International Journal of Systems Science

ISSN: 0020-7721 (Print) 1464-5319 (Online) Journal homepage: https://www.tandfonline.com/loi/tsys20

UAV path planning using artificial potential field

method updated by optimal control theory

Yong-bo Chen, Guan-chen Luo, Yue-song Mei, Jian-qiao Yu & Xiao-long Su

To cite this article: Yong-bo Chen, Guan-chen Luo, Yue-song Mei, Jian-qiao Yu & Xiao-long Su
(2016) UAV path planning using artificial potential field method updated by optimal control theory,
International Journal of Systems Science, 47:6, 1407-1420, DOI: 10.1080/00207721.2014.929191

To link to this article: https://doi.org/10.1080/00207721.2014.929191

Published online: 18 Jun 2014.

Submit your article to this journal

Article views: 3374

View related articles

View Crossmark data

Citing articles: 116 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=tsys20
International Journal of Systems Science, 2016
Vol. 47, No. 6, 1407–1420, http://dx.doi.org/10.1080/00207721.2014.929191

UAV path planning using artificial potential field method updated by optimal control theory
Yong-bo Chen, Guan-chen Luo, Yue-song Mei, Jian-qiao Yu∗ and Xiao-long Su
School of Aerospace Engineering, Beijing Institute of Technology, Beijing, P. R. China
(Received 19 March 2013; accepted 24 May 2014)

The unmanned aerial vehicle (UAV) path planning problem is an important assignment in the UAV mission planning.
Based on the artificial potential field (APF) UAV path planning method, it is reconstructed into the constrained optimisation
problem by introducing an additional control force. The constrained optimisation problem is translated into the unconstrained
optimisation problem with the help of slack variables in this paper. The functional optimisation method is applied to reform
this problem into an optimal control problem. The whole transformation process is deduced in detail, based on a discrete
UAV dynamic model. Then, the path planning problem is solved with the help of the optimal control method. The path
following process based on the six degrees of freedom simulation model of the quadrotor helicopters is introduced to verify
the practicability of this method. Finally, the simulation results show that the improved method is more effective in planning
path. In the planning space, the length of the calculated path is shorter and smoother than that using traditional APF method.
In addition, the improved method can solve the dead point problem effectively.
Keywords: path planning; functional optimisation problem; artificial potential field (APF); additional control force; optimal
control

1. Introduction plied in the multi-UAV system. Paul, Krogstad, and Grav-

Unmanned aerial vehicle (UAV) path planning is a process dahl (2008) presented a solution based on the APF method
in which the UAV finds a path from the starting point to the for collision- and obstacle-free formation flight and recon-
destination. It is important in the UAV mission planning and figuration of groups of autonomous helicopters. Similarly,
has been a research focus in recent years. In this process, Oland and Kristiansen (2013) applied the APF method to
based on the task requirements and certain restraints, the solve the problem of collision and terrain avoidance for mul-
UAV should assess the condition and information of the tiple UAVs, tracking a low-altitude trajectory. From these
planning space and move along a collision-free path (Caves, applications, it is further explained that the APF method is
2010). an effective method in the guidance and control process of
The UAV path planning problem has been dealt with the UAV.
by many methods, such as artificial potential field (APF) However, it also has its own defects. Koren and Boren-
method, A∗ algorithm, genetic algorithm and so on (Chen, stein point out four significant problems that are inherent
Zhao, & Han, 2010). Thereinto, the APF method is com- to the APF method. Thereinto, the local minimum problem
monly used in path planning for it has many advantages, is the biggest problem. Many papers may face this problem
including the simple algorithm structure, the concise math- (Koren & Borenstein, 1991). Besides the local minimum
ematical description and the convenience for real-time con- problem, the arbitrary configurations of the artificial poten-
trol (Berry, Howitt, Gu, & Postlethwaite, 2010; Khatib, tial field cannot guarantee that the planning result is optimal
1986; Koren & Borenstein, 1991; Scherer, Singh, Cham- (Park, Jeon, & Lee, 2001).
berlain, & Elgersma, 2008). For example, Srikanthakumar, In order to solve these problems, some scholars present
Liu, and Chen (2012) chose the APF method as a path several ways to improve the traditional APF method. The
planning and obstacle avoidance candidate technique for intuitionistic way is to combine some other methods with
verification study as it is a simple and widely used method. the APF method. In general, a new method which com-
But beyond that, the APF method has a strong portabil- bines two or more methods may effectively solve many
ity. By changing the source of the artificial potential field, original problems. For example, Xu, Xie, and Xie (2006)
it can also solve the collision avoidance problem of the put forward a novel algorithm combining APF with genetic
multi-agent system. Therefore, the APF method is not only algorithm (GA) for the robot path planning and obstacle
used in the single UAV path planning problem, but also ap- avoidance. Lin, Li, and Aouf (2010) present a novel flight

∗
Corresponding author. Email: [email protected]


C 2014 Taylor & Francis
1408 Y.-b. Chen et al.

route planning design technique for a group of UAVs’ au- 2. Modelling of the UAV path planning problem
tonomous navigation based on the combination of evolu- In order to simplify the algorithm, we can use the simplified
tionary algorithms and modified APF methods. Although linear particle dynamics to replace the real dynamic model
the effect of the combined algorithm may be well, it always in the path planning process. The UAV particle dynamics
causes some new problems, such as the higher operational is based on the APF method with the additional control
complexity, the larger calculated amount and so on. The force. The main idea of the APF method is to use repulsive
other way is to improve the APF method by some new potential fields emanating from the obstacles to force the
concepts or some new ways of modelling. Shim, Kim, and UAV away and an attractive potential field emanating from
Sastry (2003) established a decentralised discrete-time op- the target so as to attract the UAV.
timal control problem based on the potential field method,
and then solved the optimisation problem with the help Remark 1: The UAV in this paper means the rotary wing
of the non-linear model predictive control (NMPC). The vertical take-off and landing (VTOL) aircraft.
approach is less susceptible to the local minima problem Remark 2: The path planning process is one part of the
due to its predictive nature. However, because the NMPC guidance process. For most situations, the model of an air-
method is based on the optimisation in the finite receding craft in the guidance process is the simple particle model
horizon, the effect of this approach is directly related to the which only needs to limit the velocity and overload of the
range of the horizon. The bigger the horizon is, the better aircraft. In other words, it only needs to guarantee that the
the effect of the result will be and the larger the calculated path planning result, which is a space curve varying with
amount will be. Therefore, the choice of the horizon range time, is flyable for the UAV. Finally, the UAV can move in
is the most important and difficult parameter in this ap- the space according to the computed path. Therefore, the
proach. Analogously, Yang and Sukkarieh (2012) also used model of the UAV can be simplified into the linear particle
the model predictive control (MPC) based on the improved dynamics under some constraints.
APF method to solve the navigation problem of the UAV.
There are various forms of repulsive field functions
The distances to obstacles, angles to obstacles, obstacle size
and attractive field functions. Many entirely different field
and the UAV velocity were introduced to improve the effect
functions have been used in earlier studies. The repulsive
of the APF method. However, the improved APF method
field and attractive field are defined in simple mathematical
is not very suitable for the irregular obstacles. Luo updated
forms in the following.
the APF method by introducing the additional control force
(Luo, Yu, Zhang, & Zhang, 2012). However, the solution
algorithm for this updated method was not discussed in
2.1. Description of force fields
her paper, and it did not make a detailed analysis of the
optimisation model. Indeed, it is difficult to solve the addi- 2.1.1. The attractive field
tional control force, as the variable is time-varying. More- The attractive field ϕ is chosen as
over, it is also hard to ensure the optimality index of the     
solution. p ,  p  = 0
φ( p ) =  
A further discussion to solve these problems is estab- 0,  p  =0 , (1)
lished in this paper by remodelling the functional opti-
misation model. This model, based on the UAV particle where p = pt − p, and p is the vector from the origin to
dynamics system, uses the APF method and takes the ad- the mass centre of the UAV and pt is the vector from the
ditional control force as an independent variable. Also, origin to the target point.
the constraint conditions and the objective function are The gradient of the field function is the field force:
analysed in detail in this work. Moreover, the optimal
  
control method is applied to solve the additional control p  p  = 0
force. Fy =  p  ,   . (2)
0,  p  =0
The structure of this paper is as follows. In Section 2, the
functional optimisation model is established by introducing
the additional control force into the dynamic model. In The attractive field covers the whole planning space
Section 3, the optimal control theory is applied to solve (Aitsaadi, Achir, Boussetta, & Pujolle, 2011).
the additional control force, and the control sequence can
be obtained to achieve the real-time control. In Section
4, the path following process is presented to verify the 2.1.2. The repulsive field
reasonability of the computed path. In Sections 5 and 6, The repulsive field function ϕ i is chosen as
two different simulations are carried out. The results verify   ∗
  εe−l  pi  ,  
∗
the proposed algorithm. Some conclusions are drawn and
ϕi p∗i =  pi∗  ≤ d0 , (3)
further discussions are presented in Section 7. 0,  pi  > d0
 International Journal of Systems Science 1409

If the resultant force is applied, the UAV particle dy-

namics system will be as follows:
x(k + 1) = Ax(k) + B a(k), (5)
where x = [pT vT ]T is the motion state of the centroid (v is
the speed of the mass centre of the UAV). The variable k
denotes the variables at time step, so x(k) denotes the UAV
motion state of the mass centre at the kth step, and

I tI T
A= , B= t 2
I tI (6)
O I 2

relate to time interval t.

If a(k) is the acceleration caused by the field force, the
damping force and the additional control force, then
Figure 1. Sketch map of the repulsion.
a(k) = r(k) + u(k), (7)
where the subscript i denotes the sequence of the obstacles,
and ε and l are the alterable coefficients. p∗i is the relative where r(k) is the acceleration caused by the resultant of the
position of the UAV from the ith obstacle, and p∗i = p̃i − p. field force and the damping force Fz at kth step and u(k)
p̃i is the vector from the origin to the obstacle point which is the acceleration caused by the additional control force.
is the nearest to the UAV in the ith obstacle. The damping force Fz is given by Fz = μv, where μ is the
The repulsive field force is regional. The range of its damping coefficient.
effect is named safety distance d0 . The existence of the The reason for introducing the damping force is that it
repulsive field force depends on whether the UAV enters can prevent the ceaseless increasing velocity of the UAV
the safety space (Lee & Park, 2003; Xie, Xie, Chen, & Li, under the effect of the global attractive field.
2008; Yin & Yin, 2009). The gradient of the repulsive field
function ϕ i is the repulsive field force:
⎧ 2.3. The cost function
∗  
⎨ εl pi∗ e−l  p∗i  ,  p∗  ≤ d0 The UAV optimisation cost function is a multi-objective
Fc =  pi  i
⎩  ∗ . (4) optimisation function (Aitsaadi et al., 2011; Deb, 2005). It
0,  p  > d0
i includes the following two parts.
The repulsion can help the UAV avoid the obstacles First, in order to limit the consumption of control energy,
during the path planning process. According to the form the minimum total force is proposed. Then the cost function
of the repulsion Fc , the change of the function value is is given as follows:
shown in Figure 1. Thereinto, the values A and B are based 1
N
on the coefficients ε and l. For the repulsion, it generally Js = a(k)2 . (8)
2 k=0
demands that the smaller  p∗i  causes the bigger repulsion
Fc  (hatched area). When the UAV flies towards the obsta- Second, the objective function of distance is
cle at its maximum velocity with the effect of the artificial
potential field, the coefficients ε and l, and the safety dis-
1
N
tance d0 guarantee the UAV’s velocity reducing to zero in
Jd =  p(k + 1) − p(k)2 . (9)
the hatched area. Therefore, the value A needs to be much 2 k=0
less than the size of the safe distance d0 .
The weighted method is commonly used to transform
2.2. Modelling of the UAV particle dynamics the multi-objective optimisation problem into a single-
objective one. Thus the total objective function is
The improved APF method used in this paper is updated by
introducing the additional control force. While the obstacles
α
n−1
bring the repulsive field, the targets will bring the attractive min J = αJs + βJd =  p(k + 1) − p(k)2
field too. Then, the gradient of each field generates the u(k) 2 k=0
field force. Thus, the resultant force controls the UAV to
β
n−1
avoid the obstacles and arrive at the destination (Charifa & + a(k)2 , (10)
Bikdash, 2009). 2 k=0
1410 Y.-b. Chen et al.

where α and β are weighted coefficients, and they are sub- 2.4.4. The boundary conditions
jected to the following conditions: 0 ≤ α ≤ 1 and 0 ≤ The boundary values of the UAV path planning problem are
β ≤ 1. the starting point and the target point. The value range of k
is [0, n], and n is not fixed.

2.4. The constraints and the boundary conditions p(0) = ps , (15)

2.4.1. The performance constraints
The UAV has some performance constraints itself, and p(n) = pt , (16)
among the constraints the two major ones are: the speed
constraint and the acceleration constraint. Then where ps and pt are the coordinates of the starting point and
the target point, respectively.

v ≤ vm , (11) 2.5. The optimisation model

a ≤ am , (12) Finally, the UAV path planning problem can be recon-
structed as a functional optimisation problem:

where vm and am are the upper bound of the absolute values α β

n−1 n−1
of the UAV speed and acceleration, respectively. min J =  p(k + 1) − p(k)2 + a(k)2
u(k) 2 k=0 2 k=0
In order to simplify those constraints, the UAV ⎧
performance constraints are replaced by the control en- ⎪
⎪ p(k + 1) I tI p(k) t 2
I
⎪
⎪ = + 2 r(k)
ergy constraint considering the characteristics of the APF ⎪
⎪ v(k + 1) 0 I v(k) tI
⎪
⎪
method. Then, we have the following theorem: If u ≤ um , ⎨ t 2
I
s.t. + 2 u(k)
then (11) and (12) hold. The detailed proof is given in tI .
⎪
⎪
Appendix 1. ⎪
⎪ u(k) ≤ u
⎪
⎪
m
⎪
⎪ p(0) = p
⎩ s
p(n) = pt
2.4.2. The space constraints (17)
The UAV needs to avoid the obstacles in the planning space. In this optimisation problem, u(k) is the optimisation
In other words, the space constraints are made up of the variable. When the additional control force u(k) acts on the
obstacles. The space constraint is defined as UAV particle dynamics, the optimal path can be obtained.

p∈
/ S, (13) 3. Solution for the additional control force
According to the functional optimisation model (17), the
where S is the set of the obstacles. design variable u(k) varies with discrete time. Because of
The existence of the repulsive fields keeps the UAV the complexity of the objective function and constraint con-
away from the obstacles. In other words, the constraint ditions, the solving process of u(k) is very difficult. The
above is satisfied by establishing the repulsive fields. paper introduces the optimal control theory based on the
idea of the variational method to solve the problem.
The general form of the discrete system in the optimal
2.4.3. The dynamic constraints control theory (Wu, 2008) is
kf −1
Because of the UAV dynamic constraints, there exist dif- 
ferential and integral relationships among the force, the min J = θ [x(kf ), kf ] + ϕ(x(k), u(k), k)
u(k)
acceleration, the speed and the position. k=k0
The dynamic constraints are shown in Equation (5). As s.t. x(k + 1) = g(x(k), u(k), k)
Equation (7), they are rewritten as follows: x(k0 ) = x0 , (18)

p(k + 1) I tI p(k) t 2

I where x ∈ R n and u ∈ R m .
= + 2 r(k)
v(k + 1) O I v(k) tI First, the UAV path planning model (17) needs to be
t 2 transformed into the general form of Equation (18). The
I
+ 2 u (k) . (14) procedure is divided into two parts as discussed in the
tI following.
 International Journal of Systems Science 1411

Part 1: simplify the constraints of the optimisation

model. The velocity constraints, acceleration constraints GT (k)G(k) 0
Q1 (k) = , (26)
and boundary conditions are the given conditions which 0 0
 4 
can be used directly in the optimal control theory. The t
G(k)
force induced by the potential field r(k) is the function of S (k) = 4 3
t (27)
2
I
the UAV’s position p(k) and velocity v(k). Quasi-linearise it ,
into the following formula: 0
S1 (k) = , (28)
0
r(k) = G(k) p(k), (19) t 4
R= I, (29)
4
where G(k) is the transitive matrix that varies with p(k) and
v(k). Then, the dynamic constraint is rewritten as R1 = I, (30)
2
I + t2 G(k) tI
t 2 t 2 (k) = , (31)
p(k + 1) = p(k) + G(k) p(k) + u(k) tG(k) I
2 2 t 2
+ tI v(k) = 2 . (32)
v(k + 1) = v(k) + t u(k) + tGk p(k). (20) t

With the denotations of (25)–(32), the cost function is

Part 2: transform the cost function into the quadratic
formed into the quadratic form, and the dynamic constraint
form. According to Equation (20), we have
is formed as the state equation. By now, the optimisation
model is rewritten as
t 2 t 2
p(k + 1) − p(k) = G(k) p(k) + u(k) + tI v(k).
2 2 1  x T (k)
n−1 T
(21) min J =
Then the cost function is rewritten as u(k) 2 k=0 uT (k)
αQ(k) + βQ1 (k) αS(k) + βS1 (k) x(k)
α β
n−1 n−1
J =  p(k + 1) − p(k) +
2
a(k)2 αS T (k) + βS1T (k) αR + βR1 u(k)
2 k=0 2 k=0 ⎧
⎪ x(k + 1) = (k) x(k) + u(k)
⎪
⎨  T
n−1  2 x(0) = pTs v T (0)
α   t 2 G(k) tI p(k) t 2 
 s.t.  T . (33)
=  + u(k) ⎪ = T
v T
2 k=0 2 v(k) 2 ⎪
⎩ x(n) p t (n)
u(k) ≤ um
β
n−1
+ u(k) + r(k)2 . (22) Let us define:
2 k=0

Q2 (k) = αQ(k) + βQ1 (k), (34)

In order to transform (22) into quadratic form, we
have to partition the following quadratic form matrices as
follows:
S2 (k) = αS(k) + βS1 (k), (35)
⎡ ⎤
4 3 4
t T t t
⎢ 4 G3 (k)G(k) 2
G(k) G(k) ⎥
4 3
⎢ t GT (k) 2 t
I ⎥
2
t I
⎣ --------------------------------------
2 ⎦, (23)
t 4
GT (k) t
I
3
t
I
4
R2 (k) = αR + βR1 . (36)
4
⎡ 2
⎤
4

T Performing a variable substitution, we have

⎢ G (k)G(k) 0 0⎥
⎢ 0⎥
⎣ ---------------------
0 0 ⎦. (24)
 T
Q̃(k) = Q2 (k) − S2 (k) R2−1 S2T (k)
0 0 I
R̃ = R2
Let us define: 
(k) = (k) − 
(R2−1 )T S2T (k)

  =
t 4 T
G (k)G(k) t 3
G(k) x  (k) = x(k)
Q (k) = 4
t 3 T
2 , (25)
2
G (k) t I 2
u (k) = (R2−1 )T S2T (k)x(k) + u(k). (37)
1412 Y.-b. Chen et al.

The linear discrete optimal system is According to the minimum theorem (Liu, 2000), the
necessary conditions to achieve the optimal control are as
1  T follows:
n−1
min J = [x (k)Q̃(k)x  (k) + uT (k)R ˜ u (k)]
u(k) 2 k=0
(1) Based on the adjoint equation (the co-state
⎧ 
⎪ x (k + 1) =  (k)x  (k) +  u (k) equation),
⎪
⎨  T
x  (0) = pTs v T (0)
s.t.  T . (38) δH δg(x  (k) , u (k) , k)
⎪
⎪ x 
(n) = p T
v T
(n) λ(k) = + T
⎩  t  γ (k) ,
um ≥ u (k) − α R̃ −1 S T (k)x  (k) δx  (k) δx  (k)
(44)
we have
In the above formula, R̃ and Q̃(k) can be proved to
be positive-definite and semi-positive-definite, respectively.
λ(k) = x  (k)Q̃(k) + λ(k + 1) T
(k)
The above expression has met the standard form of the linear
δg(x  (k) , u (k) , k)
discrete system with terminal constraints and inequality + γ T (k)  (k) + γ T (k) .
constraints. δx  (k)
The inequality constraint can be transformed into an (45)
equality constraint by introducing a slack variable τ ∈ R:
(2) In the optimum trajectory, the Hamiltonian function
u (k) − α R̃ −1 S T (k) x  (k) − um + τ 2 = 0. (39) needs to get the absolute minimum. Therefore, we
have
Then, the constrained optimisation problem (38) can be
transformed into the unconstrained optimisation problem:
min H (x ∗ , u∗ , λ∗ , k) = H (x ∗ , u , λ∗ , k)
n−1 
u∗ ∈U
   1  H (x ∗ , u∗ , λ∗ , k) ≤ H (x ∗ , u , λ∗ , k) , (46)
min J = υ T x  (n) − pt + x (k)Q̃(k)x T (k)
u(k)
k=0
2
1
+ u (k)R̃u T (k) + λT (k + 1)[  (k)x  (k) where u (k) is the random control vector in U, u∗ (k) is the
2 control vector which can get the absolute minimum of the
+  (k)u (k) − x  (k + 1)] Hamilton function, x∗ (k) is the optimal state trajectory and

+ γ T (k) u (k)−αR ˜−1 S T (k)x  (k) λ∗ (k) is the optimal Lagrange multiplier.

 In the optimum trajectory, we have
− um + α12 , (40)
δH δg(x  (k) , u (k) , k)
 (k) = −γ (k)
T
. (47)
where υ, λ(k) and γ (k) are three Lagrange multipliers. δu δu (k)
The unconstrained optimisation problem can be given
as Therefore, the following expression cannot be guaran-
teed in the allowable interval of the control vector u (k):

n−1
min J = υ T [x  (n) − pt ] + [H (x  (k), u (k),
u(k) δH
k=0 = 0. (48)
λ(k + 1), γ (k), ϕ(k), k) − λ (k)x(k) + γ T (k)g(x  (k) ,
T δu (k)
u (k) , k)], (41) Based on Equations (47) and (48), the global minimum
may not be reachable. However, the local minimum is at-
where Hamiltonian function is tained in the allowable interval U. Then we can get the
  approximate solution of the control vector (Fang & Wang,
H x  (k), u (k), λ(k + 1), γ (k), ϕ(k), k 2006; Li, 2011):
1 1
= x  (k)Q̃(k)x T (k) + u (k)R̃(k)uT (k)
2 2 u (k) = u∗ (k) = [u∗x (k) u∗y (k)u∗z (k)]T
+ λT (k + 1)[  (k)x  (k) +  (k)u (k)], (42) umin (k) = [ux min (k) uy min (k) uz min (k)]T
⎧
⎪
⎪ um (u• min ≥ um )
and g(x  (k) , u (k) , k) is given by ⎪
⎪
⎨
u∗• (k) = u• min (k) (umin (k) ∈ U ) , (49)
g(x  (k) , u (k) , k) = u (k) − α R̃ −1 (k) S T (k) x  (k) ⎪
⎪
⎪
⎪
⎩
− um + α12 . (43) −um (u• min ≤ −um )
 International Journal of Systems Science 1413

where umin (k) is the optimal control u (k) without the con-
trol constraint, and the subscript · represents x, y, z.
According to the unconstrained optimal control
algorithm, let

λ(k) = P (k)x  (k). (50)

Then, P(k) can be calculated by using the correspond-

ing matrices  (k),  (k), Q̃(k) and R̃(k), via Riccati equa-
tion (Sun Jinsheng & Wang, 1994; Xiao-Qun, Wei-Min,
Jun-Qiang, Xiao-Qian, & Jun, 2012):

T 
P (k) = (k)P (k + 1) (k) + Q̃(k)
˜ Figure 2. Sketch map of the quadrotor helicopter.
T
− (k)P (k + 1)  [R + T
P (k + 1)  ]−1
T 
× P (k + 1) (k). (51) to follow the planning results which include the position
information of the UAV. Obviously, the position informa-
Therefore, the control force umin (k) is given by tion is time-varying. After this process, the practicability
of the method is verified. The sketch map of the quadrotor
umin (k) = −R̃ −1 (k) T
(P −1 (k + 1) helicopter is shown in Figure 2.

+ R̃ −1 (k) T −1
) 
(k)x  (k). (52) Those time-varying position results of the path planning
method are the input of the position tracker of the UAV. The
In this expression, the matrices R̃(k),  ,  (k), position tracker makes up by an inner-loop attitude con-
P(k + 1), Q̃(k) and x  (k) are all known. Therefore, the ad- troller and an outer-loop position controller. The dynamic
ditional control force u (k) is gained in the optimal control model of the quadrotor helicopter is given by Bai, Liu, Shi,
problem. In order to obtain the u(k), the linear inverse trans- and Zhong (2012). The attitude controller and the position
form of (38) is obtained. The additional control force u(k) controller are both realised with the classical proportion
makes up of a(k) according to (7). Finally, a(k) can be used integration differentiation (PID) method. The structure of
to attain the UAV’s path via (5). the whole system is shown in Figure 3.

4. Path following process 5. Simulation example 1

After the path planning process, the reasonability of the In the process of moving from the starting point to the
computed path also needs to be verified by the path follow- target, the UAV may meet many typical obstacles, such
ing process, although the path planning process is finished as buildings, mountains, fire threats, other vehicles and so
under some constraints. In the path following process, the on. They can be simplified as cuboids, floating balls and
real dynamic model of UAVs can be applied to simulate the other simple models. The complicated obstacles can be
real UAV flying state. The quadrotor helicopters are used transformed into a composition of simplified models.

Figure 3. The whole system chart.

1414 Y.-b. Chen et al.

Figure 5. Path planning result in situation 1.

Figure 4. UAV simulation task 1.

The length, width and height of the planning space are
600, 600 and 100 m, respectively. The discrete time interval
Note that, in the UAV path planning problem, it is diffi- is t = 1 s, and the mass of the UAV is m = 1 kg.
cult to deal with the non-convex obstacles. Therefore, these In the first situation, the starting point coordinates of
types of obstacles should be paid more attention to in the the UAV (A) are (0 m, 150 m, 0 m); in the second situation,
simulation process. the starting point coordinates of the UAV (B) are (0 m,
250 m, 0 m). The target point coordinates of both these two
situations are (600 m, 600 m, 40 m). The abstract scene is
5.1. Scenario description and constraint shown in Figure 4.
condition The additional control force constraint is um = 0.1 m/s2 ;
The simulation parameters and the simulation constraint the flight altitude constraint is above the ground level, i.e.,
conditions in the examples are discussed in the following. z ≥ 0 m. In this paper, both the least control energy rule and

Figure 6. Path following result in situation 1.

 International Journal of Systems Science 1415

Figure 7. Path following error in situation 1.

the shortest distance rule are used. So compromising them obvious that the path planned by the improved method
at α = 1 and β = 1 in (10). is smoother and shorter in the simulation result, which
means the improved method is more effective in path
planning.
5.2. Simulation result In order to verify the feasibility of the red line, the path
5.2.1. Situation 1 following process is finished by the frame of Section 4. The
path following result is shown in Figure 6, where the red
First, point A is chosen as the starting point to compare
line means the real path following result, and the blue line
the updated path planning method with the traditional APF
means the UAV path planning result. During the whole path
method. The simulation result is shown in Figure 5, where
following process, the path following error of the UAV is
the blue path is planned by the traditional APF method
shown in Figure 7, and the velocity and the acceleration are
while the red path is planned by the improved APF method.
shown in Figure 8. Therefore, the feasibility of the planning
The total distance of the blue path is 827.95 m, and
result is verified.
the total distance of the red path is 820.25 m. It is

Figure 8. Flying data in situation 1.

1416 Y.-b. Chen et al.

Then, the path following error of the UAV is shown in

Figure 11, and the velocity and the acceleration are shown
in Figure 12. Therefore, the feasibility of the planning result
is verified.

5.3. Simulation result analysis

The simulated results show that the path planning problem
can be transformed into an optimal control problem via
solving the additional control force by using the optimal
control method. The optimal control path planning algo-
rithm ensures that all the UAVs arrive at terminals from
the starting points without the restriction of the ‘dead area’
problem in every potential field configuration form. At the
same time, the optimal additional control force can ensure
that the index of the planning path using the improved APF
method is better than the one obtained using the traditional
APF method. Furthermore, the curve is smoother. Then, the
path following error and flying data show that the planning
Figure 9. Path planning result in situation 2.
paths are flyable under some constraint conditions.

5.2.2. Situation 2 6. Simulation example 2

It is difficult to avoid the ‘dead point’ problem when the Except the above simulation scenario, the digital map is
traditional APF method is used. In this situation, the attrac- widely used in many similar problems. In order to fur-
tion and repulsion balance, and the speed of the vehicle are ther prove the effectiveness of the improved method, the
zero. In the second scenario, point B is chosen as the start- interpolation function is used to build the digital map in
ing point. Then the UAV will drop into the dead point when simulation example 2.
the traditional method is used, as shown in the blue path,
compared with the path planned via the updated method,
which is shown in the red path (Figure 9). 6.1. Scenario description and constraint
The path following result of situation 2 is shown in condition
Figure 10, where the red line means the real path following The simulation parameters and the simulation constraint
result, and the blue line means the UAV path planning result. conditions in the examples are discussed in the following.

Figure 10. Path following result in situation 2.

 International Journal of Systems Science 1417

Figure 11. Path following error in situation 2.

The length, width and height of the planning space are 6.2. Simulation result
5000, 6000 and 1000 m, respectively. The discrete time Both the traditional APF method and the improved APF
interval is t = 1 s, and the mass of the UAV is m = 1 kg. method are used in the digital map scene. The simula-
The starting point coordinates of the UAV (A) are tion result is shown in Figure 14, where the Figure 14(a)
(2000 m, 2500 m, 500 m) in the second situation. The is planned by the traditional APF method while the Fig-
target point coordinates are (6000 m, 7000 m, 0 m). The ure 14(b) is planned by the improved APF method.
new scene is shown in Figure 13. Then, the path following process is introduced to verify
The additional control force constraint is um = 0.1 m/s2 ; the rationality of the planning result. All the errors are below
the flight altitude constraint is above the ground level. In 2 m, and the flying data are reasonable under the constraint
this paper, both the least control energy rule and the shortest conditions.
distance rule are used. So compromising them at α = 1 and
β = 1 in (10).
6.3. Simulation result analysis
The new simulated results show some similar conclusions
as the simulation example 1. These results show that the

Figure 12. Flying data in situation 1.

1418 Y.-b. Chen et al.

proves that the improved method is better than the tradi-

tional APF method.

7. Conclusion
In this paper, the additional control force is introduced into
the APF method for the UAV path planning problem. Since
the additional control force can be regarded as the optimi-
sation variable, the path planning problem is transformed
into an optimisation problem. Then, by introducing cer-
tain slack variables, the original constrained optimisation
problem is converted into an unconstrained optimisation
problem, which can be easily solved via optimal control.
The whole process is deduced in detail in this paper, based
on a discrete UAV dynamic model. The optimal control law
is obtained via Riccati equation approximately. Therefore,
Figure 13. UAV simulation task 2.
the updated method is more effective and convenient to
achieve. The optimal simulation results have verified the
updated method, demonstrating its strong ability in path
planning.

Funding
This work was supported by the National Natural Science Foun-
dation of China [grant number 61350010].

Notes on contributors
Yongbo Chen was born in 1990. He received
his BS degree in Beijing Institute of Tech-
nology in 2012. He is currently working to-
wards a doctoral degree at the School of
Aerospace Engineering, Beijing Institute of
Technology, Beijing, China. His research in-
terest is UAV path planning.

Guanchen Luo was born in 1984. She re-

ceived her BS and MS degrees in flight
vehicle engineering from Beijing Institute
of Technology, Beijing, China in 2006 and
2008, respectively. Since 2008, she is a PhD
candidate at the School of Aerospace Engi-
neering at Beijing Institute of Technology,
Beijing, China. Her research interests in-
clude UAV path planning, flight dynamics
and control, and robust control.
Figure 14. UAV path planning result in simulation example 2.
(a) Path planning result using the traditional APF method. (b) Path Yuesong Mei was born in 1978. He received
planning result using the improved method. his PhD degree in flight vehicle design from
Beijing Institute of Technology in 2007. He
is now a lecturer in School of Aerospace En-
UAV used the traditional APF method meets with the dead gineering, Beijing Institute of Technology.
point problem. However, the improved path planning algo- His research interests include flight vehicle
system design, flight dynamics and control.
rithm ensures that the UAV successfully solves the dead
point problem. Therefore, the new simulation scene further
 International Journal of Systems Science 1419

Jianqiao Yu was born in 1972. He received aerial vehicles. Proceedings of the Institution of Mechani-
BS, MS and PhD degrees from Beijing Insti- cal Engineers, Part G: Journal of Aerospace Engineering,
tute of Technology in 1994, 1997 and 2007, 224(11), 1229–1242.
respectively, and now is a professor there. Liu, B. (2000). The modern control theory. Beijing: China
His main research interests include flight Machine Press.
dynamics and control, cooperative control, Luo, G.-c., Yu, J.-q., Zhang, S.-y., & Zhang, W. (2012). Artificial
flight vehicle system design and robust con- potential field based receding horizon control for path plan-
trol. ning. In G.-H. Yang, W.-D. Zhang (Eds.), The 24th Chinese
Control and Decision Conference (CCDC) (pp. 3665–3669).
Xiaolong Su was born in 1989. He re- Piscataway, NJ: IEEE.
ceived his BS degree from Beijing Institute Oland, Espen, & Kristiansen, Raymond. (2013). Collision and
of Technology in 2012. He is a postgrad- terrain avoidance for UAVs using the potential field method.
uate in flight vehicle design from Beijing In IEEE Aerospace Conference (pp. 1–7). Washington, DC:
Institute of Technology. His research inter- IEEE Computer Society.
ests include flight dynamics and control and Park, M.G., Jeon, J.H., & Lee, M.C. (2001). Obstacle avoidance
flight vehicle system design. for mobile robots using artificial potential field approach with
simulated annealing. In M. H. Lee, H. Haneda (Eds.), Pro-
ceedings of the IEEE International Symposium on Industrial
Electronics (pp. 1530–1535). Piscataway, NJ: IEEE Service
References Center.
Aitsaadi, N., Achir, N., Boussetta, K., & Pujolle, G. (2011). Paul, Tobias, Krogstad, Thomas R., & Gravdahl, Jan Tommy.
Artificial potential field approach in WSN deployment: (2008). Modelling of UAV formation flight using 3D poten-
Cost, QoM, connectivity, and lifetime constraints. Computer tial field. Simulation Modelling Practice and Theory, 16(9),
Networks, 55, 84–105. 1453–1462.
Bai, Y.-q., Liu, H., Shi, Z.-y., & Zhong, Y.-s. (2012). Robust flight Scherer, S., Singh, S., Chamberlain, L., & Elgersma, M. (2008).
control of quadrotor unmanned air vehicles. Robot, 34(5), Flying fast and low among obstacles: Methodology and ex-
519–524. periments. The International Journal of Robotics Research,
Berry, A., Howitt, J., Gu, D.-W., & Postlethwaite, I. (2010). En- 27, 549–574.
abling the operation of multiple micro-air-vehicles in increas- Shim, D.H., Kim, H.J., & Sastry, S. (2003). Decentralized non-
ingly complex obstacle-rich environments. In J. Rankin (Ed.), linear model predictive control of multiple flying robots. In
AIAA Infotech@ Aerospace 2010, Atlanta, Georgia (p. 1–14). D. A. Lawrence, T. Parisini (Eds.), Proceedings of the 42nd
Reston, VA: American Institute of Aeronautics and Astronau- IEEE Conference on Decision and Control (pp. 3621–3626).
tics Inc.. Piscataway, NJ: IEEE Operations Center.
Caves, A.D.J. (2010). Human-automation collaborative RRT for Srikanthakumar, S., Liu, C., & Chen, W.H. (2012). Optimization-
UAV mission path planning. Cambridge, MA: Massachusetts based safety analysis of obstacle avoidance systems for un-
Institute of Technology. manned aerial vehicles. Journal of Intelligent & Robotic Sys-
Charifa, S., & Bikdash, M. (2009). Comparison of geometrical, tems, 65(1–4), 219–231.
kinematic, and dynamic performance of several potential field Sun Jinsheng, L.J., & Wang Zhiquan. (1994). Discrete fault-
methods. In A. Wilson, B. Marshal, M. L. Smith (Eds.), South- tolerant control system design based on Riccati equations.
eastcon, 2009. SOUTHEASTCON’09. IEEE: (pp. 18–23). Pis- Journal of Nanjing University of Aeronautics & Astronautics,
cataway, NJ: IEEE Service Center. 26S1, 48–52.
Chen, Y., Zhao, X., & Han, J. (2010). Review of 3D path planning Wu, S.Z. (2008). Optimal control theory and application. Beijing:
methods for mobile robot. Robot, 32, 568–576. China Machine Press.
Deb, K. (2005). Multi-objective optimization. Search Methodolo- Xiao-Qun, C., Wei-Min, Z., Jun-Qiang, S., Xiao-Qian, Z., & Jun,
gies, 273–316. Z. (2012). Parameter identification of map chaotic system
Fang, Q., & Wang, X. (2006). Application of the maximum prin- with discrete variational method. Acta Physica Sinica, 61(2):
ciple with state constraints to reservoir operation. Acta Auto- 020507. p. 1–7.
matica Sinica, 32, 767–773. Xie, L.-j., Xie, G.-r., Chen, H.-w., & Li, X.-l. (2008). Solution
Khatib, O. (1986). Real-time obstacle avoidance for manipula- to reinforcement learning problems with artificial potential
tors and mobile robots. The International Journal of Robotics field. Journal of Central South University of Technology, 15,
Research, 5, 90–98. 552–557.
Koren, Y., & Borenstein, J. (1991). Potential field methods and Xu, Xinying, Xie, Jun, & Xie, Keming. (2006). Path planning
their inherent limitations for mobile robot navigation. In Pro- and obstacle-avoidance for soccer robot based on artificial
ceedings of the IEEE International Conference on Robotics potential field and genetic algorithm. In T.-Y. Chai, Meng
and Automation (pp. 1398–1404). Piscataway, NJ: IEEE. Max Q.-H. (Eds.), The Sixth World Congress on Intelligent
Lee, M.C., & Park, M.G. (2003). Artificial potential field based Control and Automation (pp. 3494–3498). Piscataway, NJ:
path planning for mobile robots using a virtual obstacle con- IEEE.
cept. In H. Hashimoto (Ed.), Proceedings of the IEEE/ASME Yang, K., & Sukkarieh, S. (2012). Model predictive unified plan-
International Conference on Advanced Intelligent Mechatron- ning and control of rotary-wing unmanned aerial vehicle. In
ics (pp. 735–740). Piscataway, NJ: IEEE. N.-S. Hur, C.-G. Kang (Eds.), The 12th International Confer-
Li, B. (2011). Optimal control problems with constraints on the ence on Control, Automation and Systems (pp. 1974–1979).
state and control and their applications. Harbin City, Hei- Piscataway, NJ: IEEE.
longjiang Province, China: Curtin University. Yin, L., & Yin, Y.-x. (2009). Simulation research on path planning
Lin, C.L., Li, Y.H., & Aouf, N. (2010). Potential-field-based evo- based on dynamic artificial potential field. Journal of System
lutionary route planner for the control of multiple unmanned Simulation, 21, s3325–s3329.
1420 Y.-b. Chen et al.

Appendix 1. Proof of theorem Then we can obtain

Due to the arbitrariness of the field of the APF method, the sound μ
e− m t (C1 + C2 + um ) C1 + C2 + um
APF and the sound distance between the UAV and the obstacles v ≤ − + . (A4)
can guarantee the following conditions: μ μ
⎧ C1 +C2 +um
⎪ F c  ≤ C1
 If t → ∞, then v → = const. Therefore, v(k) is
⎪
⎪  
μ
⎪
⎨ F y ≤ C2
⎪ bounded.
C1 + C2 We also have
< vm , (A1)
⎪
⎪ μ
⎪
⎪ 1    
 F y + F c + F z +u ≤ 1 ( F y  + F c  + F z  + u)
⎩ C1 + C2 + μvm < am
⎪
m m m
1
where μ is the drag coefficient. The initial velocity v is 0 m/s. ≤ (C1 + C2 + μvm + um ). (A5)
m
Based on the triangle inequality, we have
Therefore, the acceleration constraint is also bounded.
   
 F c + F y  ≤ F c  +  F y  ≤ C1 + C2 . (A2) Suppose u is limited by

If the discrete time interval is short enough, the result of ||u|| ≤ um ≤ min(μvm − C1 − C2 , mam − C1 − C2 − μvm ).
the discrete time system and the continuous time system will (A6)
be very close. In order to simplify the proof procedure, we use
the continuous time system to replace the discrete time system. Let um ≤ min(μvm − C1 − C2 , mam − C1 − C2 − μvm ). Then,
According to the UAV particle dynamics, we have (11) and (12) are satisfied simultaneously.
Above all, the sound APF and the sound distance between the
dv 1 UAV and the obstacles can ensure the constraints of a and v.
= (F y + F c + F z + u). (A3)
dt m Theorem is proved.