A Distributed Method for UAV Swarm Path Planning

Wee Liang Boon Paw Yew Chai

Infocomm Technology/ Engineering Cluster
Engineering Cluster
Singapore Institute of Technology
Singapore Institute of Technology
Singapore
Singapore
[email protected]
[email protected]

Murteza Kaizar Zakir Faheem Bin Fawzi

Engineering Cluster Engineering Cluster
Singapore Institute of Technology Singapore Institute of Technology
Singapore Singapore
[email protected] [email protected]

Abstract— In Unmanned Aerial Vehicle (UAV) swarm traditional methods of centralized control of swarm, the
2022 International Conference on Computer and Drone Applications (IConDA) | 978-1-6654-9235-5/22/$31.00 ©2022 IEEE | DOI: 10.1109/ICONDA56696.2022.10000376

applications there is a need for autonomous mapping and route current work uses a distributed approach where the task
planning. These applications include collaborative crop assigned to the swarm are decided in a distributed structure.
monitoring, aerial transportation, offshore delivery, facility In addition, a novel feature of our strategy is the highly
monitoring, disaster relieve as well as search and rescue. scalability of the swarm. This means that new members can
Current UAV operations are heavily dependent on human be easily added or removed, and the swarm will continue to
operators that require each UAV to be controlled by a dedicated function and achieve its mission.
safety pilot as well as a separate payload operator which limits
the efficient use of UAVs required in swarm applications. There II. LITERATURE REVIEW
is thus an imperative to mature mapping and path planning
technologies to allow UAV swarm operations. This paper Navigation capability allows the UAV to determine its
explores mapping and path planning strategies that allow higher current location as well as calculates its next flight path to
levels of autonomy for a UAV swarm to operate reliably. In this complete its mission. Practice navigation capability can be
work we developed novel strategies to facilitate swarm mapping, achieved through two approaches: (a) Centralized, and (b)
and swarm path planning. The methods developed can be easily Distributed approaches [1, 2, 3, 4, 5, 6]. In the following
scaled to adapt to different UAV swarm sizes. The developed sections we consider centralized approaches in Section A and
strategies are implemented and tested using simulations. Results B and this is followed by a discussion on distributed
and conclusions drawn from these simulation experiments are approaches in Sections C.
presented.
A. Graph Methods
Keywords— UAV, Path Planning, Swarm Graph methods can be applied when mission terrain is
unknown, and no-fly areas zones are enforced. The Voronoi
I. INTRODUCTION
Graph Search done by Bortoff [4] utilises vertices and edges
Swarm concepts have received increasing attention in to solve a UAV path planning.
recent years. The idea of controlling a swarm of UAVs started
with the initial concept of pre-programmed flight where the By assigning vertices to a state space and connecting them
exact position and location of each drone in the swarm is through edges, providing that the aircraft can fit through, and
established prior to flight and there is no change to the assigning the cost of flight to each edge, we can produce an
preplanned trajectory after the UAV swarm have started the
optimal flight path as shown in Fig 1.
mission. This method is only suitable for entertainment
applications or used in mass aerial display at special events.
The swarm is not adaptable and there is a need for human
pilot on standby to perform immediate manual control in case
of unforeseen events or flight deviations from the safety
protocol. Practical applications require autonomous systems
to adapt to the changing environment with minimal to no-
human intervention. The key difference between an
autonomous UAV operating alone versus operating in a
swarm is that there is a possibility for UAV swarms to
collaboratively perform tasks by sharing information and data
through inter-UAV communications. The ability to
communicate allows UAVs operating in a swarm to share
data and gain access to processed information collected by
other UAVs.
In this work we explore autonomous navigation strategies
that require minimal swarm communications to achieve a
given task. Our aim is to achieve autonomous navigation and Fig 1 Illustration of a path planning problem [4]
control with minimal data transfer between agents in the
swarm. This will facilitate experimentation with sensors, and
computational hardware currently available. Unlike

978-1-6654-9235-5/22/$31.00 ©2022 IEEE 12

Authorized licensed use limited to: NORTHWESTERN POLYTECHNICAL UNIVERSITY. Downloaded on April 29,2023 at 09:09:50 UTC from IEEE Xplore. Restrictions apply.
 To perform path planning, the 3D world space is
discretized into a 2D representation with 0 assigned to sea
level terrain as shown in Fig 3(b). Red circles, see Fig 3(a),
can also be implemented to show no fly zone so that the UAV
will avoid planning a path through it.
Constraints are given to the GA to allow the system to
find the best path that is also associated with the cheapest
path. In UAV path planning, there are many different
characteristics to find the optimal path. A possible cost
function created by V Roberge, M Tarbouchi, and G Labonte
[7] is illustrated below in Equation (1):

Fig 2 Voronoi graph with potential flight paths [4]

(1)
B. Genetics Algorithm Optimisation Approach where corrects any long paths, corrects path
with a high average altitude, corrects path
Genetic Algorithm (GA) methods [7] require high
through a danger zone, corrects any path requiring
computational power to calculate the path that produced the
lowest cost. GA path planning is where a UAV autonomously more power, corrects path colliding with the
computes the best path from the start to the end of the ground, correct path requiring more fuel, and
mission. This is best used in UAVs that are constantly flying corrects path with any sharp edges that requires
in different types of terrain. the UAV to make sharp turns. With all that into account the
GA will be able to produce an adequate flight path for any
scenario.

C. On-line Planning Approaches

Both Voronoi Graph and GA method are applied during
pre-flight phase where the central computer uses known
terrain information. These methods face serious limitations
when applied to path planning for swarm drones operating in
a dynamic environment.
To mitigate the limitations of centralized planning, the
control and decision making for swarm operation has to make
use of the information from the onboard sensors [5]. The
structure are defined in three stages: Data, Control, and
Process. The decisions are made either through an intelligent
or non-intelligent mode as discussed below and shown in Fig.
(a) 4.

Fig 4 Decision flowchart of an autonomous system [6]

Pre-Planned Approach
Based on Ground Control Station (GCS), the human
operator will engage the task or pre-programmed for the
UAVs. UAVs will change from one state to another state
based on the programming. These are usually executed by
conventional control algorithms via the controls and sensors.
UAVs received sequential tasks and relied on directions from
the GCS to perform the decision making. This method is
(b)
Fig 3 Navigation workspace in 3D and 2D [7]
essentially centralized type of control where the GCS

13

Authorized licensed use limited to: NORTHWESTERN POLYTECHNICAL UNIVERSITY. Downloaded on April 29,2023 at 09:09:50 UTC from IEEE Xplore. Restrictions apply.
 received information from all drones in the swarm in its
decision-making process [6].

Adhoc Approach
The UAVs has high level of decision-making skills. In
contrast to the pre-planned mode, this requires each drone to
be able to make their own decision. Based on a given situation
the drone has to perform on-board processing and decide
whether to move in a particular direction when approaching
an obstacle or another neighbouring drones. The algorithms
to be implemented on drones for distributed processing must
be computationally efficient. Examples of some algorithms
currently investigated are illustrated in Fig 5.
These include artificial intelligence/machine learning
approaches, formal logic, expert systems, and other
distributed path planning methods [5].
Fig 6 Illustration of search environment with walls and obstacles.

IV.PROBLEM FORMULATION
In this work, we shall assume each drone in the swarm is
equipped with time-of-flight (tof) sensors that allows it to
detect obstacles and other agents in its vicinity. In practice,
small drones such as Crazyflie [8] can carry 13 of these tof
sensors to provide 360 degrees coverage around the drone
[9]. The measurements from these tof sensors will allow the
drone to determine if a particular direction is passable at
each time step of the mission. Based on this assumption, we
shall discretize the direction of movement of agent to one of
8 directions where tof information is available. In addition,
Fig 5 Distributed Path Planning Algorithm [5]
each drone in the swarm could be connected to a Wi-Fi-
router where information is exchanged between the drones
The discussion of earlier work is mostly based on and a common database.
centralized planning with a shift towards distributed planning
starting with the ad hoc approach. In the following sections, The distance of obstacles detected by the tof sensors of
we propose a distributed navigation strategy for a swarm to each drone is represented by the vector " ∈ #$ . The
collaboratively plan paths to complete a mission task. We individual component of the vector " is represented by " % ∈
start by developing the notations in Section 3. This is # which is the distance to an obstacle for agent i in the
followed by our formulation of the distributed path planning direction & ∈ '1,2, . .8, of the associated tof sensor, Fig 8
problem in Section 4. A novel strategy for distributed path illustrate the 8 directions of the tof sensors. Note that in our
planning that solve the swarm navigation problem is formulation, each agent can only move in any one of the 8
developed in Section 5 together with proposed metrics to directions or remain stationary at its current location.
quantify performance. Simulations results are presented in
Section 6, we present our analysis of the results and key
findings of the proposed distributed strategy. A summary of
the key contributions is provided in Section 6.
III. NOTATIONS
The position of each agent in the swarm is defined by its
coordinates ∈ , where , ! ) represents the 2D
coordinates and represents the world environment of all
agents. In addition, we define a subset ∈ which
represents a partial knowledge of the work at each time step
of the mission. Fig 6 provides an illustration of the search
space.

Fig 7 Illustration of partial map indicating revisit rate.

14

Authorized licensed use limited to: NORTHWESTERN POLYTECHNICAL UNIVERSITY. Downloaded on April 29,2023 at 09:09:50 UTC from IEEE Xplore. Restrictions apply.
 discovered the metric will be saturated. The exit condition for
the search phase is thus developed using the detection of a
saturation in the discovery rate # , and this condition can be
defined mathematically as follows in Equation (4):

lim # 9: ; # 9 ; 1:: 0
→8

(3)
To account for noise in a practical system, we can extend
the conditions with a buffer in the actual implementation of
the exit condition. Fig. 9 provides an illustration of the
discovery rate for a swarm size of 125 agents.

Fig 8 Illustration of an 8 direction tof sensor

V. A DISTRIBUTED PATH PLANNING STRATEGY

The strategy developed can be classified according to the
different tasks to be accomplished by the swarm. For
example, in search and rescue applications, there are two
types of tasks to be accomplished in a mission. During the
initial phase, the task for each agent in the swarm is to explore
new grounds in search for victims as well as to map the search
area for rescue workers.
A. Exploration Task
The task of exploration is formulated as a cost
minimization problem at any time instant for each agent. The
cost function in this case is the revisit rate of adjacent cells
that an agent can take with the constraint that it cannot move
into a cell that is occupied. As discussed in the problem
formulation, the revisit map is developed collaboratively as Fig 9 Discovery rate with 125 agents.
the swarm explores the search area. Suppose - ∈ '1,2, . .8,
represents the best direction to be explored by an agent i, then While our proposed strategy may not always provide the
the mathematical formulation to find the lowest cost may be shortest path to the home position, it does guarantee a feasible
written as follows respectively in Equation (2) and Equation solution from any point in the cost map to the home position.
(3): The key advantage of the proposed strategy lies in its
,% efficiency and scalability to any swarm size. We shall
- min , for each agent i. compare the performance of our approach to an optimal
%∈'1,$,
method to illustrate the benefit of our collaborative strategy.
(1)
This minimization problem is subject to the following
constraint: C. Performance Metric
23 The performance metric depends on the task and the
" 40 (2)
desired outcome, and hence in this section we develop the
2
where " 3 is the distance of agent i to an obstacle in direction metric for the search phase. In the search phase the objectives
are to cover the maximum area in the shortest possible time.
- .This translates to the requirement that the selected
This can be represented as follows in Equation (5) and
direction corresponds to the direction with the least revisit
Equation (6):
rate subjected to the constraint that the distance to an obstacle
is greater than zero.
>?@A "BCDEF@?@"
B. Discovery rate and exit conditions <=
GE9AH C@A?Dℎ A?@A
In practice the exploration phase requires a condition to (4)
determine that the search task has been completed when all
the valid cells in the world environment have been <J 9B<@ 9A&@K 9E DE< H@9@ C@A?Dℎ 9AC&
discovered. We define a scalar metric # ∈ # that measures
the total number of cells discovered by the swarm. This (5)
metric is a function of time and as the search progresses, the
measure will increase. Once all the cells have been

15

Authorized licensed use limited to: NORTHWESTERN POLYTECHNICAL UNIVERSITY. Downloaded on April 29,2023 at 09:09:50 UTC from IEEE Xplore. Restrictions apply.
 The metric <= provides a measure of the percentage of TABLE I. PERFORMANCES OF PROPOSED COLLABORATION STRATEGY
search area covered and the measure < J provides a measure S/N Swarm size <= 1L: < J 1L:
of the speed. In practice, these two metrics are mutually
exclusive. When the time allocated to the search task is 1 25 98 (3.2) 971 (157)
constraint, then <= provides a good measure of the
performance. On the other hand, when there is no time 2 50 99 (2.7) 697 (99)
constraint, the search measure will always reach 100% and 3 100 99.8 (1.2) 501 (72)
the time taken will be a better performance measure.
However, when looking at different swarm sizes we could 4 150 99.8 (1.1) 433 (71)
effectively use both metrics to determine the robustness of the
5 200 99.7 (1.3) 422 (69)
developed algorithms to different swarm sizes.

VI. SIMULATION AND DISCUSSION Plots of the respective metric are shown in Fig 11 and Fig
12. We observed in Fig 11 that a smaller swarm size is
We have evaluated the performance and robustness of the subjected to a larger deviation in the search area discovered
proposed collaborative swarm navigation and control by the swarm. This indicates that for a given search area, a
algorithm for different swarm sizes. The simulation is set up small swarm may not always cover the entire search area, as
to mimic a search and rescue mission where the search area there are inconsistencies in search coverage with small
is defined by an enclosed area with partitions for rooms number of agents. As the swarm size increases, the
located within the area. Obstacles are generated and discovered area approaches 100%. Note that in our
randomly distributed within the search area to simulate the simulation the obstacles are randomly generated and there
unknown features that could be present in the search area. may be some scenarios where entry to a search area is blocked
The entry point of the swarm is also randomly selected to by obstacles and some areas are inaccessible, thus 100%
facilitate testing of the robustness of the algorithms. coverage is not attainable in practice.
Monte-Carlo runs are conducted with different swarm
sizes of 25, 50, 100, 150, and 200 agents. Obstacles and
starting location of the agents are randomly generated using
uniform distribution random number generators. The
distribution of search time for a Monte-Carlo run with 50
agents which is shown in Fig 10 below. This histogram
plot indicates a normal distribution with a mean of 697 and
standard deviation of 99.

Fig 11 Discovered Area vs swarm size

Intuitively, the time needed to complete a search should

be reduced as the swarm size is increased. This trend is also
observed in the experimental data shown in Fig 12, where we
observed that there is initially a steep drop in the search time
as swarm size is increased. However, further increase in
swarm size after 100 only produces marginal reduction in the
Fig 10 Histogram of search time for swarm size of 50 agents
search time. This observation indicates that the performance
cannot be improved indefinitely by simply increasing the
swarm size. In practice, a larger swarm size also increases the
The standard deviation of the performance metric
time needed to recover the agents at the end of each mission.
provides an indication of the repeatability of the performance
In addition, there are other issues to consider when operating
under different conditions. A smaller deviation indicates the
large swarms, and the examples of factors to consider include
consistency of the swarm performance and is therefore more
maintenance and operating costs. The simulation results
desirable. Table 1 below summarizes the mean and standard
suggest an optimal swarm size in relation to the size of the
deviations of the performance metrics using 50 Monte-Carlo
search area.
runs for swarm sizes of 25, 50, 100, 150, and 200.

16

Authorized licensed use limited to: NORTHWESTERN POLYTECHNICAL UNIVERSITY. Downloaded on April 29,2023 at 09:09:50 UTC from IEEE Xplore. Restrictions apply.
 REFERENCES

[1] L. Yang, J. Qi, J. Xiao and X. Yong, “A Literature

Review of UAV 3D Path Planning,” in 11th World
Congress on Intelligent Control and Automation,
Shenyang, China, 2014.
[2] Y. Zhao, Z. Zheng and Y. Liu, “Survey on
computational-intelligence-based UAV path planning,”
Knowledge-Based Systems, vol. 158, pp. 54-64, 2018.
[3] S. Aggarwal and N. Kumar, “Path planning techniques
for unmanned aerial vehicles: A review, solutions,”
Computer Communications, vol. 149, pp. 270-299,
2020.
[4] S. A. Bortoff, “Path Planning for UAVs,” in
Proceedings of the American Control Conference,
Chicago, IIinois, 2000.
[5] S. Plathottam and P. Ranganathan, “Next Generation
Fig 12 Time counter for search vs swarm size
Distributed and Networked Autonomous Vehicles: A
Review,” International Conference on Communication
Systems and Networks, 2018.
VII. CONCLUSION
[6] S. Lacroix, R. Alami, T. Lemaire, H. Gautier and J.
Crucial to the successful implementation of any swarm Gancet, “Decision Making in Multi-UAVs Systems:
system is the ability to provide communications between Architecture and Algorithms,” Multiple Heterogeneous
agents in a swarm to accomplish a given task. A novel Unmanned Aerial Vehicles, pp. 15-48, 2007.
approach for collaborative swarm navigation and control is
developed in this work. The outcome of our approach is a [7] V. Roberge, M. Tarbouchi and G. Labonte,
highly scalable swarm where new agents can be added, or “Comparison of Parallel Genetic Algorithm and,” IEEE
existing agents can be removed. Our proposed strategy will TRANSACTIONS ON INDUSTRIAL INFORMATICS,
allow the swarm to achieve its mission despite changes in the vol. 9, pp. 132-141, 2013.
swarm size. We have also demonstrated that our method uses [8] “Crazyflie 2.1,” Bitcraze, [Online]. Available:
minimal information transfer between agents. Our proposed https://www.bitcraze.io/products/crazyflie-2-1/.
strategy can be implemented in UAVs with very limited [Accessed 1 July 2022].
computational power because the information exchange is
low. For example, our approach can also be realized with [9] Padmal, “CrazyFlieToFDeck,” [Online]. Available:
commercially available drones that use Wi-Fi for https://github.com/CloudyPadmal/CrazyFlieToFDeck.
communications. We have shown through Monte-Carlo [Accessed 1 July 2022].
simulations on the performance and robustness of our swarm
control strategy.

17

Authorized licensed use limited to: NORTHWESTERN POLYTECHNICAL UNIVERSITY. Downloaded on April 29,2023 at 09:09:50 UTC from IEEE Xplore. Restrictions apply.