Andrzej Kos
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Andrzej Kos has a total of 28 co-authors. Here are the top co-authors by number of shared publications:
- Ravi Raj
Department of Mining, Industrial and ICT Engineering (EMIT), Manresa School of Engineering (EPSEM), Universitat Politècnica de Catalunya (UPC), BarcelonaTech, Barcelona, Spain
13 shared publications - Piotr Kocanda
Department of Electronics, AGH University of Science and Technology, al. Mickiewicza 30, C3-512, 30-059 Kraków, Poland
4 shared publications - Gilbert De Mey
Department of Electronics and Information Systems (ELIS), Ghent University, Technologiepark 126, 9052 Ghent, Belgium
3 shared publications Pawel Marzec
3 shared publications- Adam Gołda
AGH University of Science and Technology, Faculty of Computer Science, Electronics and Telecommunications, Department of Electronics, al. A. Mickiewicza 30, 30-059 Krakow, Poland
2 shared publications Adama Samake
2 shared publications
Latest Publications
- Open Access
- 31 July 2026
- Published by Springer Nature in Scientific Reports
- Vol. 16 (1)
Understanding the Markov decision process for reinforcement learning in robotics perception
- 31 July 2026
- Published by Springer Nature in Scientific Reports
- Vol. 16 (1)
Open AccessUnderstanding the Markov decision process for reinforcement learning in robotics perception
- Open Access
Recent Developments and Applications of Drone Swarm: Techniques, Strategies, and Challenges
The dynamic and complex environment, together with challenging assignments, requires that unmanned aerial vehicle (UAV) systems evolve toward cooperation, autonomy, and cognition. UAV swarms illustrate a revolutionary development in aerial robotics, which utilizes coordinated autonomy to improve operational efficiency. This study offers a detailed examination of UAV swarm systems, the latest developments, and their different applications. The main domains, such as intelligent path planning, work allocation, coordinated control, and safety issues, are analyzed, focusing on the integration of Artificial Intelligence (AI) and Deep Learning (DL) to enhance decision-making and agility. We address the constraints and potential advances in the field of swarm intelligence to facilitate additional research endeavors. The ongoing advancement of drone swarm technologies and its exploration of military uses highlight the increasing importance of anti-drone swarm strategies. Therefore, studying these strategies will have substantial practical importance in preventing and countering drone swarm combat. Thus, this article provides detailed drone swarm applications and the importance of anti-drone swarm techniques in strategic operations. Furthermore, this comprehensive study of the literature aims to offer innovative perspectives on the latest advances in UAV swarm intelligence technology. Future research trends and challenges are discussed to find the research gap.
Open AccessRecent Developments and Applications of Drone Swarm: Techniques, Strategies, and Challenges
The dynamic and complex environment, together with challenging assignments, requires that unmanned aerial vehicle (UAV) systems evolve toward cooperation, autonomy, and cognition. UAV swarms illustrate a revolutionary development in aerial robotics, which utilizes coordinated autonomy to improve operational efficiency. This study offers a detailed examination of UAV swarm systems, the latest developments, and their different applications. The main domains, such as intelligent path planning, work allocation, coordinated control, and safety issues, are analyzed, focusing on the integration of Artificial Intelligence (AI) and Deep Learning (DL) to enhance decision-making and agility. We address the constraints and potential advances in the field of swarm intelligence to facilitate additional research endeavors. The ongoing advancement of drone swarm technologies and its exploration of military uses highlight the increasing importance of anti-drone swarm strategies. Therefore, studying these strategies will have substantial practical importance in preventing and countering drone swarm combat. Thus, this article provides detailed drone swarm applications and the importance of anti-drone swarm techniques in strategic operations. Furthermore, this comprehensive study of the literature aims to offer innovative perspectives on the latest advances in UAV swarm intelligence technology. Future research trends and challenges are discussed to find the research gap.
- 19 January 2026
- Published by Polish Academy of Sciences Chancellery in International Journal of Electronics and Telecommunications
- p. 1-8
Mitigating the impact of thermal reflections in object detection using Vision-Language Models
Thermal imaging is increasingly employed for navigation in challenging conditions such as dense smoke or fog. However, the limited availability of thermal images compared to RGB data makes training deep learning models, such as Convolutional Neural Networks (CNNs), significantly more difficult and often yields unsatisfactory results. Vision-Language Models (VLMs), due to their ability to perform tasks without extensive retraining or with only a small number of training samples, hold the potential to overcome current limitations in thermal imaging applications. This paper introduces a method leveraging VLMs to reduce the impact of reflections in thermal images on object detection accuracy, with a particular focus on human detection. The proposed approach improves the F1-score from 0.83 to 0.97 on a dedicated evaluation dataset, outperforming a baseline solution based solely on the widely used YOLOv11 model. Furthermore, we investigate the effects of quantization on various open-source VLMs, analyzing their performance, processing speed, and memory requirements.
- 19 January 2026
- Published by Polish Academy of Sciences Chancellery in International Journal of Electronics and Telecommunications
- p. 1-8
Mitigating the impact of thermal reflections in object detection using Vision-Language Models
Thermal imaging is increasingly employed for navigation in challenging conditions such as dense smoke or fog. However, the limited availability of thermal images compared to RGB data makes training deep learning models, such as Convolutional Neural Networks (CNNs), significantly more difficult and often yields unsatisfactory results. Vision-Language Models (VLMs), due to their ability to perform tasks without extensive retraining or with only a small number of training samples, hold the potential to overcome current limitations in thermal imaging applications. This paper introduces a method leveraging VLMs to reduce the impact of reflections in thermal images on object detection accuracy, with a particular focus on human detection. The proposed approach improves the F1-score from 0.83 to 0.97 on a dedicated evaluation dataset, outperforming a baseline solution based solely on the widely used YOLOv11 model. Furthermore, we investigate the effects of quantization on various open-source VLMs, analyzing their performance, processing speed, and memory requirements.
- 19 January 2026
- Published by Polish Academy of Sciences Chancellery in International Journal of Electronics and Telecommunications
- p. 1-6
Optimal thermal placement of electronic functional blocks using symbolic algebra. A new application of Groebner Analysis
The paper presents a new idea and algorithm for arranging topography of functional blocks (hereinafter referred to as components) on an integrated circuit substrate, e.g. a digital processor, power circuit etc., in such a way as to minimize the mutual thermal interactions of individual components. Our analytical method finds the global optimum, which distinguishes it from numerical methods that can only find local minima. This leads to uniform temperature distribution, and therefore full use of the thermal properties of the electronic system, minimizing the maximum temperature on the substrate, and consequently allows for increasing the throughput. Presented approach is universal and allows for solving many similar global minimum search problems.
- 19 January 2026
- Published by Polish Academy of Sciences Chancellery in International Journal of Electronics and Telecommunications
- p. 1-6
Optimal thermal placement of electronic functional blocks using symbolic algebra. A new application of Groebner Analysis
The paper presents a new idea and algorithm for arranging topography of functional blocks (hereinafter referred to as components) on an integrated circuit substrate, e.g. a digital processor, power circuit etc., in such a way as to minimize the mutual thermal interactions of individual components. Our analytical method finds the global optimum, which distinguishes it from numerical methods that can only find local minima. This leads to uniform temperature distribution, and therefore full use of the thermal properties of the electronic system, minimizing the maximum temperature on the substrate, and consequently allows for increasing the throughput. Presented approach is universal and allows for solving many similar global minimum search problems.
- Open Access
- 22 October 2025
- Published by Polish Academy of Sciences Chancellery in International Journal of Electronics and Telecommunications
- p. 1-10
An improved human pose estimation using Deep Neural Network for the optimization of human-robot interactions
Research shows that mobile support robots are becoming increasingly valuable in various situations, such as monitoring daily activities, providing medical services, and supporting elderly people. For interpreting human conduct and intention, these robots largely depend on human activity recognition (HAR). However, previous awareness of human appearance (human recognition) and recognition of humans for monitoring (human surveillance) are necessary to enable HAR to work with assistance robots. Al-so However, multimodal human behavior recognition is constrained by costly hardware and a rigorous setting, making it challenging to effectively balance inference accuracy and system expense. Naturally, a key problem in human pose or behavior detection is the ability to extract additional purposeful interpretations from easily accessible live videos. In this paper, we employ human pose detection to address the problem and provide well-crafted assessment measures to show demonstrate the effectiveness of our approach, which utilizes deep neural networks (DNNs) This article proposes a human intention detection system that anticipates human intentions in human- and robot-centered scenarios by utilizing the incorporation of visual information as well as input features, including human positions, head orientations, and critical skeletal key points. Our goal is to aid human-robot interactions by helping mobile robots through realtime human pose prediction using the recognition of 18 distinct key points in the body's structure. The effectiveness of this strategy is demonstrated by the suggested study using Python, and the results of simulations verify the reliability and accuracy of this method.
- 22 October 2025
- Published by Polish Academy of Sciences Chancellery in International Journal of Electronics and Telecommunications
- p. 1-10
Open AccessAn improved human pose estimation using Deep Neural Network for the optimization of human-robot interactions
Research shows that mobile support robots are becoming increasingly valuable in various situations, such as monitoring daily activities, providing medical services, and supporting elderly people. For interpreting human conduct and intention, these robots largely depend on human activity recognition (HAR). However, previous awareness of human appearance (human recognition) and recognition of humans for monitoring (human surveillance) are necessary to enable HAR to work with assistance robots. Al-so However, multimodal human behavior recognition is constrained by costly hardware and a rigorous setting, making it challenging to effectively balance inference accuracy and system expense. Naturally, a key problem in human pose or behavior detection is the ability to extract additional purposeful interpretations from easily accessible live videos. In this paper, we employ human pose detection to address the problem and provide well-crafted assessment measures to show demonstrate the effectiveness of our approach, which utilizes deep neural networks (DNNs) This article proposes a human intention detection system that anticipates human intentions in human- and robot-centered scenarios by utilizing the incorporation of visual information as well as input features, including human positions, head orientations, and critical skeletal key points. Our goal is to aid human-robot interactions by helping mobile robots through realtime human pose prediction using the recognition of 18 distinct key points in the body's structure. The effectiveness of this strategy is demonstrated by the suggested study using Python, and the results of simulations verify the reliability and accuracy of this method.
- Open Access
- 3 September 2025
- Published by Polish Academy of Sciences Chancellery in Bulletin Of The Polish Academy Of Sciences-Technical Sciences
- p. 155044
Fast data transmission with FPGAs for implementation in ALICE FIT detector in CERN
In high-energy physics experiments, massive amounts of data need to be processed. The limitations of peripherals in field-programmable gate arrays (FPGAs), such as analog-to-digital converters (ADCs), make it impractical to design setups based solely on a single FPGA chip, creating a need for high-speed, low-latency methods for inter-chip data transmission, where multi-gigabit transceivers (MGTs) are essential. This paper aims to assess the feasibility of using MGTs in the readout electronics of the fast interaction trigger (FIT) detector within the A Large Ion Collider Experiment (ALICE) at CERN. It examines two established protocols, i.e. Aurora 8b/10b and Aurora 64b/66b, focusing on their throughput and latency. Additionally, it presents a custom protocol based on 64b/66b encoding, specifically designed to reduce latency, and compares its performance with standard solutions. Transmission tests were conducted between two Xilinx boards, ZCU102 and ZCU106, using both the Aurora protocols and the custom protocol. Furthermore, transmission quality was evaluated based on the type and length of the transmission medium as well as on the lengths of the pseudo-random binary sequence (PRBS).
- 3 September 2025
- Published by Polish Academy of Sciences Chancellery in Bulletin Of The Polish Academy Of Sciences-Technical Sciences
- p. 155044
Open AccessFast data transmission with FPGAs for implementation in ALICE FIT detector in CERN
In high-energy physics experiments, massive amounts of data need to be processed. The limitations of peripherals in field-programmable gate arrays (FPGAs), such as analog-to-digital converters (ADCs), make it impractical to design setups based solely on a single FPGA chip, creating a need for high-speed, low-latency methods for inter-chip data transmission, where multi-gigabit transceivers (MGTs) are essential. This paper aims to assess the feasibility of using MGTs in the readout electronics of the fast interaction trigger (FIT) detector within the A Large Ion Collider Experiment (ALICE) at CERN. It examines two established protocols, i.e. Aurora 8b/10b and Aurora 64b/66b, focusing on their throughput and latency. Additionally, it presents a custom protocol based on 64b/66b encoding, specifically designed to reduce latency, and compares its performance with standard solutions. Transmission tests were conducted between two Xilinx boards, ZCU102 and ZCU106, using both the Aurora protocols and the custom protocol. Furthermore, transmission quality was evaluated based on the type and length of the transmission medium as well as on the lengths of the pseudo-random binary sequence (PRBS).
- Top Cited PapersOpen Access
An Extensive Study of Convolutional Neural Networks: Applications in Computer Vision for Improved Robotics Perceptions
Convolutional neural networks (CNNs), a type of artificial neural network (ANN) in the deep learning (DL) domain, have gained popularity in several computer vision applications and are attracting research in other fields, including robotic perception. CNNs are developed to autonomously and effectively acquire spatial patterns of characteristics using backpropagation, leveraging an array of elements, including convolutional layers, pooling layers, and fully connected layers. Current reviews predominantly emphasize CNNs’ applications in various contexts, neglecting a comprehensive perspective on CNNs and failing to address certain recently presented new ideas, including robotic perception. This review paper presents an overview of the fundamental principles of CNNs and their applications in diverse computer vision tasks for robotic perception while addressing the corresponding challenges and future prospects for the domain of computer vision in improved robotic perception. This paper addresses the history, basic concepts, working principles, applications, and the most important components of CNNs. Understanding the concepts, benefits, and constraints associated with CNNs is crucial for exploiting their possibilities in robotic perception, with the aim of enhancing robotic performance and intelligence.
Top Cited PapersOpen AccessAn Extensive Study of Convolutional Neural Networks: Applications in Computer Vision for Improved Robotics Perceptions
Convolutional neural networks (CNNs), a type of artificial neural network (ANN) in the deep learning (DL) domain, have gained popularity in several computer vision applications and are attracting research in other fields, including robotic perception. CNNs are developed to autonomously and effectively acquire spatial patterns of characteristics using backpropagation, leveraging an array of elements, including convolutional layers, pooling layers, and fully connected layers. Current reviews predominantly emphasize CNNs’ applications in various contexts, neglecting a comprehensive perspective on CNNs and failing to address certain recently presented new ideas, including robotic perception. This review paper presents an overview of the fundamental principles of CNNs and their applications in diverse computer vision tasks for robotic perception while addressing the corresponding challenges and future prospects for the domain of computer vision in improved robotic perception. This paper addresses the history, basic concepts, working principles, applications, and the most important components of CNNs. Understanding the concepts, benefits, and constraints associated with CNNs is crucial for exploiting their possibilities in robotic perception, with the aim of enhancing robotic performance and intelligence.
- Open Access
- 1 October 2024
- Published by Springer Nature in Scientific Reports
- Vol. 14 (1) , 1-17
Intelligent mobile robot navigation in unknown and complex environment using reinforcement learning technique
The usage of mobile robots (MRs) has expanded dramatically in the last several years across a wide range of industries, including manufacturing, surveillance, healthcare, and warehouse automation. To ensure the efficient and safe operation of these MRs, it is crucial to design effective control strategies that can adapt to changing environments. In this paper, we propose a new technique for controlling MRs using reinforcement learning (RL). Our approach involves mathematical model generation and later training a neural network (NN) to learn a policy for robot control using RL. The policy is learned through trial and error, where MR explores the environment and receives rewards based on its actions. The rewards are designed to encourage the robot to move towards its goal while avoiding obstacles. In this work, a deep Q-learning (QL) agent is used to enable robots to autonomously learn to avoid collisions with obstacles and enhance navigation abilities in an unknown environment. When operating MR independently within an unfamiliar area, a RL model is used to identify the targeted location, and the Deep Q-Network (DQN) is used to navigate to the goal location. We evaluate our approach using a simulation using the Epsilon-Greedy algorithm. The results show that our approach outperforms traditional MR control strategies in terms of both efficiency and safety.
- 1 October 2024
- Published by Springer Nature in Scientific Reports
- Vol. 14 (1) , 1-17
Open AccessIntelligent mobile robot navigation in unknown and complex environment using reinforcement learning technique
The usage of mobile robots (MRs) has expanded dramatically in the last several years across a wide range of industries, including manufacturing, surveillance, healthcare, and warehouse automation. To ensure the efficient and safe operation of these MRs, it is crucial to design effective control strategies that can adapt to changing environments. In this paper, we propose a new technique for controlling MRs using reinforcement learning (RL). Our approach involves mathematical model generation and later training a neural network (NN) to learn a policy for robot control using RL. The policy is learned through trial and error, where MR explores the environment and receives rewards based on its actions. The rewards are designed to encourage the robot to move towards its goal while avoiding obstacles. In this work, a deep Q-learning (QL) agent is used to enable robots to autonomously learn to avoid collisions with obstacles and enhance navigation abilities in an unknown environment. When operating MR independently within an unfamiliar area, a RL model is used to identify the targeted location, and the Deep Q-Network (DQN) is used to navigate to the goal location. We evaluate our approach using a simulation using the Epsilon-Greedy algorithm. The results show that our approach outperforms traditional MR control strategies in terms of both efficiency and safety.
- Open Access
- 19 August 2024
- Published by MDPI AG in Electronics
- Vol. 13 (16) , 3285
Study of Human–Robot Interactions for Assistive Robots Using Machine Learning and Sensor Fusion Technologies
In recent decades, the potential of robots’ understanding, perception, learning, and action has been widely expanded due to the integration of artificial intelligence (AI) into almost every system. Cooperation between AI and human beings will be responsible for the bright future of AI technology. Moreover, for a perfect manually or automatically controlled machine or device, the device must perform together with a human through multiple levels of automation and assistance. Humans and robots cooperate or interact in various ways. With the enhancement of robot efficiencies, they can perform more work through an automatic method; therefore, we need to think about cooperation between humans and robots, the required software architectures, and information about the designs of user interfaces. This paper describes the most important strategies of human–robot interactions and the relationships between several control techniques and cooperation techniques using sensor fusion and machine learning (ML). Based on the behavior and thinking of humans, a human–robot interaction (HRI) framework is studied and explored in this article to make attractive, safe, and efficient systems. Additionally, research on intention recognition, compliance control, and perception of the environment by elderly assistive robots for the optimization of HRI is investigated in this paper. Furthermore, we describe the theory of HRI and explain the different kinds of interactions and required details for both humans and robots to perform different kinds of interactions, including the circumstances-based evaluation technique, which is the most important criterion for assistive robots.
- 19 August 2024
- Published by MDPI AG in Electronics
- Vol. 13 (16) , 3285
Open AccessStudy of Human–Robot Interactions for Assistive Robots Using Machine Learning and Sensor Fusion Technologies
In recent decades, the potential of robots’ understanding, perception, learning, and action has been widely expanded due to the integration of artificial intelligence (AI) into almost every system. Cooperation between AI and human beings will be responsible for the bright future of AI technology. Moreover, for a perfect manually or automatically controlled machine or device, the device must perform together with a human through multiple levels of automation and assistance. Humans and robots cooperate or interact in various ways. With the enhancement of robot efficiencies, they can perform more work through an automatic method; therefore, we need to think about cooperation between humans and robots, the required software architectures, and information about the designs of user interfaces. This paper describes the most important strategies of human–robot interactions and the relationships between several control techniques and cooperation techniques using sensor fusion and machine learning (ML). Based on the behavior and thinking of humans, a human–robot interaction (HRI) framework is studied and explored in this article to make attractive, safe, and efficient systems. Additionally, research on intention recognition, compliance control, and perception of the environment by elderly assistive robots for the optimization of HRI is investigated in this paper. Furthermore, we describe the theory of HRI and explain the different kinds of interactions and required details for both humans and robots to perform different kinds of interactions, including the circumstances-based evaluation technique, which is the most important criterion for assistive robots.
- 24 June 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-6
Learning the Dynamics of Human Patterns for Autonomous Navigation
The state of autonomous driving, robot navigation, and athletic activities is very significant in modern society due to humankind's ongoing advancement. When it comes to mobile robots or automated vehicles operating within congested areas, mobile communication systems are essential components for autonomous navigation aimed at preventing accidents involving people. This article presents a deep learning (DL) and hybrid multiprocessor-based approach for estimating human posture using biological signal processing. The main objective of this article is to develop an algorithm to recognize human poses for the purpose of protecting human beings from collision with higher accuracy to assist robots or autonomous vehicles in safe and efficient navigation in crowded environments. This study presents a transformer-based framework that uses input characteristics such as human locations, face orientations, and crucial 2D skeleton points using integrated within-the-wild sensory feedback to anticipate human potential paths in human-centric situations. We will identify 18 distinct locations within the human skeleton for pose estimation in real time, which will help autonomous navigation operations for mobile robots or autonomous vehicles avoid collisions with humans in a crowded environment. This proposed research implements the Python programming language to demonstrate the performance of this approach, and the outcomes of simulations validate its reliability and accuracy.
- 24 June 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-6
Learning the Dynamics of Human Patterns for Autonomous Navigation
The state of autonomous driving, robot navigation, and athletic activities is very significant in modern society due to humankind's ongoing advancement. When it comes to mobile robots or automated vehicles operating within congested areas, mobile communication systems are essential components for autonomous navigation aimed at preventing accidents involving people. This article presents a deep learning (DL) and hybrid multiprocessor-based approach for estimating human posture using biological signal processing. The main objective of this article is to develop an algorithm to recognize human poses for the purpose of protecting human beings from collision with higher accuracy to assist robots or autonomous vehicles in safe and efficient navigation in crowded environments. This study presents a transformer-based framework that uses input characteristics such as human locations, face orientations, and crucial 2D skeleton points using integrated within-the-wild sensory feedback to anticipate human potential paths in human-centric situations. We will identify 18 distinct locations within the human skeleton for pose estimation in real time, which will help autonomous navigation operations for mobile robots or autonomous vehicles avoid collisions with humans in a crowded environment. This proposed research implements the Python programming language to demonstrate the performance of this approach, and the outcomes of simulations validate its reliability and accuracy.