You are on page 1of 9

148 IEEE/CAA JOURNAL OF AUTOMATICA SINICA, VOL. 8, NO.

1, JANUARY 2021

Modeling and Trajectory Tracking Control for


Flapping-Wing Micro Aerial Vehicles
Wei He, Senior Member, IEEE, Xinxing Mu, Liang Zhang, and Yao Zou

Abstract—This paper studies the trajectory tracking problem multiple disciplines, e.g., control, computer science,
of flapping-wing micro aerial vehicles (FWMAVs) in the mechanics and informatics, etc.
longitudinal plane. First of all, the kinematics and dynamics of the
FWMAV are established, wherein the aerodynamic force and
Researchers have made great efforts in developing
torque generated by flapping wings and the tail wing are FWMAVs over the past few years [8], [9]. By imitating
explicitly formulated with respect to the flapping frequency of the different creatures, FWMAVs are generally categorized into
wings and the degree of tail wing inclination. To achieve insect-scale micro aerial vehicles (MAVs) and bird-scale
autonomous tracking, an adaptive control scheme is proposed micro aerial vehicles (MAVs). Currently, typical representa-
under the hierarchical framework. Specifically, a bounded tives of the insect-scale FWMAVs are the “Robobee” [4] and
position controller with hyperbolic tangent functions is designed
to produce the desired aerodynamic force, and a pitch command “nano hummingbird” [10]. Wood’s team from Harvard
is extracted from the designed position controller. Next, an University invents an insect-scale FWMAV “Robobee” by
adaptive attitude controller is designed to track the extracted using high-performance piezoelectric actuators, and such an
pitch command, where a radial basis function neural network is FWMAV is capable of hovering at a fixed height [4].
introduced to approximate the unknown aerodynamic However, “Robobee” cannot be powered sufficiently for long-
perturbation torque. Finally, the flapping frequency of the wings duration maneuvers by the limitation of load capacity.
and the degree of tail wing inclination are calculated from the
designed position and attitude controllers, respectively. In terms AeroVironment, Inc. develops an insect-scale FWMAV “nano
of Lyapunov’s direct method, it is shown that the tracking errors hummingbird” [10], which can finish some specific tasks. e.g.,
are bounded and ultimately converge to a small neighborhood vertical take-off and landing, hovering, free flight forward and
around the origin. Simulations are carried out to verify the backward. For the bird-like FWMAVs, researchers first built a
effectiveness of the proposed control scheme. bat-like “Bat Bot2” by imitating the limb structure and flight
Index Terms—Flapping-wing micro aerial vehicles (FWMAVs), mode of a bat [11]. To facilitate the flight, “Bat Bot2” adopts
modeling, neural networks, trajectory tracking. a boundary control method to make its wings flutter in
accordance with the desired trajectory. Festo designs a light-
I. Introduction weighted bird-like FWMAV “Smartbird” by taking a
IOMIMETIC robots are popular research subjects in streamlined shape and wing twist into account, which is able
B many fields of robotics [1]–[3]. In recent decades,
flapping-wing micro aerial vehicles (FWMAVs), as a type of
to fly stably by remote control [12]. He’s team from the
University of Science and Technology Beijing developed a
innovative flying robots, have received considerable attention bird-like FWMAV “USTBird” driven by 2 servo motors,
in the aviation community [4], [5]. Characterized by good which changes course by asymmetric flapping of its wings.
agility, ability for concealment and high flight efficiency, they Although these FWMAVs show good flight performance,
have extensive potential applications in military and civil having an autonomous flight mode is not taken into account.
fields including disaster investigation, intelligence collection, To achieve autonomous flight, it is crucial to determine the
anti-terrorism reconnaissance and so forth [6], [7]. The driving behaviors of the FWMAVs. For the insect-scale
research of the FWMAVs is comprehensive and integrates MAVs, a flapping frequency is set to around 20–600 Hz for
Manuscript received February 28, 2020; revised April 28, 2020; accepted
providing sufficient thrust to allow them to hover [13]. To
June 3, 2020. This work was supported in part by the National Natural adjust their flight attitude, they are generally steered by
Science Foundation of China (61933001, 62061160371), Joint Funds of fluttering the wings asymmetrically [14]. The flapping
Equipment Pre-Research and Ministry of Education of China (6141A02033 frequency of most bird-like MAVs is below 10 Hz for level
339), and Beijing Top Discipline for Artificial Intelligent Science and
Engineering, University of Science and Technology Beijing. Recommended flight. Their lift mainly results from the aerodynamic force
by Associate Editor Dongpu Cao. (Corresponding author: Wei He.) generated by the forward flight speed, and in the meantime,
Citation: W. He, X. X. Mu, L. Zhang, and Y. Zou, “Modeling and their flight orientation is altered by adjusting both the flapping
trajectory tracking control for flapping-wing micro aerial vehicles,” and tail wings [15]. For dynamic analysis and modeling, the
IEEE/CAA J. Autom. Sinica, vol. 8, no. 1, pp. 148–156, Jan. 2021.
The authors are with the Institute of Artificial Intelligence, University of
power curves associated with different flight modes and
Science and Technology Beijing, and also with the School of Automation and speeds of several certain species have been extracted in [16],
Electrical Engineering, University of Science and Technology Beijing, [17]. Meantime, some relevant analysis has been given
Beijing 100083, China (e-mail: weihe@ieee.org; muxinxing@126.com; 1881
1329798@163.com; zouyao@ustb.edu.cn).
therein. In addition, the leading-edge vortex effect on the tail
Color versions of one or more of the figures in this paper are available wing in the aerodynamic performance has been formulated in
online at http://ieeexplore.ieee.org. [18]. Besides, the external disturbances have an important
Digital Object Identifier 10.1109/JAS.2020.1003417 influence on the performances of FWMAVs and is considered

Authorized licensed use limited to: Tamkang Univ.. Downloaded on January 17,2024 at 14:50:53 UTC from IEEE Xplore. Restrictions apply.
HE et al.: MODELING AND TRAJECTORY TRACKING CONTROL FOR FWMAVs 149

in practice [19], [20]. In this paper, we focus on the modeling of the FWMAV. Moreover, a hierarchical framework is
of bird-like FWMAVs, which is characterized by nonlinearity, introduced to exploit the cascaded structure of the established
parametric coupling, and uncertainty. model for control scheme development.
Many intelligent methods have been developed and applied 2) This paper considers the unknown aerodynamic
in various fields [21]–[24], and many control approaches are perturbation of flapping wings on the torque generated by the
applied in the FWMAVs to achieve autonomous flight tail wing. A radial basis function neural network is introduced
[25]–[27]. In [25], a DelFly Explorer equipped with a 0.98 to estimate and compensate for this perturbation and for
gram autopilot unit and a 4.0 gram onboard stereo vision improving tracking accuracy.
system is established, and a stereo vision algorithm is 3) This paper designs a bounded position controller with
developed to achieve autonomous obstacle avoidance. In [28], hyperbolic tangent functions to guarantee a bounded
the location of the trailing edges at the wing roots is adjusted aerodynamic force. Also, this design effectively alleviates the
to generate control torques through an angular rate feedback coupling between the closed-loop position and attitude error
control scheme for the insect-scale tailless FWMAV, which systems, and thus facilitates the stability analysis greatly.
achieves stable vertical flight and hovering. In [26], the The remaining sections of this paper are organized as
authors propose a hybrid control policy which combines follows. Section II states the model and control objective of
model-based nonlinear control with model-free reinforcement the FWMAV. Section III presents the main results of this
learning. The model-free reinforcement learning policy is paper including the control scheme development and the
trained to accomplish fast evasive maneuvers on the stability analysis. Section IV performs simulations to validate
hummingbird robot. In [27], neural network based control and highlight the performance of the proposed control
strategies with full state and output feedback are developed scheme. Finally, Section V makes conclusions and draws
for the attitude and position tracking of the FWMAV. In future works.
particular, a model-based nonlinear controller based on hybrid
control strategy is proposed therein to guarantee the stability II. Problem Statements
of the closed-loop FWMAV system.
A. System Models
Although multiple mathematical models and control
strategies have been proposed for the FWMAVs, most of them Generally, an FWMAV is modeled in three reference
just consider aerodynamic force and torque as control inputs frames (shown in Fig. 1) as follows:
and do not establish their aerodynamic relationship with Inertia frame: S i = {oi xi yi zi }, origin oi is fixed at a point on
respect to actual physical actuators. Therefore, it is important the ground, axes xi and zi point to the north and east, and axis
to use these control strategies in practice for reliable flight. In yi is determined by the right hand rule.
this paper, we develop a bird-like FWMAV, which has 2 Track frame: S t = {ot xt yt zt }, origin ot is located at the center
degrees of freedom (DOFs) of flapping frequency and tail of mass (c.m.) of the fuselage, axis xt is along the velocity
wing inclination to achieve automatous trajectory tracking on direction, axis zt is perpendicular to the plane containing axis
the vertical plane. It is important to formulate the relationship xt and pointing to the right, and axis yt is determined by the
of the aerodynamic force and torque with respect to flapping right hand rule.
frequency and tail wing inclination for real applications. Body frame: S b = {ob xb yb zb }, origin ob coincides with ot ,
Moreover, we also consider the airflow effect of flapping axis xb points towards the head, axis zb is perpendicular to the
wings in the aerodynamics of the tail wing in the modeling of plane of symmetry containing axis xb and points to the right,
the FWMAVs. Based on the established model, a neural and axis yb is determined by the right hand rule.
network control scheme is proposed under the hierarchical
yb yt
framework such that trajectory tracking of the FWMAV on
the longitudinal plane is achieved. Specifically, a position yi
controller with control input saturation in [29] is synthesized,
and the frequency of flapping wings and a pitch command are xb
extracted from the synthesized position controller. Further, an
attitude controller is synthesized by introducing a radial basis Ob
function neural network for the pitch to track commands, Oi τ Ot V xt
where the neural network is used to eliminate the aerodynamic
zb xi
effect caused by the airflow of flapping wings. Finally, the tail zi zt
wing inclination is extracted from the synthesized attitude
controller. With Lyapunov’s direct method, it is shown that Fig. 1. Reference frames of a typical FWMAV.
tracking errors driven by the proposed control scheme are
bounded and ultimately converge to a neighborhood around For the convenience of modeling, some reasonable
the origin. The main contributions of this paper are assumptions are made as follows:
enumerated as follows: Assumption 1: The aerodynamic force of the tail wing and
1) This paper formulates the aerodynamic force and torque the torque generated by the drag of the tail wing are so small
generated by the actual flapping frequency of flapping wings that it is negligible.
and tail wing inclination when constructing the system model Assumption 2: The increments of thrust and lift are

Authorized licensed use limited to: Tamkang Univ.. Downloaded on January 17,2024 at 14:50:53 UTC from IEEE Xplore. Restrictions apply.
150 IEEE/CAA JOURNAL OF AUTOMATICA SINICA, VOL. 8, NO. 1, JANUARY 2021

proportional to the flapping frequency. TABLE I


Remark 1: Studies in [18], [30] indicate that the bird’s tail Model Parameter Definitions
has nearly no effect on lift and drag. By considering the short Symbol Description Unit
arm with respect to c.m., the torque generated by the drag of AR Aspect ratio –
the tail wing can be negligible. However, the lift generated by
L Wing length m
the tail wing still contributes to the pitch torque arising from
lec Distance from the center of gravity to
its long arm with respect to c.m.. Generally, the aerodynamic the center of lift of flapping wings m
force of the FWMAV is related to the flapping frequency lt Distance from the center of gravity to m
the tail wing’s hinge
[31]–[33]. There exists almost a linear relationship between
lh Distance from the tail wing’s hinge to the
the flapping frequency and the average of lift, and the thrust is m
aerodynamic center of the tail wing
almost proportional to the flapping frequency [34]. In this S Area of flapping wings m2
paper, the FWMAV’s tail is designed with a biological tail St Area of the tail wing m2
structure, therefore, Assumptions 1 and 2 are reasonable. ρ kg/m3
Air density
Under Assumptions 1 and 2, and according to Newton’s m Total mass kg
second law, the longitudinal motion of the FWMAV can be
g Acceleration of gravity m/s2
formulated as follows [35], [36]:
ν Correction coefficient –
M q̈ + G = uc (1)
τ Torque of axis zb N·m
where θ Pitch rad
   
[ ]  m 0 0  [ ]  x  α Angle of attack rad
Mi 0   p  
M= =  0 m 0  , q = =  y  γ Flightpath slope rad
0 J   θ  
0 0 J θ m/s
    V Free stream velocity
[ ]  0  [ ]  F xi 
Gi   Fi   J Moment of inertia about the axis zb m4
G= =  mg  , uc = =  Fyi  .
0   τ  CD Drag coefficient –
0 τ
CL Lift coefficient –
Specifically, p = [x, y]T is the real trajectory of c.m. of the kx Frequency increment coefficient at axis xb –
fuselage in the inertia frame, and other parameter definitions ky
are given in Table I. Frequency increment coefficient at axis yb –
Moreover, under Assumption 2, the aerodynamic force R12 Rotation matrix from 2 to 1 –
components generated by flapping wings in the track frame f Flapping frequency Hz
are expressed as θt Tail wing inclination rad
ρ F·i , T ·i The aerodynamics forces of the flapping wings N
F xk = k x f − C D (α)S V 2 and the tail wing in the inertia frame
2 F·t , T ·t The aerodynamics forces of the flapping wings
ρ and the tail wing in the track frame N
Fyk = ky f + C L (α)S V 2 (2) The aerodynamics forces of the flapping wings
2 F·b , T ·b N
and the tail wing in the body frame
where the aerodynamic parameters are also defined in Table I.
According to [37], the lift and drag coefficients are
determined by perpendicular to the body axis on the plane of symmetry of
the tail wing, is expressed as
C D (α) = 1.55 cos(2.04α − 9.82◦ ) + 1.92 ρ
C L (α) = 1.58 sin(2.13α − 7.2◦ ) + 0.225. (3) T yb = C L (θt )S t v2b (6)
2
Correspondingly, the aerodynamic force expressed in the where vb = V cos α is the air speed along the direction of axis
inertia frame satisfies xb . It can be known from Fig. 2. The pitch torque in the body
frame is expressed as
Fi = Rit (−γ)Ft (4)
τ = −Fyb lec − (T yb + δ)(lh + lt cos θt ) (7)
where γ = θ − α , F = [F xi , Fyi ]T , Ft = [F xk , Fyk ]T and
[ ] where δ is the aerodynamic perturbation on the tail wing
cos γ − sin γ
R (−γ) =
it
. resulting from the flutter of flapping wings.
sin γ cos γ
Remark 2: Due to complex aerodynamic characteristics, it is
Moreover, the aerodynamic force (2) expressed in the body imperative to determine the aerodynamic perturbation δ . A
frame satisfies potential formulation was presented in [38] from the fixed-
wing perspective. In particular
Fb = Rbt (α)Ft (5)
 
where Fb = [F xb , Fyb and ]T 2C L (α) ∂C∂α
L (α)
 ν2 ( π4 )2 + ν2 
[ ] δ= ln π 2 2 − ln  QS (8)
cos α sin α AR π3 (4) +ν ν2
R (α) =
bt
.
− sin α cos α where Q = 0.5ρV 2 and ν is the correction term. Note that it is
In addition, the aerodynamic force of the tail wing, which is related to the attack angleα and the free stream velocity V . In

Authorized licensed use limited to: Tamkang Univ.. Downloaded on January 17,2024 at 14:50:53 UTC from IEEE Xplore. Restrictions apply.
HE et al.: MODELING AND TRAJECTORY TRACKING CONTROL FOR FWMAVs 151

this paper, the aerodynamic perturbation δ is assumed to be where u p = Rit (−γd )Ft is the position controller to be
unavailable and is approximated by a neural network. specified, R e = Rit (−γ) − Rit (−γd ), and γd is the command
flightpath slope. Define ∆x = xd (ti+1 ) − x(ti ) and
yt
yb up ∆y = yd (ti+1 ) − y(ti ). Then the command flightpath slope can
yi Fyt Fi be evaluated by γd = arctan(∆y/∆x).
Define the position tracking error e p1 = p1 − pd and the
Fyb
xb velocity tracking error e p2 = p2 − ṗd . By virtue of (9), their
Fxb V dynamics are derived as
αd
Tyb lt γ xt
ė p1 = e p2
Ot(b) Fxt e t − Gi ) − p̈d .
ė p2 = Mi−1 (u p + RF (10)
θt
Oi lh l ec xi
The position controller with saturation is synthesized in the
Gi following form:
u p = Gi + Mi p̈d − Mi l1 tanh(k1 e p1 + k2 e p2 )
Fig. 2. Aerodynamic force analysis diagram. − Mi l2 tanh(k2 e p2 ) (11)
where l1, l2 , k1 and k2 are positive control parameters. By
B. Control Objective substituting (11) into (10), the following closed-loop position
In this paper, we focus on the solution to the trajectory error dynamics can be derived
tracking control problem of the FWMAV on the longitudinal ė p1 = e p2
plane. In particular, given a reference trajectory pd = [xd , yd ]T , ė p2 = −l1 tanh(k1 e p1 +k2 e p2 )
we intend to design flapping frequency f and tail wing e t.
− l2 tanh(k2 e p2 ) + Mi−1 RF (12)
inclination θt via an adaptive control scheme such that
limt→∞ |p(t) − pd (t)∥ ≤ ϵ , where ϵ > 0 is a sufficiently small Since the position controller u p has been determined by
constant. Since the FWMAV system (1) is of a cascaded (11), the flapping frequency f and the pitch command θd can
structure consisting of an outer position loop and an inner be extracted accordingly. Note from (2) that the aerodynamic
attitude loop (see Fig. 3 for illustration), the control scheme force [F xk , Fyk ]T is nonlinear with respect to the flapping
design is under a hierarchical framework such that the frequency f and the command attack angle αd . To solve this
position and attitude loops are studied in sequence. Also, note nonlinear equation, we introduce the trust-region-dogleg
from (7) that, the pitch torque τ is subject to an unknown algorithm [42] such that f and αd are evaluated. Moreover,
aerodynamic perturbation δ . We introduce a neural network the pitch command can be determined by θd = αd + γd.
technique to approximate and compensate for these
B. Attitude Controller Synthesis
perturbations which improves tracking accuracy [39]–[41].
Define θ1 = θ and θ2 = θ̇ . Based on (1) and (7), their
f p
Mi p + Gi = Fi dynamics are derived as
θ̇1 = θ2
θt θ
Jθ = τ θ̇2 = J −1 (ur + ∆δ ) (13)
where ur = −Fyb lec − T yb (lh + lt cos θt ) is the attitude controller
Fig. 3. Cascaded structure of the FWMAV system.
and ∆δ = δ(lh + lt cos θt ) is the lumped uncertainty caused by
unknown δ . Define the pitch tracking error er1 = θ1 − θd and
III. Main Results
its changing rate error er2 = θ2 − αr, where αr = θ̇d − k3 er1 with
In this section, an adaptive control scheme under the k3 being a positive constant. It follows from (13) that their
hierarchical framework is developed to solve the concerned dynamics satisfy
tracking control problem of the FWMAV. In particular, a
position controller with input saturations is synthesized for ėr1 = −k3 er1 + er2
position tracking to pd . Then, the flapping frequency f and ėr2 = J −1 (ur + ∆δ ) − α̇r . (14)
the pitch command θd are extracted from the synthesized Due to the unavailability of ∆δ, a radial basis function
position controller. Next, the attitude controller introducing neural network with l nodes is introduced to approximate it. In
the neural network adaptation is synthesized for attitude particular,
tracking to θd . Finally, the tail wing inclination θt is extracted
from the synthesized attitude controller. Besides these, a ∆δ = (W ∗ )T S (z) + ϵ (15)
detailed stability analysis is also performed. where W∗ ∈ Rl is the weight vector, z = [α, V, θt ]T ,
S (z) : R → R is the radial basis function vector with each
3 l
A. Position Controller Synthesis component being in a Gaussian function form [43], and ϵ is
Define p1 = p and p2 = ṗ . It follows from (1) that the bounded estimation error. To achieve pitch tracking, the
ṗ1 = p2 attitude controller is synthesized in the following form:
e t − Gi )
ṗ2 = Mi−1 (u p + RF (9) b T S (z)
ur = −er1 − k4 er2 + J α̇r − W (16)

Authorized licensed use limited to: Tamkang Univ.. Downloaded on January 17,2024 at 14:50:53 UTC from IEEE Xplore. Restrictions apply.
152 IEEE/CAA JOURNAL OF AUTOMATICA SINICA, VOL. 8, NO. 1, JANUARY 2021

Position loop

Desired trajectory pd ep1 ep2 f Actual position p


Position controller
Actual attitude θ

Command attitude θd er1 er2 θt


Attitude controller

Adaptive law

Attitude loop

Fig. 4. Block diagram of control scheme.

b is the estimation of
where k4 is a positive control gain and W controller.

W . Moreover, the adaptive law with a σ-modification for W b Theorem 1: Consider the attitude error system (18). The
is designed as synthesized attitude controller (16) with the adaptive law (17)
guarantees that the attitude tracking error er1, its changing rate
ḃ = ξ(S (z)er2 − σW)
W b (17) error er2 and the estimation error W e are bounded and
where ξ and σ are positive adaptive parameters. In addition, ultimately converge to small neighborhoods around origin,
by considering (15) and substituting the synthesized attitude respectively.
controller (16) into (14), the following closed-loop pitch error Proof: Consider a Lyapunov function candidate as
dynamics can be derived:
1 J 1 e 2
ėr1 = −k3 er1 + er2 V1 = e2r1 + e2r2 + ∥W∥ . (19)
2 2 2ξ
e T S (z) + ϵ)
ėr2 = J −1 (−er1 − k4 er2 + W (18) Its derivative along (18) satisfies
e ∗ b
where W = W − W is the estimation error.
e T S (z)+ϵ)
V̇1 = er1 (− k3 er1 +er2 )+er2 (− er1 −k4 er2 +W
Since the attitude controller ur has been determined by (16),
the tail wing inclination θt can be extracted. Due to their 1 e T ḃ
− W W
nonlinear relation, the trust-region-dogleg algorithm is still ξ
available here to obtain θt . 1 e T ḃ
= − k3 e2r1 − k4 e2r2 − W (W − ξS (z)er2 )+er2 ϵ. (20)
Remark 3: Intuitively, Fig. 4 illustrates the overall control ξ
scheme implementation progress. Specifically, a brief By substituting the adaptive law (17) into (20), it follows
description is made as follows: that
1) Calculate the position controller u p via (11).
2) Extract the flapping frequency f and the command angle eTW
V̇1 = −k3 e2r1 − k4 e2r2 + σW b + er2 ϵ
of attack αd via (1) and the trust-region-dogleg algorithm. e T (W ∗ − W)
e + er2 ϵ
≤ − k3 e2r1 − k4 e2r2 + σW
Then, calculate the pitch command θd = αd + γd.
3) Calculate the weight estimation W b via (17). Then k4 2 σ e 2 σ ∗ 2 1 2
evaluate the attitude controller ur via (16). ≤ − k3 e2r1 − er2 − ∥W∥ + ∥W ∥ + ϵ
2 2 2 2k4
4) Extract the tail wing inclination θt from the determined
attitude controller ur . ≤ − µ1 V 1 + C 1 (21)
where µ1 = µ2 /µ3 with µ2 = min{k3 , k4 /2, σ/2} and µ3 = max
C. Stability Analysis {1/2, J/2, 1/2ξ}, and C1 = σ∥W ∗ ∥2 /2 + ϵ 2 /2k4.
In this subsection, we focus on the stability analysis of the
Next, by multiplying both sides of (21) with eµ1 t and
closed-loop system (12) and (18), which is examined to be of
performing analytical operations, it can be derived that
a cascaded structure. We carry out the analysis based on a
hierarchical framework. In particular, we show that the d
(V1 eµ1 t ) ≤ C1 eµ1 t . (22)
attitude tracking error is bounded and ultimately converges to dt
a small neighborhood of origin. Further, we show that the According to comparison lemma, it follows that
position tracking error is also bounded and ultimately ( )
C1 −µ1 t C1
converges to a small neighborhood of origin. Moreover, these V1 (t) ≤ V1 (0) − e + . (23)
neighborhoods can be tuned arbitrarily small by adjusting the µ1 µ1
control parameters. Before proceeding with the main results, a This implies that V1 is bounded and ultimately converges to
useful lemma is presented. the compact set
Lemma 1 [44]: Given x ∈ S = {x ∈ Rn | ∥x∥ < x̄, x̄ > 0}, there { }
C1
exists a positive constant χ( x̄) such that χ( x̄)∥x∥ ≤ ∥ tanh(x)∥ ≤ ΩV = V1 ∈ R+ |V1 ≤ .
∥x∥. µ1
The next theorem indicates that the ultimately bounded e
This, according to (19), further implies that er1, er2 and W
attitude tracking is achieved with the synthesized attitude are bounded and converge to the compact sets

Authorized licensed use limited to: Tamkang Univ.. Downloaded on January 17,2024 at 14:50:53 UTC from IEEE Xplore. Restrictions apply.
HE et al.: MODELING AND TRAJECTORY TRACKING CONTROL FOR FWMAVs 153

{ √ }
2C1 − k2 ΘT Θ − l2 k1 eTp2 tanh(η2 )
Ωer1 = er1 ∈ R||er1 | ≤
µ1 l2 k1
{ √ } ≤ −k2 ΘT Θ − tanh(η2 )T tanh(η2 )
2C1 k2
Ωer2 = er2 ∈ R||er2 | ≤
Jµ1 = − tanh(η)T D tanh(η) (27)
{ √ }
e ∈ Rl |∥W∥
e ≤ 2C1 ξ where
ΩW = W  2 
µ1  l1 k2 I2l1 l2 k 2 I2 
respectively. Note that the compact sets Ωer1 and Ωer2 can be D =  l2 k1 
(l2 k2 +
2
l1 l2 k2 I2 )I2 
made arbitrarily small by sufficiently increasing the control k2
parameters k3 and k4. ■ with I2 being a 2 × 2 unit matrix. There exists an invertible
In view of (19) and (23), it is also trivial to show that for matrix
any ς > 0 and τ1 > 0 , a sufficiently large µ1 is sufficient to  
guarantee ∥er1 (t)∥ ≤ ς , t ∈ [τ1 , ∞). By implementing this result,  l2 

 I − I
the following theorem states that the ultimately bounded T =  2
l1 
2

02 I2
position tracking is fulfilled by the synthesized position
controller (11). such that
Theorem 2: Consider the position error system (10). The  2 
 l1 k2 I2 02 
synthesized position controller (11) guarantees that the T T DT =  l2 k1  .
 (28)
position tracking error e p1 and the velocity tracking error e p2 02 I2
k2
are bounded and ultimately converge to small neighborhoods
It thus follows that
around the origin, respectively.
Proof: It is not hard to validate that the closed-loop position −k2 ΘT Θ − l2 k1 eTp2 tanh(k2 e p2 ) ≤ −χ2 ∥η∥2 (29)
error system (12) satisfies the Lipschitz condition. Therefore,
l2 k 1 2
it is unlikely for the evolved state trajectory to escape to where χ2 = min{l12 k2 χ2 , χ }. In addition, it also follows
infinity within [0, τ1 ]. We next focus on the time interval k2
that
[τ1 , ∞) . Consider a Lyapunov function candidate as ( )T
w k1 e p1 +k2 e p2 e t∥
l1 k2 tanh(η1 )+l2 k2 tanh(η2 ) + k1 e p2 ∥Mi−1 RF
V2 = l1 tanh(s)T ds ( )
0 k1 e t∥
w k2 e p2 k1 ≤ l1 k2 ∥η1 ∥ + (l2 k2 + )∥η2 ∥ ∥Mi−1 RF
+ l2 tanh(s)T ds + ∥e p2 ∥2 . (24) k2
0 2 e t∥
≤ χ3 ∥η∥ ∥M −1 RF (30)
Let η1 = k1 e p1 + k2 e p2 , η2 = k2 e p2 and η = [ηT1 , ηT2 ]T . Given a i

compact set {η ∈ R4 |∥η∥ < η, η > 0}, according to Lemma 1, where χ3 = max{l1 k2 , (l2 k2 + k1 /k2 }. Combining (29) and (30),
there exists a positive constant χ = χ(η) such that we obtain that
χ(η)∥η∥ < ∥ tanh(η)∥ < ∥η∥. In such a case, it follows that e t ∥∥η∥.
V̇2 ≤ −χ2 ∥η∥2 + χ3 ∥Mi−1 RF (31)
1 k1
V2 > χ(l1 ηT1 η1 + l2 ηT2 η2 ) + 2 ηT2 η2 > χ1 ∥η∥2 (25) Next, note from (11) that, the aerodynamic force ∥Ft ∥ can be
2 2k2 upper bounded by a positive constant F t . Besides, according
1 1 k1 to the properties
√ of√the rotation matrix, we obtain that
where χ1 = min{ χl1 , (χl2 + 2 )}. For t ∈ [τ1 , ∞), the e
∥RFt ∥ ≤ 2 2∥Ft ∥ ≤ 2 2 F t . Hence, for t ∈ [t¯1 , ∞), V̇2 (t)
2 2 k2
further satisfies
derivative of V2 (t) along (12) satisfies √
2 2 Ft
V̇2 = l1 tanh(η1 )T (k1 ė p1 +k2 ė p2 )+l2 k2 tanh(η2 )T ė p2 V̇2 ≤ −χ2 ∥η∥2 + χ3 ∥η∥. (32)
m
+ k1 eTp2 ė p2 Therefore, η is bounded and ultimately √ converges to the
χ3 2 2 F t
= − l12 k2 tanh(η1 )T tanh(η1 )− l1 l2 k2 tanh(η1 )T tanh(η2 ) compact set Ωη = {η ∈ R4 |∥η∥ ≤ }. Moreover, in
χ2 m
order to guarantee ∥η∥ < η , it is trivial that
− l1 l2 k2 tanh(η2 )T tanh(η1 ) − l22 k2 tanh(η2 )T tanh(η2 ) √
χ3 2 2 F t
− l2 k2 eTp2 tanh(η2 )+[l1 k2 tanh(η1 )+l2 k2 tanh(η2 ) < η.
χ2 m
e t
+k1 e p2 ]T Mi−1 RF This implies that
{ mηχ2 (l1 + l2 )m }
= −k2 ΘT Θ−l2 k1 eTp2 tanh(η2 ) F t < min √ , √ (33)
2 2χ3 2 2
e
+[l1 k2 tanh(η1 )+l2 k2 tanh(η2 )+k1 e p2 ]T Mi−1 RF (26) (l1 + l2 )m
where F t < √ directly results from (12). Finally, by
where Θ = l1 tanh(η1 ) + l2 tanh(η2 ). Moreover, it is trivial to 2 2
show that considering the definition of η, the tracking errors e p1 and e p2

Authorized licensed use limited to: Tamkang Univ.. Downloaded on January 17,2024 at 14:50:53 UTC from IEEE Xplore. Restrictions apply.
154 IEEE/CAA JOURNAL OF AUTOMATICA SINICA, VOL. 8, NO. 1, JANUARY 2021

Vertical position (m)


are bounded and ultimately converge to the compact sets
 √ 


 2 2χ3 F t 


4
Ωe p1 = 
e ∈ R |∥e p1 ∥ ≤
2

 p1 χ2 mk1   2
 √ 


 2 2χ3 F t 

 Desired Proposed PD
Ωe p2 =  0
e p2 ∈ R |∥e p2 ∥ ≤ χ2 mk2 
2
 
 0 20 40 60 80 100 120 140 160
Horizontal position (m)
(a)
respectively. ■
Remark 4: Note from the proof of Theorem 2 that, the 0.06
position controller (11) guarantees that the aerodynamic force
0.04

Δδ (N·m)
Ft is upper bounded without dependence on the position-loop
states. This effectively alleviates the attitude coupling effect in 0.02
the position loop analysis. Estimation of Δδ Δδ
0
IV. Simulations 0 2 4 6 8 10 12 14 16 18 20
t (s)
In this section, simulations are undertaken to validate the (b)
proposed control scheme. Consider an FWMAV with its
physical parameters listed in Table II. It is commanded to Fig. 5. Position tracking by the PD and the proposed control and the
track a reference trajectory in the following form: estimation of perturbation torque ∆δ.

xd (t) = 8.34t
5.8
1 1 f for proposed f for PD
yd (t) = − . 5.6
0.2 + 100e−1.6t 0.2 + 100
f (Hz)
5.4
TABLE II 5.2
Physical Parameters
5.0
Parameter Value Parameter Value 0 2 4 6 8 10 12 14 16 18 20
t (s)
m 0.06 kg ρ 1.29 (kg/m3 ) (a)
S 0.032 m2 St 0.013 m2
10
J 0.35 m4 lec 0.01 m θt for proposed θt for PD
0 10
θt (deg)

lt 0.05 m lh 0.172 m
0
AR 6 ν 10 −10
−10 −20
kx 0.1009 ky 0.001 0 0.5 1.0
−20
0 2 4 6 8 10 12 14 16 18 20
t (s)
The initial conditions of the concerned FWMAV are cho- (b)
sen as V(0) = 8.34 m/s, [x(0), y(0)]T = [0, 0]T , γ0 = 0, θ0 = 7◦,
θt0 = −10◦ and f0 = 5 Hz. Fig. 6. Control inputs of the PD and the proposed control.
To highlight the control performance of the proposed
control scheme, we compare it with a traditional PD control 16
θd for proposed 12.8
scheme with the position and attitude controllers designed by 14 12.4
θd for PD
12
θ (deg)

the PD method. Since the closed-loop FWMAV is in a θ for proposed 12.0


cascaded structure, the stability of the position loop is 10 θ for PD 8 9 10
dependent on that of the attitude loop. Towards this end, we 8
need to choose larger attitude control parameters than position 6
0 2 4 6 8 10 12 14 16 18 20
ones in order for faster convergence of the attitude tracking. t (s)
For the PD control scheme, the control parameters are chosen (a)
as k p f = 0.1, kd f = 1 in the position loop and k pθ = 20, kdθ = 8 0.6
in the attitude loop. For the proposed control scheme, the
0.4
controller parameters are specified as l1 = 1, l2 = 1, k1 = 0.1,
eθ (deg)

k2 = 0.2, k3 = 20 and k4 = 8. Moreover, the parameters of the 0.2

introduced neural network are as follows: l = 64 . σ = 0.1, 0


b = 0 . The simulation results are illustrated in eθ for proposed eθ for PD
ξ = 1 and W(0) −0.2
0 2 4 6 8 10 12 14 16 18 20
Figs. 5–8. t (s)
Firstly, Fig. 5 (a) shows the response curves of position (b)
tracking driven by two different controllers with the curve in
Fig.5 (b) showing the estimation of ∆δ with the radial basis Fig. 7. Attitude tracking by the PD and the proposed control and the
function neural network. In addition, Fig. 6 shows that the tracking errors.

Authorized licensed use limited to: Tamkang Univ.. Downloaded on January 17,2024 at 14:50:53 UTC from IEEE Xplore. Restrictions apply.
HE et al.: MODELING AND TRAJECTORY TRACKING CONTROL FOR FWMAVs 155

0.01 scheme for longitudinal motion to a general three-dimensional


ex for proposed ex for PD
0 case with both longitudinal and lateral motions. Also, we will
ex (m)

−0.01 apply the proposed control scheme to a practical FWMAV


−0.02
platform for verification.
−0.03
0 2 4 6 8 10 12 14 16 18 20 References
t (s)
(a) [1] C. Li, Q. Shi, Z. H. Gao, M. C. Ma, Q. Huang, H. Ishii, A. Takanishi,
and T. Fukuda, “Bioinspired phase-shift turning action for a biomimetic
0.3 robot,” IEEE/ASME Trans. Mechatron., vol. 25, no. 1, pp. 84–94, Feb.
2020.
0.2
[2] Q. Shi, C. Li, K. Li, Q. Huang, H. Ishii, A. Takanishi, and T. Fukuda,
ey (m)

0.1 “A modified robotic rat to study rat-like pitch and yaw movements,”
IEEE/ASME Trans. Mechatron., vol. 23, no. 5, pp. 2448–2458, Oct.
0 2018.
ey for proposed ey for PD
−0.1 [3] W. He, T. T. Wang, X. Y. He, L. J. Yang, and O. Kaynak, “Dynamical
0 2 4 6 8 10 12 14 16 18 20 modeling and boundary vibration control of a rigid-flexible wing sys-
t (s) tem,” IEEE/ASME Trans. Mechatron., 2020. DOI: 10.1109/TMECH.
(b) 2020.2987963
[4] N. T. Jafferis, E. F. Helbling, M. Karpelson, and R. J. Wood,
Fig. 8. Position tracking errors by the PD and proposed control. “Untethered flight of an insect-sized flapping-wing microscale aerial
vehicle,” Nature, vol. 570, no. 7762, pp. 491–495, Jun. 2019.
control inputs of both controls are bounded. However, it can [5] X. B. Ji, X. C. Liu, V. Cacucciolo, M. Imboden, Y. Civet, A. El
Haitami, S. Cantin, Y. Perriard, and H. Shea, “An autonomous
be observed at the zoom-in figure that tail wing inclination of untethered fast soft robotic insect driven by low-voltage dielectric
the proposed control has less fluctuation at the beginning of elastomer actuators,” Science Robotics, vol. 4, no. 37, pp. eaaz6451, Dec.
2019.
simulation than one with PD control. This reflects a better
[6] G. Dimitriadis, “Finite wings,” in Introduction to Nonlinear
control performance. Moreover, it can be seen from Figs. 7 Aeroelasticity, G. Dimitriadis. Chichester, UK: John Wiley & Sons, Ltd,
and 8 that the attitude and position tracking errors of both the 2017, pp. 503–553.
PD control and the proposed control are bounded and [7] D. Mackenzie, “A flapping of wings,” Science, vol. 335, no. 6075,
pp. 1430–1433, Mar. 2012.
ultimately converge to small neighborhoods of origin.
[8] B. Zhu, J. Z. Zhu, and Q. W. Chen, “A bio-inspired flight control
However, by careful comparison, the ultimate position strategy for a tail-sitter unmanned aerial vehicle,” Sci. China Inform.
tracking errors of the proposed control are smaller than that Sci., vol. 63, no. 7, pp. 170203, May 2020.
using PD control. To show this more clearly, Table III shows [9] H. Y. Li, L. J. Wang, H. P. Du, and A. Boulkroune, “Adaptive fuzzy
backstepping tracking control for strict-feedback systems with input
a data comparison. delay,” IEEE Trans. Fuzzy Syst., vol. 25, no. 3, pp. 642–652, Jun. 2017.
[10] M. Keennon, K. Klingebiel, H. Won, and A. Andriukov, “Tailless
TABLE III flapping wing propulsion and control development for the nano
Performance Comparison hummingbird micro air vehicle,” in The American Helicopter Society
68th Annual Forum, Fort Worth, USA, 2012.
e x (×10−6 ) (m) ey (×10−6 ) (deg) eθ (×10−3 ) (deg)
Strategy [11] A. Ramezani, S. J. Chung, and S. Hutchinson, “A biomimetic robotic
Mean Std Mean Std Mean Std platform to study flight specializations of bats,” Sci. Robot., vol. 2, no. 3,
PD –18 080 99 530 169 600 89 710 –1855 968.8 pp. eaal2505, Feb. 2017.
[12] B. Wong, “Lab bench-new robot designs are for the birds,” Electronic
Proposed –2.78 312.4 0.44 2 141 0.529 1.414
Design, vol. 59, no. 6, pp. 14, 2011.
[13] D. S. Farner, “Dimensional relationships for flying animals. Crawford
H. Greenewalt,” Auk, vol. 78, no. 4, pp. 653–654, Oct. 1961.
V. Conclusion
[14] Q. V. Nguyen, W. L. Chan, and M. Debiasi, “Hybrid design and
In this paper, the trajectory tracking problem of FWMAVs performance tests of a hovering insect-inspired flapping-wing micro
is studied on the longitudinal plane. The system model is aerial vehicle,” J. Bionic Eng., vol. 13, no. 2, pp. 235–248, Jun. 2016.
established by formulating the aerodynamic force and torque [15] A. Hedenström, “Aerodynamics, evolution and ecology of avian flight,”
Trends Ecol. Evol., vol. 17, no. 9, pp. 415–422, Sep. 2002.
generated by flapping wings and the tail wing. A neural
[16] B. W. Tobalske, T. L. Hedrick, K. P. Dial, and A. A. Biewener,
network based control scheme is proposed in terms of a “Comparative power curves in bird flight,” Nature, vol. 421, no. 6921,
hierarchical framework. In particular, a saturated position pp. 363–366, Jan. 2003.
controller is developed for position tracking, which guarantees [17] C. G. Yang, C. Z. Chen, N. Wang, Z. J. Ju, J. Fu, and M. Wang,
“Biologically inspired motion modeling and neural control for robot
a bounded aerodynamic force. Then, the flapping frequency of learning from demonstrations,” IEEE Trans. Cogn. Dev. Syst., vol. 11,
flapping wings and a pitch command are extracted from the no. 2, pp. 281–291, Jun. 2019.
position controller. Subsequently, an adaptive attitude [18] A. L. R. Thomas, “On the aerodynamics of birds’ tails,” Phil. Trans. R.
Soc. Lond. B: Biol. Sci., vol. 340, no. 1294, pp. 361–380, Jun. 1993.
controller is designed for the pitch to track the command
[19] G. P. He, T. T. Su, T. M. Jia, L. Zhao, and Q. L. Zhao, “Dynamics
controller, and a radial basis function neural network is analysis and control of a bird scale underactuated flapping-wing
implemented to approximate the aerodynamic perturbation vehicle,” IEEE Trans. Control Syst. Technol., vol. 28, no. 4, pp. 1233–
torque. It is demonstrated that the tracking errors are bounded 1242, Jul. 2020.
[20] A. R. Shanmugam and C. H. Sohn, “Systematic investigation of a
and ultimately converge to a small neighborhood around flapping wing in inclined stroke-plane hovering,” J. Braz. Soc. Mech.
origin. Sci. Eng., vol. 41, no. 8, pp. 347, Jul. 2019.
In the future, we will extend the proposed tracking control [21] G. Xie, A. Q. Shangguan, R. Fei, W. J. Ji, W. G. Ma, and X. H. Hei,

Authorized licensed use limited to: Tamkang Univ.. Downloaded on January 17,2024 at 14:50:53 UTC from IEEE Xplore. Restrictions apply.
156 IEEE/CAA JOURNAL OF AUTOMATICA SINICA, VOL. 8, NO. 1, JANUARY 2021

“Motion trajectory prediction based on CNN-LSTM sequential model,” pp. 3686–3698, Dec. 2019.
Sci. China Inform. Sciences, 2020. DOI: 10.1007/s11432-019-2761-y [42] J. Z. Zhang and C. X. Xu, “Trust region dogleg path algorithms for
[22] M. Bortolini, M. Faccio, F. G. Galizia, M. Gamberi, and F. Pilati, unconstrained minimization,” Ann. Oper. Res., vol. 87, pp. 407–418,
“Design, engineering and testing of an innovative adaptive automation Apr. 1999.
assembly system,” Assembly Autom., vol. 40, no. 3, pp. 531–540, Feb. [43] H. J. Yang and J. K. Liu, “An adaptive RBF neural network control
2020. method for a class of nonlinear systems,” IEEE/CAA J. Autom. Sinica,
[23] H. Qiao, M. Wang, J. H. Su, S. X. Jia, and R. Li, “The concept of vol. 5, no. 2, pp. 457–462, Mar. 2018.
‘attractive region in environment’ and its application in high-precision [44] B. Zhu and W. Huo, “Nonlinear control for a model-scaled helicopter
tasks with low-precision systems,” IEEE/ASME Trans. Mechatron., with constraints on rotor thrust and fuselage attitude,” Acta Autom.
vol. 20, no. 5, pp. 2311–2327, Oct. 2015. Sinica, vol. 40, no. 11, pp. 2654–2664, Nov. 2014.
[24] W. He, S. X. Nie, T. T. Meng, and Y. J. Liu, “Modeling and vibration
control for a moving beam with application in a drilling riser,” IEEE
Trans. Control Syst. Technol., vol. 25, no. 3, pp. 1036–1043, May 2017. Wei He (S’09–M’12–SM’16) received the B.Eng.
[25] S. Tijmons, C. De Wagter, B. Remes, and G. de Croon, “Autonomous and the M.Eng. degrees from the College of
door and corridor traversal with a 20-gram flapping wing MAV by Automation Science and Engineering, South China
onboard stereo vision,” Aerospace, vol. 5, no. 3, pp. 69, Jun. 2018. University of Technology (SCUT), China in 2006
and 2008, respectively, and the Ph.D. degree from
[26] F. Fei, Z. Tu, J. Zhang, and X. Y. Deng, Learning extreme
Department of Electrical and Computer Engineering,
hummingbird maneuvers on flapping wing robots. 2019. [Online].
Available: arXiv: 1902.09626. the National University of Singapore (NUS),
Singapore in 2011. He is currently working as a Full
[27] W. He, Z. C. Yan, C. Y. Sun, and Y. Chen, “Adaptive neural network Professor at the School of Automation and Electrical
control of a flapping wing micro aerial vehicle with disturbance Engineering, University of Science and Technology
observer,” IEEE Trans. Cybernet., vol. 47, no. 10, pp. 3452–3465, Oct. Beijing, China. He has co-authored 3 books published in Springer and
2017. published over 100 international journal and conference papers. He was
[28] H. V. Phan, T. Kang, and H. C. Park, “Design and stable flight of a 21g awarded a Newton Advanced Fellowship from the Royal Society, UK. He
insect-like tailless flapping wing micro air vehicle with angular rates was a recipient of the IEEE SMC Society Andrew P. Sage Best Transactions
feedback control,” Bioinspir. Biomim., vol. 12, no. 3, pp. 036006, Apr. Paper Award. He is serving the Chair of IEEE SMC Society Beijing Capital
2017. Region Chapter. From 2018, he has been the Chair of Technical Committee
[29] A. Ailon, “Simple tracking controllers for autonomous VTOL aircraft on Autonomous Bionic Robotic Aircraft (TC-ABRA), IEEE Systems, Man
with bounded inputs,” IEEE Trans. Autom. Control, vol. 55, no. 3, and Cybernetics Society. He is serving as an Associate Editor of IEEE
pp. 737–743, Mar. 2010. Transactions on Robotics, IEEE Transactions on Neural Networks and
[30] V. A. Tucker, “Pitching equilibrium, wing span and tail span in a Learning Systems, IEEE Transactions on Control Systems Technology, IEEE
gliding Harris’ hawk, parabuteo unicinctus,” J. Exp. Biol., vol. 165, Transactions on Systems, Man, and Cybernetics: Systems, SCIENCE CHINA
no. 4, pp. 21–41, Jan. 1992. Information Sciences, IEEE/CAA Journal of Automatica Sinica, Assembly
Automation, and an Editor of Journal of Intelligent Robotic Systems. His
[31] I. Fenercioglu and O. Cetiner, “Effect of unequal flapping frequencies current research interests include robotics, distributed parameter systems, and
on flow structures,” Aerosp. Sci. Technol., vol. 35, pp. 39–53, May
intelligent control systems.
2014.
[32] M. V. OL, J. D. Eldredge, and C. J. Wang, “High-amplitude pitch of a
flat plate: an abstraction of perching and flapping,” Int. J. Micro Air Xinxing Mu received the B.Eng. and M.Eng.
Veh., vol. 1, no. 3, pp. 203–216, Nov. 2009. degrees from the School of Electrical Engineering,
[33] J. H. Kim, C. Y. Park, S. M. Jun, D. K. Chung, H. C. Hwang, P. Beran University of Jinan in 2015 and 2017, respectively.
and D. Mrozinski, “Flight test measurement and assessment of a He is currently working toward the Ph.D. degree in
flapping micro air vehicle,” Int. J. Aeronaut. Space Sci., vol. 13, no. 2, control science and engineering with the School of
pp. 238–249, Jun. 2012. Automation and Electrical Engineering, University
[34] D. Mueller, H. A. Bruck, and S. K. Gupta, “Measurement of thrust and of Science and Technology Beijing. His current
lift forces associated with drag of compliant flapping wing for micro air research interests include FWMAV, intelligent
vehicles using a new test stand design,” Exp. Mech., vol. 50, no. 6, control, and system modeling.
pp. 725–735, Jul. 2010.
[35] A. Banazadeh and N. Taymourtash, “Adaptive attitude and position
control of an insect-like flapping wing air vehicle,” Nonlinear Dyn., Liang Zhang received the B.Eng. degree in
vol. 85, no. 1, pp. 47–66, Jul. 2016. intelligence science and technology from the School
[36] S. M. Nogar, A. Serrani, A. Gogulapati, J. J. McNamara, M. W. of Automation and Electrical Engineering,
Oppenheimer, and D. B. Doman, “Design and evaluation of a model- University of Science and Technology Beijing in
based controller for flapping-wing micro air vehicles,” J. Guid. Control 2019. He is currently working toward the Ph.D.
Dyn., vol. 41, no. 12, pp. 2513–2528, Dec. 2018. degree with the School of Automation and Electrical
[37] M. H. Dickinson, F. O. Lehmann, and S. P. Sane, “Wing rotation and Engineering, University of Science and Technology
the aerodynamic basis of insect flight,” Science, vol. 284, no. 5422, Beijing. His research interests include neural network
pp. 1954–1960, Jun. 1999. control, vibration control, machine learning, and
[38] M. Pachter, J. J. D’Azzo, and A. W. Proud, “Tight formation flight flexible structures.
control,” J. Guid. Control Dyn., vol. 24, no. 2, pp. 246–254, Mar.-Apr.
2001.
Yao Zou received the B.S. degree in automation
[39] H. W. Lin, B. Zhao, D. R. Liu, and C. Alippi, “Data-based fault tolerant from Dalian University of Technology (DUT) in
control for affine nonlinear systems through particle swarm optimized 2010, and the Ph.D. degree in control science and
neural networks,” IEEE/CAA J. Autom. Sinica, vol. 7, no. 4, pp. 954– engineering from Beijing University of Aeronautics
964, Jul. 2020. and Astronautics in 2016. He was a Post-Doctoral
[40] B. Xu, Y. X. Shou, J. Luo, H. Y. Pu, and Z. K. Shi, “Neural learning Research Fellow with the Department of Precision
control of strict-feedback systems using disturbance observer,” IEEE Instrument, Tsinghua University from 2017 to 2018.
Trans. Neural Netw. Learn. Syst., vol. 30, no. 5, pp. 1296–1307, May He is currently an Associate Professor with the
2019. School of Automation and Electrical Engineering,
[41] S. L. Dai, S. D. He, M. Wang, and C. Z. Yuan, “Adaptive neural control University of Science and Technology Beijing. His
of underactuated surface vessels with prescribed performance current research interests include nonlinear control, unmanned aerial vehicle
guarantees,” IEEE Trans. Neural Netw. Learn. Syst., vol. 30, no. 12, control, and multiagent control.

Authorized licensed use limited to: Tamkang Univ.. Downloaded on January 17,2024 at 14:50:53 UTC from IEEE Xplore. Restrictions apply.

You might also like