Tesi etd-09292025-233611
Link copiato negli appunti
Tipo di tesi
Dottorato
Autore
IORI, FRANCESCO
URN
etd-09292025-233611
Titolo
From Reactive to Learned Models: Adaptive Control and Sparse Representations for Robust Robotic Systems
Settore scientifico disciplinare
ING-IND/34
Corso di studi
Istituto di Biorobotica - PHD IN BIOROBOTICA
Relatori
relatore Prof. FALOTICO, EGIDIO
Parole chiave
- Robust Robotic Systems
- learning-based control
- brain-inspired computation
- Sparse Distributed Representations
- human-robot handover
- trajectory modulation
- robust performance
Data inizio appello
18/05/2026;
Disponibilità
parziale
Riassunto analitico
Achieving robust and adaptive behavior remains a fundamental challenge in robotics, where systems must operate reliably despite environmental variability, unexpected perturbations, and the inherent uncertainties of real-world deployment. This thesis addresses this challenge through a systematic exploration of the representation-control interdependence in robotic systems, demonstrating how data-driven approaches can enhance robustness while maintaining the interpretability and safety requirements essential for practical applications. The work progresses from analytical frameworks with learned parameters to fully learned representations, revealing how different levels of learning integration affect system robustness and adaptability.
The thesis presents five interconnected contributions across three main research areas. Two novel approaches are presented for human-robot handover, addressing the critical challenge of maintaining safe and coordinated interaction under perturbations. The first contribution introduces a reactive control framework that enables real-time trajectory modulation based on human partner behavior, demonstrating through user studies that humans consistently value adaptive robot behavior over pure efficiency. The second advances this work through an anticipatory control architecture that detects human engagement to enable proactive trajectory modulation, achieving robust performance under extreme disturbances while operating with minimal perceptual requirements. This anticipatory approach eliminates the need for perturbation-specific training data, representing a significant advance toward deployable human-robot collaboration.
The third contribution explores industrial deployment of learning-based control through behavioral cloning for automated insole packing. This work reveals fundamental limitations in compartmentalized system design, where independently developed perception and control modules struggle to achieve robust integration, providing crucial insights that motivate the subsequent theoretical investigations.
The final contributions address representation learning from first principles. A novel framework for learning Sparse Distributed Representations through mutual information maximization is developed, demonstrating how sparsity emerges naturally from information-theoretic objectives without requiring explicit architectural constraints. This work bridges decades of neuroscience research with modern information theory, providing both theoretical foundations and practical algorithms for brain-inspired computation. The thesis concludes by demonstrating how these sparse representations enable efficient memory-based control, implementing a sparse memory architecture that achieves competitive performance with significant computational advantages, thus completing the representation-control cycle explored throughout the work.
Together, these contributions establish a comprehensive framework for understanding and implementing robust robotic systems that effectively balance adaptability, interpretability, and computational efficiency in real-world applications.
The thesis presents five interconnected contributions across three main research areas. Two novel approaches are presented for human-robot handover, addressing the critical challenge of maintaining safe and coordinated interaction under perturbations. The first contribution introduces a reactive control framework that enables real-time trajectory modulation based on human partner behavior, demonstrating through user studies that humans consistently value adaptive robot behavior over pure efficiency. The second advances this work through an anticipatory control architecture that detects human engagement to enable proactive trajectory modulation, achieving robust performance under extreme disturbances while operating with minimal perceptual requirements. This anticipatory approach eliminates the need for perturbation-specific training data, representing a significant advance toward deployable human-robot collaboration.
The third contribution explores industrial deployment of learning-based control through behavioral cloning for automated insole packing. This work reveals fundamental limitations in compartmentalized system design, where independently developed perception and control modules struggle to achieve robust integration, providing crucial insights that motivate the subsequent theoretical investigations.
The final contributions address representation learning from first principles. A novel framework for learning Sparse Distributed Representations through mutual information maximization is developed, demonstrating how sparsity emerges naturally from information-theoretic objectives without requiring explicit architectural constraints. This work bridges decades of neuroscience research with modern information theory, providing both theoretical foundations and practical algorithms for brain-inspired computation. The thesis concludes by demonstrating how these sparse representations enable efficient memory-based control, implementing a sparse memory architecture that achieves competitive performance with significant computational advantages, thus completing the representation-control cycle explored throughout the work.
Together, these contributions establish a comprehensive framework for understanding and implementing robust robotic systems that effectively balance adaptability, interpretability, and computational efficiency in real-world applications.
File
| Nome file | Dimensione |
|---|---|
Ci sono 1 file riservati su richiesta dell'autore. |
|