Harokopio University
A unified framework for efficient and adaptive autonomous systems
Abstract
dc:descriptionThis dissertation proposes a unified design framework for autonomous systems that are capable of adapting to changing conditions, operating under real-world constraints, and remaining human-centered. It argues that conventional approaches to artificial intelligence often rely, during the design of autonomous systems, on assumptions such as remote access to high computational power and a separation between the application itself and the actual operating conditions of the system. The computational resources available during autonomous operation are often limited compared with those used during the development and training of the corresponding intelligent mechanisms, while low-latency requirements restrict reliance on remote processing. At the same time, autonomous systems must be understandable and usable by human operators, as well as capable of adapting to changing operating conditions. To address these challenges, this dissertation introduces a flexible three-pillar framework that connects adaptive learning, efficient deployment, and natural human-system interaction, without imposing a single architecture for every application. The proposed framework is organized around three main pillars: human-centered interaction through natural language interfaces, efficient deployment through distributed and resource-aware execution, and adaptive decision-making through learning and control methods. The main contribution of the dissertation is not limited to these pillars individually, but lies in the way they can be selected and combined according to the requirements of each application. In particular, the framework highlights the connection between linguistic reasoning and adaptive action planning, the importance of deploying intelligent functions efficiently under constraints, and the role of natural interaction interfaces as a means through which users can supervise, configure, and adapt autonomous systems. These ideas are evaluated through targeted studies corresponding to the three pillars of the framework. With respect to human-centered interaction, the dissertation examines a perception-to-action process in which action planning is based on structured scene representations and spatial information. In parallel, a comparative evaluation of contemporary language-based reasoning systems reveals limitations in spatial understanding and highlights the need for explicit control mechanisms. With respect to efficient deployment, different approaches to optimization, processing distribution, and privacy protection are investigated, with the aim of deploying intelligent functions under constraints related to computational power, communication, latency, energy consumption, and data management. With respect to adaptive decision-making, learning and control methods are evaluated in navigation and robotic control scenarios, demonstrating how behavioral policies derived from the training of reinforcement learning models can support autonomous operation in simulated environments with obstacles and task-specific requirements, using limited sensory information and feedback. Finally, an implementation is presented that combines the three pillars in search-and-rescue missions, showing how interaction, deployment, and adaptation can be connected within a common application. Overall, the studies presented for each pillar show that adaptive intelligence, efficient deployment, and human-centered interaction can be combined in ways that support the development of autonomous systems that are more practical, more sensitive to privacy concerns, and more closely aligned with real-world application conditions.
Degree
thesis:*- Grantor dc:publisher
- Harokopio University
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Chronis, Christos
- Χρόνης, Χρήστος
Subjects
dc:subject × 16- Προσαρμοστική μάθηση
- Ευφυΐα με Περιορισμούς Πόρων
- Φυσική Αλληλεπίδραση Ανθρώπου-Συστήματος
- Αυτόνομα συστήματα
- Ανθρωποκεντρική τεχνητή νοημοσύνη
- Adaptive learning
- Resource-Constrained Intelligence
- Natural Human-System Interaction
- Autonomous systems
- Human-Centered Artificial Intelligence
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Επιστήμες Μηχανικού και Τεχνολογία
- Υπολογιστές, Υλικό (hardware) και Αρχιτεκτονική
- Electrical Engineering, Electronic Engineering, Information Engineering
- Engineering and Technology
- Computer science, Hardware and Architecture
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/62364