Eduard Zelenay
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Eduard Zelenay has a total of 9 co-authors. Here are the top co-authors by number of shared publications:
- František Duchoň
Institute of Robotics and Cybernetics, Faculty of Electrical Engineering and Information Technology, Slovak University of Technology in Bratislava, Ilkovicova 3, 812 19 Bratislava, Slovakia
2 shared publications - Maroš Kocúr
Faculty of Electrical Engineering and Information Technology, Slovak University of Technology, Bratislava, Slovakia
2 shared publications - Martin Lukáč
Faculty of Electrical Engineering and Information Technology, Slovak University of Technology, Bratislava, Slovakia
2 shared publications - Izabela Trepáčová
Institute of Robotics and Cybernetics, Faculty of Electrical Engineering and Information Technology, Slovak University of Technology in Bratislava, Ilkovicova 3, 812 19 Bratislava, Slovakia
1 shared publication - Jakub Ivan
Institute of Robotics and Cybernetics, Faculty of Electrical Engineering and Information Technology, Slovak University of Technology in Bratislava, Ilkovicova 3, 841 04 Bratislava, Slovakia
1 shared publication - Martin Dekan
Institute of Robotics and Cybernetics, Faculty of Electrical Engineering and Information Technology, Slovak University of Technology in Bratislava, Bratislava, Slovakia
1 shared publication
Latest Publications
- Open Access
- 21 September 2026
- Published by Springer Nature in International Journal of Intelligent Robotics and Applications
- p. 1-17
SMaRTAban: a voice-controlled LLM agent for quadruped mobile robotics with integrated vision
Large Language Models (LLMs) have emerged as powerful tools for enabling intuitive human-robot interaction through natural language. However, existing approaches often lack evaluated voice control in languages other than English, integrated vision, and comprehensive safety mechanisms. We present SMaRTAban (Smart Multi-legged Robotic Transformer Agent Artaban), a system that combines a LangGraph ReAct agent with the quadruped mobile robot Artaban for natural language control. Our system integrates: (1) voice commands with automatic language detection via Whisper, evaluated on English and Slovak, (2) on-demand scene interpretation via GPT-4o vision, (3) a safety-focused system prompt with interruptible tool execution, and (4) a React-based GUI for transparent operation monitoring. A pilot voice-control evaluation on 120 samples showed robust semantic understanding across English and Slovak (BERTScore F1: 0.87–1.00), despite transcription challenges in Slovak. We further conducted 4,800 accepted tool-calling trials across two models, six agent and prompting configurations, two languages, and 10 command scenarios. The strongest configuration combined iterative ReAct execution, an explicit system prompt, and few-shot prompting with three task-specific movement examples, achieving a tool-call F1 of 0.9977 and 96.5% exact-sequence success with GPT-4o. A separate performance evaluation showed that remote transcription and model processing dominated response time, whereas local ROS communication contributed only a few milliseconds.
- 21 September 2026
- Published by Springer Nature in International Journal of Intelligent Robotics and Applications
- p. 1-17
Open AccessSMaRTAban: a voice-controlled LLM agent for quadruped mobile robotics with integrated vision
Large Language Models (LLMs) have emerged as powerful tools for enabling intuitive human-robot interaction through natural language. However, existing approaches often lack evaluated voice control in languages other than English, integrated vision, and comprehensive safety mechanisms. We present SMaRTAban (Smart Multi-legged Robotic Transformer Agent Artaban), a system that combines a LangGraph ReAct agent with the quadruped mobile robot Artaban for natural language control. Our system integrates: (1) voice commands with automatic language detection via Whisper, evaluated on English and Slovak, (2) on-demand scene interpretation via GPT-4o vision, (3) a safety-focused system prompt with interruptible tool execution, and (4) a React-based GUI for transparent operation monitoring. A pilot voice-control evaluation on 120 samples showed robust semantic understanding across English and Slovak (BERTScore F1: 0.87–1.00), despite transcription challenges in Slovak. We further conducted 4,800 accepted tool-calling trials across two models, six agent and prompting configurations, two languages, and 10 command scenarios. The strongest configuration combined iterative ReAct execution, an explicit system prompt, and few-shot prompting with three task-specific movement examples, achieving a tool-call F1 of 0.9977 and 96.5% exact-sequence success with GPT-4o. A separate performance evaluation showed that remote transcription and model processing dominated response time, whereas local ROS communication contributed only a few milliseconds.
- Open Access
- 6 August 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (15) , 7849
Large Language Models: From Internal Architecture and Distributed Training Optimisation to Adaptation Strategies
This paper presents a unified technical survey of Large Language Models (LLMs), connecting three layers of the modelling pipeline that existing surveys address in isolation: internal architecture, distributed training optimisation, and downstream adaptation. Its organising principle is the dependency between these layers—how a choice at one constrains what remains feasible at the next. The survey examines fundamental mechanisms (tokenisation, scaled dot-product attention, activation functions, and normalisation, including RMSNorm and pre- versus post-normalisation placement), then the engineering of training at scale: data, tensor, and pipeline parallelism, hybrid schemes, mixed-precision training with BF16 and FP8, ZeRO-Offload memory management, activation checkpointing, and compute-optimal scaling laws together with the conditions under which they fail. The adaptation section covers supervised and instruction fine-tuning, a comparison of parameter-efficient methods (LoRA, QLoRA, adapters, prefix and prompt tuning), Reinforcement Learning from Human Feedback with its reward-hacking failure mode, alternatives including DPO, KTO and Constitutional AI, Retrieval-Augmented Generation beyond the basic pipeline, and decoding strategies. Practical configuration guidance is given for 7B, 70B and trillion-parameter regimes. Dedicated treatments of Mixture-of-Experts architectures, long-context modelling, and hardware-aware co-design close the survey, with open challenges classified by origin and severity.
- 6 August 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (15) , 7849
Open AccessLarge Language Models: From Internal Architecture and Distributed Training Optimisation to Adaptation Strategies
This paper presents a unified technical survey of Large Language Models (LLMs), connecting three layers of the modelling pipeline that existing surveys address in isolation: internal architecture, distributed training optimisation, and downstream adaptation. Its organising principle is the dependency between these layers—how a choice at one constrains what remains feasible at the next. The survey examines fundamental mechanisms (tokenisation, scaled dot-product attention, activation functions, and normalisation, including RMSNorm and pre- versus post-normalisation placement), then the engineering of training at scale: data, tensor, and pipeline parallelism, hybrid schemes, mixed-precision training with BF16 and FP8, ZeRO-Offload memory management, activation checkpointing, and compute-optimal scaling laws together with the conditions under which they fail. The adaptation section covers supervised and instruction fine-tuning, a comparison of parameter-efficient methods (LoRA, QLoRA, adapters, prefix and prompt tuning), Reinforcement Learning from Human Feedback with its reward-hacking failure mode, alternatives including DPO, KTO and Constitutional AI, Retrieval-Augmented Generation beyond the basic pipeline, and decoding strategies. Practical configuration guidance is given for 7B, 70B and trillion-parameter regimes. Dedicated treatments of Mixture-of-Experts architectures, long-context modelling, and hardware-aware co-design close the survey, with open challenges classified by origin and severity.