This paper presents a novel concept of a collaborative
multifunctional agro-robotic system for intensive farming,
designed to address key challenges in modern agriculture,
including labor shortages, increasing production costs, and
the need for sustainable and efficient crop management. The
proposed system is based on a multi-agent architecture in
which multiple mobile agro-robots operate cooperatively
within a shared field environment. Each robot is capable to
operate independently or cooperatively as part of the
multi/robots team. Robot is equipped with a modular
structure, including an omnidirectional mobile platform, a
Cartesian manipulator for ground-level operations, and
advanced perception systems for autonomous harvesting. A
key feature of the system is its collaborative capability,
enabling both robot-to-robot and human–robot interaction.
Robots communicate wirelessly with each other and a central
base station, allowing dynamic task allocation, path
optimization, and real-time coordination. In the event of a
robot failure, neighboring units autonomously redistribute
tasks, ensuring continuity of operation and system
robustness. Additionally, the system incorporates human
operators as support agents, where robots provide intuitive
visual and audio signals to request assistance for battery
replacement or unloading harvested produce. The paper
further presents methods for field mapping, localization,
and navigation using a combination of aerial imaging,
sensor fusion, and AI-based perception. Optimization
strategies for multi-robot scheduling are introduced to
minimize total operation time and energy consumption under
real-world constraints. Experimental considerations
demonstrate that the proposed system enables scalable,
cost-effective deployment in row-based crop production such
as peppers, strawberries, and vineyards. The results
indicate that collaborative agro-robotic systems have
strong potential to significantly enhance productivity,
reduce dependency on manual labor, and support the
transition toward smart and sustainable agriculture.

This paper presents a novel concept of a compliant,
multifunctional agro-robotic arm intended for application
in intensive agricultural production. The proposed system
is designed as a lightweight, tendon-driven, redundant
manipulator with seven degrees of freedom (7-DOF), capable
of operating in unstructured outdoor environments. The
robot is optimized for tasks such as harvesting, pruning,
and planting, while ensuring adaptability, low cost, and
energy efficiency. A compliant mechanical structure and
cable-driven actuation system are introduced to improve
robustness and safety in interaction with plants and
environment. Key performance indicators (KPIs) are defined
to evaluate system effectiveness in real agricultural
conditions. The concept aims to provide an affordable and
scalable solution for small and medium-sized farms for
intensive agricultural production.

Autonomous systems in precision agriculture rely on
accurate scene understanding, navigation, and object
detection. Structured environments, such as raised beds,
are common in agriculture, where precise geometric analysis
is critical for robot operations. One of the most important
tasks for this use case is plane segmentation and parameter
extraction. Due to real-world agricultural conditions such
as uneven terrain, random occlusions, wind, and sun,
various methods have been developed to focus on speed,
accuracy, and robustness of the solution. This paper
compares four geometry-based methods across three
paradigms: consensus-fitting, region-growing, and
topology-driven. RANSAC is evaluated through two
independent implementations — PCL in C++ and Open3D in
Python — resulting in five tested implementations in total.
Depth images coming from an OAK-D stereo camera are sourced
to a point cloud and used as an input. As a starting point,
an artificial raised bed of blueberries is constructed and
used for experiments and testing. Standard metrics,
including root mean square (RMS), planarity, roughness,
inlier count, and execution time, are used for
benchmarking. The evaluation showcases the strengths and
weaknesses of tested methods for the agricultural use case.

This research explores the development of a sophisticated
edge-computing architecture designed for autonomous visual
perception tasks of mobile manipulators in precision
agriculture. By moving high-intensity computational
workloads from remote cloud servers to a localized,
GPU-accelerated infrastructure, the system achieves a high
degree of responsiveness and reliability, fundamental for
real-time robotic systems. The proposed solution integrates
an NVIDIA Jetson module with an RBKairos mobile platform,
using hardware-specific deep learning optimizations and
precise manipulation logic. Experimental validation
indicates that this design substantially mitigates temporal
instabilities in the control loop and enhances the system
energy autonomy by 12%, allowing greater duration of
autonomous field operations.

To facilitate ever-growing need for customized production,
modern automation requires extensive development of
advanced adaptive systems for robotic grasping. Among many
approaches, the application of underactuated mechanisms has
proven to solve many challenges that the mass customization
paradigm imposes. They enable simpler design of actuation
mechanisms which can be controlled using less complex
algorithms. As a rule, control systems of underactuated
grasping devices can be modeled as discrete-event systems
and a variety of control methods have been developed for
their operation. Amongst these, Petri Net-based methods
provide an effective framework for modeling and controlling
such systems with event-driven behavior and sensor-based
decision making. This paper presents the design of Petri
Net-inspired control system for an adaptive two-finger
underactuated robotic gripper. This system is presented in
the form of control interpreted Petri Net with defined
actions and conditions, and its real-world applicability is
experimentally verified using different grasping scenarios.

The increasing use of AI-based computer vision and
industrial robotics in automation has created new demands
on production systems. Automated production lines that
integrate object detection and classification with robotic
manipulation require reliable, deterministic, and flexible
PLC-based control solutions. This paper demonstrates the
development of a sustainable and innovative industrial
solution capable of performing a pick-and-place process by
executing data acquisition, decision-making through
neural-network inference, and the generation of actuator
control commands. The proposed solution integrates an
industrial PLC with a NPU (Neural Processing Unit) module
to control a robotic arm. The entire process, from
perception to control, is implemented on industrial-grade
edge devices. The system is demonstrated through a
representative industrial use case in which six classes of
objects are detected, classified, and sorted into
designated positions, serving as an example of a broader
class of assembly, sorting, packaging, and
material-handling processes. The solution was tested and
achieved over 96% accuracy in a six-class object-detection
task, while successfully controlling an anthropomorphic
robotic arm. The results indicate that the proposed
solution is highly adaptive, sustainable, and versatile,
enabling easy integration into new production lines as well
as the upgrading of existing ones.

This paper addresses the problem of determining the
position and orientation of 3D objects within a robot's
workspace.
The primary objective is to enable the robot to map a scene
and
store the spatial arrangement of objects, allowing for its
later
reconstruction. The solution is implemented through two
distinct
approaches: the first integrates the MATLAB environment
with the
ChatGPT model for multimodal image analysis and descriptive
identification, while the second relies on a locally trained
Convolutional Neural Network for shape classification. In
both
approaches, MATLAB is used for image processing,
segmentation,
and precise computation of object coordinates and
orientation
angles. The paper provides a comparative analysis of these
approaches in terms of accuracy, response time, and
practical
applicability in real-world robotic systems.

This paper presents the development of a convolutional
neural network architecture for semantic segmentation in
autonomous mobile robots. The model comprises a mix of
traditional CNNs and transformer-based attention modules.
ResNet18 and ResNet34 are employed as backbone models with
fast attention modules integrated at the end of each ResNet
level. The SUNRGBD dataset is used for training and
evaluation, while the final network is tested on a real
mobile robot TIAGo. An extensive parameter-tuning procedure
is conducted to determine optimal parameters and assess the
model's sensitivity to changes in parameters. The proposed
approach aims to improve segmentation performance while
meeting real-time computational constraints.

This paper presents a setup for a human-centered training
environment in assembly manufacturing. Building on a
previously developed neuroergonomic, cobot-assisted
workstation, the approach extends a factory-validated
system toward training-oriented applications. After
outlining the baseline architecture, the paper proposes an
enhanced use of its modular components - cognitive
monitoring, motion recognition, adaptive interfaces, and
collaborative robotics - to enable closed-loop,
individualized training. A methodology for leveraging
multimodal worker-state indicators to dynamically adjust
instruction and assistance is introduced. The framework for
implementation and validation in industrial settings is
discussed, with focus on learning efficiency, operational
performance, and ergonomics. The proposed approach is
expected to accelerate skill acquisition, improve workforce
flexibility, and enhance overall production efficiency.

Human pose estimation has become a critical component in
the development of autonomous robotic systems, enabling
machines to perceive and interpret human body positions in
real time. Applications range from autonomous vehicles,
where accurate detection of pedestrian motion, whether
walking, running, or falling is essential for safe
navigation, to humanoid robots capable of learning through
human motion imitation, as well as medical rehabilitation
systems that monitor patient movements during therapy. This
paper presents a systematic review of deep learning methods
for human pose estimation, with particular emphasis on
their applicability in autonomous robotic systems. We
analyze and compare state-of-the-art approaches, focusing
on key performance parameters such as accuracy, inference
speed, and computational requirements, providing a
structured overview of the current landscape in this
rapidly evolving field. One of the significant challenges
in training pose estimation models is the availability of
large and diverse datasets. In this context, we discuss the
potential of synthetic data generation as a promising
direction, offering several advantages over real-world data
collection, the ability to produce large quantities of
training samples, controlled variation of environmental
conditions such as lighting, contrast, and viewing angles,
as well as precise ground truth annotations that enable
reliable model evaluation. Finally, initial results from a
real-world implementation of a pose estimation system are
presented and discussed in the context of existing
approaches, highlighting practical potential and directions
for future research.

Ovaj rad pruža sveobuhvatnu analizu arhitektonskog sistema
za kreiranje naprednih virtuelnih okruženja namenjenih
prevazilaženju jaza između simulacije i stvarnosti u domenu
robotske montaže. Kroz prizmu montaže složenih
kontaktno-intenzivnih zadataka, kao što je manipulacija i
umetanje različitih komunikacionih kablova u serverska
kućišta, rad detaljno opisuje tehničke, softverske, i
hardverske prepreke koje otežavaju razvoj robotskih rešenja
bez upotrebe realne opreme. Virtualna okruženja su
konstruisana kao skalabilni sistem koji verno preslikava
realne sklopove kao i njihove međusobne interakcije.
Skalabilno bezbedno virtuelno okruženje omogućavaju trening
i testiranje neuralnih mreža adaptivne vizualno-motorne
politike pre nego što se rešenje testira na pravom
robotskom sistemu. Rad opisuje sve potrebne korake da se
simulira i obuči polisa veštačke inteligencije da bi
robotski sistem izvršavao zadatke u sub-milimetarskoj
preciznosti.