Методи роєвого штучного інтелекту в завданнях автономної навігації БПЛА
This paper presents a comparative analysis of nine swarm intelligence (SI) methods in terms of their suitability for onboard AI platforms in autonomous unmanned aerial vehicle (UAV) swarms. A set of key criteria is defined, including computational complexity, scalability, latency, robustness to agen...
Saved in:
| Date: | 2025 |
|---|---|
| Main Authors: | , , , |
| Format: | Article |
| Language: | English |
| Published: |
The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute"
2025
|
| Subjects: | |
| Online Access: | https://journal.iasa.kpi.ua/article/view/343081 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Journal Title: | System research and information technologies |
| Download file: |  |
Institution
System research and information technologies| _version_ | 1867334453875965952 |
|---|---|
| author | Zgurovsky, Michael Zaychenko, Yuriy Tytarenko, Andrii Kuzmenko, Oleksii |
| author_facet | Zgurovsky, Michael Zaychenko, Yuriy Tytarenko, Andrii Kuzmenko, Oleksii |
| author_institution_txt_mv | [
{
"author": "Michael Zgurovsky",
"institution": "Educational and Scientific Complex “Institute for Applied System Analysis” of the National Technical University of Ukraine “Igor Sikorsky Kyiv Polytechnic Institute”, Kyiv"
},
{
"author": "Yuriy Zaychenko",
"institution": "Educational and Research Institute for Applied System Analysis of the National Technical University of Ukraine “Igor Sikorsky Kyiv Polytechnic Institute”, Kyiv"
},
{
"author": "Andrii Tytarenko",
"institution": "Educational and Research Institute for Applied System Analysis of the National Technical University of Ukraine “Igor Sikorsky Kyiv Polytechnic Institute”, Kyiv"
},
{
"author": "Oleksii Kuzmenko",
"institution": "Educational and Research Institute for Applied System Analysis of the National Technical University of Ukraine “Igor Sikorsky Kyiv Polytechnic Institute”, Kyiv"
}
] |
| author_sort | Zgurovsky, Michael |
| baseUrl_str | http://journal.iasa.kpi.ua/oai |
| collection | OJS |
| datestamp_date | 2025-11-09T00:01:30Z |
| description | This paper presents a comparative analysis of nine swarm intelligence (SI) methods in terms of their suitability for onboard AI platforms in autonomous unmanned aerial vehicle (UAV) swarms. A set of key criteria is defined, including computational complexity, scalability, latency, robustness to agent loss, and adaptability. Decentralized Behavior Trees (BTs) are identified as the most balanced approach for the reactive behavior layer, while the global swarm optimization method GBestPSO proves effective for high-level planning. A hybrid two-layer cognitive architecture is proposed that integrates BTs and GBestPSO, with functional separation between layers and communication based on DDS/RTPS protocols. The architecture exhibits high autonomy, fault tolerance, modularity, and suitability for real-time embedded systems operating in dynamic or adversarial environments. The results were partially supported by the National Research Foundation of Ukraine, grant No. 2025.06/0022 “AI platform with cognitive services for coordinated autonomous navigation of distributed systems consisting of a large number of objects”. |
| doi_str_mv | 10.20535/SRIT.2308-8893.2025.3.11 |
| first_indexed | 2025-11-09T02:11:03Z |
| format | Article |
| fulltext |
M.Z. Zgurovsky, Yu.P. Zaychenko, A.M. Tytarenko, O.V. Kuzmenko, 2025
Системні дослідження та інформаційні технології, 2025, № 3 137
TIДC
МЕТОДИ, МОДЕЛІ ТА ТЕХНОЛОГІЇ ШТУЧНОГО
ІНТЕЛЕКТУ В СИСТЕМНОМУ АНАЛІЗІ
ТА УПРАВЛІННІ
UDC 004.8:681.5:629.735
DOI: 10.20535/SRIT.2308-8893.2025.3.11
METHODS OF SWARM ARTIFICIAL INTELLIGENCE
IN AUTONOMOUS NAVIGATION TASKS OF UAVS
M.Z. ZGUROVSKY, Yu.P. ZAYCHENKO, A.M. TYTARENKO, O.V. KUZMENKO
Abstract. This paper presents a comparative analysis of nine swarm intelligence
(SI) methods in terms of their suitability for onboard AI platforms in autonomous
unmanned aerial vehicle (UAV) swarms. A set of key criteria is defined, including
computational complexity, scalability, latency, robustness to agent loss, and adapta-
bility. Decentralized Behavior Trees (BTs) are identified as the most balanced ap-
proach for the reactive behavior layer, while the global swarm optimization method
GBestPSO proves effective for high-level planning. A hybrid two-layer cognitive
architecture is proposed that integrates BTs and GBestPSO, with functional separa-
tion between layers and communication based on DDS/RTPS protocols. The archi-
tecture exhibits high autonomy, fault tolerance, modularity, and suitability for real-
time embedded systems operating in dynamic or adversarial environments. The results
were partially supported by the National Research Foundation of Ukraine, grant
No. 2025.06/0022 “AI platform with cognitive services for coordinated autonomous
navigation of distributed systems consisting of a large number of objects”.
Keywords: swarm intelligence, UAV, autonomous navigation, behavior trees,
GBestPSO, ROS 2, DDS, cognitive architecture.
INTRODUCTION
In the current environment of increasing complexity and dynamism in both mili-
tary and civilian settings, autonomous swarms of unmanned aerial vehicles
(UAVs) are increasingly being seen as an effective tool for reconnaissance, patrol,
escort, and rapid response missions [1; 2]. Their advantages include high mobil-
ity, the ability to cover large areas, and the ability to operate in a decentralized
mode [3]. However, to achieve true autonomy, each agent in the swarm must have
the cognitive ability to perceive the environment, assess the situation, predict the
consequences of its actions, and interact with other agents without centralized
control [4].
Building such a system requires the deployment of an onboard AI platform
with cognitive services capable of functioning in conditions of partial or complete
loss of communication, in an informationally complex environment, or under the
influence of electronic warfare [5]. Architecturally, this platform includes several
functional components: a cognitive core, a sensory-analytical layer, a navigation
controller, a communication module, a security module, a digital twin, and recon-
M.Z. Zgurovsky, Yu.P. Zaychenko, A.M. Tytarenko, O.V. Kuzmenko
ISSN 1681–6048 System Research & Information Technologies, 2025, № 3 138
figuration services [6]. The cognitive core plays a decisive role in this architec-
ture, implementing the agent’s intellectual subjectivity. It is based on the swarm
artificial intelligence (AI) method, which forms mechanisms of perception, plan-
ning, coordination, and adaptation [7].
In this context, a key scientific and practical task arises – the choice of a
swarm AI method suitable for implementation on board a drone. Such a method
must meet strict limitations on computing resources (limited processor
performance, small memory capacity), ensure real-time operation (low decision-
making latency), support scaling to tens or hundreds of agents, be resistant to
swarm element losses, and be adaptive to dynamic environmental changes [8].
In addition, the method must be integrated into the open-source modular
platform Robot Operating System (ROS) using the Data Distribution Service
(DDS) [9; 10], and must support working with simulation digital twins, which are
critical for validating and training behavioural models in a safe virtual
environment [11].
This challenge has no trivial solution. Traditional approaches — such as
centralized planning, ant algorithms, and reinforcement learning methods —
although highly effective in certain aspects, are usually overly resource-intensive,
dependent on stable connectivity, or too inert to adapt to environmental changes
[12; 13].
Therefore, this research focuses on finding, comparing, and justifying the
most acceptable swarm AI method for embedding into an onboard AI platform
with cognitive services, capable of not only ensuring the autonomous behaviour
of an individual drone, but also implementing a holistic, adaptive, self-
reconfiguring swarm system of a new generation.
CRITERIA FOR THE CHOICE OF SWARM METHODS OF AN ON-BOARD AI
PLATFORM WITH COGNITIVE SERVICES
The successful implementation of autonomous swarm navigation for drones re-
quires careful selection of a swarm artificial intelligence method that not only
provides the necessary behavioural complexity but also meets the limitations of
the onboard AI platform. The basis for this selection is a set of criteria that con-
sider both the technical and functional requirements for an autonomous multin-
drone system. Below is a list of key criteria and their rationale:
1. Computational complexity. This criterion assesses how much the swarm
AI method loads the processor when the number of agents increases. Quadratic or
cubic complexity (e.g., )( 2nO , )( 3nO ) significantly limits scalability and per-
formance when running on embedded processors, especially in environments with
limited computing resources (e.g., STM32, Raspberry Pi CM4, Jetson Nano).
Methods with linear or logarithmic complexity ( )(nO , )log( nnO ) scale better
and are more suitable for decentralized autonomy [8; 18].
2. Memory Requirements. Methods that require large buffers, tables, and
historical data are not suitable for implementation on compact boards without ad-
ditional GPU or expanded memory. The total amount of memory required to store
internal variables, sensor maps, behavioural patterns, etc. is considered [6].
3. Latency. Latency is critical for real-time applications: the system must
respond to changes in the environment or commands with a delay of no more than
Methods of swarm artificial intelligence in autonomous navigation tasks of UAVS
Системні дослідження та інформаційні технології, 2025, № 3 139
50–70 ms. Methods with high latency do not allow the drone to manoeuvre safely
or make timely decisions in combat or rescue operations [13].
4. Scalability. This indicator determines how effectively the method main-
tains performance as the number of agents increases. For example, the Boids
method scales well to 100+ agents, while ACO scales to no more than 30. High
scalability is a prerequisite for complex scenarios with dozens or hundreds of
drones [3].
5. Robustness to Agent Loss. In military and unstable environments,
drones may be lost. The swarm AI method must automatically adapt to a decrease
in the number of swarm members. Centralized approaches or Leader-Follower
models are vulnerable to such losses, while decentralized behaviour trees (BTs)
remain functional even when individual nodes are lost [17].
6. Environmental Adaptivity. True autonomy requires the ability to re-
spond to changes in the environment: obstacles, changes in terrain, dynamic
threats. Methods with low levels of sensory integration (e.g., classical PSO) have
limited adaptability. In contrast, behavioural trees or DRL with sensory context
integration are capable of dynamically reconfiguring behaviour [14].
7. Onboard Real-Time Suitability. This criterion refers to the ability of the
algorithm to operate without the need for an external computing centre or cloud
services. The method must function autonomously on the drone’s hardware plat-
form, considering real time, energy efficiency, and resource limitations [9]. Algo-
rithms that already have known libraries under ROS 2, RTOS, or support DDS
protocols are particularly valued.
These seven criteria form the basis of a multi-criteria model for evaluating
and selecting a swarm AI method. They not only formalize the decision-making
process, but also directly influence the architecture of the AI platform, its stabil-
ity, adaptability, and viability in a real-world environment.
ANALYSIS AND SELECTION OF SWARM METHODS FOR USE IN AN AI
PLATFORM WITH COGNITIVE SERVICES
Developing an effective AI platform for an autonomous swarm of unmanned ae-
rial vehicles involves choosing a swarm method that not only meets functional
requirements (adaptation, coordination, decentralization) but also complies with
the strict limitations of the real environment: limited computing resources, the
need for real-time decision-making, loss of communication, or individual agents.
That is why it is crucial to compare existing methods in terms of computational
complexity, memory requirements, latency, scalability, resilience to agent loss,
adaptability, and suitability for onboard implementation (Table 1).
Classic swarm optimization methods, particularly Method 1 and its global
modification (Method 9), demonstrate moderate complexity ( )( 2nO ) and scale well
to 50 agents. They are suitable for search and global positioning tasks, but depend on
one or more leader agents, which is a vulnerability in real-world environments.
Method 2, on the other hand, shows better adaptation to complex environ-
ments through a pheromone amplification mechanism, but has significantly
higher complexity and latency, which makes it unsuitable for real-time tasks on
typical onboard processors.
M.Z. Zgurovsky, Yu.P. Zaychenko, A.M. Tytarenko, O.V. Kuzmenko
ISSN 1681–6048 System Research & Information Technologies, 2025, № 3 140
Both methods (1 and 2) are suitable for centralized or periodic optimization
tasks but are limited in scenarios where continuous adaptation to the external en-
vironment is required.
Methods 3 and 5 are characterized by the lowest computational complexity
( )(nO ), high scalability (100+ agents), and low latency ( 30 ms), which makes
them technically attractive. However, they exhibit limited cognitive ability: Boids
does not support goal-oriented planning, and stigmergy requires precise signal
tuning and has limited adaptation to unpredictable scenarios.
T a b l e 1 . Comparative table of swarm AI methods
N Swarm AI
method
Computa-
tional
complexity
Memory
Require
ments Latency Scalability
Robustness
to Agent
Loss
Environ-
mental
Adaptivity
Onboard
Real-Time
Suitability
1
Particle Swarm
Optimization
(PSO) [16]
Medium
(O(n²))
Medium < 50 ms
Good (up
to ~50
agents)
Medium Limited Yes
2
Ant Colony
Optimization
(ACO) [3]
High (O(n²)
log n) High 100–200
ms
Limited
(up to 30
agents)
High Good Limited
3
Boids
(Reynolds’
Rules) [21]
Low
(O(n))
Low < 20 ms High (100
+ agents) Low Low Yes
4
Consensus-
based Algo-
rithms [19]
Medium
(O(n log n))
Medium 50–100
ms
Medium
(up to 50
agents)
Medium Medium Yes
5
Stigmergy-
based Models
[20]
Low
(O(n))
Low < 30 ms High (100+
agents) High Good Yes
6
Deep Reinfor--
cement Learn-
ing (DRL) [22]
High
(O(n3))
High > 200 ms Limited (20–
30 agents) Medium High No
7
Decentralized
Behaviour
Trees (BTs) [18]
Medium
(O(n log n))
Medium 30–70 ms
High
(50–100
agents)
High High No
8
Leader-
Follower
[17]
Low
(O(n))
Low < 20
ms
Limited
(~10–20
drones in
classic im-
plementation)
Low–
medium
(loss of
leader →
critical risk)
Limited
(mainly in
modified
versions)
Yes, espe-
cially with
low data
intensity
9
Global swarm
optimization
method
(GBestPSO)
[16,23]
Medium
(O(n²))
Medium < 50
ms
Good (up to
50 agents)
Medium
(dependence
on leader)
Limited Limited
Consensus-based algorithms (Method 5) ensure consistency within the
swarm with moderate complexity and delays. However, they also do not allow
modeling complex agent behavior or changing the mission structure during its
execution. In turn, the Leader-Follower method, despite its simple implementation
and low latency, has a fatal flaw – critical dependence on the leader. If the leader is
lost, coordinated behavior is destroyed, which is unacceptable for combat or
emergency scenarios.
The deep reinforcement learning method (Method 6), on the contrary, dem-
onstrates the highest level of adaptability: models can self-learn and generalize
knowledge about new environments. However, the inference delay exceeds 200
Methods of swarm artificial intelligence in autonomous navigation tasks of UAVS
Системні дослідження та інформаційні технології, 2025, № 3 141
ms even on easy tasks, and the hardware requirements include a GPU (Jetson Xa-
vier, Orin, or Nvidia RTX), which is beyond the capabilities of typical onboard
platforms.
The most balanced method in terms of all criteria is the decentralized
behavior tree method (Method 7). It has moderate computational complexity
( )log( nnO ), operates with an average latency of 30–70 ms even on weak proces-
sors, scales up to 100 agents, demonstrates high adaptability, resilience to losses,
and is suitable for implementation in ROS 2 with DDS protocols. BTs provides
autonomous decision-making logic for each agent based on local context, sensor
information, and a predefined behavior structure. In addition, it supports tree
reconfiguration during a mission, which is especially important for tasks with
variable goal or role structures in a swarm system.
Therefore, although Method 7 is the best choice for building the basic cogni-
tive architecture of the agent, it is advisable to use a hybrid approach in which the
global swarm optimization method (Method 9) plays a supporting role in global
or semi-global optimization tasks. This division of functions allows BTs to be
responsible for reactive behavior, real-time decision-making, and fault tolerance,
while GBestPSO is responsible for strategic tasks, such as finding optimal agent
locations, distributing subtasks, or optimizing configuration at the beginning of a
mission.
This combination allows for the implementation of a two-level cognitive ar-
chitecture: behavioral level – BTs, planning level – GBestPSO. Interaction be-
tween these levels can be carried out via DDS messages, which will ensure de-
terministic exchange between autonomous agents without loss of synchronization.
Figure 1 illustrates the two-level cognitive architecture of an autonomous
swarm system that combines two complementary approaches: behavior trees
(BTs) at the behavioral level and global swarm optimization GBestPSO at the
planning level. The architecture is designed to
ensure autonomous, adaptive, and fault-
tolerant behavior of the UAV swarm in dy-
namic environments and in the event of com-
munication loss.
The lower block of the figure reflects the
logic of the behavioral level, constructed in
the form of a decision tree. The ROOT node
initiates the execution of the tree,
CONDITION defines the situational condi-
tions (for example, the presence of an obstacle
or loss of communication), and ACTION
represents the drone’s reactive actions in re-
sponse to these conditions. This structure al-
lows each agent to make decisions autono-
mously based on the local context. This level
is implemented locally on board the drone and
provides the ability to react immediately, dy-
namically reconfigure the tree, and operate in
real time.
The upper block represents the planning
level — global swarm optimization using the GBestPSO method, in which agents
Fig. 1. Two-level cognitive architec-
ture: planning level — GBestPSO,
behavioral level — BTs
M.Z. Zgurovsky, Yu.P. Zaychenko, A.M. Tytarenko, O.V. Kuzmenko
ISSN 1681–6048 System Research & Information Technologies, 2025, № 3 142
collectively search for optimal locations or configurations. Here, GLOBAL BEST
means the best position found by all agents, which is transmitted to others as a
strategic reference point. This benchmark is transmitted to the behavioral level in
the form of mission conditions via DDS/RTPS protocols. In case of communica-
tion loss, the planning level can be disabled, and the behavioral level will con-
tinue to operate using the last received gBest.
The communication between levels is provided through DDS messages,
which guarantees reliable, deterministic communication without centralized con-
trol. This integration allows for a flexible balance between the local autonomy of
each drone and the global coordination of the entire swarm, ensuring high system
efficiency in complex scenarios, including military or rescue operations.
Therefore, the proposed hybrid approach combines the modularity, adapta-
bility, and speed of BTs with the global optimization capabilities of GBestPSO,
creating a flexible, scalable, and technically feasible next-generation AI platform
for UAV swarms.
THE BASIC ARCHITECTURE OF THE HYBRID APPLICATION
OF METHODS BTS + GBESTPSO
To respond to the challenges of autonomous swarm navigation in conditions of
limited resources and a highly dynamic environment, a hybrid two-level architec-
ture has been proposed that combines the declarative-reactive approach of behav-
ioral trees (BTs) with the evolutionary-optimization strategy of the global particle
swarm (GBestPSO). Such integration allows for the simultaneous achievement of
real-time performance, adaptability, scalability, and initial strategic coordination
of swarm behavior.
The structure of hybrid architecture is divided into two functional levels:
Planning level (GBestPSO-Level): implements global or semi-global
optimization at the beginning of the mission or periodically in task replanning
mode. It determines the selection of scenarios for current tasks, the strategic
positions of agents, task clustering, and the distribution of roles or target points.
Behavioral level (BT-Level): provides reactive decision-making by each
agent in real time [18], including analysis of the local context, avoidance of ob-
stacles, adaptation to role changes, loss of communication, or loss of an agent.
Communication between levels is carried out via DDS/RTPS protocols [9] in
the form of semantically described messages (Table 2).
Data exchange between subsystems of two levels is organized via
DDS/RTPS protocols [9], which allow semantically structured messages to be
transmitted between agents. For example, the Swarm Optimizer module, which
operates at the GBestPSO level, periodically calculates the globally best swarm
location strategy (gBest) and transmits it via DDS as an optimized configuration
or task. At the behavioral level, such information is interpreted as external mis-
sion conditions, which are automatically reflected in the updated behavioral logic
through the Condition → Action tree branches. Importantly, even if communica-
tion with the planning level is lost, the behavioral level continues to autono-
mously execute the mission based on the last received strategic guideline. Mutual
adaptation is ensured by QoS DDS channels, which provide priority delivery of
important messages, such as changes in behavior tree statuses, sensor events, or
strategic instructions. The above-mentioned relationships between modules can be
represented by the following list:
Methods of swarm artificial intelligence in autonomous navigation tasks of UAVS
Системні дослідження та інформаційні технології, 2025, № 3 143
T a b l e 2 . Main modules and functions of the basic hybrid architecture 7. BTs + 9.
GBestPSO
Level Module Functions
Behavior
Manager
Interprets behavior of trees in real time, activates local
actions, reconfigures the tree in response to external changes
Perception
Module
Processing sensor data, recognizing situations,
building a local map
Communication
Module
DDS/RTPS exchange of states, positions, parts
of the BT tree
Local Planner Low-level planning (SLAM, obstacle avoidance, actions)
BT level
Adaptation
Module
Dynamic tree reconfiguration in case of environmental changes
or losses
Swarm
Optimizer
Implements a global PSO algorithm; forms
a gBest position that sets a strategic landmark
Global Task
Allocator
Distribution of subtasks, goals, or roles among agents
based on the gBest position
GBestPSO
level
Mission Planner
(optional) Scenario planning for complex missions or dynamic restarts
Both
levels
Telemetry and
Diagnostics
Status exchange, logging, communication
with ground station
The Swarm Optimizer module (at the GBestPSO level) periodically cal-
culates the global best strategy (gBest) and transmits it via DDS in the form of
tasks or configurations.
The behavioral level (BT-Level) perceives these tasks as “external
mission conditions” that are reflected in the corresponding branches of the
behavior tree.
Change in gBest → change in agent behavior through the reaction of
Condition → Action nodes.
Reaction to loss of communication: BT-Level works autonomously, with
the last accepted gBest, allowing the agent to continue the mission without
external optimization.
Mutual adaptation is provided through QoS DDS channels, which priori-
tize behavioral commands, sensor events, tree statuses, and strategic instructions.
The software and hardware implementation of this architecture is based on
ROS 2 (Foxy or Humble versions) and RTOS for real-time systems such as
STM32H7. The communication layer is implemented using Fast DDS [9] or Cy-
clone DDS with QoS support. At the library level, BehaviorTree.CPP is used for
BT modules and our own implementation of the GBestPSO algorithm based on
the classical approach [14]. Jetson Orin Nano, Xavier NX, Raspberry Pi CM4 with
Coral TPU, as well as lightweight STM32H7 controllers are used as hardware
platforms. The main components of the software and hardware architecture are:
Operating system: ROS 2 (Foxy/Humble), RTOS (on STM32H7).
Communication: Fast DDS / Cyclone DDS (with QoS support).
Libraries: BehaviorTree.CPP (BT-level), own implementation of GBestPSO
(based on [14]).
Platforms: Jetson Orin Nano / Xavier NX, Raspberry Pi CM4 with Coral
TPU, STM32H7.
Among the advantages of the proposed architecture, its autonomy, flexibil-
ity, and fault tolerance should be noted. The behavioral level ensures local adapt-
ability and the ability to respond to changes in the environment without the need
for global control. The planning level, in turn, ensures strategic coordination be-
M.Z. Zgurovsky, Yu.P. Zaychenko, A.M. Tytarenko, O.V. Kuzmenko
ISSN 1681–6048 System Research & Information Technologies, 2025, № 3 144
tween agents, forming a single target configuration of the system. The use of DDS
protocols eliminates dependence on a central node, which significantly increases
fault tolerance. In the event of a loss of the global reference point (gBest), the sys-
tem automatically switches to fully local mode without losing functionality. The
modular structure of this architecture ensures its scalability and adaptability to
different mission classes: from reconnaissance and escort to combat or search and
rescue operations. Thus, the main advantages of hybrid architecture are:
BT-Level provides local adaptability and autonomy.
GBestPSO-Level adds strategic coordination and global coordination.
Interconnection via DDS eliminates dependence on a central node.
High fault tolerance: if gBest is lost, the architecture switches to fully lo-
cal mode.
Modularity support allows architecture to be scaled for different mission
classes (reconnaissance, escort, attack).
Thus, the proposed hybrid architecture, combining behavioral trees (BTs)
and global particle swarm optimization (GBestPSO), demonstrates an effective
combination of local autonomy and strategic coordination in swarm navigation
tasks. Its two-level structure allows it to simultaneously achieve adaptability, re-
sponsiveness, and energy efficiency while maintaining a high degree of resilience
to failures and communication losses. The behavioral level provides an immediate
response to environmental dynamics, while the planning level defines global
landmarks and roles, which increases the integrity of swarm behavior in complex
conditions. The use of DDS/RTPS protocols allows for reliable, distributed com-
munication between agents without centralized dependency. Such architecture not
only has high application potential in the military and civilian spheres but also
provides a scalable foundation for the further evolution of swarm intelligence sys-
tems towards the self-learning and self-coordinated platforms of the future.
ALGORITHM
The movement of each drone is described by two main phases:
1. Phase 1: Approach to the Target. Drones move toward target *Y ac-
cording to the velocity equation with the local best position. The influence of
cognitive ( 1C ) and social ( 2C ) components depends on the distance to the leader
( LY ) and target, respectively:
)]()([)()))(), ((()()1( 11 tXtYtrtXtYdCtvtv ijLiLijij j
,)]()([)()))() ((( *
2
*
2 tXtYtrtXtYdC ijji (1)
where coefficients 1C and 2C are linearly dependent on distance, which allows
adjusting the influence of the leader and the target on the behavior of the swarm:
min1
min1max1
1 ))()((
)0()0(
)))(), ((( CtXtY
XY
CC
tXtYdC ijLj
iL
iL
; (2)
min2
*
*
min2max2*
2 ))()((
)0()0(
)))(), ((( CtXtY
XY
CC
tXtYdC ijj
i
i
. (3)
2. Phase 2: Firing Ring for Killing. After reaching the target’s circumfer-
ence with a radius of 0r , the drones switch to polar coordinates and begin to rotate
Methods of swarm artificial intelligence in autonomous navigation tasks of UAVS
Системні дослідження та інформаційні технології, 2025, № 3 145
around the target. The coordinates of the leader LY are described using the radius
0r and angle )(tL :
.))((sin)(
,))((cos)(
02
01
trty
trty
L
L
L
L
(4)
The angular position is updated using the following formula:
tttt LLL )()()1( , (5)
where )(tL — angular velocity of rotation.
Each drone is controlled by its own BT, which cyclically processes the tick.
The tree consists of control nodes (Selector, Sequence) and execution nodes
(Condition, Action). Behavior Tree diagram:
Root: initiates tree execution.
Sequence: main sequence of actions for each drone.
Parallel: used for simultaneous execution of tasks independent of the
main movement cycle, such as modeling the impact of electronic warfare and
drone loss.
o Action: Model_Jamming() — simulates the probability of communi-
cation loss ( 10P ).
Action: Model_Drone_Loss() — simulates the probability of drone loss ( 2P ).
Selector (Main mission logic): Checks the mission status and switches
between phases.
o Branch 1: Approach to the Target.
Condition: Is_Far_From_Target( 0r ) – checks whether the dis-
tance to the target is 0r .
Action: Run_LBestCombinedPSO_Approach_Phase() – calcu-
lates the new speed and position.
Action: Elect_New_Leader() – this node is activated in case of
Failure of the Is_Leader_Connected() node, which implements the self-
organization mechanism.
o Branch 2: Firing Ring for Killing.
Condition: Is_Close_To_Target( 0r ) – checks whether the dis-
tance to the target is 0r .
Sequence: Firing_Ring_Sequence.
Action: Convert_To_Polar() – transition to polar coordinates.
Decorator: Sync_Attack() – uses the Decorator
Decorator: Sync_Attack() – uses the Decorator mechanism for
synchronization. The node allows the execution of the child node only after
*_
Y
Magentsready .
Action: Attack_Target() – attack the target.
A key element of the system is its ability to be adaptive:
Loss of leader. If the leader drone loses connection with other agents
(simulated by the Model_Jamming node), this initiates Failure in the leader ver-
ification conditions. The Selector in BT automatically switches to the branch that
triggers self-organization. A new leader is selected based on the criterion
)()( tXtY iLnew , where )), ((min)), (( ** YtXdYtXd i
i
i .
M.Z. Zgurovsky, Yu.P. Zaychenko, A.M. Tytarenko, O.V. Kuzmenko
ISSN 1681–6048 System Research & Information Technologies, 2025, № 3 146
Synchronization. The Decorator node ensures that the ring firing phase
does ot start until enough drones )( YM reach the target area, which is critical for
coordinating the attack.
Inertial component. In the equation, the velocity )(tvij acts as an inertial com-
ponent, preventing sudden changes in trajectory and ensuring smooth movement.
EXAMPLE
Let us consider a practical simulation experiment using a two-level cognitive ar-
chitecture consisting of a group of drones and integrating behavioral trees (BTs)
and global optimization using the GBestPSO method. According to this
experiment:
The behavioral level (BT-level) functioned locally on board each drone
and provided reactive adaptation to sensory events such as obstacle detection, role
changes, or loss of communication. This level was implemented using the
BehaviorTree.CPP library with a built-in dynamic tree reconfiguration
mechanism. Condition → Action nodes allowed for the rapid transformation of
behavioral logic in response to changes in external mission conditions received
from the planning level.
The planning level based on GBestPSO was responsible for forming the
strategic configuration of the swarm through global or local optimization. It
periodically calculated the globally best position or scenario (gBest), which was
broadcast as strategic guidelines to each agent. Data transfer between levels was
carried out using DDS/RTPS protocols, which guaranteed quality of service (QoS),
buffering of critical messages, and resistance to temporary communication losses.
The simulation experiment was conducted in a configuration of 5–10
drones equipped with Jetson Orin Nano or Raspberry Pi CM4 paired with
STM32H7 (Table 3).
T a b l e 3 . Quantitative characteristics of the simulation experiment
No Parameter Value
1 Number of drones
in the swarm group
5–10 (Jetson Orin Nano or Raspberry Pi CM4
+ STM32H7)
2 Coverage area 100 × 100 m
3 Number of control points 6
4 Mission type Search and patrol with reconfiguration elements
5 BTs algorithm 6–10 conditional nodes, 3–4 adaptive branches
6 GBestPSO algorithm Classic method [23]
7 Number of PSO iterations before start 30
8 BT module response delay ~45 ms (on Jetson Nano)
9 gBest transmission delay via DDS ~10–15 ms
10 Tree adaptation time when changing gBest < 150 ms
11 Simulated communication losses up to 30% of DDS packets
12 Mission success rate > 95% (all drones completed the route
or selected fallback actions)
The mission covered an area of 100 × 100 m, contained six control points,
and was of a search-and-patrol nature with elements of reconfiguration. The BTs
algorithm included 6 to 10 conditional nodes and 3–4 adaptive branches responsi-
ble for changing behavior in conditions of uncertainty. The GBestPSO algorithm
was implemented in its classical form [23] with typical parameters: space dimen-
Methods of swarm artificial intelligence in autonomous navigation tasks of UAVS
Системні дослідження та інформаційні технології, 2025, № 3 147
sion 2d , inertial weight 8.0 , attraction coefficients 497.121 CC . Be-
fore the start of the mission, 30 PSO iterations were performed to find the initial
strategic configuration. The BT module’s response delay to an event was about
45 ms, gBest transmission via DDS was 10–15 ms, and behavior tree adaptation
when gBest changed took up to 150 ms. Even with a simulated loss of up to 30%
of DDS messages, the mission success rate exceeded95%.
In critical situations, the system demonstrated high adaptability. In the event
of a single drone failure, other agents automatically restructured the behavior tree
to compensate for the loss. If the planning level lost communication, the BT mod-
ules continued to operate autonomously based on the last strategic reference
point. When obstacles were detected, the SLAM module formed a new route, and
the corresponding branch of the behavior tree activated avoidance actions. The
total mission time exceeded the baseline scenario by no more than 12%. The av-
erage delay between events and response was about 90 ms. The positioning error
in areas without GNSS remained within 2.8 m, and swarm synchronization was
maintained even with the loss of up to 40% of DDS messages.
The diagram (Fig. 2) shows a typical scenario for the implementation of a
UAV swarm mission for the proposed architecture with hybrid integration of BTs
and GBestPSO. This implementation is viable and technically effective for com-
plex search and patrol missions in conditions of limited communication and a dy-
namic environment.
Fig. 2. Scenario for implementing the mission of a swarm of UAVs for the hybrid
architecture of BTs and GBestPSO
Y
, m
X, m
1
2
2
3
3
1 —
2 —
3 —
M.Z. Zgurovsky, Yu.P. Zaychenko, A.M. Tytarenko, O.V. Kuzmenko
ISSN 1681–6048 System Research & Information Technologies, 2025, № 3 148
The following images are used in Fig. 2: blue dots (WP1–WP6) – mission
route control points; red circle – electronic warfare (EW) zone; gray squares —
obstacles on the route (buildings, infrastructure objects); lines with markers —
trajectories of three drones moving autonomously using behavioral trees (BTs), re-
sponding to obstacles and conditions broadcast from the global level of GBestPSO.
This example demonstrates the realistic implementation of a two-level cog-
nitive architecture for swarm drones in a simulation environment with digital twin
elements under threatening conditions and limited connectivity. It confirms the
feasibility of the BTs + GBestPSO hybrid approach in complex, dynamic, or hos-
tile conditions.
CONCLUSIONS
1. The article analyzes modern swarm artificial intelligence methods from
the perspective of their suitability for implementation as part of an onboard AI
platform for autonomous unmanned aerial vehicles. Considering the requirements
for limited computing resources, real-time operation, resistance to agent loss, and
adaptability, a list of key criteria for selecting a swarm AI method was formulated.
2. A comparative analysis of nine leading methods allowed us to identify
decentralized behavioral trees (BTs) as the most balanced approach for the basic
cognitive architecture of an agent. BTs combine low latency, resilience to losses,
adaptability to the environment, and suitability for onboard implementation in
ROS 2. At the same time, the GBestPSO method, as a classic global swarm
optimization tool, proved to be useful for performing strategic tasks at the higher
level of the system, for the initial configuration of the swarm, task distribution,
and search for optimal agent locations.
3. The hybrid architecture proposed in the article, which combines BTs and
GBestPSO into a two-level structure, allows for a compromise between local
responsiveness and global coordination. This architecture is characterized by high
autonomy, fault tolerance, scalability, and real-time adaptability. It provides a
flexible division of responsibilities: BTs for immediate behavioral response,
GBestPSO for planning optimization. The communication foundation,
implemented through DDS/RTPS, eliminates dependence on central control and
guarantees deterministic data exchange between agents.
4. A practical simulation experiment using digital twins and the BTs +
GBestPSO hybrid architecture demonstrated its effectiveness in complex dynamic
conditions. Even with the loss of up to 30% of DDS network messages, the
system maintained coordinated swarm behavior, and the response time between a
sensor event and tree restructuring averaged ~90 ms. Adaptation to new obstacles,
role changes, resistance to agent failures, and the ability to operate without a global
reference point demonstrated the high viability of the architecture. Mission success
exceeded 95%, confirming the practical feasibility of the proposed approach.
5. Thus, the results of the study not only confirm the theoretical validity of
hybrid architecture but also demonstrate its real implementation in conditions
close to combat or emergency rescue situations. The proposed system can serve as
a basis for building a new generation of swarm AI platforms with the potential for
further evolution towards self-learning, self-coordinated, and safe-to-use
autonomous systems in real environments.
Methods of swarm artificial intelligence in autonomous navigation tasks of UAVS
Системні дослідження та інформаційні технології, 2025, № 3 149
REFERENCES
1. D. Floreano, R.J. Wood, “Science, technology and the future of small autonomous
drones,” Nature, 521(7553), pp. 460–466, 2015. doi: https://doi.org/10.1038/nature14542
2. M.J. Krieger, J.-B. Billeter, “The call of duty: Self-organised task allocation in a
population of up to twelve mobile robots,” Robotics and Autonomous Systems, 30(1–
2), pp. 65–84, 2000. doi: 10.1016/S0921-8890(99)00065-2
3. M. Brambilla, E. Ferrante, M. Birattari, M. Dorigo, “Swarm robotics: a review from
the swarm engineering perspective,” Swarm Intelligence, 7, pp. 1–41, 2013. doi:
10.1007/s11721-012-0075-2
4. G.-Z. Yang et al., “The grand challenges of Science Robotics,” Science Robotics,
3(14), eaar7650, 2018. doi: 10.1126/scirobotics.aar7650
5. Y. Mekdad, A. Arış, L. Babun, A. El Fergougui, M. Conti, R. Lazzeretti, S. Uluagac,
“A survey on security and privacy issues of UAVs,” Computer Networks, vol. 224,
109626, 2023. doi: 10.1016/j.comnet.2023.109626
6. B. Barricelli, E. Casiraghi, D. Fogli, “A Survey on Digital Twin: Definitions, Char-
acteristics, Applications, and Design Implications,” IEEE Access, pp. 167653–
167671, 2019. doi: 10.1109/ACCESS.2019.2953499
7. V. Trianni, A. Campo, Fundamental Collective Behaviors in Swarm Robotics,
pp. 1377–1394, 2015. doi: 10.1007/978-3-662-43505-2_71
8. H. Hamann, Swarm Robotics: A Formal Approach. Springer, Cham, 2018. doi:
https://doi.org/10.1007/978-3-319-74528-2
9. Y. Maruyama, S. Kato, T. Azumi, Exploring the performance of ROS2, pp. 1–10,
2016. doi: 10.1145/2968478.2968502
10. “Data Distribution Service (DDS) Specification. Version 1.4.”, Object Management
Group (OMG). 2015. Available: https://www.omg.org/spec/DDS
11. S. Boschert, R. Rosen, “Digital twin—The simulation aspect,” in Mechatronic Fu-
tures, pp. 59–74. Springer, 2016. doi: 10.1007/978-3-319-32156-1_5
12. M. Dorigo, M. Birattari, M. Brambilla, “Swarm robotics and artificial swarm intelli-
gence, in Handbook of Artificial Intelligence, 2014.
13. M. Hüttenrauch, A. Šošić, G. Neumann, “Deep Reinforcement Learning for Swarm
Systems,” Journal of Machine Learning Research, vol. 20, pp. 1–31, 2019.
14. M. Colledanchise, P. Ögren, Behavior Trees in Robotics and AI: An Introduction.
CRC Press, 2018. doi: https://doi.org/10.1201/9780429489105
15. M. Dorigo, V. Maniezzo, A. Colorni, “Ant system: Optimization by a colony of co-
operating agents,” IEEE Transactions on Evolutionary Computation, 1(1), pp. 53–66,
1996. https://doi.org/10.1109/4235.585892
16. R.C. Eberhart, J. Kennedy, “A new optimizer using particle swarm theory,” Pro-
ceedings of the Sixth International Symposium on Micro Machine and Human Sci-
ence, pp. 39–43. IEEE, 1995. doi: https://doi.org/10.1109/MHS.1995.494215
17. W. Li, H. Dong, H. Gao, “Leader-follower formation control of multi-UAVs with
collision avoidance,” IEEE Transactions on Industrial Electronics, 64(4),
pp. 3184–3196, 2017. doi: https://doi.org/10.1109/TIE.2016.2638818
18. A. Marzinotto, M. Colledanchise, C. Smith, P. Ögren, “Towards a unified behav-
ior trees framework for robot control,” 2014 IEEE International Conference on Ro-
botics and Automation (ICRA), pp. 5420–5427. doi: https://doi.org/10.1109/
ICRA.2014.6907656
19. R. Olfati-Saber, J.A. Fax, R.M. Murray, “Consensus and cooperation in networked
multi-agent systems,” Proceedings of the IEEE, 95(1), pp. 215–233, 2007. doi:
https://doi.org/10.1109/JPROC.2006.887293
20. V. Parunak, M. Purcell, R. O’Connell, “Digital Pheromones for Autonomous Coor-
dination of Swarming UAV’s,” in 1st UAV Conference, 2002. doi: 10.2514/6.2002-
3446
M.Z. Zgurovsky, Yu.P. Zaychenko, A.M. Tytarenko, O.V. Kuzmenko
ISSN 1681–6048 System Research & Information Technologies, 2025, № 3 150
21. C.W. Reynolds, “Flocks, herds and schools: A distributed behavioral model,” Pro-
ceedings of the 14th Annual Conference on Computer Graphics and Interactive
Techniques (SIGGRAPH ‘87), pp. 25–34. doi: https://doi.org/10.1145/37401.37406
22. O. Vinyals et al., “Grandmaster level in StarCraft II using multi-agent reinforcement
learning,” Nature, 575(7782), pp. 350–354, 2019. doi: https://doi.org/10.1038/
s41586-019-1724-z
23. M. Zgurovsky, Y. Zaychenko, The Fundamentals of Computational Intelligence:
System Approach. Springer, 2016. doi: https://doi.org/10.1007/978-3-319-35162-9
Received 24.08.2025
INFORMATION ON THE ARTICLE
Michael Z. Zgurovsky, ORCID: 0000-0001-5896-7466, Educational and Scientific Complex
“Institute for Applied System Analysis” of the National Technical University of Ukraine
“Igor Sikorsky Kyiv Polytechnic Institute”, Ukraine, e-mail: zgurovsm@hotmail.com
Yuriy P. Zaychenko, ORCID: 0000-0001-9662-3269, Educational and Research Institute
for Applied System Analysis of the National Technical University of Ukraine “Igor Sikor-
sky Kyiv Polytechnic Institute”, Ukraine, e-mail: zaychenkoyuri@ukr.net
Andrii M. Tytarenko, ORCID: 0000-0002-8265-642X, Educational and Research Insti-
tute for Applied System Analysis of the National Technical University of Ukraine “Igor
Sikorsky Kyiv Polytechnic Institute”, Ukraine, e-mail: titarenkoan@gmail.com
Oleksii V. Kuzmenko, ORCID: 0000-0003-1581-6224, Educational and Research Insti-
tute for Applied System Analysis of the National Technical University of Ukraine “Igor
Sikorsky Kyiv Polytechnic Institute”, Ukraine, e-mail: oleksii.kuzmenko@ukr.net
МЕТОДИ РОЄВОГО ШТУЧНОГО ІНТЕЛЕКТУ В ЗАВДАННЯХ АВТОНОМНОЇ
НАВІГАЦІЇ БПЛА / М.З. Згуровський, Ю.П. Зайченко, А.М. Титаренко, О.В. Кузь-
менко
Анотація. Подано порівняльний аналіз дев’яти методів ройового інтелекту
(РІ) з точки зору їхньої придатності для бортових платформ ШІ в автономних
роях безпілотних літальних апаратів (БПЛА). Визначено набір ключових кри-
теріїв, включаючи обчислювальну складність, масштабованість, затримку,
стійкість до втрати агентів та адаптивність. Децентралізовані дерева поведінки
(ДП) визначені як найбільш збалансований підхід для реактивного рівня пове-
дінки, тоді як глобальний метод оптимізації рою GBestPSO виявляється ефек-
тивним для високорівневого планування. Запропоновано гібридну двошарову
когнітивну архітектуру, яка інтегрує ДП та GBestPSO, із функціональним роз-
діленням між шарами та зв’язком на основі протоколів DDS/RTPS. Архітекту-
ра демонструє високу автономність, відмовостійкість, модульність та придат-
ність для вбудованих систем реального часу, що працюють у динамічних або
змагальних середовищах. Результати частково підтримано Національним фон-
дом досліджень України, грант № 2025.06/0022 «Платформа штучного інтеле-
кту з когнітивними сервісами для скоординованої автономної навігації розпо-
ділених систем, що складаються з великої кількості об’єктів».
Ключові слова: ройовий інтелект, БПЛА, автономна навігація, дерева поведі-
нки, GBestPSO, ROS 2, DDS, когнітивна архітектура.
|
| id | journaliasakpiua-article-343081 |
| institution | System research and information technologies |
| keywords_txt_mv | keywords |
| language | English |
| last_indexed | 2025-11-09T02:11:03Z |
| publishDate | 2025 |
| publisher | The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute" |
| record_format | ojs |
| resource_txt_mv | journaliasakpiua/f2/a3801f1cae1471bc0981e22603d30bf2.pdf |
| spelling | journaliasakpiua-article-3430812025-11-09T00:01:30Z Methods of swarm artificial intelligence in autonomous navigation tasks of UAVs Методи роєвого штучного інтелекту в завданнях автономної навігації БПЛА Zgurovsky, Michael Zaychenko, Yuriy Tytarenko, Andrii Kuzmenko, Oleksii swarm intelligence unmanned aerial vehicles (UAVs) Autonomous Navigation behavior trees (BT) GBestPSO ROS 2 DDS cognitive architecture ройовий інтелект БПЛА автономна навігація дерева поведінки GBestPSO ROS 2 DDS когнітивна архітектура This paper presents a comparative analysis of nine swarm intelligence (SI) methods in terms of their suitability for onboard AI platforms in autonomous unmanned aerial vehicle (UAV) swarms. A set of key criteria is defined, including computational complexity, scalability, latency, robustness to agent loss, and adaptability. Decentralized Behavior Trees (BTs) are identified as the most balanced approach for the reactive behavior layer, while the global swarm optimization method GBestPSO proves effective for high-level planning. A hybrid two-layer cognitive architecture is proposed that integrates BTs and GBestPSO, with functional separation between layers and communication based on DDS/RTPS protocols. The architecture exhibits high autonomy, fault tolerance, modularity, and suitability for real-time embedded systems operating in dynamic or adversarial environments. The results were partially supported by the National Research Foundation of Ukraine, grant No. 2025.06/0022 “AI platform with cognitive services for coordinated autonomous navigation of distributed systems consisting of a large number of objects”. Подано порівняльний аналіз дев’яти методів ройового інтелекту (РІ) з точки зору їхньої придатності для бортових платформ ШІ в автономних роях безпілотних літальних апаратів (БПЛА). Визначено набір ключових критеріїв, включаючи обчислювальну складність, масштабованість, затримку, стійкість до втрати агентів та адаптивність. Децентралізовані дерева поведінки (ДП) визначені як найбільш збалансований підхід для реактивного рівня поведінки, тоді як глобальний метод оптимізації рою GBestPSO виявляється ефективним для високорівневого планування. Запропоновано гібридну двошарову когнітивну архітектуру, яка інтегрує ДП та GBestPSO, із функціональним розділенням між шарами та зв’язком на основі протоколів DDS/RTPS. Архітектура демонструє високу автономність, відмовостійкість, модульність та придатність для вбудованих систем реального часу, що працюють у динамічних або змагальних середовищах. Результати частково підтримано Національним фондом досліджень України, грант № 2025.06/0022 "Платформа штучного інтелекту з когнітивними сервісами для скоординованої автономної навігації розподілених систем, що складаються з великої кількості об’єктів". The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute" 2025-09-29 Article Article Peer-reviewed Article application/pdf https://journal.iasa.kpi.ua/article/view/343081 10.20535/SRIT.2308-8893.2025.3.11 System research and information technologies; No. 3 (2025); 137-150 Системные исследования и информационные технологии; № 3 (2025); 137-150 Системні дослідження та інформаційні технології; № 3 (2025); 137-150 2308-8893 1681-6048 en https://journal.iasa.kpi.ua/article/view/343081/331018 |
| spellingShingle | ройовий інтелект БПЛА автономна навігація дерева поведінки GBestPSO ROS 2 DDS когнітивна архітектура Zgurovsky, Michael Zaychenko, Yuriy Tytarenko, Andrii Kuzmenko, Oleksii Методи роєвого штучного інтелекту в завданнях автономної навігації БПЛА |
| title | Методи роєвого штучного інтелекту в завданнях автономної навігації БПЛА |
| title_alt | Methods of swarm artificial intelligence in autonomous navigation tasks of UAVs |
| title_full | Методи роєвого штучного інтелекту в завданнях автономної навігації БПЛА |
| title_fullStr | Методи роєвого штучного інтелекту в завданнях автономної навігації БПЛА |
| title_full_unstemmed | Методи роєвого штучного інтелекту в завданнях автономної навігації БПЛА |
| title_short | Методи роєвого штучного інтелекту в завданнях автономної навігації БПЛА |
| title_sort | методи роєвого штучного інтелекту в завданнях автономної навігації бпла |
| topic | ройовий інтелект БПЛА автономна навігація дерева поведінки GBestPSO ROS 2 DDS когнітивна архітектура |
| topic_facet | swarm intelligence unmanned aerial vehicles (UAVs) Autonomous Navigation behavior trees (BT) GBestPSO ROS 2 DDS cognitive architecture ройовий інтелект БПЛА автономна навігація дерева поведінки GBestPSO ROS 2 DDS когнітивна архітектура |
| url | https://journal.iasa.kpi.ua/article/view/343081 |
| work_keys_str_mv | AT zgurovskymichael methodsofswarmartificialintelligenceinautonomousnavigationtasksofuavs AT zaychenkoyuriy methodsofswarmartificialintelligenceinautonomousnavigationtasksofuavs AT tytarenkoandrii methodsofswarmartificialintelligenceinautonomousnavigationtasksofuavs AT kuzmenkooleksii methodsofswarmartificialintelligenceinautonomousnavigationtasksofuavs AT zgurovskymichael metodiroêvogoštučnogoíntelektuvzavdannâhavtonomnoínavígacííbpla AT zaychenkoyuriy metodiroêvogoštučnogoíntelektuvzavdannâhavtonomnoínavígacííbpla AT tytarenkoandrii metodiroêvogoštučnogoíntelektuvzavdannâhavtonomnoínavígacííbpla AT kuzmenkooleksii metodiroêvogoštučnogoíntelektuvzavdannâhavtonomnoínavígacííbpla |