Microsystem Technologies (2020) 26:3533–3546

[Link]

TECHNICAL PAPER

A real-time path planning algorithm for cable-driven parallel robots

in dynamic environment based on artificial potential guided RRT
Jiajun Xu1 • Kyoung-Su Park1

Received: 17 November 2019 / Accepted: 26 June 2020 / Published online: 3 July 2020
 Springer-Verlag GmbH Germany, part of Springer Nature 2020

Abstract
This paper deals with the collision-free path planning of cable-driven parallel robots (CDPRs) in a dynamic three-
dimensional environment. The proposed algorithm is based on the artificial potential field (APF) approach which provides
a simple and effective path planning method. However, the APF approach is easy to cause the robot oscillation and hard to
reach the goal when it is applied in a dynamic environment. To overcome these two problems, the rapidly exploring
random tree (RRT) was used to provide a random disturbance in order to reduce oscillations and reach the goal in less time.
Due to the particularity of CDPRs, the proposed algorithm consider the wrench feasible workspace (WFW) and various
collision conditions caused by the cable. The proposed algorithms were evaluated with three kinds of simulations.
According to the simulation results, the proposed algorithms found an optimized path with less time cost was reduced about
27% compared with the APF approach. Furthermore, the oscillation phenomenon in the APF approach was effectively
mitigated by 49%. Finally, the experimental result demonstrates the validity of the proposed algorithm.

1 Introduction and Gouttefarde 2014). The research on WFW has made

great progress that a variety of workspaces such as stiffness
Cable-driven parallel robots (CDPRs) is a special type of feasible workspace and collision-free workspace were
parallel robots which mainly transmit the force and motion proposed (Gouttefarde et al. 2010; Wang et al. 2016;
from the driving device to the end-effector by the cable Bolboli et al. 2019).
(Gosselin 2014). As a result of cable actuation, CDPRs The other issue about CDPRs is the cable interferences
have several advantages such as a large workspace, high during the path planning in case of a cluttered environment.
motion speed, and high payload capacity (Heo et al. 2018). This phenomenon becomes more serious in a dynamic
These advantages make CDPRs widely used in material environment. Accordingly, researchers have developed
handling, industrial production, virtual reality and aircraft several path planning algorithms to avoid cable interfer-
spraying, etc. (Martin et al. 2018; Wu et al. 2015; Michelin ences (Youssef and Otis 2020; Bak et al. 2019; Lahouar
et al. 2014). Therefore, many researchers have focused on et al. 2009). Youssef and Otis (2020). developed a new
utilizing the cable to exert the advantage of CDPRs. approach to avoid interference between two cables in real-
However, CDPRs also have some problems induced by time by moving the attachment point of the cable, which
a cable actuation. Because the cable can only be pulled, takes advantage of the properties of reconfigurable CDPRs.
one fundamental issue about CDPRs is to guarantee that Bak et al. (2019). proposed a modified goal-biased RRT
the cable tensions remain positive all the time. This special algorithm which suggested the post-processing algorithm
characteristic determines the wrench feasible workspace to reduce entire path cost and completed the collision-free
(WFW) in which any wrench can be generated at the end- path planning for CDPRs in a cluttered environment.
effector while satisfying the cable tension limits (Nguyen Lahouar et al. (2009). presented a grid method to find a
way to avoid the obstacle. However, these algorithms are
implemented in a static environment without moving
& Kyoung-Su Park objects.
pks6348@[Link]
To achieve collision-free planning in a dynamic envi-
1
Department of Mechanical Engineering, Gachon University, ronment for CDPRs. The artificial potential field (APF)
1342 Seongnamdaero, Sujeong-gu, approach and the rapidly exploring random tree (RRT)
Seongnam-si 461-701, Gyeonggi-do, Korea

123
3534 Microsystem Technologies (2020) 26:3533–3546

algorithm can be used and improved (Mirjalili et al. 2020). solved for a dynamic environment. To do so, the WFW
The APF approach was proposed by Khatib (1985), the analysis is carried out. The simulation and experiment
principle of APF is to create a virtual potential field where results show the effectiveness of the proposed algorithm.
a goal point produces an attractive field while the obstacle This paper is arranged as follows. Section 2 introduces the
produces a repulsion field. The APF approach can be used problem definition and background. Section 3 represents
to solve the real-time planning problems of the robot. the various constraints which include workspace and col-
Orozco-Rosas et al. (2015) proposed a novel method called lision detection. In Sect. 4, the dynamic APG-RRT algo-
Bacteria Potential Field (BPF) based on APF, which uses rithm is introduced. Finally, simulation and experiments
the Bacterial Evolutionary Algorithm (BEA) to find the are presented in Sect. 5 to verify the proposed algorithm.
shortest path according to different criteria. Yan et al.
(2018) proposed an improved artificial potential field
method by establishing a new from a repulsive force, but 2 Background
this method has only been used in a static environment.
LaValle (1998) proposed RRT in 1998, which is a 2.1 Problem formulation and considerations
sampling-based path planning algorithm. RRT is consid-
ered probabilistically complete which means the algorithm The setup and considerations of the path-planning algo-
converges to the feasible solution as the number of itera- rithms addressed in this paper are described in this sec-
tions increases. The RRT exploration process is completely tion. Path planning for CDPRs is a challenging task due to
random but the computation time can be reduced by lim- the interference of the cable. Thence, the following
iting the sampling region or establish two trees from the assumptions and considerations about the system and
start point and the goal point (Kang et al. 2019; Klemm environment are presented.
et al. 2015). Xiang et al. (2020) represented a dynamic
1. The cable is simplified to a straight line model for
point-to-point trajectory planning method for CDPRs using
faster obstacle detection and the cable does not deform
the RRT algorithm. Zhang et al. (2018) completed the path
due to external forces.
planning and control for CDPRs by optimal RRT* method.
2. The cable outgoing point where usually exist the pulley
In addition, the winding path generated by the original
is considered a fixed point and the effect of the pulley
RRT algorithm can be optimized with the iteration process.
on the length of the cable is ignored.
The asymptotically optimal RRT* and RRT*FN was pro-
3. The end effector is a regular cube and the moving
posed to optimize the path with iteration increases (Kara-
obstacle is treated as a sphere. In the real world, these
man and Frazzoli 2011; Adiyatov et al. 2013). However,
assumptions need to be premeditated based on the
these two methods required more time to perform calcu-
actual application.
lations and make them difficult to apply in a dynamic
environment. The primary task of the path planning problem is to find
For the dynamic environment, some researchers have a feasible path from the initial point to the goal point. In a
proposed RRT*FND and RRTX to solve path planning dynamic environment, path planning demands a shorter
problems (Adiyatov et al. 2017; Otte and Frazzoli 2016). time cost while satisfying the relevant constraints. The
These two algorithms carried out a quick re-planning when configuration space is represented as X 2 Rd , where x 2 X
the obstacle is moving through repairing and refining the is a particular configuration of the robot, and d is the
tree. However, both algorithms need to build the tree dimension of state space. Xobs 2 X is the set of states rep-
before the end effector moves to reuse the information of resenting obstacles in the configuration space which are
the tree for re-planning. For CDPRs, pre-establishing a tree forbidden regions for the [Link] ¼ XXobs is the set of
not only consumes time but also makes it difficult to repair states that makes up the collision-free region. The Tree
due to the cable will also cause the branches to break. represents the tree of RRT growing from xstart 2 Xfree to
Based on the work and problems mentioned, this paper xgoal 2 Xfree , where xstart and xgoal represents the initial and
presents a simple and fast path planning algorithm based on goal state, respectively. V represents the vertices that
APF and improved by RRT. Although some researchers consist of feasible nodes and these nodes are intercon-
have also considered the combination of the APF method nected to form edges E. The Tree is a topological graph
and RRT to enhance the algorithm, only the static envi- composed of V and [Link] collision-free path is generated
ronment has been studied (Zhang et al. 2019; Tahir et al. by finding the parent node of the end node constantly.
2018; Agarwal et al. 2018). The comparison between
current and previous studies with different criteria is shown
in Table 1. In this paper, a three-dimensional dynamic
environment with the moving obstacle is considered and

123
Microsystem Technologies (2020) 26:3533–3546 3535

Table 1 The comparison of the different method

Reference Degrees of Number of Workspace Dynamic Cable Obstacle Experiment
freedom cables analysis environment collision avoidance

Bak et al. (2019) 6 8 Yes No Yes Yes No

Lahouar et al. 4 4 Yes No Yes Yes No
(2009)
Xiang et al. 3 3 Yes No No No No
(2020)
Zhang et al. 3 4 Yes No Yes Yes No
(2018)
Current study 6 8 Yes Yes Yes Yes Yes

2.2 The artificial potential field approach where ero is the unit vector from the robot to the obstacle.
When vro  0 the robot moves away from obstacles and
The APF has two components: the goal’s potential and the does not necessary to avoid the obstacle. The obstacle
obstacles’ potential. The gravitational field centered on the avoidance behavior occurs in the case of vro [ 0 which the
xgoal which makes the extension node approach the xgoal as robot moves toward the obstacle.
soon as possible. The gravitational field Uatt and the
gravitational function Fatt can be expressed as 2.3 Rapidly-exploring random tree algorithm
 2  2
Uatt ðx; vÞ ¼ nx xcurr  xgoal  þnv vcurr  vgoal  ð1Þ
The basic RRT begins with a tree rooted at the only start
Fatt ðx; vÞ ¼ rUatt ðx; vÞ ¼ rx Uatt ðx; vÞ  rv Uatt ðx; vÞ node xstart and incrementally extends the tree by randomly
ð2Þ searches all over the workspace. First, the algorithm
    determines if the distance between xstart and xgoalis less
Fatt ¼ 2nx xcurr  xgoal exrg þ 2nv vcurr  vgoal evrg ð3Þ
than a specific threshold. Then, the algorithm randomly
where nx and nv are positive attractive [Link] is the samples a node xrand in configuration space. Second, the
unit vector from the current position xcurr to the goal NearestVertex procedure returns a node xnear which is the
position. evrg is also a unit vector with the direction of the nearest neighbor node of xrand. The new node xnew could be
relative velocity between the goal point and the robot. obtained by a specific step from xnearto xrandwhich makes
The obstacles’ potential generates the repulsive force to up the edge of the tree. Finally, the line segment from
actuate the robot move away from the obstacles. The xnearto xnew will be checked if it is satisfied collision-free.
repulsive field in a dynamic environment changes with the The above process will be repeated until the distance
obstacle moving and its expression is as follows between xnew and xgoal is smaller than a given value called
8 2  2 3 error tolerance, which is the end of this algorithm. The
>
>  2 kxcurr xgoal k
>
> 1 1 41  e R
5 whole process of the algorithm is shown in Algorithm 1.
>
< g1 q  q
0
Urep ðx; vÞ ¼
>
>
>
> þg2 vro ; q  q0 and vro [ 0
>
:
0q [ q0 or vro  0
ð4Þ
where g1 and g2 are positive repulsive coefficient, q rep-
resent the shortest distance between the robot, R is the
robot’s radius and the obstacle, and q0 is the radius of
obstacles influence, the obstacle avoidance behavior will
occur when the robot enters the range of obstacles.
The relative velocity component of the robot and the
obstacle is expressed as follows
vro ¼ðv  vobs ÞT ero ð5Þ
• Distance: Given two states xnew and xgoal, it returns the
distance between these two states by a distance

123
3536 Microsystem Technologies (2020) 26:3533–3546

is determined by the length and geometry of the connecting

Discrete framespace
link. However, the workspace of CDPRs mainly depends
on the property of the cable which sustains only positive
tension. Hence, a pose of the end-effector which causes the
Given Orientation Angle
cable to generate negative tensile force must be avoided.
,Wrench W
Based on the feasible wrench set, the wrench feasible
workspace (WFW) was defined by researchers (Bosscher
Calculate AT,Pseudo inverse et al. 2006).
A+T, Null space AT In all proposed methods, researchers did not consider the
maximum workspace of CDPRs in which the cable obtains
the worse wrench set. As a result of the maximum work-
LMI Solver space is extremely inclusive, the workspace analysis makes
it possible to provide designers with analytical insights and
the maximum workspace can be restricted according to
LMI =Feasible NO different criteria.
The relationship between the cable tension and external
force/torque acting on end-effector is given by
YES
2 3
  f1
Save feasible point u1  um 6 .. 7
4 . 5 ¼ w ð6Þ
b 1  u 1    b m  um
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} fm
AT |fflffl{zfflffl}
f
Wrench feasible workspace
where um and bm represent the unit vector of the m’ th cable
Fig. 1 Flowchart of wrench feasible workspace and the position vector of the m’ th cable attachment point
on the end-effector. AT is the structure matrix, f denotes
function. Euclidean distance is usually used in this cable force and the wrench w denotes external force/torque
algorithm. acting at the end-effector.
• RandomSample: Given a configuration space, it The wrench feasible workspace of an n-DOF CDPRs
returns a random node xrand. In many cases, the xgoal with r degrees of redundancy can be defined as the set of
can be chosen as xrand with a certain probability to poses of the end-effector at which
speed up the convergence. 8w 2 Rn ; 9ðf [ 0Þ 2 Rm 3 AT f ¼ w ð7Þ
• NearestVertex: Given the tree and xrand, it returns the
state xnear that is the nearest node to xrand on the graph In general, the structure matrix AT is square when
tree. CDPRs are under-constrained, which means the cable
• ExtendTowards: Given xnear and xrand, it returns a new tension is unique and the distribution depends entirely on
node xnew that is located on the line segment from xnear the external force. However, when the redundancy r is not
to xrand and distance between xnear and xnew is equal to zero, there is an infinite number of solutions for the cable
the step size. tension distribution. The solution of Eq. (6) can be
• CollisionFree: Given xnear and xnew, it returns true if obtained by the least square minimum norm as follows
the line segment between xnear and xnew is collision- f ¼ |fflfflffl
A þT
ffl} þ |ffl{zffl}
ffl{zfflfflfflw HkH ð8Þ
free. Conversely, it returns false if this path has f ext f int
collision.
AþT ¼ AT ðA  AT Þ1 ð9Þ

where AþT is the Moore–Penrose matrix inverse of AT , H

3 Modeling of cable-driven parallel robot denotes the null space of the structure matrix, and kH 2 Rr
and collision detection is the r-dimensional column vector called the cable tension
general solution coefficient. According to the definition of
3.1 Wrench feasible workspace wrench feasible workspace in (7) and the solution in (8),
the workspace problem can be defined as
One of the characteristics of CDPRs is that they have a
large workspace. The workspace of the rigid parallel robot 8w 2 Rn ; 9kH 2 Rr 3 AþT w þ HkH [ 0 ð10Þ

123
Microsystem Technologies (2020) 26:3533–3546 3537

lm ¼ am  r  Rbm ð12Þ
where am donate the position vector of the m’ th cable
Pulley outlet point, r represents the end-effector position vector in
the global coordinate system. R is the rotation matrix
associated with the rotation of the end-effector. The colli-
End-effector sion detection can be implemented by calculating the
Cable shortest distance between the point and the line segment.
Fame In addition to external interference, the self-interfer-
ences of CDPRs is also an important factor for path plan-
ning. The self-interferences of CDPRs involve the
collisions of the cable to cable and the cable to the end-
effector. Although the Gilbert–Johnson–Keerthi (GJK)
algorithm can directly check for self-interferences (Gilbert
Winch-servo et al. 1988), but the GJK algorithm consumes time for
system iteration. In this paper, the method proposed by Aref and
Taghirad (2008) is used to check the self-interferences of
CDPRs.
Fig. 2 The redundant constraint CDPR with eight cables

4 Artificial potential guided RRT in dynamic

Equation (10) can be represented in terms of linear environment
matrix inequality (LMI). LMI approaches are convex
optimization problem and the Eq. (10) is described as The oscillation phenomenon is one of the shortcomings for
AðxÞ ¼ A0 þ Ax [ 0 ð11Þ the APF method in a dynamic environment, which due to
the constant movement of the obstacle to make the robot
where x 2 Rr is variable and A0 ,A are given constant cannot get rid of the collision region. Compared with local
symmetric real matrices. The projective method is an minima traps, this phenomenon is caused by uncontrollable
efficient interior-point method that can be used to solving movement obstacles and it is difficult to eliminate by
the feasibility problem (Gahinet and Nemirovski 1997) and adjusting parameters (Ge and Cui 2000). Therefore, the
feasp function of Matlab is used for the strict feasibility of random perturbation with a certain probability is added to
LMI. the algorithm and the whole process is shown in Algorithm
As shown in Fig. 1, the discretization method is used to 2.
solve the wrench feasible workspace. The space of CDPRs
frame is discreted to check the feasibility for each pose of
the end-effector. Then, the Moore–Penrose matrix inverse
and null space AT are calculated to formulate the LMI.
Finally, the LMI solver returns the feasible points of
CDPRs, these feasible points make up the wrench feasible
workspace.

3.2 Collision detection

The collision detection is one of the most important tasks

for CDPRs path planning. In a dynamic environment, the
CDPRs is assumed that has sensors to sense the position
and velocity of the moving obstacle. Therefore, the shortest
distance between the end-effector and the moving obstacle
can be obtained directly by the sensor. The kinematics of
CDPRs can be used to calculate the shortest distance StepSize defines a determined step size at each sampling
between the cable and the moving obstacle. interval, which also describes the speed of the end-effector.
The cable vector lm with the straight-line model is When the CDPRs are not within the influence region of the
expressed as follows moving obstacle, it means that both end-effector and the

123
3538 Microsystem Technologies (2020) 26:3533–3546

Table 2 The coordinates of cable outlet point contrast, when the end-effector or cables enter the influence
Cable outlet point x (mm) y (mm) z (mm)
region of the obstacle, the DynamicExtendTree function
will cause the end-effector to move to a new point xnew.
P1 - 427.08 524.32 864.74 The DynamicExtendTree function is shown in Function 1.
P2 439.72 525.94 860.87
P3 - 435.14 - 521.91 839.17
P4 437.89 - 523.83 859.46
P5 - 535.04 466.89 79.72
P6 531.98 451.43 80.32
P7 - 543.04 - 463.81 71.49
P8 529.05 - 448.63 84.36

Function 1 describes the local obstacle avoidance

strategy of the end-effector. Firstly, CertainProbability is
a user-defined probability aim to reduce oscillations. In
practically, a certain probability can take a random value.
However, the algorithm needs to consume more computing
Fig. 3 The wrench feasible workspace time to find an optimized sub-path with the probability
increases, which will reduce the real-time performance of
the algorithm. Therefore, we take the value of 0.5 in this
Table 3 Simulation parameters paper to verify the algorithm.
When the probability is met, ExtendTowards is affec-
Parameters Value
ted by both attractive and repulsive forces which cause the
End-effector start point /mm [200, 350, 300] end-effector to move with the step size in the direction of
End-effector goal point/mm [- 200, - 350, 600] the resultant force. In contrast, when the probability is not
Obstacle start point/mm [200, 350, 600] met, the NewPoint will generate some random points to
Obstacle start point/mm [- 200, - 350, 300] obtain the series of new points xnew like RRT algorithm,
nx 20 and these new points are checked by ConstrainsFeasible
nv 10
g1 30
g2 10
q0 /mm 70
Obstacle radius/mm 40
Sampling interval/s 0.2
Obstacle speed m/s 0.086

cable are far away from the moving obstacle. In this case,
the end effector only affected by the attractive force and
moves directly to the goal point. ExtendTowards is
affected by the attractive force which makes Tree grows
with the current moment T to the goal point directly. In Fig. 4 The direct situation

123
Microsystem Technologies (2020) 26:3533–3546 3539

Fig. 5 Distance change in the direct situation Fig. 8 Distance change in APF approach situation

Fig. 6 Position change in the direct situation Fig. 9 CDPR position change in APF approach situation

and CollisionFree. ConstrainsFeasible is used to check if

xnew is within the wrench feasible workspace and Colli-
sionFree is used to check various collisions include the
external interference the self-interference. Finally, a series
of new points that meet the above conditions are derived to
optimize the path and FindCheapest will find the cheapest
point from these new points. The cheapest point can be
defined according to different evaluation indexes. In this
paper, time priority is used as an evaluation index which
means FindCheapest always finds the new point which is
closest to the goal point.
Although the path obtained by the proposed algorithm is
Fig. 7 The APF approach situation
relatively optimal, it does not require a complex calculation

123
3540 Microsystem Technologies (2020) 26:3533–3546

Fig. 10 Relative position between CDPR and the moving obstacle in APF approach situation: a t = 1.6 s; b t = 5.2 s; c t = 7.6 s

Fig. 11 The dynamic APG-RRT situation

Fig. 13 CDPR position change in APG-RRT situation

obstacle. In Function 1, there is only one loop is executed

to avoid the moving obstacle. Therefore, the time com-
plexity of our method is O(n) which supports the real-time
operation. The fast calculation process guarantees the real-
time property of the algorithm and it is critical to the
CDPRs in a dynamic environment.

5 Simulation and experimental results

In this section, several simulations and experiments are

carried out based on a typical redundantly restrained CDPR
with eight cables in Fig. 2. The redundantly restrained
CDPR means m [ n ? 1 in this special configuration and
Fig. 12 Distance change in APG-RRT situation the cable tension has an infinite number of solutions. The
redundant actuation makes CDPRs have high stability, but
process. The proposed method only executes when the end- the overmuch cables are more likely to interfere with the
effector or the cable close to the influence radius of the surrounding environment. Therefore, it is essential to make

123
Microsystem Technologies (2020) 26:3533–3546 3541

Fig. 14 Relative position between CDPR and the moving obstacle in dynamic APF-RRT situation: a t = 2.2 s; b t = 3.2 s; c t = 5.2 s

Table 4 Simulation comparison

Simulation Step size Average Average path SD of path Average path SD of path Average arrival
(mm) oscillation time calculation time (s) calculation time (s) cost (m) cost (m) time (s)
(s)

Direct 80 0.8 0.0038 0.00072 1.28 0.021 3.2

situation
APF 20 10.2 0.0168 0.0026 1.88 0.112 18.8
approach
Dynamic 20 6.2 0.8486 0.3854 1.36 0.394 13.6
APG-RRT

DJI Tello
Drone

Calibration
Checkerboard

Logitech 4K
Webcam 2 Logitech 4K
Webcam 1

Fig. 16 The experimental environment of CDPR

Fig. 15 The set of moving obstacle path

The wrench feasible workspace cannot form a complete

path planning simulation with this typical CDPR. The end-
coverage of the space even with the worst wrench set,
effector is considered as a cube with a size of
which is determined by the characteristic of the structure
65 9 65 9 65 mm and the weight is 0.4 kg. The coordi-
matrix.
nates of cable outlet point Pi(i = 1,…,m) as shown in
The simulation is divided into three cases for compari-
Table 2.
son. The moving obstacle is in the same motion state which
Figure 3 shows the wrench feasible workspace which is
is a diagonal path with constant speed during these three
calculated according to Fig. 1. The red line in Fig. 3 is
simulations. These three simulations include the direct
linked by the cable outlet point and the blue region is the
situation, APF approach situation, and dynamic APG-RRT
wrench feasible workspace which covers 71% of space.

123
3542 Microsystem Technologies (2020) 26:3533–3546

Fig. 17 The experimental results of the APF approach: a The position c

changes of the drone; b The position changes of the end-effector;
c The distance changes between the moving obstacle and the CDPR

Table 5 Hardware parameters

Hardware parameters Value

Drone size/mm 158 9 158 9 50

End-effector speed m/s 0.5
Web camera fps 30
Object detection fps 2
End-effector initial position/mm [0, 0, 445]
CDPR step size/mm 50

situation. The parameters used in the simulation are shown (a)

in Table 3.

5.1 Simulation of direct situation

The collision in a dynamic environment depends largely on

the relative speed of the end effector and the obstacle. The
direct situation means that the end-effector can quickly
reach the goal point without entering the affected region of
the obstacle when it moves at a faster speed and Fig. 4
illustrates this situation.
In Fig. 4, the yellow part is a spherical obstacle and the
red square is the end-effector. The end effector has a step
size of 80 mm during each sampling interval. Figure 5
illustrates the distance change of the end-effector and the
obstacle. The black line shows the shortest distance
between the end-effector and the obstacle. Other lines are
the distance between the cable and the obstacle.
(b)
The path planning in a dynamic environment depends
largely on the relative velocity of the obstacle and the
robot. The end effector can reach the target point directly
because of the obstacle has no chance to affect the robot
even at 1 s in Fig. 5. Therefore, the robot position changes
without avoidance behavior as shown in Fig. 6.

5.2 Simulation of APF approach

It can be known from Function 1 that Algorithm 2

degenerates into the APF approach when a certain proba-
bility is gradually approached to zero. This situation is
shown in Fig. 7.
Figure 7 shows the obstacle avoidance case using the
APF approach when the step size of the end effector is
20 mm in each sampling interval. Although CDPR can
avoid moving obstacles, the moving obstacle still influence
(c)
the robot, which causes the end-effector to produce sig-
nificant oscillation. Figure 8 illustrates that the oscillation
phenomenon due to the robot is affected by local minima.

123
Microsystem Technologies (2020) 26:3533–3546 3543

Fig. 18 The double webcams shot results of the APF approach: a 18 s; b 20 s; c 44 s; d 45 s

The end- effector is always within the influence radius of In addition, the random disturbance causes the robot to
the moving obstacle during 1.6–7.6 s, which leads to the generate a serious of sub-paths to execute the
robot oscillates between the obstacle and the goal point. FindCheapest function which will find an optimal sub-
Furthermore, this phenomenon ends with the obstacle dis- path based on the time priority. The optimized sub-path is
appear in the workspace. The final trajectory and relative closer to the goal point compared to the APF approach and
position are shown in Figs. 8 and 9, respectively. The the total time cost is reduced.
relative position between the CDPR and the obstacle at
specific moments is shown in Fig. 10, which demonstrates 5.4 Evaluation for three kinds of approach
the oscillation phenomenon caused by potential force.
In this section, batch evaluation is implemented to evaluate
5.3 Simulation of dynamic APG-RRT the rate of the proposed algorithm. The moving obstacle
paths with different angles is depicted in Fig. 15. These
It can be seen from Figs. 8 and 9 that although the APF moving obstacles move at a constant speed and the total
method can avoid the collision between CDPR and the time from the start point to the target time is 10 s. A total
moving obstacle, the end effector will be dominated by the of 1000 diagonal paths with the center at point (x = 0 mm,
movement obstacle and fall into the long-term oscillation. y = 0 mm, z = 445 mm) are calculated numerically. The
To alleviate this oscillation, the proposed dynamic APF- simulation environment is MATLAB 2020a with an Intel
RRT algorithm adds a certain probability to make the end- i7-9750H processor and 32 GB RAM system. The results
effector get rid of the moving obstacle and introduce of the batch evaluation are presented in Table 4.
optimization conditions to make the end-effector reach the Comparing the APF approach with Dynamic APG-RRT
goal with a faster time. Figure 11 shows the simulation from Table 4, the oscillation time is reduced by 49% and
using dynamic APG-RRT. the arrival time is reduced by 27% using the proposed
Same as the APF approach situation, the step size of the algorithm. The average path calculation time of the pro-
end effector is 20 mm in each sampling interval in simu- posed algorithm is increased because of the random pro-
lation 3. However, due to the certain probability is 0.5, it cess. However, the increase of average path calculation
can be seen from Fig. 11 that the end-effector quickly gets time is less than 1 s, which guarantee the real-time prop-
rid of the obstacles of the moving obstacle. Compared with erty. Moreover, according to the direct situation, it is
Fig. 9, Figs. 12 and 13 illustrate the proposed algorithm necessary for the end-effector to obtain a faster speed for
with a random disturbance makes the end-effector not the path planning of CDPRs in a dynamic environment.
always follow the direction of the resultant force caused by However, when the influence of the obstacle is unavoid-
APF which provides the robot a chance to get rid of the able, the proposed algorithm can effectively alleviate the
control of local minima. Moreover, the relative position oscillation and drive the end effector to reach the goal in a
between CDPR and the obstacle in Fig. 14 reveals the short time.
optimized sub-path obtained by Function 1 to avoid colli-
sion and reduce the oscillation phenomenon.

123
3544 Microsystem Technologies (2020) 26:3533–3546

b Fig. 19 The experimental result of the dynamic APG-RRT approach:

a The position changes of the drone; b The position changes of the
end-effector; c The distance changes between the moving obstacle
and the CDPR

5.5 Experimental results and discussion

The proposed APG-RRT algorithm is compared with the

APF to verify the effectiveness during the experiment.
A DJI Tello drone is controlled manually as the moving
obstacle to identify the real-time property. The drone
approaches the end-effector from above. To obtain the 3-d
coordinate of the drone, a vision system with double web
cameras is developed based on the Faster-RCNN algorithm
(a) (Ren et al. 2015). The experimental environment of CDPR
is illustrated in Fig. 16 and the hardware parameters are
shown in Table [Link] a result of our CDPR prototype is
controlled by TwinCAT3 PLC, the automation device
specification (ADS) communication protocol is utilized to
build the link between the algorithm and the industrial PC.
Figures 17 and 18 illustrate the collision avoidance
process with the APF approach. The experimental result is
similar to the APF simulation, the robot moves with the
direction of potential force which is related to the distance
between the moving obstacle and CDPR. Although the
robot avoids the moving obstacle in Fig. 17c, the avoid-
ance strategy leads to the end-effector close to the edge of
the workspace. Therefore, the random disturbance provided
by APG-RRT makes the robot to find another sub-path to
avoid the obstacle.
Figures 19 and 20 show the collision avoidance process
(b) with the APG-RRT approach. Figure 19c demonstrates the
feasibility of the proposed algorithm. Compared to
Figs. 18, 20 shows that the collision-free workspace can be
utilized by the APG-RRT approach, which reduces the
oscillation as the previous simulation. The robot completes
a subpath to avoid the obstacle and the edge of the
workspace.

6 Conclusions

In this research, a real-time path planning algorithm to find

a collision-free path of CDPRs in a dynamic environment
was suggested based on the APF approach. The method to
consider various interferences caused by the cable and
wrench feasible workspace during path planning was sug-
gested. Every collision including the cable to the cable, the
(c) cable to the end-effector, the cable to the obstacle was
detected in the proposed algorithm. The proposed algo-
rithms were evaluated with three kinds of simulations.
According to the simulation results, the proposed

123
Microsystem Technologies (2020) 26:3533–3546 3545

Fig. 20 The double webcams shot results of the APG-RRT approach: a 9 s; b 10 s; c 12 s; d 13 s

algorithms found an optimized path with less time cost was Gahinet P, Nemirovski A (1997) The projective method for solving
reduced by 27% compared with the APF approach. Fur- linear matrix inequalities. Math Program 77(1):163–190
Ge SS, Cui YJ (2000) New potential functions for mobile robot path
thermore, the oscillation phenomenon in the APF approach planning. IEEE Trans Rob Autom 16(5):615–620
was effectively mitigated by 49%. The experiment is car- Gilbert EG, Johnson DW, Keerthi SS (1988) A fast procedure for
ried out with a drone as the moving obstacle. The experi- computing the distance between complex objects in three-
mental results also demonstrate the versatility of the dimensional space. IEEE J Rob Autom 4(2):193–203
Gosselin C (2014) Cable-driven parallel mechanisms: state of the art
proposed algorithm. and perspectives. Mech Eng Rev 1(1):DMS0004–DMS0004
Gouttefarde M, Daney D, Merlet JP (2010) Interval-analysis-based
Acknowledgements This work was supported by the National determination of the wrench-feasible workspace of parallel
Research Foundation of Korea(NRF) grant funded by the Korea cable-driven robots. IEEE Trans Rob 27(1):1–13
government (MSIT) (2018R1A2B6006891). Heo JM, Park BJ, Park JO, Kim CS, Jung J, Park KS (2018)
Workspace and stability analysis of a 6-DOF cable-driven
parallel robot using frequency-based variable constraints. J Mech
References Sci Technol 32(3):1345–1356
Kang G, Kim YB, Lee YH, Oh HS, You WS, Choi HR (2019)
Sampling-based motion planning of manipulator with goal-
Adiyatov O, Varol HA (2013) Rapidly-exploring random tree based oriented sampling. Intell Serv Rob 12:265–273
memory efficient motion planning. In: 2013 IEEE International Karaman S, Frazzoli E (2011) Sampling-based algorithms for optimal
Conference on Mechatronics and Automation. IEEE, motion planning. Int J Rob Res 30(7):846–894
pp 354–359 Khatib O (1985) Real-time obstacle avoidance for manipulators and
Adiyatov O, Varol HA (2017) A novel RRT*-based algorithm for mobile robots. In: Proceedings. 1985 IEEE International Con-
motion planning in dynamic environments. In: 2017 IEEE ference on Robotics and Automation Vol. 2. IEEE, pp 500–505
International Conference on Mechatronics and Automation Klemm S, Oberländer J, Hermann A, Roennau A, Schamm T, Zollner
(ICMA). IEEE, pp 1416–1421 JM, Dillmann R (2015) RRT*-Connect: Faster, asymptotically
Agarwal S, Gaurav AK, Nirala MK, Sinha S (2018) Potential and optimal motion planning. In: 2015 IEEE International Confer-
sampling based RRT star for real-time dynamic motion planning ence on Robotics and Biomimetics (ROBIO). IEEE,
accounting for momentum in cost function. In International pp 1670–1677
Conference on Neural Information Processing. Springer, Cham, Lahouar S, Ottaviano E, Zeghoul S, Romdhane L, Ceccarelli M
pp 209–221 (2009) Collision free path-planning for cable-driven parallel
Aref MM, Taghirad HD (2008) Geometrical workspace analysis of a robots. Rob Auton Syst 57(11):1083–1093
cable-driven redundant parallel manipulator: KNTU CDRPM. Lavalle SM (1998) Rapidly-exploring random trees: a new tool for
In: 2008 IEEE/RSJ International Conference on Intelligent path planning
Robots and Systems. IEEE, pp 1958–1963 Martin A, Caro S, Cardou P (2018) Design of a cable-driven parallel
Bak JH, Hwang SW, Yoon J, Park JH, Park JO (2019) Collision-free robot with grasping device. Procedia Cirp 70:290–295
path planning of cable-driven parallel robots in cluttered Michelin M, Baradat C, Nguyen DQ, Gouttefarde M (2015)
environments. Intell Serv Rob 12:243–253 Simulation and control with xde and matlab/simulink of a
Bolboli J, Khosravi MA, Abdollahi F (2019) Stiffness feasible cable-driven parallel robot (cogiro). In: cable-driven parallel
workspace of cable-driven parallel robots with application to robots. Springer, Cham, pp 71–83
optimal design of a planar cable robot. Rob Auton Syst Mirjalili S, Dong JS, Lewis A (2020) Ant Colony optimizer: theory,
114:19–28 literature review, and application in AUV path planning. In:
Bosscher P, Riechel AT, Ebert-Uphoff I (2006) Wrench-feasible nature-inspired optimizers. Springer, Cham, pp 7–21
workspace generation for cable-driven robots. IEEE Trans Rob
22(5):890–902

123
3546 Microsystem Technologies (2020) 26:3533–3546

Montiel O, Orozco-Rosas U, Sepúlveda R (2015) Path planning for and Programming for Autonomous Robots (SIMPAR). IEEE,
mobile robots using Bacterial Potential Field for avoiding static pp 166–173
and dynamic obstacles. Expert Syst Appl 42(12):5177–5191 Xiang S, Gao H, Liu Z, Gosselin C (2020) Dynamic point-to-point
Nguyen DQ, Gouttefarde M (2014) Study of reconfigurable sus- trajectory planning for three degrees-of-freedom cable-sus-
pended cable-driven parallel robots for airplane maintenance. In: pended parallel robots using rapidly exploring random tree
2014 IEEE/RSJ International Conference on Intelligent Robots search. J Mech Rob 12(4):041007
and Systems. IEEE, pp 1682–1689 Yan P, Yan Z, Zheng H, Guo J (2018). Real time robot path planning
Otte M, Frazzoli E (2016) RRTX: asymptotically optimal single- method based on improved artificial potential field method. In:
query sampling-based motion planning with quick replanning. 2018 37th Chinese Control Conference (CCC). IEEE,
Int J Rob Res 35(7):797–822 pp 4814–4820
Ren S, He K, Girshick R, Sun J (2015) Faster r-cnn: towards real-time Youssef K, Otis MJD (2020) Reconfigurable fully constrained cable
object detection with region proposal networks. In: Advances in driven parallel mechanism for avoiding interference between
neural information processing systems, pp 91–99 cables. Mech Mach Theory 148:103781
Tahir Z, Qureshi AH, Ayaz Y, Nawaz R (2018) Potentially guided Zhang B, Shang W, Cong S (2018) Optimal RRT* planning and
bidirectionalized RRT* for fast optimal path planning in synchronous control of cable-driven parallel robots. In: 2018 3rd
cluttered environments. Rob Auton Syst 108:13–27 International Conference on Advanced Robotics and Mecha-
Wang B, Zi B, Qian S, Zhang D (2016) Collision free force closure tronics (ICARM). IEEE, pp 95–100
workspace determination of reconfigurable planar cable driven Zhang Z, Wu D, Gu J, Li F (2019) A path-planning strategy for
parallel robot. In: 2016 Asia-Pacific Conference on Intelligent unmanned surface vehicles based on an adaptive hybrid dynamic
Robot Systems (ACIRS). IEEE, pp 26–30 stepsize and target attractive force-RRT algorithm. J Marine Sci
Wu Y, Cheng HH, Fingrut A, Crolla K, Yam Y, Lau D (2018) CU- Eng 7(5):132
brick cable-driven robot for automated construction of complex
brick structures: from simulation to hardware realisation. In: Publisher’s Note Springer Nature remains neutral with regard to
2018 IEEE International Conference on Simulation, Modeling, jurisdictional claims in published maps and institutional affiliations.

123