp turns during the
movement, the followers are required to provide a large veloc-
ity to maintain the RTI formation accordingly. However, this
to the formation pattern unfeasible. Hence, in this subsection,
(42) the formation feasibility will be investigated under the input
saturation constraint. Before providing the main results, we
introduce the following lemma which will be utilized later.
Lemma 4 ([41]). Given R ∈ SO(3) and the associated
adjoint map AdR ∈ L(so(3); so(3)), where L(U; V) denote
Then, the matrix representation of the adjoint map is given
saturation constraint in the following theorem.
Theorem 2. Let ḡ = (R̄, p̄) denote the desired formation
pattern. The saturation constraints for the leader’s velocity
x
are given by kωL k ≤ αL , β L ≤ vL ≤ β L , where αL > 0,
β L > β L > 0. Similarly, the saturation constraints for the
follower’s velocity are given by kωF k ≤ αF , β F ≤ vFx ≤ β F ,
where αF = αL , 0 < β F < β L < β L < β F . Assume that the
leader’s velocity has satisfied the saturation constraints. Then,
the statements given below illustrate the conditions when the
follower’s velocity satisfies the saturation constraint from the
perspective of the desired formation pattern. 1)
1) Regarding the angular velocity ωF , there always holds
kωF k ≤ αF for arbitrary desired formation pattern ḡ.
 8

2) Regarding the linear velocity vFx , there holds β F ≤

vFx ≤ β F if the desired relative position p̄ satisfies
(
c1 , for kωL k = 0,
kp̄k = (44)
c2 , for kωL k 6= 0,
where c1 is an arbitrary positive scalar, and c2 is given
by
βF − βL βL − βF
 
c2 = min , . (45)
αL αL
Proof. According to Lemma 2 and (24)(25), the formulation
of the follower’s angular and linear velocities are able to be
expressed as
∧
ωF = (R̄T ωL R̄)∨ , (46) Fig. 2. Schematic diagram demonstrating the upper and lower bounds of the
follower’s saturation constraint.
∧
vF = R̄T (ωL p̄ + vL ). (47)
1) The angular velocity ωF can be rewritten as where kωL k ≤ αL and kvL k ≤ β L are utilized. Similarly,
ωF = ∧ ∨
(AdR̄T ωL ) = [AdR̄T ]ωL . (48) we can obtain that the inequality (56) holds as long as kpk
satisfies
Then, it follows from Lemma 4 that kvL kmin − β F β − βF
kpk ≤ = L . (59)
ωF = R̄T ωL , (49) kωL kmax αL
Based on (58) and (59), once kωL k = 6 0, the follower’s linear
which becomes a rotation transformation with respect to ωL .
velocity satisfies the constraint β F ≤ vFx ≤ β F if the desired
Then, with the property of R̄R̄T = I3 , we have
q q relative position p̄ satisfies
kωF k = ωL T R̄R̄T ω = ωL T ω = kω k ≤ α = α ,
βF − βL βL − βF
L L L L F
 
(50) kpk = min , . (60)
αL αL
which indicates that kωF k ≤ αF always holds no matter the
choice of ḡ. In summary, the linear velocity constraint β F ≤ vFx ≤ β F
2) Regarding the linear velocity, it can be obtained from holds if the condition in (44) is satisfied.
(47) that
vFx = kωL p̄ + vL k. (51) Remark 5. It is seen from Theorem 2 that the angular
Due to vL > 0, we further have speed bounds are identical (αF = αL ), while the linear
speed bounds are not (β F < β L , β F > β L ). This can
− kωL kkp̄k + kvL k ≤ vFx ≤ kωL kkp̄k + kvL k (52) be intuitively explained with the aid of Figure 2, where the
orientations of all fixed-wing UAVs are identical, and the
For the case of kωL k = 0, it follows that
position vector p points laterally. In order to keep RTI, the
x
vFx = kvL k = vL , (53) followers rotate along with the leader at the same angular
velocity, i.e., ωF = ωL . However, regarding the linear
which indicates vFx is not related to kp̄k any more. Given that
x velocities, due to the anticlockwise rotation, an additional
β L ≤ vL ≤ β L , there naturally holds
forward velocity is induced for the right follower, yet an
β F < β L < vFx < β L < β F , (54) additional backward velocity for the left follower. Then, it
is obtained that kvFouts k = kvL + ωL × pk > kvL k and
demonstrating the saturation constraint of vFx is always sat- kvFins k = kvL − ωL × pk < kvL k. Thus, the linear speed
isfied, so that the desired relative position p̄ can be chosen range of the followers should be larger than the leader’s.
arbitrarily. For the case of kωL k 6= 0, based on (52), the
constraint β F ≤ vFx ≤ β F is satisfied if there holds Remark 6. Given the fact that the feasibility of p̄ is always
ensured no matter how kωL k varies in (0, αL ], the conditions
kωL kkp̄k + kvL k ≤ β F , (55) of kp̄k proposed in (45) may be conservative to some extent.
−kωL kkp̄k + kvL k ≥ β F . (56) For example, if the angular speed kωL k is small, that is,
kωL k = ǫ where 0 < ǫ ≪ αL , the position norm kp̄k might
It can be derived from (55) that kpk should satisfy not necessarily be chosen as small as (45). However, in such
β F − kvL k cases, if it is guaranteed that kωL k merely varies in a small
kpk ≤ . (57) neighborhood around 0, the value of αL could be turned down
kωL k
to make the condition (45) less conservative. Essentially, the
Note that (57) is able to hold if kpk satisfies range of kωL k greatly influences the upper bound of kp̄k,
β F − kvL kmax β − βL so that we can appropriately choose αL so as to reduce the
kpk ≤ = F , (58) conservativeness as much as possible.
kωL kmax αL
 9

Following Lemma 5, the problem to be solved is trans-

formed to trajectory tracking. To this end, we define the
relative configuration of gF with respect to gC , that is
−1
gCF = gC gF . (65)
It is obtained that the time derivative of gCF is
∧
ġCF = gCF ξCF , (66)
Fig. 3. Leader-follower formation with the aid of a virtual leader ∧
where ξCF is the relative velocity defined by
∧
ξCF ∧
= ξF∧ − Adg−1 ξC . (67)
CF
V. RTI FORMATION CONTROLLER DESIGN
Lemma 6. The follower gF tracks the virtual leader gC if the
In this section, we design the formation controller for the relative configuration gCF is stabilized to the identity matrix
feasible formations of the fixed-wing UAVs. Inspired by [42], I4 .
which deals with the formation tracking of the nonholonomic
mobile robots, we introduce a virtual leader related to the Proof. The proof is trivial based on (64) and (65).
desired formation pattern, and then convert leader-follower
Then, the control task is further converted to stabilization
formation into trajectory tracking. The challenge of designing
of the relative configuration gCF by designing the relative
a formation controller arises from the underactuation of the ∧
velocity ξCF . The following lemma provides a stabilization
fixed-wing UAVs, i.e., six DOFs controlled by only four
control with logarithmic feedback.
inputs. Therefore, to solve this problem, we design an addi-
tional angular velocity to compensate the lack of certain linear Lemma 7 ([43]). Consider the system ġ = gξ ∧ in the Lie
velocities. Then, based on the logarithmic feedback in the Lie group SE(3) and let kp > 0 be a control gain. Then, the
group SE(3), the formation controller is proposed accordingly. control law ξ ∧ = −kp logSE(3) (g) almost globally stabilizes
Firstly, we define a virtual leader gC , whose configuration the state g at I4 from the initial condition g(0) = (R(0), p(0))
is given by satisfying tr(R(0)) 6= −1, where logSE(3) is the logarithmic
gC = gL ḡ, (61) map in the Lie group SE(3).
∧
where ḡ is the desired formation pattern. By taking the time Based on Lemma 7, the relative velocity ξCF can be
∧
derivative of gC , we have designed to be ξCF = −kp logSE(3) (gCF ). By substituting
it into (67), we can obtain the formulation of the follower’s
∧ ∧
ġC = ġL ḡ = gL (ξL )ḡ = gC Adḡ−1 ξL , velocity, that is

where the kinematics (3) and the adjoint map (4) are utilized. ∧
ξF∧ = −kp logSE(3) (gCF ) + Adg−1 ξC . (68)
CF
Define the velocity of the virtual leader as follows
Considering the fact that the virtual leader is regarded merely
∧ ∧ as a middle variable, we are supposed to cancel the terms
ξC = Adḡ−1 ξL . (62)
∧
related with the configuration gC and velocity ξC . By substi-
Then, the kinematics of the virtual leader is expressed as tuting (61) and (62) into (68), it follows that
∧
ġC = gC ξC . (63) ξF∧ = −kp logSE(3) (gL ḡ)−1 gF + Ad(g−1 gC ḡ) Adḡ−1 ξL ∧


.
F

Lemma 5. The follower gF achieves the formation pattern For simplicity, we define a relative configuration
ḡ with respect to the leader gL if the follower gF tracks the −1
virtual leader gC . gLF = gL gF , (69)

Proof. Once the follower gF tracks the virtual leader gC , there and then ξF∧ can be further expressed as
holds
ξF∧ = −kp logSE(3) (ḡ −1 gLF ) + Adg−1 ξL
∧
. (70)
gF = gC . (64) LF

Although we have derived the control input in (70), it cannot

By substituting (61) into (64), we obtain the condition (18), directly serve as the formation controller. This is because (70)
which indicates gF achieves the formation pattern ḡ with is only applicable to fully actuated systems, while the fixed-
respect to gL . wing UAVs are restricted by the nonholonomic constraints.
Thus, in the following, we will make the control input become
Remark 7. As shown in Figure 3, the virtual leader gC given nonholonomic constrained.
by (61) is defined in the body-fixed frame of the real leader For simplicity, we refer to (70) as the standard velocity, and
gL . Since the formation pattern ḡ is fixed, the virtual leader let a vector Ξ ∈ R6 denote it, that is
will be rigidly attached to the real leader and move along T  ∨
Ξ = Ω Λ , −kp logSE(3) (ḡ −1 gLF ) + Adg−1 ξL ∧

with it. In this way, ḡ is achieved if the follower gF tracks the ,
LF
virtual leader gC . (71)
 10

where Ω = [Ωx Ωy Ωz ]T ∈ R3 and Λ = [Λx Λy Λz ]T ∈ R3 Concerning a desired formation pattern ḡ given by Theorem 1,
can be regarded as the standard angular velocity and linear design the following controller
velocity, respectively. Due to the nonholonomic constraints,
ωF = Ω + ka ΩAD , (76a)
the linear velocities v y and v z of the fixed-wing UAVs are
always zero. Then, the standard velocity components Λy and vF = Λ N H . (76b)
Λz derived from (71) cannot be provided to the fixed-wing
Then, the controller given in (76) makes the follower gF
UAVs through the input channels of v y and v z . That is to say, realize the formation pattern ḡ with respect to the leader
the real linear velocity of the fixed-wing UAVs is actually
gL , where Ω, ΩAD , ΛN H are defined in (71), (75), (72),
T
ΛN H = Λ x 0 0 . respectively, and ka is a positive control gain.

(72)
Note that the nonholonomic constraints restrict the direction Proof. According to Lemma 5 and Lemma 6, the follower gF
of the linear velocity. This motivates us that if the direction realizes the formation pattern ḡ with respect to the leader gL ,
of the real linear velocity ΛN H is tuned to be aligned with if the relative configuration gCF converges to I4 . Thus, we
that of the standard linear velocity Λ, then the fixed-wing have to prove that the controller (76) stabilizes the kinematics
UAVs can be controlled by Ξ given in (71) to accomplish (66) at I4 .
the formation task. Therefore, inspired by [44], we construct For the sake of illustration, we express the kinematics
new feedback terms by employing Λy and Λz , and let these (66) under the exponential coordinate. Referring to [43], the
terms be additional angular velocities, which can rotate the exponential coordinates of gCF are defined by
real linear velocity ΛN H to the standard one Λ indeed. The XCF = (logSE(3) (gCF ))∨ . (77)
details of the design process are given below.
We firstly construct a rotation matrix Rz around the z-axis Based on Lemma 4 in [43], the time derivative of XCF is
of the body-fixed frame F b . Define a vector n = [Λx Λy 0]T ,
ẊCF = [BX ]ξCF , (78)
and it can be verified that
where [BX ] is a matrix related to XCF . Substituting (67) into
n · e3 = 0,
(78), we have
where the symbol “·” represents the dot-product and e3 = ∧ ∨
ẊCF = [BX ](ξF − (Adg−1 ξC ) ), (79)
[0 0 1]T . Then, we define another vector n⊥ by the cross- CF

product of e3 and n, that is which is the formulation of the kinematics (66) under the
T exponential coordinates XCF . According to the definition of
n⊥ = e3 × n = −Λy Λx 0 .

XCF in (77), once gCF = I4 , there holds XCF = 0. Then,
Thus, {n, n⊥ , e3 } constitute a set of orthogonal vectors in what we need to prove is that the system (79) is stabilized at
R3 . Based on such a group of vectors, we can construct the XCF = 0 by the controller ξF given in (76).
following orthogonal matrix Based on the components in (76a)(76b), the controller ξF
h i can be written in a vector form as
n n⊥
Rz = knk kn⊥ k
e3 , (73)     
ωF Ω ka ΩAD

ξF = = + , (80)
vF Λ ΛAD
which satisfies RTz Rz = I3 and detRz = 1. Therefore, Rz ∈
SO(3) is a rotation matrix around the z-axis. where ΛAD is defined by ΛAD = −[0 Λy Λz ]T . Define
Next, similarly, we can construct a rotation matrix Ry ΞAD = [ka ΩAD ΛAD ]T , and with Ξ given in (71), the
around the y-axis. Define a vector m = [Λx 0 Λz ]T , which controller ξF can be further written as
is orthogonal to the unit vector e2 = [0 1 0]T . Then, m⊥
can be given by ξF = Ξ + ΞAD . (81)

m⊥ = e2 × m = Λz 0 −Λx .
 T Substituting (81) into (79), we obtain the following closed-
loop system
Thus, the rotation matrix around the y-axis can be constructed ∧ ∨
by ẊCF = [BX ](Ξ + ΞAD − (Adg−1 ξC ) ). (82)
h i CF
m m⊥
Ry = kmk e2 km ⊥k . (74) Then, substituting (71) into (82), it follows that
In fact, Rz and Ry represent the rotation matrices from the ẊCF =[BX ](−kp (logSE(3) (ḡ −1 gLF ))∨ + (Adg−1 ξL
∧ ∨
)
LF
vector ΛN H to Λ. Hence, based on the logarithm of Rz and ∧ ∨
+ ΞAD − (Adg−1 ξC ) ) (83)
Ry , we design the following angular velocity CF

Utilizing the definitions in (69)(61)(62)(65)(77), we further

ΩAD = (logSO(3) (Ry ))∨ + (logSO(3) (Rz ))∨ . (75) formulate the closed-loop system (83) to be
With the help of the additional angular velocity ΩAD , the ∧ ∨
ẊCF =[BX ](−kp XCF + (Adg−1 ξC )
formation controller is given below. CF
∧ ∨
∧
+ ΞAD − (Adg−1 ξC ) )
Theorem 3. Let (gL , ξL ) denote the configuration and veloc- CF

ity of the leader, and let (gF , ξF∧ ) denote those of the follower. = − kp [BX ]XCF + [BX ]ΞAD . (84)
 11

With the definition of [BX ] in [43], it can be derived that

[BX ]XCF = XCF . Then, the closed-loop system (84) can be
written as 50

ẊCF = −kp XCF + [BX ]ΞAD . (85) 40

30

According to the proof of Theorem 1 in [44], XCF = 0 is

z
20

the exponentially stable equilibrium of the closed-loop system 10

(85). This indicates that the follower gF realizes the formation 0

pattern ḡ with respect to the leader gL under the controller ξF 100

80 250
200
60
in (76). 40 100
150
20 50
y 0 0 x
Remark 8. The additional angular velocity ΩAD in (75)
can be further understood from two aspects. On the one (a) Line reference trajectory
hand, intuitively, ΩAD is derived from the constructed rotation
matrices Rz and Ry , which represent the rotation from the
real linear velocity vector ΛN H to the standard velocity vector
200
Λ, so that ΩAD provides an extra rotation that compensates
the lack of linear velocities v y and v z . On the other hand, 150

theoretically, the additional angular velocity ΩAD (which

z
100
describes the rotation around x−, y−, z−axis) is the feedback
of the standard linear velocities Λy and Λz (which describes 50
150
the translation along y−, z−axis), demonstrating the idea 0 100
of feedback coupling in the control theory of underactuated 200
150
50

systems. Note that Λy and Λz originally cannot be imported 100

to the system due to the nonholonomic constraints. But with the
help of ΩAD , the underactuated directions becomes coupled
with the actuated directions, leading to the fact that the
underactuated states are able to be controlled by the feedback
in the actuated directions.
VI. S IMULATION EXAMPLES
This section provides two numerical simulation examples to
verify the effectiveness of the proposed results.
Fig. 4. Communication topology of 5 fixed-wing UAVs
The first example shows the formation of 5 fixed-wing
UAVs interacted by a directed acyclic graph as shown in
Figure 4. The reference trajectory of the leader is chosen as
two cases: a line and a helix. Note that the line reference
trajectory implies the leader’s angular velocity is zero, i.e.,
ωL = 03×1 . According to Corollary 1, the desired relative
rotation matrix of each UAV can be chosen as R̄ = I3 ,
indicating that all of the UAVs pointing to the same direction
in the formation. Then, under the line reference trajectory, we
set the desired formation pattern to be ḡ = (I3 , p̄), where
I3 represents the rotation matrix and the position vector p̄
composes a wedge-shaped formation pattern. Regarding the
helix reference trajectory, the position vector p̄ also forms
a wedge shape as the above, while the rotation matrix R̄
is not the identity matrix but given in Theorem 1. That is
to say, in this case, the orientations of the fixed-wing UAVs
50
0 0
x
y
(b) Helix reference trajectory
Fig. 5. Trajectories of 5 fixed-wing UAVs in RTI formation
cannot be identical but should be decided by the nonholonomic
constraints.
Figure 5 and Figure 6 demonstrate the simulation results.
The trajectories of the fixed-wing UAVs are shown in Figure 5,
where the leader is marked in blue and the followers are
in green. It can be seen from this figure, especially from
Figure 5(b), that the whole formation shape can rotate and
translate along with the reference trajectory, like a single rigid
body, demonstrating that the formation pattern is RTI. Figure 6
illustrates the relative configurations of each follower with
respect to the leader, which are expressed in the leader’s body-
fixed frame. As we can see, the relative attitude angles φ̄, θ̄, ψ̄
in these two cases are different. When the leader moves along
a straight line, φ̄, θ̄, ψ̄ can all be zero as shown in Figure 6(a),
which is guaranteed by Corollary 1. But once the leader has
nonzero angular velocity, as shown in Figure 6(b), θ̄ and ψ̄ are
not zero anymore, but decided by (27) and (28), respectively,
which results from the nonholonomic constraints of the fixed-
wing UAVs.
In the second example, we provide the simulation of 10
fixed-wing UAVs under a more complicated reference trajec-
tory, where the leader’s angular velocity ωL is a piecewise
continuously differentiable function as given in Table I, and
the leader’s linear velocity vL is set to be constant. It can be
observed that such a reference trajectory contains the straight
line motion, 2-D rotation and 3-D rotation. The simulation
time is chosen as T = 100s, and the results are presented
in Figure 7, where Figure 7(a) provides the trajectories of 10
fixed-wing UAVs in 3D, and the projections to x − z plane
 12

0.5 0.5 2 the formation problem in more practical scenarios, such as

No.0
0 1
No.1
No.2
obstacle-cluttered environments, measurement noises, and so
0 -0.5 0

-1 -1
No.3
No.4 on.
-0.5 -1.5 -2
0 10 20 0 10 20 0 10 20

20 50 20
R EFERENCES
0
0 [1] R. W. Beard and T. W. McLain, Small Unmanned UAV: Theory and
0 -20
-20
Practice. Princeton, New Jersey: Princeton University Press, 2012.
-40
[2] R. W. Beard, T. W. McLain, D. B. Nelson, D. Kingston, and D. Jo-
-40
0 10 20
-50
0 10 20
-60
0 10 20 hanson, “Decentralized cooperative aerial surveillance using fixed-wing
miniature UAVs,” Proceedings of the IEEE, vol. 94, no. 7, pp. 1306–
(a) Line reference trajectory 1324, Jul. 2006.
[3] P. B. Sujit, S. Saripalli, and J. B. Sousa, “Unmanned aerial vehicle path
1 0.5 2 following: A survey and analysis of algorithms for fixed-wing unmanned
0
No.0
No.1
aerial vehicless,” IEEE Control Systems Magazine, vol. 34, no. 1, pp.
0 1
-0.5
No.2
No.3
42–59, Feb. 2014.
-1
-1
0 No.4 [4] X. Wang, L. Shen, Z. Liu, S. Zhao, Y. Cong, Z. Li, S. Jia, H. Chen,
-2 -1.5 -1
Y. Yu, Y. Chang, and Y. Wang, “Coordinated flight control of miniature
0 10 20 0 10 20 0 10 20
fixed-wing UAV swarms: methods and experiments,” Science China
20 50 20
Information Sciences, vol. 62, Nov. 2019, (Article No. 212204).
0
[5] Z. Liu, X. Wang, L. Shen, S. Zhao, Y. Cong, J. Li, D. Yin, S. Jia,
0
0 -20
and X. Xiang, “Mission-oriented miniature fixed-wing UAV swarms:
-20
-40
A multilayered and distributed architecture,” IEEE Transactions on
-40 -50 -60
Systems, Man, and Cybernetics: Systems, vol. 52, no. 3, pp. 1588–1602,
0 10 20 0 10 20 0 10 20 Mar. 2022.
[6] K.-K. Oh, M.-C. Park, and H.-S. Ahn, “A survey of multi-agent
(b) Helix reference trajectory formation control,” Automatica, vol. 53, pp. 424–440, Mar. 2015.
[7] X. Wang, Y. Yu, and Z. Li, “Distributed sliding mode control for leader-
Fig. 6. Relative configurations of 5 fixed-wing UAVs in RTI formation follower formation flight of fixed-wing unmanned aerial vehicles subject
to velocity constraints,” International Journal of Robust and Nonlinear
Control, vol. 31, no. 6, pp. 2110–2125, Apr. 2021.
TABLE I [8] Q. Zhang and H. H. T. Liu, “Robust nonlinear close formation control
C OMPONENTS OF THE LEADER ’ S ANGULAR VELOCITY ωL of multiple fixed-wing aircraft,” Journal of Guidance, Control, and
Dynamics, vol. 44, no. 3, pp. 572–586, Mar. 2021.
x
ωL y
ωL z
ωL [9] T. Z. Muslimov and R. A. Munasypov, “Consensus-based cooperative
control of parallel fixed-wing UAV formations via adaptive backstep-
*
0 ≤ t ≤ 20 0 −0.15 sin a(t) 0 ping,” Aerospace Science and Technology, vol. 109, Feb. 2021, (Article
No. 106416).
20 < t ≤ 30 0 0 −0.25 sin b(t)* [10] H. M. Guzey, “Hybrid consensus-based formation control of fixed-wing
30 < t ≤ 50 0 0 0 MUAVs,” Cybernetics and Systems, vol. 48, no. 2, pp. 71–83, 2017.
[11] X. Wang, S. Baldi, X. Feng, C. Wu, H. Xie, and B. D. Schutter, “A
* *
50 < t ≤ 55 0.1 sin c(t) 0.15 sin c(t) 0.2 sin c(t)* fixed-wing UAV formation algorithm based on vector field guidance,”
IEEE Transactions on Automation Science and Engineering, 2022, (To
t > 55 0.1 0.15 0.2 be published, doi: 10.1109/TASE.2022.3144672).
* a(t) = 0.1πt, b(t) = 0.1π(t − 20), c(t) = 0.1π(t − 50). [12] S. Baldi, D. Sun, G. Zhou, and D. Liu, “Adaptation to unknown
leader velocity in vector-field UAV formation,” IEEE Transactions on
Aerospace and Electronic Systems, vol. 58, no. 1, pp. 473–484, Feb.
2022.
and y − z plane are given in Figure 7(b) and Figure 7(c), [13] Z. Sun, H. G. de Marina, G. S. Seyboth, B. D. O. Anderson, and
respectively. It is illustrated in these figures that the formation C. Yu, “Circular formation control of multiple unicycle-type agents with
nonidentical constant speeds,” IEEE Transactions on Control Systems
pattern can move along the reference trajectory, exhibiting the Technology, vol. 27, no. 1, pp. 192–205, Jan. 2019.
RTI characteristics as a single rigid body. [14] I. Bayezit and B. Fidan, “Distributed cohesive motion control of
flight vehicle formations,” IEEE Transactions on Industrial Electronics,
vol. 60, no. 12, pp. 5763–5772, Dec. 2013.
VII. C ONCLUSION [15] H. Chen, X. Wang, L. Shen, Z. Li, and Z. Liu, “Formation reconfig-
uration for fixed-wing UAVs: A coordinated path following approach,”
In this paper, we have investigated the RTI formation Journal of Intelligent & Robotic Systems, vol. 102, no. 1, May 2021,
problem of the fixed-wing UAVs by proposing the formation (Artical No. 27).
feasibility and designing the control strategy. Particularly, the [16] Y. Wang, M. Shan, and D. Wang, “Formation reconstruction and
trajectory replanning for multi-UAV patrol,” IEEE Transactions on
fixed-wing UAV is modelled by a rigid body in 3D, whose Mechatronics, vol. 26, no. 2, pp. 719–729, Apr. 2021.
kinematics evolves in the Lie group SE(3). The novelty of this [17] V. R. Challa and A. Ratnoo, “On maneuverability of fixed-wing un-
paper lies in the roto-translation invariance of the whole forma- manned aerial vehicle formations,” Journal of Guidance, Control, and
Dynamics, vol. 44, no. 7, pp. 1327–1344, Jul. 2021.
tion pattern, which characterizes an overall rigid-body motion. [18] Y. Wang, M. Shan, and D. Wang, “Motion capability analysis for multi-
Furthermore, the formation feasibility has been presented un- ple fixed-wing UAV formations with speed and heading rate constraints,”
der the nonholonomic and input saturation constraints, which IEEE Transactions on Control of Network Systems, vol. 7, no. 2, pp.
977–989, Jun. 2020.
lays the foundation to the formation control. In addition, we [19] Z. Sun, H. G. de Marina, B. D. O. Anderson, and C. Yu, “Collaborative
have employed the idea of feedback coupling in the design target-tracking control using multiple fixed-wing unmanned aerial vehi-
of control strategy so as to handle the underactuation of cles with constant speeds,” Journal of Guidance, Control, and Dynamics,
vol. 44, no. 2, pp. 238–250, Feb. 2021.
fixed-wing UAVs, which achieves the objective of controlling [20] Z. Chen, “On dubins paths to a circle,” Automatica, vol. 117, Jul. 2020,
more DOFs with fewer inputs. Future works will focus on (Article No. 108996).
 13

300

250

200

150

z
100
300
50

200
0
0 50 100 150 200 250 300 350 400 450 500 550
z

100 x

(b) Trajectories in x − z plane

0
350 300

300
250
250
200
200
550
150 500
150

z
450
100 400
350
100
50 300
250
0 200 50
y 150
-50 100
x
50 0
-100 0 350 300 250 200 150 100 50 0 -50 -100
y

(a) Trajectories in 3-D space (c) Trajectories in y − z plane

Fig. 7. Trajectories of 10 fixed-wing UAVs in RTI formation

[21] A. K. Das, R. Fierro, V. Kumar, J. P. Ostrowski, J. Spletzer, and
C. J. Taylor, “A vision-based formation control framework,” IEEE
Transactions on Robotics and Automation, vol. 18, no. 5, pp. 813–825,
Oct. 2002.
[22] W. Li, “Notion of control-law module and modular framework of
cooperative transportation using multiple nonholonomic robotic agents
with physical rigid-formation-motion constraints,” IEEE Transactions on
Cybernetics, vol. 46, no. 5, pp. 1242–1248, May 2016.
[23] X. Peng, Z. Sun, K. Guo, and Z. Geng, “Mobile formation coordination
and tracking control for multiple nonholonomic vehicles,” IEEE Trans-
actions on Mechatronics, vol. 25, no. 3, pp. 1231–1242, Jun. 2020.
[24] Y.-W. Chen, M.-L. Chiang, and L.-C. Fu, “Three-dimensional
maneuver control of multiagent systems with constrained input,”
IEEE Transactions on Cybernetics, 2022, (To be published, doi:
10.1109/TCYB.2022.3165767).
[25] P. Tabuada, G. J. Pappas, and P. Lima, “Motion feasibility of multi-
agent formations,” IEEE Transactions on Robotics, vol. 21, no. 3, pp.
387–392, Jun. 2005.
[26] Z. Sun and B. D. O. Anderson, “Formation feasibility on coordination
control of networked heterogeneous systems with drift terms,” in Pro-
ceedings of IEEE Conference on Decision and Control, Las Vegas, USA,
Dec. 2016, pp. 3462–3467.
[27] F. Morbidi and E. Bretagne, “A new characterization of mobility for
distance-bearing formations of unicycle robots,” in Proceedings of
IEEE/RSJ International Conference on Intelligent Robots and Systems,
Madrid, Spain, Oct. 2018, pp. 4833–4839.
[28] L. J. Colombo and D. V. Dimarogonas, “Motion feasibility conditions
for multiagent control systems on lie groups,” IEEE Transactions on
Control of Network Systems, vol. 7, no. 1, pp. 493–502, Mar. 2020.
[29] M. J. Whitzer, “Coordinating multi-robot teams in the presence of
motion constraints and heterogeneity,” Ph.D. dissertation, University of
Pennsylvania, 2020.
[30] L. Consolini, F. Morbidi, D. Prattichizzo, and M. Tosques, “On a class of
hierarchical formations of unicycles and their internal dynamics,” IEEE
Transactions on Automatic Control, vol. 57, no. 4, pp. 845–859, Apr.
2012.
[31] W. Yao and M. Cao, “Path following control in 3D using a vector field,”
Automatica, vol. 117, Jul. 2020, (Article No. 108957).
[32] S. Zhao, “Affine formation maneuver control of multiagent systems,”
IEEE Transactions on Automatic Control, vol. 63, no. 12, pp. 4140–
4155, Dec. 2018.
[33] X. He, J. Zhai, and Z. Geng, “Roto-translation invariant formation
of multiple underactuated planar rigid bodies,” IEEE Transactions on
Cybernetics, vol. 52, no. 12, pp. 12 818–12 831, Dec. 2022.
[34] C. K. Verginis, A. Nikou, and D. V. Dimarogonas, “Robust formation
control in SE(3) for tree-graph structures with prescribed transient and
steady state performance,” Automatica, vol. 103, pp. 538–548, May
2019.
[35] J. Thunberg, J. Goncalves, and X. Hu, “Consensus and formation control
on SE(3) for switching topologies,” Automatica, vol. 66, pp. 109–121,
Apr. 2016.
[36] M. Wang and A. Tayebi, “Hybrid feedback for global tracking on matrix
lie groups SO(3) and SE(3),” IEEE Transactions on Automatic Control,
vol. 67, no. 6, pp. 2930–2945, Jun. 2022.
[37] G. Liu, G. Zhang, Y. Guan, Y. Yang, and X. Chen, “Geometry of adjoint-
invariant submanifolds of SE(3),” IEEE Transactions on Robotics,
vol. 37, no. 2, pp. 699–705, Apr. 2021.
[38] X. Wang, C. Yu, and Z. Lin, “A dual quaternion solution to attitude
and position control for rigid-body coordination,” IEEE Transactions on
Robotics, vol. 28, no. 5, pp. 1162–1170, Oct. 2012.
[39] A. M. Bloch, Nonholonomic Mechanics and Control. New York, NY:
Springer, 2003.
[40] X. Peng, J. Sun, and Z. Geng, “The geometric convexity on SE(3)
and its application to the formation tracking in multivehicle systems,”
International Journal of Control, vol. 92, no. 3, pp. 528–539, Mar. 2019.
[41] F. Bullo and A. D. Lewis, Geometric Control of Mechanical Systems:
Modeling, Analysis, and Design for Simple Mechanical Control Systems.
New York, NY: Springer, 2005.
[42] X. He and Z. Geng, “Trajectory tracking of nonholonomic mobile robots
by geometric control on special Euclidean group,” International Journal
of Robust and Nonlinear Control, vol. 31, no. 12, pp. 5680–5707, Aug.
2021.
[43] F. Bullo and R. Murray, “Proportional derivative (PD) control on the
Euclidean group,” in Proceedings of European Control Conference,
Rome, Italy, Sep. 1995, pp. 1091–1097.
[44] X. He, Z. Sun, Z. Geng, and A. Robertsson, “Exponential set-point
stabilization of underactuated vehicles moving in three-dimensional
space,” IEEE/CAA Journal Automatica Sinica, vol. 9, no. 2, pp. 270–
282, Feb. 2022.