Proceedings of 2023 IEEE

International Conference on Mechatronics and Automation

August 6 - 9, Harbin, Heilongjiang, China

Trajectory Tracking Control for Flapping-wing UAV

Based on Model-Free Predictive Control
Xuguang Liu1, Yuean Wu1, Changping Du1, 2*, Jiahao Yu2, Yao Zheng1, 2
2023 IEEE International Conference on Mechatronics and Automation (ICMA) | 979-8-3503-2084-8/23/$31.00 ©2023 IEEE | DOI: 10.1109/ICMA57826.2023.10215628

1
Department of Aeronautics and Astronautics, Zhejiang University, Hangzhou, Zhejiang, China
2
Deqing Institute of Turbomachinery and Propulsion Systems, Zhejiang University, Huzhou, Zhejiang, China
*Corresponding Author: [email protected]

Abstract - This paper proposes a trajectory tracking now exploring various advanced control theories to optimize
algorithm that uses Model Predictive Control (MPC) to address the traditional flight control system.
the issue of poor tracking performance of the Flapping-wing For instance, Wang et al. [4] used a nonlinear incremental
UAV. However, due to the difficulty in constructing a prediction dynamic inverse method combined with differential flat
model, the high computational complexity, and the poor real-time
feedforward to design a trajectory tracker for quadrotor UAVs,
performance of MPC, we propose a Model-Free Predictive
Control (MFPC) algorithm that uses a dynamic linearized model which improves the tracking performance of highly
as the prediction model. It greatly reduces complexity and maneuvering trajectories and stabilizes faster than the
running time while satisfying control performance. To address traditional feedforward PID. Adlakha et al. [5] improved the
the effects of the inaccuracy of the prediction model and tracking performance of the current iteration, which injects the
disturbances such as crosswind on the tracking performance of tracking errors of previous iterations into the learning structure
MFPC, we optimize the feedback correction law of MFPC using consisting of a learning filter and a robust filter to generate
the L1 algorithm idea and significantly improve the robustness of learning signals. Additionally, Zhou et al. [6] proposed a fuzzy
the resulting MFPC-L1. Simulation results demonstrate that
adaptive backstepping control law, which suppresses time-
MFPC-L1 can reduce the computation time for a given trajectory
by approximately 46% compared to the classical MPC.
varying external disturbances to increase the stability and
Moreover, it reduces the tracking errors of circular and S-shaped robustness of the trajectory-tracking control.
trajectories by over 96% compared to PID and Model-Free Researchers have achieved significant progress in
Adaptive Control (MFAC) algorithm. In the presence of improving the performance of fixed-wing UAV flight control
crosswind, MFPC-L1 reduces the tracking errors by systems. For instance, Hernandez et al. [7] proposed a flight
approximately 85% compared to the MFPC-PD algorithm based control system architecture with an optimized cascade
on the lateral offset PD feedback correction law without L1 controller that enables the system to track time-parameterized
algorithm optimization. Therefore, the simulation results indicate 3D-position trajectories. Yang et al. [8] designed an adaptive
that MFPC-L1 has strong robustness in the trajectory-tracking
trajectory controller for fixed-wing UAVs with uncertain
control of the Flapping-wing UAV.
structural parameters. Furthermore, some researchers have
Index Terms – Flapping-wing UAV; Model-Free Adaptive proposed a new design concept by abandoning the traditional
Control; Trajectory Tracking; L1 Guidance; Model Predictive inner-loop and outer-loop control architecture. Zhao et al. [9]
Control; proposed a full-loop composite fast non-singular terminal
sliding mode trajectory tracking control method, significantly
improving the anti-interference capability by avoiding the
I. INTRODUCTION
complex multi-loop decoupling problem. Similarly, Manzoor
In recent years, the UAV research community has shifted et al. [10] proposed an offset-free MPC trajectory tracking
its focus from traditional fixed-wing and rotary-wing UAVs to control solution.
newer UAV designs that are better suited for various Rotary-wing and fixed-wing UAVs are well-known for
scenarios, such as military intelligence and reconnaissance, their simple aerodynamic analysis and accurate modeling.
due to their small size, stealth, and high maneuverability. However, the flapping-wing UAV is more challenging to
Regardless of the type of UAV, accurate tracking of mission model accurately due to the complex aerodynamic forces
trajectories is an essential indicator in most application generated by the flapping of the wings, making it difficult for
scenarios [1-3]. Thus, improving the UAV trajectory tracking traditional control algorithms to achieve optimal performance.
performance depends on designing more accurate UAV flight To address this issue, researchers have proposed various
control systems. new trajectory-tracking methods. For example, Ji et al. [11]
Conventional control systems for UAVs generally employ proposed an adaptive global fast terminal sliding mode control
an inner-loop and outer-loop PID control structure, where the algorithm based on Radial Basis Neural Networks to ensure
outer loop is the position control loop and the inner loop is the finite time tracking of the desired trajectory. Xu et al. [12]
attitude control loop. The traditional PID control algorithm designed a fuzzy control strategy for large flapping-wing
can hardly meet these increasing requirements. Researchers are UAV, taking into account the priorities of attitude stability,
trajectory tracking, and flight safety. Lee et al. [13] applied

979-8-3503-2084-8/23/$31.00 ©2023 IEEE 1365

Authorized licensed use limited to: Tamkang Univ.. Downloaded on January 17,2024 at 14:58:58 UTC from IEEE Xplore. Restrictions apply.
MPC to conduct experiments, which significantly improved  pn   c c s s c − c s c s c + s s   u 
control performance compared with traditional algorithms.     
Despite the success of these methods, there are still some  pe  =  c s s s s + c c c s s − s c   v  (2)
 p   −s s c c c  w
limitations. For example, models trained by neural networks  d    
may not accurately reflect the atypical states of the system.
Reinforcement learning and iterative learning algorithms are  u   rv − qw   f x 
difficult to optimize online in real time, and the fuzzy control      
strategy heavily relies on expert rule bases, which can lead to  v  =  pw − ru  +  f y  / m (3)
 w   qu − pv   f 
system divergence if the rule base is not comprehensive.      z
This paper proposes a trajectory tracking algorithm based
Based on (1) and force analysis, the three-axis dynamics
on MFPC adapted to the flapping-wing UAV. While
equations in the body frame can be derived.
maintaining the classical inner-loop and outer-loop control
architecture, MFPC is added as a feedforward, using actual
 f x   −mg sin    kmotor f − Va2 
velocity and attitude information collected from the control      
object to perform online data-driven control. With the  f y  =  mg cos  sin   + 0.5 S prop Cprop  0 
 f   mg cos  cos    0 
 z    
application of MFPC, this data-driven control method can take (4)
 C X ( ) + C X q ( )cq / 2Va + C X  e ( ) e 
full advantage of the MFAC online optimization and combines  
the high accuracy of MPC with a lower computational cost +0.5Va S  CY0 + CY  + CYp bp / 2Va + CYr bq / 2Va + CYa  a + CY r  r 
2

 
compared to the traditional MPC method. To address the 
 CZ ( ) + CZq ( )cq / 2Va + CZ e ( ) e 

potential discrepancies between the mathematical model
optimized by the data-driven method and the actual model, we     1 sin  tan  cos  tan   p 
designed a feedback control algorithm with an L1-like guiding     
law to replace the lateral offset PD feedback control algorithm   =  0 cos  − sin   q  (5)
    0 sin  / cos cos  / cos  
commonly used in the feedback correction part of the     r 
traditional MPC, further improving the accuracy and anti-
interference of the trajectory tracking algorithm.  p   1 pq −  2 qr   3l +  4 n 
   2   
 q  =  5 pr −  6 ( p − r )  +  m / J y  (6)
2
Our proposed algorithm shows promising results in
simulations and experiments, demonstrating its potential for  r    pq −  qr    l +  n 
   7 1   4 8 
practical application in flapping-wing UAV control.
The parameters in (6) are shown in Table I, where the
II. MPC-BASED TRAJECTORY TRACKING CONTROL
inertia of each axial direction is given by J x , J y and J z .
ALGORITHM
The moment l m n  in (6) is calculated as follows
T
Unlike fixed-wing UAVs and rotary-wing UAVs,
flapping-wing UAV generate the required lift and thrust by [16].
fluttering their wings, resulting in a highly complex flow field
 l   −kTp (k f ) 
2
that poses significant challenges in achieving accurate models.
 
To simplify the construction of the prediction model, the m =  0 
following assumptions are made [15]:  n   0


   
Assumption 1: The aerodynamic force and drag torque of
 b C + C  + C bp / 2V + C br / 2V + C  + C   
the elevator are negligible.   l0 l lp a lr a l a a l r r 

Assumption 2: The increase in thrust and lift is  
+0.5 Va2 S  c Cm0 + Cm  + Cmq cq / 2Va + Cm  e   (7)
 e 
proportional to the flutter frequency.  

 b Cn0 + Cn  + Cn p bp / 2Va + Cnr br / 2Va + Cn  a + Cn  r  
Based on the consumptions above, we can obtain (1):   a r 

 Fxk = k x f −  CD ( ) SV 2 / 2

 (1) Model Predictive Control (MPC) was first proposed in the
 Fzk = k z f −  CL ( ) SV / 2
2
 1960s and typically consists of three parts: prediction model,
rolling optimization, and feedback correction. To utilize this
Where f is the flutter frequency of the wing,  is the air algorithm for flapping-wing UAV trajectory tracking, each
density, CD , CL represents the drag and lift coefficients component must be designed as illustrated in Fig. 1.
The prerequisite for applying MPC is to build an accurate
respectively, S is the wing area, and Q is the airspeed in the
prediction model for flapping-wing UAV, while transfer
body frame. functions, state space descriptions, step responses, and
distributed parameter systems can all be used as the prediction
model [17].

1366

Authorized licensed use limited to: Tamkang Univ.. Downloaded on January 17,2024 at 14:58:58 UTC from IEEE Xplore. Restrictions apply.
 ( , y ( k − ny ) , u(k ), )
u (k ) y (k + 1)
UAV y(k + 1) = f y(k ), , u ( k − nu ) (8)
Where, y (k )  R m and u (k )  R m denote the output and
MPC
Controller input of the system respectively, while representing an
+ Feedback
e (k + 1 k ) + yd (k + 1) unknown nonlinear vector-valued function. The positive
Reference
+ Correction −
y (k + 1 k )
integer n y and nu indicate the output and input dimensions,
u (k ) Rolling Predictive respectively. A lemma is derived assuming that the nonlinear
Optimization Model
system satisfies the generalized Lipschitz condition and that
the vector-valued function f ( ) has a continuous partial
Fig. 1 MPC Algorithm Structure
derivative for each component [18].
The complexity of the obtained prediction model is Lemma 1: For the system described in (8), there exists a
evident. The traditional MPC control method used as the bounded time-varying parameter  c (k )  R mm , known as the
prediction model has high computational complexity and poor pseudo-partial derivative matrix, such that when u(k )  0 ,
real-time performance, which makes it difficult to adapt to the the system can be transformed into the following data model.
complex and rapidly changing environment during the flight
process.  y(k + 1) = c (k )u(k ) (9)
TABLE I
PARAMETERS
Based on this dynamic linearization model, the control
Parameter Expression system is designed and stability analysis is carried out. The
control algorithm is typically designed by setting the penalty
 J x J z − J xz2 function first as follows.

1 J xz ( J x − J y + J z ) /  2
J (u(k )) = y* (k + 1) − y(k + 1) +  u(k ) − u(k − 1)
2
(10)

2 J xz2 /  + J z ( J z − J y ) /  By substituting (9), it can be observed that the penalty

function is minimized when u (k ) satisfies the following
3 Jz / 
equation.
4 J xz / 
u(k ) = u(k −1)
5 (Jz − Jx ) / J y +  c ( k ) ( y* ( k + 1) − y( k ) ) / ( +  c ( k ) )
2
(11)
6 J xz / J y
Equation (11) represents the control law of the Model-
Free Adaptive algorithm. In order to achieve real-time rolling
7 J /  + Jx ( Jx − Jy ) / 
2
xz
optimization, rolling calculations for  c (k ) , which can be
8 Jx /  performed using a similar solution process as above.
Additionally, the estimation algorithm for the pseudo-partial
derivative matrix is given as follows.
III. MODEL-FREE ADAPTIVE CONTROL ALGORITHM
ˆ c (k ) = ˆ c (k − 1)
Model-Free Adaptive Control (MFAC) is a data-driven
approach that employs input and output data of the controlled ( )
+  y(k ) − ˆ c (k − 1)u(k − 1) uT (k − 1) / ( + u(k − 1) )
2
(12)
system to design and analyze controllers. It allows for adaptive
control of unknown nonlinear systems, including parameter Here,  has the same meaning as  and ranges from
adaptive control and structural adaptive control. One of the
(0, 2] , and ˆ (k ) represents the estimated value of  (k ) .
main advantages of MFAC is its independence from the c c

mathematical model of the controlled system and various However, the calculation of the input u (k ) requires obtaining
theoretical issues, making the controller design and analysis the  c (k ) calculated first by (9). If u (k ) is uncertain, the
more straightforward [18]. The key concept of MFAC is to use actual difference cannot be obtained. Therefore, a pseudo-
a pseudo-partial derivative matrix to create a dynamic partial derivative estimation algorithm in the form of (12) is
linearization model that describes the system dynamics at the necessary to estimate  c (k ) and use it for calculating the
operating point.
actual value. The flowchart of Algorithm I for this method is
Assumption 3: The system can be represented by a
shown below.
nonlinear discrete-time model as follows.
This method offers several advantages, including low
computational complexity, monotonic convergence, and

1367

Authorized licensed use limited to: Tamkang Univ.. Downloaded on January 17,2024 at 14:58:58 UTC from IEEE Xplore. Restrictions apply.
bounded input-output stability of the control error, compared Outer-loop Control
System
Reference

with other data-driven methods. However, due to the Lateral Subsystem

+
extremely high requirements for fast response of the MFAC −
algorithm itself, large deviations may lead to slow Model
Parameter
Matrix  Dynamic x y Memory of
MPC
convergence or even divergence. Algorithm
Linearization
Model
Position &
Pose

d UAV
IV. DESIGN OF MFPC TRAJECTORY TRACKING ALGORITHMS
+ Vd
Section 2 of the paper highlights the importance of precise +
Inner-loop Control
System
Actuator

modeling of the controlled object. The problem that it is Vertical Subsystem

d

difficult to carry out accurate physical modeling of flapping-

wing UAV, the application of the MFAC algorithm can reduce Fig. 2 Structure of Model-Free System
the computational complexity well, thereby ensuring real-time Referring to (9), the results of the dynamic linearization of
performance. the lateral system state can be obtained as follows.
However, it relies solely on feedback control and dynamic
optimization, degrading performance applied to complex  x  11 12   V 
 y  =    (13)
systems like flapping-wing UAV. To address these limitations,    21 22    
a new MFPC algorithm is proposed by leveraging the strengths
of both algorithms. On the basis of (14), a further differential equation of the
ALGORITHM I horizontal lateral subsystem is expressed as follows.
MODEL-FREE ADAPTIVE TRAJECTORY CONTROL ALGORITHM
Input: u(k ) = [V (k ),  (k )]T , V ( k ) is the airspeed,  ( k ) is the course 1 0   12 
angle; x + = Ax + Bu =   x +  11 u (14)
0 1  21 22 
Output: y(k ) = [ pn (k ), pe (k )]T , Coordinates of the current plane
position of the airframe in the inertial system
B. Design of the Rolling Optimization
The target system is converted into a linear system by
1: Set initial value u  [V (0),  (0)]T , y  [ pn (0), pe (0)]T ; implementing dynamic linearization at the operating point, as
outlined in the previous section. A quadratic optimization
2: and c (k )  ˆ c (0) , y(0) , y (1) , u (0) ; algorithm [19] is utilized for the rolling optimization
3: for i = 1: k component of MPC. The cost function of the optimization
algorithm is given by the equation below.
4: ˆc (k )  getnewPJM ( y(k ),ˆc (k − 1), u(k −1))
k +t p

5: u(k )  getnewInput ( y (k + 1),ˆc (k ), y(k ))

* J ( x(k ), u ) = min  ( xT (k )Qx(k ) + u T (k ) Ru (k )) (15)
k

6: y(k )  uavsystem(u(k ))
Where k is the same as in the previous section, and t p
7: y(k ) = y(k ) − y(k − 1)
represents the current time step, usually selected as 5 to 10. It
8: u (k ) = u (k ) − u (k − 1) should be noted that as the dynamic linearization model differs
from the actual model, a larger prediction step may affect
9: end for
control performance due to model mismatch. Therefore, the
10: Finally get the input set u (k ) . prediction step 5 is adopted in this algorithm. The cost
coefficient matrix Q and R can be adjusted to suit the specific
A. Design of the system control task [20].
The system design of our algorithm incorporates the idea
of inner-loop and outer-loop control architecture used in u (k ) = min J ( x(k ), u ) (16)
u
traditional fixed-wing UAVs. Specifically, our algorithm only
focuses on the position control part of the outer-loop, while the The optimal control input of the dynamic linearization
attitude control part of the inner-loop still uses the classical model is the final solution u (k ) by minimizing (17). However,
PID controller. Fig. 2 shows the block diagram of the system due to the limited accuracy of the model, the optimal control
architecture. input based on the prediction model may result in significant
Fig. 2 shows that this method specifically focuses on the errors when applied to the actual system. Therefore, the
design of the lateral subsystem. In the subsystem, ex and e y feedback correction module of the MPC algorithm plays a
represents the position error the in x and y direction, crucial role in smaller errors and achieving satisfactory control
respectively, and the output of the subsystem is the airspeed performance.
and the heading angle of the airframe. C. Design of the L1 Feedback Calibration Module

1368

Authorized licensed use limited to: Tamkang Univ.. Downloaded on January 17,2024 at 14:58:58 UTC from IEEE Xplore. Restrictions apply.
 Different types of PID control algorithms are usually used Convert the reference point selected in the lateral offset
to achieve the feedback correction law. The most common PD-based correction algorithm from the ideal position point at
method for the lateral tracking scenario is the PD control the next instant to the circle center (or virtual circle center) of
algorithm based on the lateral offset. the given task path and the ideal trajectory point at distance L1
In Fig. 3, the airframe is located at the point ( x(k ), y(k )) from the current position point. To solve the above problem,
at time k, while the ideal position is ( xd (k ), yd (k )) . the value of L1 should be much greater than the difference d
between the current position of the vehicle and the target
trajectory. The difference is defined below.

( x − xd ) + ( y − yd ) − Rd
2 2
d = R − Rd = (19)

where 1 , 2 ,  are all small angles, while L1 is much

greater than d, so (20) can be obtained.

an = 2V 2 sin  2V 2 ( (1 +  ) cos2 + sin2 ) / L1 (20)

Fig. 3 Lateral Offset Feedback Correction
The corresponding trajectory errors and control laws are Simplifying (20) and yields.
presented below.
an = 2 ( pV / L1 ) d + 2 pVd / L1 + V 2 / Rd
2
(21)
ex (k ) = xd (k ) − x(k )

 (17)
 e y ( k ) = yd ( k ) − y ( k ) p = cos2 = 1 − 0.25 ( L1 / Rd )
2
 (22)

 K pvx K pvy   ex (k − 1)  One of the optimization objectives of the L1 feedback

u (k ) = u (k ) + 
 K p x K p y   e (k − 1) 
 y 
correction algorithm is the heading angle. an represents the
(18) calculated desired centripetal acceleration in (21).
 K dvx K dvy   ex (k − 1) 
+
 K d x K d y   e (k − 1) 
 y 

The u (k ) on the right-hand side of (18) is the optimal

control input obtained from the rolling optimization
calculation, and a correction calculation is performed using the
derivative of the position error and the position error at the
previous instant to overcome the errors caused by the model
misalignment and possible unknown disturbances and other
Fig. 4 Feedback Correction Algorithm
factors. The simulation of the traditional PD feedback
After having obtained the desired centripetal acceleration,
correction algorithm based on the lateral offset shows the
it can finally be converted into the desired angular velocity of
phenomenon of long adjustment time and scattered control. It
the heading n by the following equations.
is easy to see from the analysis that this correction method
requires the target waypoints to be updated in real-time and
n = an / V = 2 ( p / L1 ) Vd + 2 pd / L1 + V / Rd
2
that the target waypoints are selected more intensively and (23)
closer to the current waypoint. At the same time, the control
input is the airspeed and heading angle of the vehicle, and  fb = n t (24)
there is a theoretical differential relationship between the
current control input and the error information, which makes it From (24), the desired heading angular velocity can be
easy to override the control while performing the optimization converted into the differential component of the heading angle,
calculation. If the actual waypoint is ahead of the target and this differential component is used as the feedback
waypoint, the control input calculated by the control system correction of the input. So far, the final control law can be
will cause the airframe to slow down and steer, causing the obtained as follows.
trajectory to fluctuate or even diverge. To address these
problems, an L1 feedback correction algorithm is proposed by  Kp 
u (k ) = u (k ) +  d
 2 ( p / L1 ) V t 
drawing on the ideas associated with the L1 guidance law. The 2

algorithm selects a new target point and error calculation (25)

method without changing the feedback correction input, and  Kd   0 
+  d + 
the selection is shown in Fig. 4.  2 pt / L1  V t / Rd 

1369

Authorized licensed use limited to: Tamkang Univ.. Downloaded on January 17,2024 at 14:58:58 UTC from IEEE Xplore. Restrictions apply.
 V. EXPERIMENTAL RESULTS Specifically, the rolling optimization of the MPC
algorithm is combined with the L1 algorithm as a feedforward
In this section, we validate the proposed algorithm using
link for feedback correction.
MATLAB and Simulink simulations. Three common target
trajectories are tracked: hovering trajectory without crosswind, Thus, significantly improve the performance of trajectory
S-shaped level flight trajectory, and hovering trajectory with tracking, as shown in Fig. 5, where the trajectory
crosswind. The simulation step size is set to 0.01 seconds and corresponding to this method closely follows the ideal
the simulation time is set to 160 seconds. The equations for the trajectory.
hovering trajectory are provided below.

 xd = 450sin( t / 50)

 yd = 450 − 450 cos( t / 50) (26)
 z = 120 + t
 d

North (m)
The initial conditions related to the flapping-wing UAV
are shown in Table II.
TABLE II
SIMULATION INITIAL CONDITIONS
Para
pn pe pd u0 v0 w0 0 0 0
meter

0.0
Value 0 0 -120 24.9 0 2.05 0 0
8
East (m)
The variables pn , pe , pd selected in Table II are the Fig. 5 Comparison of Hovering Trajectory Tracking without Crosswind

position quantity under the inertial system, u0 , v0 and w0 are B. S-shaped Level Flight Trajectory without Crosswind
the three-axis components of airspeed in the body frame, and The second task trajectory is an S-shaped level flight
the initial value of the pitch angle 0 is obtained through the trajectory, which includes an additional trajectory switch point
compared to the circular trajectory. At the trajectory switch
leveling calculation. The simulation environment was set as an point, the control law must be switched, and this switch has a
ideal environment without wind, and the tracking trajectories, significant impact on the control effect of traditional control
in this case, were hovering trajectories and S-shaped laws such as PID, which may result in system oscillation or
trajectories, and the PID control law and MFAC were used as even divergence.
control experiments to compare with this method. In addition, When mainly applying the PID control law, the tracking
the tracking trajectory is set to be a hovering trajectory in the error in the S-shaped trajectory tracking task is significantly
presence of crosswind. Considering the low robustness of the increased, as shown in Fig. 6. The increased error is mainly
conventional method in the trajectory-tracking scenario of the concentrated after the trajectory switching point. While the
flapping-wing UAV. This task only compares with the MFPC- tracking error of the S-shaped trajectory is slightly higher than
L1 method, which applies the L1 method for feedback that of the circular trajectory, it does not change much
correction, i.e., the present method, with the MFPC-PD compared to the other methods, indicating that the proposed
method based on the lateral offset PD. method is not sensitive to system switches and exhibits strong
A. Hovering Trajectory without Crosswind robustness.
In an environment without crosswind, all three methods C. Hovering Trajectory with Crosswind
can converge to the ideal trajectory. However, due to the The most common disturbance during flight is the ever-
highly nonlinear nature of flapping-wing UAV, control present and ever-changing crosswind. Therefore, experiments
performance is difficult to guarantee when using the PID are designed to track circular trajectories in an environment
control algorithm, resulting in constant fluctuations around the with crosswind, and the robustness of different methods in the
ideal trajectory as observed in Fig. 5. On the other hand, the persistent and changing noise environment can be measured by
MFAC algorithm utilizes dynamic rolling linearization of the comparing the magnitude of each tracking error in this
controlled object at each moment, and employs an adaptive environment. The circular trajectory is obtained from (26), and
PID control algorithm to obtain the control law, which greatly the additional crosswind disturbance can be described as
improved the stability and tracking performance compared to follows.
the traditional PID control algorithm.
The proposed algorithm in this paper combines the rolling Ws = 5.5cos(0.5t ) + 0.5sin(0.5t ) (27)
linear model from the MFAC method with the model
predictive control (MPC) algorithm to enhance trajectory Where Ws is the lateral airspeed in m/s. The tracking effect of
tracking performance. the conventional PID and MFAC methods is poor in an

1370

Authorized licensed use limited to: Tamkang Univ.. Downloaded on January 17,2024 at 14:58:58 UTC from IEEE Xplore. Restrictions apply.
environment where this disturbance exists. Therefore, only the As for the hovering trajectory tracking without crosswind,
two methods based on MFPC will be discussed. the results in Fig. 5 show that the trajectories of the three
As shown in Fig. 7, the RMSE of this method is only methods mentioned above, except for the PID, all fit the target
0.6428 in the presence of the crosswind as above, which is trajectory well, making it difficult to demonstrate the
much lower than that of the MFPC-PD algorithm without the improvement in trajectory accuracy of the algorithms. In
L1 algorithm optimization of 3.7982. combination with Table III, it is easy to see that the tracking
There is a significant difference in the trajectories error of the MFPC method is reduced by at least 75%
generated by the two methods in Fig. 7. compared to the MFAC algorithm. Based on this, the L1
When the lateral offset PD control law is chosen as the algorithm was applied to optimize the feedback correction law,
feedback method, it is very easy to have the next position point and the MFPC-L1 algorithm further reduced the error by
ahead of the next ideal position point under the disturbance, nearly 86% compared to the MFPC-PD algorithm, which is a
for the reference point chosen is too close to the current actual significant optimization effect.
position point. Then the control system will manipulate the In the S-shaped tracking task without crosswind, the
airframe to slow down and turn, while the next ideal position tracking errors shown in Table IV are all increased compared
point may appear ahead of the current point. Finally, the to the hovering trajectories in Table III. The MFPC-L1 method
airframe will turn in the opposite direction, and eventually, shows the smallest increase in RMSE of only 0.0916. It can be
form a trajectory that fluctuates around the ideal trajectory. concluded that this method works well for different
After applying the L1 feedback correction, the reference trajectories.
point is chosen more rationally, and a reference quantity The runtime of each algorithm was calculated and is
related to the current ideal trajectory is added, so the shown in Table V. The PID method has a runtime of only
fluctuation problem of the traditional feedback correction law 14.77s due to its simple structure, and the MFAC algorithm is
based on the lateral offset PD algorithm is well solved. In the next best. This method is expected to retain the accuracy of
addition, the robustness of the overall control system is the MPC algorithm and the speed and real-time performance
improved by introducing the new feedback correction law. of the MFAC algorithm by integrating MFAC with MPC.
Table V shows that the MPC algorithm, which is based on
D. Quantification results the idea of MFAC, is almost twice as fast as the traditional
This section uses RMSE as the tracking error metric to MPC method which is remarkable.
quantify and compare the trajectory error. The results are The analysis of the quantified results above shows that the
presented in Table III and Table IV. application of MFAC based on MPC can significantly reduce
the runtime and improve the tracking performance compared
to other traditional methods. In terms of both runtime and
tracking errors, the method performs well.
VI. CONCLUSIONS
North (m)

This paper investigates a high-accuracy trajectory tracking

control method for flapping-wing UAV in the presence of
crosswind and other disturbances. Due to the difficulty of
accurately modeling flapping-wing UAV, a trajectory tracking
control method based on the MPC algorithm is proposed. The
East (m) real-time flight information of the flapping-wing UAV is
Fig. 6 Comparison of S-shaped Level Flight Trajectory Tracking without utilized to dynamically linearize the system and obtain a linear
Crosswind model for rolling optimization. An L1 feedback correction law
is also proposed to improve the performance of trajectory
tracking and the robustness of the algorithm by combining it
with the feedback correction algorithm of MPC.
Simulation results demonstrate that while ensuring real-
time online applicability, the proposed algorithm significantly
improves the tracking performance compared to existing and
North (m)

traditional methods for different methods. The proposed

method also shows robustness in the experimental environment
with constantly changing crosswinds. Furthermore, to validate
the practical application of the proposed algorithm, it will be
implemented on a real flapping-wing UAV prototype.

East (m)
Fig. 7 Comparison of Hovering Trajectory Tracking with Crosswind

1371

Authorized licensed use limited to: Tamkang Univ.. Downloaded on January 17,2024 at 14:58:58 UTC from IEEE Xplore. Restrictions apply.
 TABLE III [5] R. Adlakha, M. Zheng, “An optimization-based iterative learning control
COMPARISON OF HOVERING TRAJECTORY TRACKING WITHOUT CROSSWIND design method for UAV’s trajectory tracking,” in Proc. 2nd International
Method RMSE Conference on Computer Science and Application Engineering, Oct.
2020, pp. 1353-1359.
PID 13.0635 [6] L. Zhou, J. Zhang, J. Dou, “A fuzzy adaptive backstepping control based
on mass observer for trajectory tracking of a quadrotor UAV,”
MFAC 2.2665 International journal of adaptive control and signal processing, vol. 32,
no. 12, pp. 1675-1693, Oct. 2018.
MFPC-PD 0.5843 [7] R. Hernandez, M. Nahon, “Trajectory tracking control of highly
maneuverable fixed-wing unmanned aerial vehicles,” in AIAA Scitech
MFPC-F1(This method) 0.0855
2020 Forum, Jan. 2020, pp. 2074.
[8] J. Yang, X. Wang, S. Baldi, “A software-in-the-loop implementation of
adaptive formation control for fixed-wing UAVs,” in Proc. IEEE/CAA
TABLE IV Journal of Automatica Sinica, vol. 6, no. 5, pp. 1230-1239, Sep. 2019.
COMPARISON OF S-SHAPED LEVEL FLIGHT TRAJECTORY TRACKING WITHOUT [9] Z. H. Zhao, L. Xiao, B. Jiang, “Fast nonsingular terminal sliding mode
CROSSWIND trajectory tracking control of a quadrotor UAV based on extended state
Method RMSE observers,” Control and Decision, vol. 37, no. 9, pp. 2201-2210, Sep.
2022.
PID 35.9897 [10] T. Manzoor, Y. Xia, “Trajectory tracking control of a VTOL unmanned
aerial vehicle using offset-free tracking MPC,” Chinese Journal of
MFAC 4.4730
Aeronautics, vol. 33, no. 7, pp. 2024-2042, Jul. 2020.
MFPC-PD 1.8596 [11] X. M. Ji, H. H. Wen, “Finite-time trajectory tracking control based on an
adaptive neural network for a quadrotor UAV,” CAAI Transactions on
MFPC-F1(This method) 0.1771 Intelligent Systems, vol. 17, no. 3, pp. 540-546, Apr. 2022.
[12] W. F. Xu, E. Z. Pan, J. T. Liu, “Flight control of a large-scale flapping-
wing flying robotic bird: System development and flight experiment,”
TABLE V Chinese Journal of Aeronautics, vol. 35, no. 2, pp. 235-249, Feb. 2022.
RUNNING TIME STATISTICS BY ALGORITHM [13] J. Lee, S. Ryu, T. Kim, “Learning-based path tracking control of a
This flapping-wing micro air vehicle,” in Proc. 2018 IEEE/RSJ International
Method PID MFAC MFPC-PD MPC Conference on Intelligent Robots and Systems (IROS), Aug. 2018, pp.
method
7096-7102.
Time [14] B. Y. Zhang, “Model-Free Adaptive Flight Control Method for Flapping
14.77 31.30 70.20 135.40 72.52 Wing Aerial Vehicle,” MS Thesis, Beijing Jiaotong University, 2021.
Consuming/s
[15] H. Wei, X. X. Mu, L. Zhang, Y. Zou, “Modeling and Trajectory
Tracking Control for Flapping-Wing Micro Aerial Vehicles,” in Proc.
IEEE/CAA Journal of Automatica Sinica, vol. 8, no. 1, pp. 148-156,
REFERENCES Nov. 2021.
[1] W. He, S. Q. Ding, C. Y. Sun, “Research Progress on Modeling and [16] R. W. Beard, T. W. McLain, “Small Unmanned UAV: Theory and
Control of Flapping-wing Air Vehicles,” Acta Automatica Sinica, vol. 43, Practice,” Princeton: Princeton University Press, 2012, pp. 20-432.
no. 5, pp. 685-696, May 2017. [17] Y. G. Xi, “Predictive Control,” Beijing: National Defense Industry
[2] J. W. Xiang, Y. Sun, T. Shen, D. C. Li, “Research Progress and Press, 2013, pp. 92-114.
Application of Flapping Wing Aerodynamics,” Engineering Mechanics, [18] Z. S. Hou, S. T. Jin, “Model-Free Adaptive Control: Theory and
vol. 36, no. 4, pp. 8-23, Apr. 2019. Application,” Beijing: Science Press, 2013, pp. 50-124.
[3] D. Chin, D. Lentink, “Flapping Wing Aerodynamics: from Insects to [19] H. Chen, “Model Predictive Control,” Beijing: Science Press, 2013, pp.
Vertebrates,” Journal of Experimental Biology, vol. 219, no. 7, pp. 920- 92-130.
932, Apr. 2016. [20] J. B. Rawlings, D. Q. Mayne, “Model Predictive Control: Theory and
[4] X. L. Zhu, D. Yue, “Stability of sampled-data systems with application to Design,” Florida: Ingram International INC, 2009.
networked control systems,” in Proc. IEEE 32nd Chinese Control Conf, [21] S. Park, J. Deyst, J. How, “A new nonlinear guidance logic for trajectory
Oct. 2013, pp. 6572-6577. tracking,” in Proc. AIAA guidance, navigation, and control conference
and exhibit, Jun. 2004, pp. 490.

1372

Authorized licensed use limited to: Tamkang Univ.. Downloaded on January 17,2024 at 14:58:58 UTC from IEEE Xplore. Restrictions apply.