Grotta, Antonio (2025) An AI-Driven Cognitive Framework for Embodied Intelligence. [Tesi di dottorato]

Documento PDF Phd_thesis_AG.pdf Visibile a [TBR] Amministratori dell'archivio Download (24MB) | Richiedi una copia

Abstract

This Thesis develops and formalizes a modular cognitive architecture for embodied interaction between artificial agents and human users, tackling one of the most profound and long-standing challenges at the intersection of cognitive science, artificial intelligence (AI) and complex systems theory: how to enable artificial systems to sense, decide and act coherently within dynamic and shared environments. Unlike conventional AI models that usually rely on monolithic data-driven architectures, this work advances a structured and interpretable framework in which perception, planning and action are organized into functionally distinct yet dynamically integrated modules operating within continuous sensorimotor loops. The conceptual foundation of the Thesis lies in the synergy between modularity and embodiment. Modularity provides a principled way to decompose intelligent behavior into specialized subsystems, e.g., perception, planning and actuation. Embodiment, by contrast, grounds these subsystems in the ongoing interaction between the agent’s body (or virtual embodiment) and the environment, emphasizing the role of physical context in shaping cognitive processes. This dual perspective shifts the design of cognitive architectures from abstract symbolic reasoning toward situated and dynamical models of intelligence. Building on these principles, the Thesis introduces an architecture structured around three tightly coupled modules. The Perception module extracts a compact, low-dimensional representation of human movement online, using learning-based methods to estimate motion phase from high-dimensional trajectories without assuming strict periodicity. The Actuation module implements a novel controller that combines Kuramoto oscillator dynamics with deep reinforcement learning, enabling the artificial agent to enhance synchronization of its behavior with one or multiple human partners. The Planning module adds a higher-level layer of adaptive decision making, using recurrent neural networks to personalize task selection over time. The framework is conceived in the domain of virtual rehabilitation and physiotherapy, an application scenario that naturally demands adaptive and safe human–AI interaction. In this context, the system interprets human movement online, improve synchronization of a group of virtual humans and adaptively recommends rehabilitation exercises based on user profiles and interaction history. Results demonstrate robust online phase estimation of human motion, synchronization in multi-user scenarios and personalized planning of exercise protocols. These findings support both the feasibility and the effectiveness of the proposed architecture in real interactive settings. Beyond its immediate application, this research contributes a general cognitive substrate that can be extended to a broad range of embodied interaction scenarios, including collaborative robotics, shared virtual and extended reality, assistive technologies and distributed multi-agent systems. Finally, the Thesis outlines several promising research directions, including the integration of additional cognitive functions such as memory, attention and predictive modeling; the use of learned world models to enhance adaptive control; and the formal study of stability and synchronization in networks of multiple cognitive agents. Together, these developments aim to extend the scope and depth of embodied cognition in artificial systems, pushing toward a new generation of intelligent architectures that are modular and deeply integrated with the dynamics of the environments and humans they interact with.

Downloads

Actions (login required)

Modifica documento